diff --git a/llvm/lib/Target/LoongArch/LoongArchTargetTransformInfo.cpp b/llvm/lib/Target/LoongArch/LoongArchTargetTransformInfo.cpp index f548a8dd0532b..5107c8def3799 100644 --- a/llvm/lib/Target/LoongArch/LoongArchTargetTransformInfo.cpp +++ b/llvm/lib/Target/LoongArch/LoongArchTargetTransformInfo.cpp @@ -111,4 +111,25 @@ bool LoongArchTTIImpl::shouldExpandReduction(const IntrinsicInst *II) const { } } -// TODO: Implement more hooks to provide TTI machinery for LoongArch. +LoongArchTTIImpl::TTI::MemCmpExpansionOptions +LoongArchTTIImpl::enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const { + TTI::MemCmpExpansionOptions Options; + + if (!ST->hasUAL()) + return Options; + + Options.MaxNumLoads = TLI->getMaxExpandSizeMemcmp(OptSize); + Options.NumLoadsPerBlock = Options.MaxNumLoads; + Options.AllowOverlappingLoads = true; + + // TODO: Support for vectors. + if (ST->is64Bit()) { + Options.LoadSizes = {8, 4, 2, 1}; + Options.AllowedTailExpansions = {3, 5, 6}; + } else { + Options.LoadSizes = {4, 2, 1}; + Options.AllowedTailExpansions = {3}; + } + + return Options; +} diff --git a/llvm/lib/Target/LoongArch/LoongArchTargetTransformInfo.h b/llvm/lib/Target/LoongArch/LoongArchTargetTransformInfo.h index e3f16c7804994..9b479f9dc0dc5 100644 --- a/llvm/lib/Target/LoongArch/LoongArchTargetTransformInfo.h +++ b/llvm/lib/Target/LoongArch/LoongArchTargetTransformInfo.h @@ -55,7 +55,8 @@ class LoongArchTTIImpl : public BasicTTIImplBase { bool shouldExpandReduction(const IntrinsicInst *II) const override; - // TODO: Implement more hooks to provide TTI machinery for LoongArch. + TTI::MemCmpExpansionOptions + enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const override; }; } // end namespace llvm diff --git a/llvm/test/CodeGen/LoongArch/expandmemcmp-optsize.ll b/llvm/test/CodeGen/LoongArch/expandmemcmp-optsize.ll index 4f7c8967c3049..36670fa801b36 100644 --- a/llvm/test/CodeGen/LoongArch/expandmemcmp-optsize.ll +++ b/llvm/test/CodeGen/LoongArch/expandmemcmp-optsize.ll @@ -38,260 +38,488 @@ entry: } define signext i32 @bcmp_size_1(ptr %s1, ptr %s2) nounwind optsize { -; LA32-LABEL: bcmp_size_1: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 1 -; LA32-NEXT: bl bcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_size_1: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.bu $a0, $a0, 0 +; LA32-UAL-NEXT: ld.bu $a1, $a1, 0 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: sltu $a0, $zero, $a0 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_size_1: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 1 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_size_1: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.bu $a0, $a0, 0 +; LA64-UAL-NEXT: ld.bu $a1, $a1, 0 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_size_1: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 1 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_size_1: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 1 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 1) ret i32 %bcmp } define signext i32 @bcmp_size_2(ptr %s1, ptr %s2) nounwind optsize { -; LA32-LABEL: bcmp_size_2: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 2 -; LA32-NEXT: bl bcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_size_2: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.hu $a0, $a0, 0 +; LA32-UAL-NEXT: ld.hu $a1, $a1, 0 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: sltu $a0, $zero, $a0 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_size_2: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 2 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_size_2: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.hu $a0, $a0, 0 +; LA64-UAL-NEXT: ld.hu $a1, $a1, 0 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_size_2: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 2 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_size_2: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 2 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 2) ret i32 %bcmp } define signext i32 @bcmp_size_3(ptr %s1, ptr %s2) nounwind optsize { -; LA32-LABEL: bcmp_size_3: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 3 -; LA32-NEXT: bl bcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_size_3: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.hu $a2, $a0, 0 +; LA32-UAL-NEXT: ld.hu $a3, $a1, 0 +; LA32-UAL-NEXT: ld.bu $a0, $a0, 2 +; LA32-UAL-NEXT: ld.bu $a1, $a1, 2 +; LA32-UAL-NEXT: xor $a2, $a2, $a3 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: or $a0, $a2, $a0 +; LA32-UAL-NEXT: sltu $a0, $zero, $a0 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_size_3: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 3 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_size_3: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.hu $a2, $a0, 0 +; LA64-UAL-NEXT: ld.hu $a3, $a1, 0 +; LA64-UAL-NEXT: ld.bu $a0, $a0, 2 +; LA64-UAL-NEXT: ld.bu $a1, $a1, 2 +; LA64-UAL-NEXT: xor $a2, $a2, $a3 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: or $a0, $a2, $a0 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_size_3: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 3 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_size_3: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 3 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 3) ret i32 %bcmp } define signext i32 @bcmp_size_4(ptr %s1, ptr %s2) nounwind optsize { -; LA32-LABEL: bcmp_size_4: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 4 -; LA32-NEXT: bl bcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_size_4: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a0, $a0, 0 +; LA32-UAL-NEXT: ld.w $a1, $a1, 0 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: sltu $a0, $zero, $a0 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_size_4: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 4 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_size_4: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.w $a0, $a0, 0 +; LA64-UAL-NEXT: ld.w $a1, $a1, 0 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_size_4: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 4 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_size_4: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 4 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 4) ret i32 %bcmp } define signext i32 @bcmp_size_5(ptr %s1, ptr %s2) nounwind optsize { -; LA32-LABEL: bcmp_size_5: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 5 -; LA32-NEXT: bl bcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_size_5: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a2, $a0, 0 +; LA32-UAL-NEXT: ld.w $a3, $a1, 0 +; LA32-UAL-NEXT: ld.bu $a0, $a0, 4 +; LA32-UAL-NEXT: ld.bu $a1, $a1, 4 +; LA32-UAL-NEXT: xor $a2, $a2, $a3 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: or $a0, $a2, $a0 +; LA32-UAL-NEXT: sltu $a0, $zero, $a0 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_size_5: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 5 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_size_5: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.w $a2, $a0, 0 +; LA64-UAL-NEXT: ld.w $a3, $a1, 0 +; LA64-UAL-NEXT: ld.bu $a0, $a0, 4 +; LA64-UAL-NEXT: ld.bu $a1, $a1, 4 +; LA64-UAL-NEXT: xor $a2, $a2, $a3 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: or $a0, $a2, $a0 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_size_5: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 5 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_size_5: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 5 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 5) ret i32 %bcmp } define signext i32 @bcmp_size_6(ptr %s1, ptr %s2) nounwind optsize { -; LA32-LABEL: bcmp_size_6: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 6 -; LA32-NEXT: bl bcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_size_6: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a2, $a0, 0 +; LA32-UAL-NEXT: ld.w $a3, $a1, 0 +; LA32-UAL-NEXT: ld.hu $a0, $a0, 4 +; LA32-UAL-NEXT: ld.hu $a1, $a1, 4 +; LA32-UAL-NEXT: xor $a2, $a2, $a3 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: or $a0, $a2, $a0 +; LA32-UAL-NEXT: sltu $a0, $zero, $a0 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_size_6: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 6 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_size_6: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.w $a2, $a0, 0 +; LA64-UAL-NEXT: ld.w $a3, $a1, 0 +; LA64-UAL-NEXT: ld.hu $a0, $a0, 4 +; LA64-UAL-NEXT: ld.hu $a1, $a1, 4 +; LA64-UAL-NEXT: xor $a2, $a2, $a3 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: or $a0, $a2, $a0 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_size_6: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 6 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_size_6: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 6 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 6) ret i32 %bcmp } define signext i32 @bcmp_size_7(ptr %s1, ptr %s2) nounwind optsize { -; LA32-LABEL: bcmp_size_7: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 7 -; LA32-NEXT: bl bcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_size_7: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a2, $a0, 0 +; LA32-UAL-NEXT: ld.w $a3, $a1, 0 +; LA32-UAL-NEXT: ld.w $a0, $a0, 3 +; LA32-UAL-NEXT: ld.w $a1, $a1, 3 +; LA32-UAL-NEXT: xor $a2, $a2, $a3 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: or $a0, $a2, $a0 +; LA32-UAL-NEXT: sltu $a0, $zero, $a0 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_size_7: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 7 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_size_7: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.w $a2, $a0, 0 +; LA64-UAL-NEXT: ld.w $a3, $a1, 0 +; LA64-UAL-NEXT: ld.w $a0, $a0, 3 +; LA64-UAL-NEXT: ld.w $a1, $a1, 3 +; LA64-UAL-NEXT: xor $a2, $a2, $a3 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: or $a0, $a2, $a0 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_size_7: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 7 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_size_7: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 7 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 7) ret i32 %bcmp } define signext i32 @bcmp_size_8(ptr %s1, ptr %s2) nounwind optsize { -; LA32-LABEL: bcmp_size_8: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 8 -; LA32-NEXT: bl bcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_size_8: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a2, $a0, 0 +; LA32-UAL-NEXT: ld.w $a3, $a1, 0 +; LA32-UAL-NEXT: ld.w $a0, $a0, 4 +; LA32-UAL-NEXT: ld.w $a1, $a1, 4 +; LA32-UAL-NEXT: xor $a2, $a2, $a3 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: or $a0, $a2, $a0 +; LA32-UAL-NEXT: sltu $a0, $zero, $a0 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_size_8: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 8 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_size_8: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.d $a0, $a0, 0 +; LA64-UAL-NEXT: ld.d $a1, $a1, 0 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_size_8: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 8 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_size_8: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 8 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 8) ret i32 %bcmp } define signext i32 @bcmp_size_15(ptr %s1, ptr %s2) nounwind optsize { -; LA32-LABEL: bcmp_size_15: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 15 -; LA32-NEXT: bl bcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_size_15: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a2, $a0, 0 +; LA32-UAL-NEXT: ld.w $a3, $a1, 0 +; LA32-UAL-NEXT: ld.w $a4, $a0, 4 +; LA32-UAL-NEXT: ld.w $a5, $a1, 4 +; LA32-UAL-NEXT: ld.w $a6, $a0, 8 +; LA32-UAL-NEXT: ld.w $a7, $a1, 8 +; LA32-UAL-NEXT: ld.w $a0, $a0, 11 +; LA32-UAL-NEXT: ld.w $a1, $a1, 11 +; LA32-UAL-NEXT: xor $a2, $a2, $a3 +; LA32-UAL-NEXT: xor $a3, $a4, $a5 +; LA32-UAL-NEXT: xor $a4, $a6, $a7 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: or $a1, $a2, $a3 +; LA32-UAL-NEXT: or $a0, $a4, $a0 +; LA32-UAL-NEXT: or $a0, $a1, $a0 +; LA32-UAL-NEXT: sltu $a0, $zero, $a0 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_size_15: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 15 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_size_15: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.d $a2, $a0, 0 +; LA64-UAL-NEXT: ld.d $a3, $a1, 0 +; LA64-UAL-NEXT: ld.d $a0, $a0, 7 +; LA64-UAL-NEXT: ld.d $a1, $a1, 7 +; LA64-UAL-NEXT: xor $a2, $a2, $a3 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: or $a0, $a2, $a0 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_size_15: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 15 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_size_15: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 15 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 15) ret i32 %bcmp } define signext i32 @bcmp_size_16(ptr %s1, ptr %s2) nounwind optsize { -; LA32-LABEL: bcmp_size_16: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 16 -; LA32-NEXT: bl bcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_size_16: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a2, $a0, 0 +; LA32-UAL-NEXT: ld.w $a3, $a1, 0 +; LA32-UAL-NEXT: ld.w $a4, $a0, 4 +; LA32-UAL-NEXT: ld.w $a5, $a1, 4 +; LA32-UAL-NEXT: ld.w $a6, $a0, 8 +; LA32-UAL-NEXT: ld.w $a7, $a1, 8 +; LA32-UAL-NEXT: ld.w $a0, $a0, 12 +; LA32-UAL-NEXT: ld.w $a1, $a1, 12 +; LA32-UAL-NEXT: xor $a2, $a2, $a3 +; LA32-UAL-NEXT: xor $a3, $a4, $a5 +; LA32-UAL-NEXT: xor $a4, $a6, $a7 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: or $a1, $a2, $a3 +; LA32-UAL-NEXT: or $a0, $a4, $a0 +; LA32-UAL-NEXT: or $a0, $a1, $a0 +; LA32-UAL-NEXT: sltu $a0, $zero, $a0 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_size_16: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 16 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_size_16: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.d $a2, $a0, 0 +; LA64-UAL-NEXT: ld.d $a3, $a1, 0 +; LA64-UAL-NEXT: ld.d $a0, $a0, 8 +; LA64-UAL-NEXT: ld.d $a1, $a1, 8 +; LA64-UAL-NEXT: xor $a2, $a2, $a3 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: or $a0, $a2, $a0 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_size_16: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 16 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_size_16: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 16 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 16) ret i32 %bcmp @@ -308,16 +536,36 @@ define signext i32 @bcmp_size_31(ptr %s1, ptr %s2) nounwind optsize { ; LA32-NEXT: addi.w $sp, $sp, 16 ; LA32-NEXT: ret ; -; LA64-LABEL: bcmp_size_31: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 31 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_size_31: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.d $a2, $a0, 0 +; LA64-UAL-NEXT: ld.d $a3, $a1, 0 +; LA64-UAL-NEXT: ld.d $a4, $a0, 8 +; LA64-UAL-NEXT: ld.d $a5, $a1, 8 +; LA64-UAL-NEXT: ld.d $a6, $a0, 16 +; LA64-UAL-NEXT: ld.d $a7, $a1, 16 +; LA64-UAL-NEXT: ld.d $a0, $a0, 23 +; LA64-UAL-NEXT: ld.d $a1, $a1, 23 +; LA64-UAL-NEXT: xor $a2, $a2, $a3 +; LA64-UAL-NEXT: xor $a3, $a4, $a5 +; LA64-UAL-NEXT: xor $a4, $a6, $a7 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: or $a1, $a2, $a3 +; LA64-UAL-NEXT: or $a0, $a4, $a0 +; LA64-UAL-NEXT: or $a0, $a1, $a0 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_size_31: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 31 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 31) ret i32 %bcmp @@ -334,16 +582,36 @@ define signext i32 @bcmp_size_32(ptr %s1, ptr %s2) nounwind optsize { ; LA32-NEXT: addi.w $sp, $sp, 16 ; LA32-NEXT: ret ; -; LA64-LABEL: bcmp_size_32: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 32 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_size_32: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.d $a2, $a0, 0 +; LA64-UAL-NEXT: ld.d $a3, $a1, 0 +; LA64-UAL-NEXT: ld.d $a4, $a0, 8 +; LA64-UAL-NEXT: ld.d $a5, $a1, 8 +; LA64-UAL-NEXT: ld.d $a6, $a0, 16 +; LA64-UAL-NEXT: ld.d $a7, $a1, 16 +; LA64-UAL-NEXT: ld.d $a0, $a0, 24 +; LA64-UAL-NEXT: ld.d $a1, $a1, 24 +; LA64-UAL-NEXT: xor $a2, $a2, $a3 +; LA64-UAL-NEXT: xor $a3, $a4, $a5 +; LA64-UAL-NEXT: xor $a4, $a6, $a7 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: or $a1, $a2, $a3 +; LA64-UAL-NEXT: or $a0, $a4, $a0 +; LA64-UAL-NEXT: or $a0, $a1, $a0 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_size_32: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 32 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 32) ret i32 %bcmp @@ -478,28 +746,44 @@ entry: } define i1 @bcmp_eq_zero(ptr %s1, ptr %s2) nounwind optsize { -; LA32-LABEL: bcmp_eq_zero: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 4 -; LA32-NEXT: bl bcmp -; LA32-NEXT: sltui $a0, $a0, 1 -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_eq_zero: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a0, $a0, 0 +; LA32-UAL-NEXT: ld.w $a1, $a1, 0 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: sltui $a0, $a0, 1 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_eq_zero: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 4 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: sltui $a0, $a0, 1 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_eq_zero: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.w $a0, $a0, 0 +; LA64-UAL-NEXT: ld.w $a1, $a1, 0 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: sltui $a0, $a0, 1 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_eq_zero: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 4 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: sltui $a0, $a0, 1 +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_eq_zero: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 4 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: sltui $a0, $a0, 1 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 4) %ret = icmp eq i32 %bcmp, 0 @@ -507,28 +791,38 @@ entry: } define i1 @bcmp_lt_zero(ptr %s1, ptr %s2) nounwind optsize { -; LA32-LABEL: bcmp_lt_zero: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 4 -; LA32-NEXT: bl bcmp -; LA32-NEXT: srli.w $a0, $a0, 31 -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_lt_zero: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: move $a0, $zero +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_lt_zero: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 4 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: slti $a0, $a0, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_lt_zero: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: move $a0, $zero +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_lt_zero: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 4 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: srli.w $a0, $a0, 31 +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_lt_zero: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 4 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: slti $a0, $a0, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 4) %ret = icmp slt i32 %bcmp, 0 @@ -536,28 +830,44 @@ entry: } define i1 @bcmp_gt_zero(ptr %s1, ptr %s2) nounwind optsize { -; LA32-LABEL: bcmp_gt_zero: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 4 -; LA32-NEXT: bl bcmp -; LA32-NEXT: slt $a0, $zero, $a0 -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_gt_zero: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a0, $a0, 0 +; LA32-UAL-NEXT: ld.w $a1, $a1, 0 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: sltu $a0, $zero, $a0 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_gt_zero: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 4 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: slt $a0, $zero, $a0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_gt_zero: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.w $a0, $a0, 0 +; LA64-UAL-NEXT: ld.w $a1, $a1, 0 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_gt_zero: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 4 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: slt $a0, $zero, $a0 +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_gt_zero: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 4 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: slt $a0, $zero, $a0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 4) %ret = icmp sgt i32 %bcmp, 0 @@ -575,260 +885,898 @@ entry: } define signext i32 @memcmp_size_1(ptr %s1, ptr %s2) nounwind optsize { -; LA32-LABEL: memcmp_size_1: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 1 -; LA32-NEXT: bl memcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_size_1: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.bu $a0, $a0, 0 +; LA32-UAL-NEXT: ld.bu $a1, $a1, 0 +; LA32-UAL-NEXT: sub.w $a0, $a0, $a1 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_size_1: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 1 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_size_1: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.bu $a0, $a0, 0 +; LA64-UAL-NEXT: ld.bu $a1, $a1, 0 +; LA64-UAL-NEXT: sub.d $a0, $a0, $a1 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_size_1: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 1 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_size_1: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 1 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 1) ret i32 %memcmp } define signext i32 @memcmp_size_2(ptr %s1, ptr %s2) nounwind optsize { -; LA32-LABEL: memcmp_size_2: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 2 -; LA32-NEXT: bl memcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_size_2: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.hu $a0, $a0, 0 +; LA32-UAL-NEXT: ld.hu $a1, $a1, 0 +; LA32-UAL-NEXT: srli.w $a2, $a0, 8 +; LA32-UAL-NEXT: slli.w $a0, $a0, 8 +; LA32-UAL-NEXT: or $a0, $a0, $a2 +; LA32-UAL-NEXT: srli.w $a2, $a1, 8 +; LA32-UAL-NEXT: slli.w $a1, $a1, 8 +; LA32-UAL-NEXT: or $a1, $a1, $a2 +; LA32-UAL-NEXT: lu12i.w $a2, 15 +; LA32-UAL-NEXT: ori $a2, $a2, 4095 +; LA32-UAL-NEXT: and $a0, $a0, $a2 +; LA32-UAL-NEXT: and $a1, $a1, $a2 +; LA32-UAL-NEXT: sub.w $a0, $a0, $a1 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_size_2: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 2 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_size_2: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.h $a0, $a0, 0 +; LA64-UAL-NEXT: ld.h $a1, $a1, 0 +; LA64-UAL-NEXT: revb.2h $a0, $a0 +; LA64-UAL-NEXT: revb.2h $a1, $a1 +; LA64-UAL-NEXT: bstrpick.d $a0, $a0, 15, 0 +; LA64-UAL-NEXT: bstrpick.d $a1, $a1, 15, 0 +; LA64-UAL-NEXT: sub.d $a0, $a0, $a1 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_size_2: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 2 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_size_2: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 2 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 2) ret i32 %memcmp } define signext i32 @memcmp_size_3(ptr %s1, ptr %s2) nounwind optsize { -; LA32-LABEL: memcmp_size_3: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 3 -; LA32-NEXT: bl memcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_size_3: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.bu $a2, $a0, 2 +; LA32-UAL-NEXT: ld.hu $a0, $a0, 0 +; LA32-UAL-NEXT: ld.bu $a3, $a1, 2 +; LA32-UAL-NEXT: ld.hu $a1, $a1, 0 +; LA32-UAL-NEXT: lu12i.w $a4, 15 +; LA32-UAL-NEXT: ori $a4, $a4, 3840 +; LA32-UAL-NEXT: and $a5, $a0, $a4 +; LA32-UAL-NEXT: or $a2, $a5, $a2 +; LA32-UAL-NEXT: slli.w $a2, $a2, 8 +; LA32-UAL-NEXT: slli.w $a0, $a0, 24 +; LA32-UAL-NEXT: or $a0, $a2, $a0 +; LA32-UAL-NEXT: and $a2, $a1, $a4 +; LA32-UAL-NEXT: or $a2, $a2, $a3 +; LA32-UAL-NEXT: slli.w $a2, $a2, 8 +; LA32-UAL-NEXT: slli.w $a1, $a1, 24 +; LA32-UAL-NEXT: or $a1, $a2, $a1 +; LA32-UAL-NEXT: sltu $a2, $a0, $a1 +; LA32-UAL-NEXT: sltu $a0, $a1, $a0 +; LA32-UAL-NEXT: sub.w $a0, $a0, $a2 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_size_3: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 3 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_size_3: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.bu $a2, $a0, 2 +; LA64-UAL-NEXT: ld.hu $a0, $a0, 0 +; LA64-UAL-NEXT: ld.bu $a3, $a1, 2 +; LA64-UAL-NEXT: ld.hu $a1, $a1, 0 +; LA64-UAL-NEXT: slli.d $a2, $a2, 16 +; LA64-UAL-NEXT: or $a0, $a0, $a2 +; LA64-UAL-NEXT: slli.d $a2, $a3, 16 +; LA64-UAL-NEXT: or $a1, $a1, $a2 +; LA64-UAL-NEXT: revb.2w $a0, $a0 +; LA64-UAL-NEXT: addi.w $a0, $a0, 0 +; LA64-UAL-NEXT: revb.2w $a1, $a1 +; LA64-UAL-NEXT: addi.w $a1, $a1, 0 +; LA64-UAL-NEXT: sltu $a2, $a0, $a1 +; LA64-UAL-NEXT: sltu $a0, $a1, $a0 +; LA64-UAL-NEXT: sub.d $a0, $a0, $a2 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_size_3: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 3 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_size_3: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 3 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 3) ret i32 %memcmp } define signext i32 @memcmp_size_4(ptr %s1, ptr %s2) nounwind optsize { -; LA32-LABEL: memcmp_size_4: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 4 -; LA32-NEXT: bl memcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_size_4: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a0, $a0, 0 +; LA32-UAL-NEXT: ld.w $a1, $a1, 0 +; LA32-UAL-NEXT: srli.w $a2, $a0, 8 +; LA32-UAL-NEXT: lu12i.w $a3, 15 +; LA32-UAL-NEXT: ori $a3, $a3, 3840 +; LA32-UAL-NEXT: and $a2, $a2, $a3 +; LA32-UAL-NEXT: srli.w $a4, $a0, 24 +; LA32-UAL-NEXT: or $a2, $a2, $a4 +; LA32-UAL-NEXT: and $a4, $a0, $a3 +; LA32-UAL-NEXT: slli.w $a4, $a4, 8 +; LA32-UAL-NEXT: slli.w $a0, $a0, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: or $a0, $a0, $a2 +; LA32-UAL-NEXT: srli.w $a2, $a1, 8 +; LA32-UAL-NEXT: and $a2, $a2, $a3 +; LA32-UAL-NEXT: srli.w $a4, $a1, 24 +; LA32-UAL-NEXT: or $a2, $a2, $a4 +; LA32-UAL-NEXT: and $a3, $a1, $a3 +; LA32-UAL-NEXT: slli.w $a3, $a3, 8 +; LA32-UAL-NEXT: slli.w $a1, $a1, 24 +; LA32-UAL-NEXT: or $a1, $a1, $a3 +; LA32-UAL-NEXT: or $a1, $a1, $a2 +; LA32-UAL-NEXT: sltu $a2, $a0, $a1 +; LA32-UAL-NEXT: sltu $a0, $a1, $a0 +; LA32-UAL-NEXT: sub.w $a0, $a0, $a2 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_size_4: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 4 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_size_4: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.w $a0, $a0, 0 +; LA64-UAL-NEXT: ld.w $a1, $a1, 0 +; LA64-UAL-NEXT: revb.2w $a0, $a0 +; LA64-UAL-NEXT: addi.w $a0, $a0, 0 +; LA64-UAL-NEXT: revb.2w $a1, $a1 +; LA64-UAL-NEXT: addi.w $a1, $a1, 0 +; LA64-UAL-NEXT: sltu $a2, $a0, $a1 +; LA64-UAL-NEXT: sltu $a0, $a1, $a0 +; LA64-UAL-NEXT: sub.d $a0, $a0, $a2 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_size_4: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 4 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_size_4: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 4 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 4) ret i32 %memcmp } define signext i32 @memcmp_size_5(ptr %s1, ptr %s2) nounwind optsize { -; LA32-LABEL: memcmp_size_5: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 5 -; LA32-NEXT: bl memcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_size_5: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a2, $a0, 0 +; LA32-UAL-NEXT: ld.w $a3, $a1, 0 +; LA32-UAL-NEXT: srli.w $a4, $a2, 8 +; LA32-UAL-NEXT: lu12i.w $a5, 15 +; LA32-UAL-NEXT: ori $a5, $a5, 3840 +; LA32-UAL-NEXT: and $a4, $a4, $a5 +; LA32-UAL-NEXT: srli.w $a6, $a2, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: and $a6, $a2, $a5 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a2, $a2, 24 +; LA32-UAL-NEXT: or $a2, $a2, $a6 +; LA32-UAL-NEXT: or $a2, $a2, $a4 +; LA32-UAL-NEXT: srli.w $a4, $a3, 8 +; LA32-UAL-NEXT: and $a4, $a4, $a5 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: and $a5, $a3, $a5 +; LA32-UAL-NEXT: slli.w $a5, $a5, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: or $a3, $a3, $a4 +; LA32-UAL-NEXT: bne $a2, $a3, .LBB26_2 +; LA32-UAL-NEXT: # %bb.1: # %loadbb1 +; LA32-UAL-NEXT: ld.bu $a0, $a0, 4 +; LA32-UAL-NEXT: ld.bu $a1, $a1, 4 +; LA32-UAL-NEXT: sub.w $a0, $a0, $a1 +; LA32-UAL-NEXT: ret +; LA32-UAL-NEXT: .LBB26_2: # %res_block +; LA32-UAL-NEXT: sltu $a0, $a2, $a3 +; LA32-UAL-NEXT: sub.w $a0, $zero, $a0 +; LA32-UAL-NEXT: ori $a0, $a0, 1 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_size_5: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 5 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_size_5: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.bu $a2, $a0, 4 +; LA64-UAL-NEXT: ld.wu $a0, $a0, 0 +; LA64-UAL-NEXT: ld.bu $a3, $a1, 4 +; LA64-UAL-NEXT: ld.wu $a1, $a1, 0 +; LA64-UAL-NEXT: slli.d $a2, $a2, 32 +; LA64-UAL-NEXT: or $a0, $a0, $a2 +; LA64-UAL-NEXT: slli.d $a2, $a3, 32 +; LA64-UAL-NEXT: or $a1, $a1, $a2 +; LA64-UAL-NEXT: revb.d $a0, $a0 +; LA64-UAL-NEXT: revb.d $a1, $a1 +; LA64-UAL-NEXT: sltu $a2, $a0, $a1 +; LA64-UAL-NEXT: sltu $a0, $a1, $a0 +; LA64-UAL-NEXT: sub.d $a0, $a0, $a2 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_size_5: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 5 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_size_5: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 5 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 5) ret i32 %memcmp } define signext i32 @memcmp_size_6(ptr %s1, ptr %s2) nounwind optsize { -; LA32-LABEL: memcmp_size_6: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 6 -; LA32-NEXT: bl memcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_size_6: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a3, $a0, 0 +; LA32-UAL-NEXT: ld.w $a4, $a1, 0 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: lu12i.w $a2, 15 +; LA32-UAL-NEXT: ori $a6, $a2, 3840 +; LA32-UAL-NEXT: and $a5, $a5, $a6 +; LA32-UAL-NEXT: srli.w $a7, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a7 +; LA32-UAL-NEXT: and $a7, $a3, $a6 +; LA32-UAL-NEXT: slli.w $a7, $a7, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a7 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a6 +; LA32-UAL-NEXT: srli.w $a7, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a7 +; LA32-UAL-NEXT: and $a6, $a4, $a6 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB27_3 +; LA32-UAL-NEXT: # %bb.1: # %loadbb1 +; LA32-UAL-NEXT: ld.hu $a0, $a0, 4 +; LA32-UAL-NEXT: ld.hu $a1, $a1, 4 +; LA32-UAL-NEXT: srli.w $a3, $a0, 8 +; LA32-UAL-NEXT: slli.w $a0, $a0, 8 +; LA32-UAL-NEXT: or $a0, $a0, $a3 +; LA32-UAL-NEXT: srli.w $a3, $a1, 8 +; LA32-UAL-NEXT: slli.w $a1, $a1, 8 +; LA32-UAL-NEXT: or $a1, $a1, $a3 +; LA32-UAL-NEXT: ori $a2, $a2, 4095 +; LA32-UAL-NEXT: and $a3, $a0, $a2 +; LA32-UAL-NEXT: and $a4, $a1, $a2 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB27_3 +; LA32-UAL-NEXT: # %bb.2: +; LA32-UAL-NEXT: move $a0, $zero +; LA32-UAL-NEXT: ret +; LA32-UAL-NEXT: .LBB27_3: # %res_block +; LA32-UAL-NEXT: sltu $a0, $a3, $a4 +; LA32-UAL-NEXT: sub.w $a0, $zero, $a0 +; LA32-UAL-NEXT: ori $a0, $a0, 1 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_size_6: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 6 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_size_6: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.hu $a2, $a0, 4 +; LA64-UAL-NEXT: ld.wu $a0, $a0, 0 +; LA64-UAL-NEXT: ld.hu $a3, $a1, 4 +; LA64-UAL-NEXT: ld.wu $a1, $a1, 0 +; LA64-UAL-NEXT: slli.d $a2, $a2, 32 +; LA64-UAL-NEXT: or $a0, $a0, $a2 +; LA64-UAL-NEXT: slli.d $a2, $a3, 32 +; LA64-UAL-NEXT: or $a1, $a1, $a2 +; LA64-UAL-NEXT: revb.d $a0, $a0 +; LA64-UAL-NEXT: revb.d $a1, $a1 +; LA64-UAL-NEXT: sltu $a2, $a0, $a1 +; LA64-UAL-NEXT: sltu $a0, $a1, $a0 +; LA64-UAL-NEXT: sub.d $a0, $a0, $a2 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_size_6: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 6 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_size_6: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 6 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 6) ret i32 %memcmp } define signext i32 @memcmp_size_7(ptr %s1, ptr %s2) nounwind optsize { -; LA32-LABEL: memcmp_size_7: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 7 -; LA32-NEXT: bl memcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_size_7: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a3, $a0, 0 +; LA32-UAL-NEXT: ld.w $a4, $a1, 0 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: lu12i.w $a2, 15 +; LA32-UAL-NEXT: ori $a2, $a2, 3840 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB28_3 +; LA32-UAL-NEXT: # %bb.1: # %loadbb1 +; LA32-UAL-NEXT: ld.w $a0, $a0, 3 +; LA32-UAL-NEXT: ld.w $a1, $a1, 3 +; LA32-UAL-NEXT: srli.w $a3, $a0, 8 +; LA32-UAL-NEXT: and $a3, $a3, $a2 +; LA32-UAL-NEXT: srli.w $a4, $a0, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a4 +; LA32-UAL-NEXT: and $a4, $a0, $a2 +; LA32-UAL-NEXT: slli.w $a4, $a4, 8 +; LA32-UAL-NEXT: slli.w $a0, $a0, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: or $a3, $a0, $a3 +; LA32-UAL-NEXT: srli.w $a0, $a1, 8 +; LA32-UAL-NEXT: and $a0, $a0, $a2 +; LA32-UAL-NEXT: srli.w $a4, $a1, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: and $a2, $a1, $a2 +; LA32-UAL-NEXT: slli.w $a2, $a2, 8 +; LA32-UAL-NEXT: slli.w $a1, $a1, 24 +; LA32-UAL-NEXT: or $a1, $a1, $a2 +; LA32-UAL-NEXT: or $a4, $a1, $a0 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB28_3 +; LA32-UAL-NEXT: # %bb.2: +; LA32-UAL-NEXT: move $a0, $zero +; LA32-UAL-NEXT: ret +; LA32-UAL-NEXT: .LBB28_3: # %res_block +; LA32-UAL-NEXT: sltu $a0, $a3, $a4 +; LA32-UAL-NEXT: sub.w $a0, $zero, $a0 +; LA32-UAL-NEXT: ori $a0, $a0, 1 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_size_7: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 7 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_size_7: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.w $a2, $a0, 0 +; LA64-UAL-NEXT: ld.w $a3, $a1, 0 +; LA64-UAL-NEXT: revb.2w $a2, $a2 +; LA64-UAL-NEXT: addi.w $a4, $a2, 0 +; LA64-UAL-NEXT: revb.2w $a3, $a3 +; LA64-UAL-NEXT: addi.w $a5, $a3, 0 +; LA64-UAL-NEXT: bne $a4, $a5, .LBB28_3 +; LA64-UAL-NEXT: # %bb.1: # %loadbb1 +; LA64-UAL-NEXT: ld.w $a0, $a0, 3 +; LA64-UAL-NEXT: ld.w $a1, $a1, 3 +; LA64-UAL-NEXT: revb.2w $a2, $a0 +; LA64-UAL-NEXT: addi.w $a0, $a2, 0 +; LA64-UAL-NEXT: revb.2w $a3, $a1 +; LA64-UAL-NEXT: addi.w $a1, $a3, 0 +; LA64-UAL-NEXT: bne $a0, $a1, .LBB28_3 +; LA64-UAL-NEXT: # %bb.2: +; LA64-UAL-NEXT: move $a0, $zero +; LA64-UAL-NEXT: ret +; LA64-UAL-NEXT: .LBB28_3: # %res_block +; LA64-UAL-NEXT: addi.w $a0, $a3, 0 +; LA64-UAL-NEXT: addi.w $a1, $a2, 0 +; LA64-UAL-NEXT: sltu $a0, $a1, $a0 +; LA64-UAL-NEXT: sub.d $a0, $zero, $a0 +; LA64-UAL-NEXT: ori $a0, $a0, 1 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_size_7: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 7 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_size_7: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 7 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 7) ret i32 %memcmp } define signext i32 @memcmp_size_8(ptr %s1, ptr %s2) nounwind optsize { -; LA32-LABEL: memcmp_size_8: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 8 -; LA32-NEXT: bl memcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_size_8: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a3, $a0, 0 +; LA32-UAL-NEXT: ld.w $a4, $a1, 0 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: lu12i.w $a2, 15 +; LA32-UAL-NEXT: ori $a2, $a2, 3840 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB29_3 +; LA32-UAL-NEXT: # %bb.1: # %loadbb1 +; LA32-UAL-NEXT: ld.w $a0, $a0, 4 +; LA32-UAL-NEXT: ld.w $a1, $a1, 4 +; LA32-UAL-NEXT: srli.w $a3, $a0, 8 +; LA32-UAL-NEXT: and $a3, $a3, $a2 +; LA32-UAL-NEXT: srli.w $a4, $a0, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a4 +; LA32-UAL-NEXT: and $a4, $a0, $a2 +; LA32-UAL-NEXT: slli.w $a4, $a4, 8 +; LA32-UAL-NEXT: slli.w $a0, $a0, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: or $a3, $a0, $a3 +; LA32-UAL-NEXT: srli.w $a0, $a1, 8 +; LA32-UAL-NEXT: and $a0, $a0, $a2 +; LA32-UAL-NEXT: srli.w $a4, $a1, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: and $a2, $a1, $a2 +; LA32-UAL-NEXT: slli.w $a2, $a2, 8 +; LA32-UAL-NEXT: slli.w $a1, $a1, 24 +; LA32-UAL-NEXT: or $a1, $a1, $a2 +; LA32-UAL-NEXT: or $a4, $a1, $a0 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB29_3 +; LA32-UAL-NEXT: # %bb.2: +; LA32-UAL-NEXT: move $a0, $zero +; LA32-UAL-NEXT: ret +; LA32-UAL-NEXT: .LBB29_3: # %res_block +; LA32-UAL-NEXT: sltu $a0, $a3, $a4 +; LA32-UAL-NEXT: sub.w $a0, $zero, $a0 +; LA32-UAL-NEXT: ori $a0, $a0, 1 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_size_8: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 8 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_size_8: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.d $a0, $a0, 0 +; LA64-UAL-NEXT: ld.d $a1, $a1, 0 +; LA64-UAL-NEXT: revb.d $a0, $a0 +; LA64-UAL-NEXT: revb.d $a1, $a1 +; LA64-UAL-NEXT: sltu $a2, $a0, $a1 +; LA64-UAL-NEXT: sltu $a0, $a1, $a0 +; LA64-UAL-NEXT: sub.d $a0, $a0, $a2 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_size_8: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 8 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_size_8: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 8 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 8) ret i32 %memcmp } define signext i32 @memcmp_size_15(ptr %s1, ptr %s2) nounwind optsize { -; LA32-LABEL: memcmp_size_15: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 15 -; LA32-NEXT: bl memcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_size_15: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a3, $a0, 0 +; LA32-UAL-NEXT: ld.w $a4, $a1, 0 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: lu12i.w $a2, 15 +; LA32-UAL-NEXT: ori $a2, $a2, 3840 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB30_5 +; LA32-UAL-NEXT: # %bb.1: # %loadbb1 +; LA32-UAL-NEXT: ld.w $a3, $a0, 4 +; LA32-UAL-NEXT: ld.w $a4, $a1, 4 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB30_5 +; LA32-UAL-NEXT: # %bb.2: # %loadbb2 +; LA32-UAL-NEXT: ld.w $a3, $a0, 8 +; LA32-UAL-NEXT: ld.w $a4, $a1, 8 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB30_5 +; LA32-UAL-NEXT: # %bb.3: # %loadbb3 +; LA32-UAL-NEXT: ld.w $a0, $a0, 11 +; LA32-UAL-NEXT: ld.w $a1, $a1, 11 +; LA32-UAL-NEXT: srli.w $a3, $a0, 8 +; LA32-UAL-NEXT: and $a3, $a3, $a2 +; LA32-UAL-NEXT: srli.w $a4, $a0, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a4 +; LA32-UAL-NEXT: and $a4, $a0, $a2 +; LA32-UAL-NEXT: slli.w $a4, $a4, 8 +; LA32-UAL-NEXT: slli.w $a0, $a0, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: or $a3, $a0, $a3 +; LA32-UAL-NEXT: srli.w $a0, $a1, 8 +; LA32-UAL-NEXT: and $a0, $a0, $a2 +; LA32-UAL-NEXT: srli.w $a4, $a1, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: and $a2, $a1, $a2 +; LA32-UAL-NEXT: slli.w $a2, $a2, 8 +; LA32-UAL-NEXT: slli.w $a1, $a1, 24 +; LA32-UAL-NEXT: or $a1, $a1, $a2 +; LA32-UAL-NEXT: or $a4, $a1, $a0 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB30_5 +; LA32-UAL-NEXT: # %bb.4: +; LA32-UAL-NEXT: move $a0, $zero +; LA32-UAL-NEXT: ret +; LA32-UAL-NEXT: .LBB30_5: # %res_block +; LA32-UAL-NEXT: sltu $a0, $a3, $a4 +; LA32-UAL-NEXT: sub.w $a0, $zero, $a0 +; LA32-UAL-NEXT: ori $a0, $a0, 1 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_size_15: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 15 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_size_15: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.d $a2, $a0, 0 +; LA64-UAL-NEXT: ld.d $a3, $a1, 0 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB30_3 +; LA64-UAL-NEXT: # %bb.1: # %loadbb1 +; LA64-UAL-NEXT: ld.d $a0, $a0, 7 +; LA64-UAL-NEXT: ld.d $a1, $a1, 7 +; LA64-UAL-NEXT: revb.d $a2, $a0 +; LA64-UAL-NEXT: revb.d $a3, $a1 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB30_3 +; LA64-UAL-NEXT: # %bb.2: +; LA64-UAL-NEXT: move $a0, $zero +; LA64-UAL-NEXT: ret +; LA64-UAL-NEXT: .LBB30_3: # %res_block +; LA64-UAL-NEXT: sltu $a0, $a2, $a3 +; LA64-UAL-NEXT: sub.d $a0, $zero, $a0 +; LA64-UAL-NEXT: ori $a0, $a0, 1 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_size_15: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 15 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_size_15: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 15 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 15) ret i32 %memcmp } define signext i32 @memcmp_size_16(ptr %s1, ptr %s2) nounwind optsize { -; LA32-LABEL: memcmp_size_16: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 16 -; LA32-NEXT: bl memcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_size_16: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a3, $a0, 0 +; LA32-UAL-NEXT: ld.w $a4, $a1, 0 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: lu12i.w $a2, 15 +; LA32-UAL-NEXT: ori $a2, $a2, 3840 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB31_5 +; LA32-UAL-NEXT: # %bb.1: # %loadbb1 +; LA32-UAL-NEXT: ld.w $a3, $a0, 4 +; LA32-UAL-NEXT: ld.w $a4, $a1, 4 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB31_5 +; LA32-UAL-NEXT: # %bb.2: # %loadbb2 +; LA32-UAL-NEXT: ld.w $a3, $a0, 8 +; LA32-UAL-NEXT: ld.w $a4, $a1, 8 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB31_5 +; LA32-UAL-NEXT: # %bb.3: # %loadbb3 +; LA32-UAL-NEXT: ld.w $a0, $a0, 12 +; LA32-UAL-NEXT: ld.w $a1, $a1, 12 +; LA32-UAL-NEXT: srli.w $a3, $a0, 8 +; LA32-UAL-NEXT: and $a3, $a3, $a2 +; LA32-UAL-NEXT: srli.w $a4, $a0, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a4 +; LA32-UAL-NEXT: and $a4, $a0, $a2 +; LA32-UAL-NEXT: slli.w $a4, $a4, 8 +; LA32-UAL-NEXT: slli.w $a0, $a0, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: or $a3, $a0, $a3 +; LA32-UAL-NEXT: srli.w $a0, $a1, 8 +; LA32-UAL-NEXT: and $a0, $a0, $a2 +; LA32-UAL-NEXT: srli.w $a4, $a1, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: and $a2, $a1, $a2 +; LA32-UAL-NEXT: slli.w $a2, $a2, 8 +; LA32-UAL-NEXT: slli.w $a1, $a1, 24 +; LA32-UAL-NEXT: or $a1, $a1, $a2 +; LA32-UAL-NEXT: or $a4, $a1, $a0 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB31_5 +; LA32-UAL-NEXT: # %bb.4: +; LA32-UAL-NEXT: move $a0, $zero +; LA32-UAL-NEXT: ret +; LA32-UAL-NEXT: .LBB31_5: # %res_block +; LA32-UAL-NEXT: sltu $a0, $a3, $a4 +; LA32-UAL-NEXT: sub.w $a0, $zero, $a0 +; LA32-UAL-NEXT: ori $a0, $a0, 1 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_size_16: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 16 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_size_16: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.d $a2, $a0, 0 +; LA64-UAL-NEXT: ld.d $a3, $a1, 0 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB31_3 +; LA64-UAL-NEXT: # %bb.1: # %loadbb1 +; LA64-UAL-NEXT: ld.d $a0, $a0, 8 +; LA64-UAL-NEXT: ld.d $a1, $a1, 8 +; LA64-UAL-NEXT: revb.d $a2, $a0 +; LA64-UAL-NEXT: revb.d $a3, $a1 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB31_3 +; LA64-UAL-NEXT: # %bb.2: +; LA64-UAL-NEXT: move $a0, $zero +; LA64-UAL-NEXT: ret +; LA64-UAL-NEXT: .LBB31_3: # %res_block +; LA64-UAL-NEXT: sltu $a0, $a2, $a3 +; LA64-UAL-NEXT: sub.d $a0, $zero, $a0 +; LA64-UAL-NEXT: ori $a0, $a0, 1 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_size_16: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 16 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_size_16: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 16 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 16) ret i32 %memcmp @@ -845,16 +1793,50 @@ define signext i32 @memcmp_size_31(ptr %s1, ptr %s2) nounwind optsize { ; LA32-NEXT: addi.w $sp, $sp, 16 ; LA32-NEXT: ret ; -; LA64-LABEL: memcmp_size_31: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 31 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_size_31: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.d $a2, $a0, 0 +; LA64-UAL-NEXT: ld.d $a3, $a1, 0 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB32_5 +; LA64-UAL-NEXT: # %bb.1: # %loadbb1 +; LA64-UAL-NEXT: ld.d $a2, $a0, 8 +; LA64-UAL-NEXT: ld.d $a3, $a1, 8 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB32_5 +; LA64-UAL-NEXT: # %bb.2: # %loadbb2 +; LA64-UAL-NEXT: ld.d $a2, $a0, 16 +; LA64-UAL-NEXT: ld.d $a3, $a1, 16 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB32_5 +; LA64-UAL-NEXT: # %bb.3: # %loadbb3 +; LA64-UAL-NEXT: ld.d $a0, $a0, 23 +; LA64-UAL-NEXT: ld.d $a1, $a1, 23 +; LA64-UAL-NEXT: revb.d $a2, $a0 +; LA64-UAL-NEXT: revb.d $a3, $a1 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB32_5 +; LA64-UAL-NEXT: # %bb.4: +; LA64-UAL-NEXT: move $a0, $zero +; LA64-UAL-NEXT: ret +; LA64-UAL-NEXT: .LBB32_5: # %res_block +; LA64-UAL-NEXT: sltu $a0, $a2, $a3 +; LA64-UAL-NEXT: sub.d $a0, $zero, $a0 +; LA64-UAL-NEXT: ori $a0, $a0, 1 +; LA64-UAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_size_31: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 31 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 31) ret i32 %memcmp @@ -871,16 +1853,50 @@ define signext i32 @memcmp_size_32(ptr %s1, ptr %s2) nounwind optsize { ; LA32-NEXT: addi.w $sp, $sp, 16 ; LA32-NEXT: ret ; -; LA64-LABEL: memcmp_size_32: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 32 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_size_32: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.d $a2, $a0, 0 +; LA64-UAL-NEXT: ld.d $a3, $a1, 0 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB33_5 +; LA64-UAL-NEXT: # %bb.1: # %loadbb1 +; LA64-UAL-NEXT: ld.d $a2, $a0, 8 +; LA64-UAL-NEXT: ld.d $a3, $a1, 8 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB33_5 +; LA64-UAL-NEXT: # %bb.2: # %loadbb2 +; LA64-UAL-NEXT: ld.d $a2, $a0, 16 +; LA64-UAL-NEXT: ld.d $a3, $a1, 16 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB33_5 +; LA64-UAL-NEXT: # %bb.3: # %loadbb3 +; LA64-UAL-NEXT: ld.d $a0, $a0, 24 +; LA64-UAL-NEXT: ld.d $a1, $a1, 24 +; LA64-UAL-NEXT: revb.d $a2, $a0 +; LA64-UAL-NEXT: revb.d $a3, $a1 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB33_5 +; LA64-UAL-NEXT: # %bb.4: +; LA64-UAL-NEXT: move $a0, $zero +; LA64-UAL-NEXT: ret +; LA64-UAL-NEXT: .LBB33_5: # %res_block +; LA64-UAL-NEXT: sltu $a0, $a2, $a3 +; LA64-UAL-NEXT: sub.d $a0, $zero, $a0 +; LA64-UAL-NEXT: ori $a0, $a0, 1 +; LA64-UAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_size_32: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 32 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 32) ret i32 %memcmp @@ -1017,16 +2033,16 @@ entry: define i1 @memcmp_eq_zero(ptr %s1, ptr %s2) nounwind optsize { ; LA32-UAL-LABEL: memcmp_eq_zero: ; LA32-UAL: # %bb.0: # %entry -; LA32-UAL-NEXT: ld.w $a1, $a1, 0 ; LA32-UAL-NEXT: ld.w $a0, $a0, 0 +; LA32-UAL-NEXT: ld.w $a1, $a1, 0 ; LA32-UAL-NEXT: xor $a0, $a0, $a1 ; LA32-UAL-NEXT: sltui $a0, $a0, 1 ; LA32-UAL-NEXT: ret ; ; LA64-UAL-LABEL: memcmp_eq_zero: ; LA64-UAL: # %bb.0: # %entry -; LA64-UAL-NEXT: ld.w $a1, $a1, 0 ; LA64-UAL-NEXT: ld.w $a0, $a0, 0 +; LA64-UAL-NEXT: ld.w $a1, $a1, 0 ; LA64-UAL-NEXT: xor $a0, $a0, $a1 ; LA64-UAL-NEXT: sltui $a0, $a0, 1 ; LA64-UAL-NEXT: ret @@ -1089,28 +2105,66 @@ entry: } define i1 @memcmp_lt_zero(ptr %s1, ptr %s2) nounwind optsize { -; LA32-LABEL: memcmp_lt_zero: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 4 -; LA32-NEXT: bl memcmp -; LA32-NEXT: srli.w $a0, $a0, 31 -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_lt_zero: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a0, $a0, 0 +; LA32-UAL-NEXT: ld.w $a1, $a1, 0 +; LA32-UAL-NEXT: srli.w $a2, $a0, 8 +; LA32-UAL-NEXT: lu12i.w $a3, 15 +; LA32-UAL-NEXT: ori $a3, $a3, 3840 +; LA32-UAL-NEXT: and $a2, $a2, $a3 +; LA32-UAL-NEXT: srli.w $a4, $a0, 24 +; LA32-UAL-NEXT: or $a2, $a2, $a4 +; LA32-UAL-NEXT: and $a4, $a0, $a3 +; LA32-UAL-NEXT: slli.w $a4, $a4, 8 +; LA32-UAL-NEXT: slli.w $a0, $a0, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: or $a0, $a0, $a2 +; LA32-UAL-NEXT: srli.w $a2, $a1, 8 +; LA32-UAL-NEXT: and $a2, $a2, $a3 +; LA32-UAL-NEXT: srli.w $a4, $a1, 24 +; LA32-UAL-NEXT: or $a2, $a2, $a4 +; LA32-UAL-NEXT: and $a3, $a1, $a3 +; LA32-UAL-NEXT: slli.w $a3, $a3, 8 +; LA32-UAL-NEXT: slli.w $a1, $a1, 24 +; LA32-UAL-NEXT: or $a1, $a1, $a3 +; LA32-UAL-NEXT: or $a1, $a1, $a2 +; LA32-UAL-NEXT: sltu $a0, $a0, $a1 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_lt_zero: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 4 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: slti $a0, $a0, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_lt_zero: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.w $a0, $a0, 0 +; LA64-UAL-NEXT: ld.w $a1, $a1, 0 +; LA64-UAL-NEXT: revb.2w $a0, $a0 +; LA64-UAL-NEXT: addi.w $a0, $a0, 0 +; LA64-UAL-NEXT: revb.2w $a1, $a1 +; LA64-UAL-NEXT: addi.w $a1, $a1, 0 +; LA64-UAL-NEXT: sltu $a0, $a0, $a1 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_lt_zero: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 4 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: srli.w $a0, $a0, 31 +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_lt_zero: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 4 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: slti $a0, $a0, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 4) %ret = icmp slt i32 %memcmp, 0 @@ -1118,28 +2172,66 @@ entry: } define i1 @memcmp_gt_zero(ptr %s1, ptr %s2) nounwind optsize { -; LA32-LABEL: memcmp_gt_zero: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 4 -; LA32-NEXT: bl memcmp -; LA32-NEXT: slt $a0, $zero, $a0 -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_gt_zero: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a0, $a0, 0 +; LA32-UAL-NEXT: ld.w $a1, $a1, 0 +; LA32-UAL-NEXT: srli.w $a2, $a0, 8 +; LA32-UAL-NEXT: lu12i.w $a3, 15 +; LA32-UAL-NEXT: ori $a3, $a3, 3840 +; LA32-UAL-NEXT: and $a2, $a2, $a3 +; LA32-UAL-NEXT: srli.w $a4, $a0, 24 +; LA32-UAL-NEXT: or $a2, $a2, $a4 +; LA32-UAL-NEXT: and $a4, $a0, $a3 +; LA32-UAL-NEXT: slli.w $a4, $a4, 8 +; LA32-UAL-NEXT: slli.w $a0, $a0, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: or $a0, $a0, $a2 +; LA32-UAL-NEXT: srli.w $a2, $a1, 8 +; LA32-UAL-NEXT: and $a2, $a2, $a3 +; LA32-UAL-NEXT: srli.w $a4, $a1, 24 +; LA32-UAL-NEXT: or $a2, $a2, $a4 +; LA32-UAL-NEXT: and $a3, $a1, $a3 +; LA32-UAL-NEXT: slli.w $a3, $a3, 8 +; LA32-UAL-NEXT: slli.w $a1, $a1, 24 +; LA32-UAL-NEXT: or $a1, $a1, $a3 +; LA32-UAL-NEXT: or $a1, $a1, $a2 +; LA32-UAL-NEXT: sltu $a0, $a1, $a0 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_gt_zero: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 4 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: slt $a0, $zero, $a0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_gt_zero: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.w $a0, $a0, 0 +; LA64-UAL-NEXT: ld.w $a1, $a1, 0 +; LA64-UAL-NEXT: revb.2w $a0, $a0 +; LA64-UAL-NEXT: addi.w $a0, $a0, 0 +; LA64-UAL-NEXT: revb.2w $a1, $a1 +; LA64-UAL-NEXT: addi.w $a1, $a1, 0 +; LA64-UAL-NEXT: sltu $a0, $a1, $a0 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_gt_zero: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 4 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: slt $a0, $zero, $a0 +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_gt_zero: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 4 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: slt $a0, $zero, $a0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 4) %ret = icmp sgt i32 %memcmp, 0 diff --git a/llvm/test/CodeGen/LoongArch/expandmemcmp.ll b/llvm/test/CodeGen/LoongArch/expandmemcmp.ll index 8b8f32e6877cc..c1bf850baa8c3 100644 --- a/llvm/test/CodeGen/LoongArch/expandmemcmp.ll +++ b/llvm/test/CodeGen/LoongArch/expandmemcmp.ll @@ -38,312 +38,652 @@ entry: } define signext i32 @bcmp_size_1(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: bcmp_size_1: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 1 -; LA32-NEXT: bl bcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_size_1: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.bu $a0, $a0, 0 +; LA32-UAL-NEXT: ld.bu $a1, $a1, 0 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: sltu $a0, $zero, $a0 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_size_1: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 1 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_size_1: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.bu $a0, $a0, 0 +; LA64-UAL-NEXT: ld.bu $a1, $a1, 0 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_size_1: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 1 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_size_1: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 1 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 1) ret i32 %bcmp } define signext i32 @bcmp_size_2(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: bcmp_size_2: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 2 -; LA32-NEXT: bl bcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_size_2: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.hu $a0, $a0, 0 +; LA32-UAL-NEXT: ld.hu $a1, $a1, 0 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: sltu $a0, $zero, $a0 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_size_2: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 2 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_size_2: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.hu $a0, $a0, 0 +; LA64-UAL-NEXT: ld.hu $a1, $a1, 0 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_size_2: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 2 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_size_2: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 2 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 2) ret i32 %bcmp } define signext i32 @bcmp_size_3(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: bcmp_size_3: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 3 -; LA32-NEXT: bl bcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_size_3: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.hu $a2, $a0, 0 +; LA32-UAL-NEXT: ld.hu $a3, $a1, 0 +; LA32-UAL-NEXT: ld.bu $a0, $a0, 2 +; LA32-UAL-NEXT: ld.bu $a1, $a1, 2 +; LA32-UAL-NEXT: xor $a2, $a2, $a3 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: or $a0, $a2, $a0 +; LA32-UAL-NEXT: sltu $a0, $zero, $a0 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_size_3: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 3 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_size_3: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.hu $a2, $a0, 0 +; LA64-UAL-NEXT: ld.hu $a3, $a1, 0 +; LA64-UAL-NEXT: ld.bu $a0, $a0, 2 +; LA64-UAL-NEXT: ld.bu $a1, $a1, 2 +; LA64-UAL-NEXT: xor $a2, $a2, $a3 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: or $a0, $a2, $a0 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_size_3: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 3 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_size_3: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 3 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 3) ret i32 %bcmp } define signext i32 @bcmp_size_4(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: bcmp_size_4: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 4 -; LA32-NEXT: bl bcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_size_4: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a0, $a0, 0 +; LA32-UAL-NEXT: ld.w $a1, $a1, 0 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: sltu $a0, $zero, $a0 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_size_4: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 4 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_size_4: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.w $a0, $a0, 0 +; LA64-UAL-NEXT: ld.w $a1, $a1, 0 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_size_4: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 4 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_size_4: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 4 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 4) ret i32 %bcmp } define signext i32 @bcmp_size_5(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: bcmp_size_5: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 5 -; LA32-NEXT: bl bcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_size_5: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a2, $a0, 0 +; LA32-UAL-NEXT: ld.w $a3, $a1, 0 +; LA32-UAL-NEXT: ld.bu $a0, $a0, 4 +; LA32-UAL-NEXT: ld.bu $a1, $a1, 4 +; LA32-UAL-NEXT: xor $a2, $a2, $a3 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: or $a0, $a2, $a0 +; LA32-UAL-NEXT: sltu $a0, $zero, $a0 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_size_5: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 5 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_size_5: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.w $a2, $a0, 0 +; LA64-UAL-NEXT: ld.w $a3, $a1, 0 +; LA64-UAL-NEXT: ld.bu $a0, $a0, 4 +; LA64-UAL-NEXT: ld.bu $a1, $a1, 4 +; LA64-UAL-NEXT: xor $a2, $a2, $a3 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: or $a0, $a2, $a0 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_size_5: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 5 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_size_5: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 5 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 5) ret i32 %bcmp } define signext i32 @bcmp_size_6(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: bcmp_size_6: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 6 -; LA32-NEXT: bl bcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_size_6: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a2, $a0, 0 +; LA32-UAL-NEXT: ld.w $a3, $a1, 0 +; LA32-UAL-NEXT: ld.hu $a0, $a0, 4 +; LA32-UAL-NEXT: ld.hu $a1, $a1, 4 +; LA32-UAL-NEXT: xor $a2, $a2, $a3 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: or $a0, $a2, $a0 +; LA32-UAL-NEXT: sltu $a0, $zero, $a0 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_size_6: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 6 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_size_6: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.w $a2, $a0, 0 +; LA64-UAL-NEXT: ld.w $a3, $a1, 0 +; LA64-UAL-NEXT: ld.hu $a0, $a0, 4 +; LA64-UAL-NEXT: ld.hu $a1, $a1, 4 +; LA64-UAL-NEXT: xor $a2, $a2, $a3 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: or $a0, $a2, $a0 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_size_6: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 6 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_size_6: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 6 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 6) ret i32 %bcmp } define signext i32 @bcmp_size_7(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: bcmp_size_7: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 7 -; LA32-NEXT: bl bcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_size_7: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a2, $a0, 0 +; LA32-UAL-NEXT: ld.w $a3, $a1, 0 +; LA32-UAL-NEXT: ld.w $a0, $a0, 3 +; LA32-UAL-NEXT: ld.w $a1, $a1, 3 +; LA32-UAL-NEXT: xor $a2, $a2, $a3 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: or $a0, $a2, $a0 +; LA32-UAL-NEXT: sltu $a0, $zero, $a0 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_size_7: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 7 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_size_7: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.w $a2, $a0, 0 +; LA64-UAL-NEXT: ld.w $a3, $a1, 0 +; LA64-UAL-NEXT: ld.w $a0, $a0, 3 +; LA64-UAL-NEXT: ld.w $a1, $a1, 3 +; LA64-UAL-NEXT: xor $a2, $a2, $a3 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: or $a0, $a2, $a0 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_size_7: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 7 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_size_7: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 7 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 7) ret i32 %bcmp } define signext i32 @bcmp_size_8(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: bcmp_size_8: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 8 -; LA32-NEXT: bl bcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_size_8: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a2, $a0, 0 +; LA32-UAL-NEXT: ld.w $a3, $a1, 0 +; LA32-UAL-NEXT: ld.w $a0, $a0, 4 +; LA32-UAL-NEXT: ld.w $a1, $a1, 4 +; LA32-UAL-NEXT: xor $a2, $a2, $a3 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: or $a0, $a2, $a0 +; LA32-UAL-NEXT: sltu $a0, $zero, $a0 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_size_8: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 8 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_size_8: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.d $a0, $a0, 0 +; LA64-UAL-NEXT: ld.d $a1, $a1, 0 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_size_8: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 8 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_size_8: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 8 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 8) ret i32 %bcmp } define signext i32 @bcmp_size_15(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: bcmp_size_15: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 15 -; LA32-NEXT: bl bcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_size_15: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a2, $a0, 0 +; LA32-UAL-NEXT: ld.w $a3, $a1, 0 +; LA32-UAL-NEXT: ld.w $a4, $a0, 4 +; LA32-UAL-NEXT: ld.w $a5, $a1, 4 +; LA32-UAL-NEXT: ld.w $a6, $a0, 8 +; LA32-UAL-NEXT: ld.w $a7, $a1, 8 +; LA32-UAL-NEXT: ld.w $a0, $a0, 11 +; LA32-UAL-NEXT: ld.w $a1, $a1, 11 +; LA32-UAL-NEXT: xor $a2, $a2, $a3 +; LA32-UAL-NEXT: xor $a3, $a4, $a5 +; LA32-UAL-NEXT: xor $a4, $a6, $a7 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: or $a1, $a2, $a3 +; LA32-UAL-NEXT: or $a0, $a4, $a0 +; LA32-UAL-NEXT: or $a0, $a1, $a0 +; LA32-UAL-NEXT: sltu $a0, $zero, $a0 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_size_15: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 15 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_size_15: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.d $a2, $a0, 0 +; LA64-UAL-NEXT: ld.d $a3, $a1, 0 +; LA64-UAL-NEXT: ld.d $a0, $a0, 7 +; LA64-UAL-NEXT: ld.d $a1, $a1, 7 +; LA64-UAL-NEXT: xor $a2, $a2, $a3 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: or $a0, $a2, $a0 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_size_15: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 15 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_size_15: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 15 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 15) ret i32 %bcmp } define signext i32 @bcmp_size_16(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: bcmp_size_16: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 16 -; LA32-NEXT: bl bcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_size_16: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a2, $a0, 0 +; LA32-UAL-NEXT: ld.w $a3, $a1, 0 +; LA32-UAL-NEXT: ld.w $a4, $a0, 4 +; LA32-UAL-NEXT: ld.w $a5, $a1, 4 +; LA32-UAL-NEXT: ld.w $a6, $a0, 8 +; LA32-UAL-NEXT: ld.w $a7, $a1, 8 +; LA32-UAL-NEXT: ld.w $a0, $a0, 12 +; LA32-UAL-NEXT: ld.w $a1, $a1, 12 +; LA32-UAL-NEXT: xor $a2, $a2, $a3 +; LA32-UAL-NEXT: xor $a3, $a4, $a5 +; LA32-UAL-NEXT: xor $a4, $a6, $a7 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: or $a1, $a2, $a3 +; LA32-UAL-NEXT: or $a0, $a4, $a0 +; LA32-UAL-NEXT: or $a0, $a1, $a0 +; LA32-UAL-NEXT: sltu $a0, $zero, $a0 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_size_16: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 16 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_size_16: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.d $a2, $a0, 0 +; LA64-UAL-NEXT: ld.d $a3, $a1, 0 +; LA64-UAL-NEXT: ld.d $a0, $a0, 8 +; LA64-UAL-NEXT: ld.d $a1, $a1, 8 +; LA64-UAL-NEXT: xor $a2, $a2, $a3 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: or $a0, $a2, $a0 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_size_16: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 16 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_size_16: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 16 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 16) ret i32 %bcmp } define signext i32 @bcmp_size_31(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: bcmp_size_31: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 31 -; LA32-NEXT: bl bcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_size_31: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a2, $a0, 0 +; LA32-UAL-NEXT: ld.w $a3, $a1, 0 +; LA32-UAL-NEXT: ld.w $a4, $a0, 4 +; LA32-UAL-NEXT: ld.w $a5, $a1, 4 +; LA32-UAL-NEXT: ld.w $a6, $a0, 8 +; LA32-UAL-NEXT: ld.w $a7, $a1, 8 +; LA32-UAL-NEXT: ld.w $t0, $a0, 12 +; LA32-UAL-NEXT: ld.w $t1, $a1, 12 +; LA32-UAL-NEXT: xor $a2, $a2, $a3 +; LA32-UAL-NEXT: xor $a3, $a4, $a5 +; LA32-UAL-NEXT: xor $a4, $a6, $a7 +; LA32-UAL-NEXT: xor $a5, $t0, $t1 +; LA32-UAL-NEXT: ld.w $a6, $a0, 16 +; LA32-UAL-NEXT: ld.w $a7, $a1, 16 +; LA32-UAL-NEXT: ld.w $t0, $a0, 20 +; LA32-UAL-NEXT: ld.w $t1, $a1, 20 +; LA32-UAL-NEXT: ld.w $t2, $a0, 24 +; LA32-UAL-NEXT: ld.w $t3, $a1, 24 +; LA32-UAL-NEXT: ld.w $a0, $a0, 27 +; LA32-UAL-NEXT: ld.w $a1, $a1, 27 +; LA32-UAL-NEXT: xor $a6, $a6, $a7 +; LA32-UAL-NEXT: xor $a7, $t0, $t1 +; LA32-UAL-NEXT: xor $t0, $t2, $t3 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: or $a1, $a2, $a3 +; LA32-UAL-NEXT: or $a2, $a4, $a5 +; LA32-UAL-NEXT: or $a3, $a6, $a7 +; LA32-UAL-NEXT: or $a0, $t0, $a0 +; LA32-UAL-NEXT: or $a1, $a1, $a2 +; LA32-UAL-NEXT: or $a0, $a3, $a0 +; LA32-UAL-NEXT: or $a0, $a1, $a0 +; LA32-UAL-NEXT: sltu $a0, $zero, $a0 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_size_31: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 31 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_size_31: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.d $a2, $a0, 0 +; LA64-UAL-NEXT: ld.d $a3, $a1, 0 +; LA64-UAL-NEXT: ld.d $a4, $a0, 8 +; LA64-UAL-NEXT: ld.d $a5, $a1, 8 +; LA64-UAL-NEXT: ld.d $a6, $a0, 16 +; LA64-UAL-NEXT: ld.d $a7, $a1, 16 +; LA64-UAL-NEXT: ld.d $a0, $a0, 23 +; LA64-UAL-NEXT: ld.d $a1, $a1, 23 +; LA64-UAL-NEXT: xor $a2, $a2, $a3 +; LA64-UAL-NEXT: xor $a3, $a4, $a5 +; LA64-UAL-NEXT: xor $a4, $a6, $a7 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: or $a1, $a2, $a3 +; LA64-UAL-NEXT: or $a0, $a4, $a0 +; LA64-UAL-NEXT: or $a0, $a1, $a0 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_size_31: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 31 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_size_31: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 31 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 31) ret i32 %bcmp } define signext i32 @bcmp_size_32(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: bcmp_size_32: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 32 -; LA32-NEXT: bl bcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_size_32: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a2, $a0, 0 +; LA32-UAL-NEXT: ld.w $a3, $a1, 0 +; LA32-UAL-NEXT: ld.w $a4, $a0, 4 +; LA32-UAL-NEXT: ld.w $a5, $a1, 4 +; LA32-UAL-NEXT: ld.w $a6, $a0, 8 +; LA32-UAL-NEXT: ld.w $a7, $a1, 8 +; LA32-UAL-NEXT: ld.w $t0, $a0, 12 +; LA32-UAL-NEXT: ld.w $t1, $a1, 12 +; LA32-UAL-NEXT: xor $a2, $a2, $a3 +; LA32-UAL-NEXT: xor $a3, $a4, $a5 +; LA32-UAL-NEXT: xor $a4, $a6, $a7 +; LA32-UAL-NEXT: xor $a5, $t0, $t1 +; LA32-UAL-NEXT: ld.w $a6, $a0, 16 +; LA32-UAL-NEXT: ld.w $a7, $a1, 16 +; LA32-UAL-NEXT: ld.w $t0, $a0, 20 +; LA32-UAL-NEXT: ld.w $t1, $a1, 20 +; LA32-UAL-NEXT: ld.w $t2, $a0, 24 +; LA32-UAL-NEXT: ld.w $t3, $a1, 24 +; LA32-UAL-NEXT: ld.w $a0, $a0, 28 +; LA32-UAL-NEXT: ld.w $a1, $a1, 28 +; LA32-UAL-NEXT: xor $a6, $a6, $a7 +; LA32-UAL-NEXT: xor $a7, $t0, $t1 +; LA32-UAL-NEXT: xor $t0, $t2, $t3 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: or $a1, $a2, $a3 +; LA32-UAL-NEXT: or $a2, $a4, $a5 +; LA32-UAL-NEXT: or $a3, $a6, $a7 +; LA32-UAL-NEXT: or $a0, $t0, $a0 +; LA32-UAL-NEXT: or $a1, $a1, $a2 +; LA32-UAL-NEXT: or $a0, $a3, $a0 +; LA32-UAL-NEXT: or $a0, $a1, $a0 +; LA32-UAL-NEXT: sltu $a0, $zero, $a0 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_size_32: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 32 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_size_32: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.d $a2, $a0, 0 +; LA64-UAL-NEXT: ld.d $a3, $a1, 0 +; LA64-UAL-NEXT: ld.d $a4, $a0, 8 +; LA64-UAL-NEXT: ld.d $a5, $a1, 8 +; LA64-UAL-NEXT: ld.d $a6, $a0, 16 +; LA64-UAL-NEXT: ld.d $a7, $a1, 16 +; LA64-UAL-NEXT: ld.d $a0, $a0, 24 +; LA64-UAL-NEXT: ld.d $a1, $a1, 24 +; LA64-UAL-NEXT: xor $a2, $a2, $a3 +; LA64-UAL-NEXT: xor $a3, $a4, $a5 +; LA64-UAL-NEXT: xor $a4, $a6, $a7 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: or $a1, $a2, $a3 +; LA64-UAL-NEXT: or $a0, $a4, $a0 +; LA64-UAL-NEXT: or $a0, $a1, $a0 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_size_32: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 32 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_size_32: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 32 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 32) ret i32 %bcmp @@ -360,16 +700,52 @@ define signext i32 @bcmp_size_63(ptr %s1, ptr %s2) nounwind { ; LA32-NEXT: addi.w $sp, $sp, 16 ; LA32-NEXT: ret ; -; LA64-LABEL: bcmp_size_63: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 63 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_size_63: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.d $a2, $a0, 0 +; LA64-UAL-NEXT: ld.d $a3, $a1, 0 +; LA64-UAL-NEXT: ld.d $a4, $a0, 8 +; LA64-UAL-NEXT: ld.d $a5, $a1, 8 +; LA64-UAL-NEXT: ld.d $a6, $a0, 16 +; LA64-UAL-NEXT: ld.d $a7, $a1, 16 +; LA64-UAL-NEXT: ld.d $t0, $a0, 24 +; LA64-UAL-NEXT: ld.d $t1, $a1, 24 +; LA64-UAL-NEXT: xor $a2, $a2, $a3 +; LA64-UAL-NEXT: xor $a3, $a4, $a5 +; LA64-UAL-NEXT: xor $a4, $a6, $a7 +; LA64-UAL-NEXT: xor $a5, $t0, $t1 +; LA64-UAL-NEXT: ld.d $a6, $a0, 32 +; LA64-UAL-NEXT: ld.d $a7, $a1, 32 +; LA64-UAL-NEXT: ld.d $t0, $a0, 40 +; LA64-UAL-NEXT: ld.d $t1, $a1, 40 +; LA64-UAL-NEXT: ld.d $t2, $a0, 48 +; LA64-UAL-NEXT: ld.d $t3, $a1, 48 +; LA64-UAL-NEXT: ld.d $a0, $a0, 55 +; LA64-UAL-NEXT: ld.d $a1, $a1, 55 +; LA64-UAL-NEXT: xor $a6, $a6, $a7 +; LA64-UAL-NEXT: xor $a7, $t0, $t1 +; LA64-UAL-NEXT: xor $t0, $t2, $t3 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: or $a1, $a2, $a3 +; LA64-UAL-NEXT: or $a2, $a4, $a5 +; LA64-UAL-NEXT: or $a3, $a6, $a7 +; LA64-UAL-NEXT: or $a0, $t0, $a0 +; LA64-UAL-NEXT: or $a1, $a1, $a2 +; LA64-UAL-NEXT: or $a0, $a3, $a0 +; LA64-UAL-NEXT: or $a0, $a1, $a0 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_size_63: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 63 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 63) ret i32 %bcmp @@ -386,16 +762,52 @@ define signext i32 @bcmp_size_64(ptr %s1, ptr %s2) nounwind { ; LA32-NEXT: addi.w $sp, $sp, 16 ; LA32-NEXT: ret ; -; LA64-LABEL: bcmp_size_64: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 64 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_size_64: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.d $a2, $a0, 0 +; LA64-UAL-NEXT: ld.d $a3, $a1, 0 +; LA64-UAL-NEXT: ld.d $a4, $a0, 8 +; LA64-UAL-NEXT: ld.d $a5, $a1, 8 +; LA64-UAL-NEXT: ld.d $a6, $a0, 16 +; LA64-UAL-NEXT: ld.d $a7, $a1, 16 +; LA64-UAL-NEXT: ld.d $t0, $a0, 24 +; LA64-UAL-NEXT: ld.d $t1, $a1, 24 +; LA64-UAL-NEXT: xor $a2, $a2, $a3 +; LA64-UAL-NEXT: xor $a3, $a4, $a5 +; LA64-UAL-NEXT: xor $a4, $a6, $a7 +; LA64-UAL-NEXT: xor $a5, $t0, $t1 +; LA64-UAL-NEXT: ld.d $a6, $a0, 32 +; LA64-UAL-NEXT: ld.d $a7, $a1, 32 +; LA64-UAL-NEXT: ld.d $t0, $a0, 40 +; LA64-UAL-NEXT: ld.d $t1, $a1, 40 +; LA64-UAL-NEXT: ld.d $t2, $a0, 48 +; LA64-UAL-NEXT: ld.d $t3, $a1, 48 +; LA64-UAL-NEXT: ld.d $a0, $a0, 56 +; LA64-UAL-NEXT: ld.d $a1, $a1, 56 +; LA64-UAL-NEXT: xor $a6, $a6, $a7 +; LA64-UAL-NEXT: xor $a7, $t0, $t1 +; LA64-UAL-NEXT: xor $t0, $t2, $t3 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: or $a1, $a2, $a3 +; LA64-UAL-NEXT: or $a2, $a4, $a5 +; LA64-UAL-NEXT: or $a3, $a6, $a7 +; LA64-UAL-NEXT: or $a0, $t0, $a0 +; LA64-UAL-NEXT: or $a1, $a1, $a2 +; LA64-UAL-NEXT: or $a0, $a3, $a0 +; LA64-UAL-NEXT: or $a0, $a1, $a0 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_size_64: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 64 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 64) ret i32 %bcmp @@ -478,28 +890,60 @@ entry: } define i1 @bcmp_eq_zero(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: bcmp_eq_zero: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 16 -; LA32-NEXT: bl bcmp -; LA32-NEXT: sltui $a0, $a0, 1 -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_eq_zero: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a2, $a0, 0 +; LA32-UAL-NEXT: ld.w $a3, $a1, 0 +; LA32-UAL-NEXT: ld.w $a4, $a0, 4 +; LA32-UAL-NEXT: ld.w $a5, $a1, 4 +; LA32-UAL-NEXT: ld.w $a6, $a0, 8 +; LA32-UAL-NEXT: ld.w $a7, $a1, 8 +; LA32-UAL-NEXT: ld.w $a0, $a0, 12 +; LA32-UAL-NEXT: ld.w $a1, $a1, 12 +; LA32-UAL-NEXT: xor $a2, $a2, $a3 +; LA32-UAL-NEXT: xor $a3, $a4, $a5 +; LA32-UAL-NEXT: xor $a4, $a6, $a7 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: or $a1, $a2, $a3 +; LA32-UAL-NEXT: or $a0, $a4, $a0 +; LA32-UAL-NEXT: or $a0, $a1, $a0 +; LA32-UAL-NEXT: sltui $a0, $a0, 1 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_eq_zero: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 16 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: sltui $a0, $a0, 1 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_eq_zero: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.d $a2, $a0, 0 +; LA64-UAL-NEXT: ld.d $a3, $a1, 0 +; LA64-UAL-NEXT: ld.d $a0, $a0, 8 +; LA64-UAL-NEXT: ld.d $a1, $a1, 8 +; LA64-UAL-NEXT: xor $a2, $a2, $a3 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: or $a0, $a2, $a0 +; LA64-UAL-NEXT: sltui $a0, $a0, 1 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_eq_zero: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 16 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: sltui $a0, $a0, 1 +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_eq_zero: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 16 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: sltui $a0, $a0, 1 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 16) %ret = icmp eq i32 %bcmp, 0 @@ -507,28 +951,38 @@ entry: } define i1 @bcmp_lt_zero(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: bcmp_lt_zero: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 4 -; LA32-NEXT: bl bcmp -; LA32-NEXT: srli.w $a0, $a0, 31 -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_lt_zero: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: move $a0, $zero +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_lt_zero: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 4 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: slti $a0, $a0, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_lt_zero: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: move $a0, $zero +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_lt_zero: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 4 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: srli.w $a0, $a0, 31 +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_lt_zero: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 4 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: slti $a0, $a0, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 4) %ret = icmp slt i32 %bcmp, 0 @@ -536,28 +990,44 @@ entry: } define i1 @bcmp_gt_zero(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: bcmp_gt_zero: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 4 -; LA32-NEXT: bl bcmp -; LA32-NEXT: slt $a0, $zero, $a0 -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_gt_zero: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a0, $a0, 0 +; LA32-UAL-NEXT: ld.w $a1, $a1, 0 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: sltu $a0, $zero, $a0 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_gt_zero: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 4 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: slt $a0, $zero, $a0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_gt_zero: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.w $a0, $a0, 0 +; LA64-UAL-NEXT: ld.w $a1, $a1, 0 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_gt_zero: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 4 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: slt $a0, $zero, $a0 +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_gt_zero: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 4 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: slt $a0, $zero, $a0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 4) %ret = icmp sgt i32 %bcmp, 0 @@ -565,28 +1035,46 @@ entry: } define i1 @bcmp_le_zero(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: bcmp_le_zero: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 4 -; LA32-NEXT: bl bcmp -; LA32-NEXT: slti $a0, $a0, 1 -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_le_zero: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a0, $a0, 0 +; LA32-UAL-NEXT: ld.w $a1, $a1, 0 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: sltu $a0, $zero, $a0 +; LA32-UAL-NEXT: slti $a0, $a0, 1 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_le_zero: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 4 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: slti $a0, $a0, 1 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_le_zero: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.w $a0, $a0, 0 +; LA64-UAL-NEXT: ld.w $a1, $a1, 0 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: sltu $a0, $zero, $a0 +; LA64-UAL-NEXT: slti $a0, $a0, 1 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_le_zero: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 4 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: slti $a0, $a0, 1 +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_le_zero: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 4 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: slti $a0, $a0, 1 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 4) %ret = icmp slt i32 %bcmp, 1 @@ -594,30 +1082,40 @@ entry: } define i1 @bcmp_ge_zero(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: bcmp_ge_zero: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 4 -; LA32-NEXT: bl bcmp -; LA32-NEXT: addi.w $a1, $zero, -1 -; LA32-NEXT: slt $a0, $a1, $a0 -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: bcmp_ge_zero: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ori $a0, $zero, 1 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: bcmp_ge_zero: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 4 -; LA64-NEXT: pcaddu18i $ra, %call36(bcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: addi.w $a1, $zero, -1 -; LA64-NEXT: slt $a0, $a1, $a0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: bcmp_ge_zero: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ori $a0, $zero, 1 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: bcmp_ge_zero: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 4 +; LA32-NUAL-NEXT: bl bcmp +; LA32-NUAL-NEXT: addi.w $a1, $zero, -1 +; LA32-NUAL-NEXT: slt $a0, $a1, $a0 +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: bcmp_ge_zero: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 4 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(bcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: addi.w $a1, $zero, -1 +; LA64-NUAL-NEXT: slt $a0, $a1, $a0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %bcmp = call signext i32 @bcmp(ptr %s1, ptr %s2, iGRLen 4) %ret = icmp sgt i32 %bcmp, -1 @@ -635,312 +1133,1394 @@ entry: } define signext i32 @memcmp_size_1(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: memcmp_size_1: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 1 -; LA32-NEXT: bl memcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_size_1: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.bu $a0, $a0, 0 +; LA32-UAL-NEXT: ld.bu $a1, $a1, 0 +; LA32-UAL-NEXT: sub.w $a0, $a0, $a1 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_size_1: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 1 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_size_1: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.bu $a0, $a0, 0 +; LA64-UAL-NEXT: ld.bu $a1, $a1, 0 +; LA64-UAL-NEXT: sub.d $a0, $a0, $a1 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_size_1: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 1 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_size_1: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 1 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 1) ret i32 %memcmp } define signext i32 @memcmp_size_2(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: memcmp_size_2: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 2 -; LA32-NEXT: bl memcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_size_2: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.hu $a0, $a0, 0 +; LA32-UAL-NEXT: ld.hu $a1, $a1, 0 +; LA32-UAL-NEXT: srli.w $a2, $a0, 8 +; LA32-UAL-NEXT: slli.w $a0, $a0, 8 +; LA32-UAL-NEXT: or $a0, $a0, $a2 +; LA32-UAL-NEXT: srli.w $a2, $a1, 8 +; LA32-UAL-NEXT: slli.w $a1, $a1, 8 +; LA32-UAL-NEXT: or $a1, $a1, $a2 +; LA32-UAL-NEXT: lu12i.w $a2, 15 +; LA32-UAL-NEXT: ori $a2, $a2, 4095 +; LA32-UAL-NEXT: and $a0, $a0, $a2 +; LA32-UAL-NEXT: and $a1, $a1, $a2 +; LA32-UAL-NEXT: sub.w $a0, $a0, $a1 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_size_2: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 2 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_size_2: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.h $a0, $a0, 0 +; LA64-UAL-NEXT: ld.h $a1, $a1, 0 +; LA64-UAL-NEXT: revb.2h $a0, $a0 +; LA64-UAL-NEXT: revb.2h $a1, $a1 +; LA64-UAL-NEXT: bstrpick.d $a0, $a0, 15, 0 +; LA64-UAL-NEXT: bstrpick.d $a1, $a1, 15, 0 +; LA64-UAL-NEXT: sub.d $a0, $a0, $a1 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_size_2: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 2 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_size_2: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 2 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 2) ret i32 %memcmp } define signext i32 @memcmp_size_3(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: memcmp_size_3: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 3 -; LA32-NEXT: bl memcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_size_3: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.bu $a2, $a0, 2 +; LA32-UAL-NEXT: ld.hu $a0, $a0, 0 +; LA32-UAL-NEXT: ld.bu $a3, $a1, 2 +; LA32-UAL-NEXT: ld.hu $a1, $a1, 0 +; LA32-UAL-NEXT: lu12i.w $a4, 15 +; LA32-UAL-NEXT: ori $a4, $a4, 3840 +; LA32-UAL-NEXT: and $a5, $a0, $a4 +; LA32-UAL-NEXT: or $a2, $a5, $a2 +; LA32-UAL-NEXT: slli.w $a2, $a2, 8 +; LA32-UAL-NEXT: slli.w $a0, $a0, 24 +; LA32-UAL-NEXT: or $a0, $a2, $a0 +; LA32-UAL-NEXT: and $a2, $a1, $a4 +; LA32-UAL-NEXT: or $a2, $a2, $a3 +; LA32-UAL-NEXT: slli.w $a2, $a2, 8 +; LA32-UAL-NEXT: slli.w $a1, $a1, 24 +; LA32-UAL-NEXT: or $a1, $a2, $a1 +; LA32-UAL-NEXT: sltu $a2, $a0, $a1 +; LA32-UAL-NEXT: sltu $a0, $a1, $a0 +; LA32-UAL-NEXT: sub.w $a0, $a0, $a2 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_size_3: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 3 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_size_3: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.bu $a2, $a0, 2 +; LA64-UAL-NEXT: ld.hu $a0, $a0, 0 +; LA64-UAL-NEXT: ld.bu $a3, $a1, 2 +; LA64-UAL-NEXT: ld.hu $a1, $a1, 0 +; LA64-UAL-NEXT: slli.d $a2, $a2, 16 +; LA64-UAL-NEXT: or $a0, $a0, $a2 +; LA64-UAL-NEXT: slli.d $a2, $a3, 16 +; LA64-UAL-NEXT: or $a1, $a1, $a2 +; LA64-UAL-NEXT: revb.2w $a0, $a0 +; LA64-UAL-NEXT: addi.w $a0, $a0, 0 +; LA64-UAL-NEXT: revb.2w $a1, $a1 +; LA64-UAL-NEXT: addi.w $a1, $a1, 0 +; LA64-UAL-NEXT: sltu $a2, $a0, $a1 +; LA64-UAL-NEXT: sltu $a0, $a1, $a0 +; LA64-UAL-NEXT: sub.d $a0, $a0, $a2 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_size_3: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 3 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_size_3: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 3 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 3) ret i32 %memcmp } define signext i32 @memcmp_size_4(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: memcmp_size_4: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 4 -; LA32-NEXT: bl memcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_size_4: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a0, $a0, 0 +; LA32-UAL-NEXT: ld.w $a1, $a1, 0 +; LA32-UAL-NEXT: srli.w $a2, $a0, 8 +; LA32-UAL-NEXT: lu12i.w $a3, 15 +; LA32-UAL-NEXT: ori $a3, $a3, 3840 +; LA32-UAL-NEXT: and $a2, $a2, $a3 +; LA32-UAL-NEXT: srli.w $a4, $a0, 24 +; LA32-UAL-NEXT: or $a2, $a2, $a4 +; LA32-UAL-NEXT: and $a4, $a0, $a3 +; LA32-UAL-NEXT: slli.w $a4, $a4, 8 +; LA32-UAL-NEXT: slli.w $a0, $a0, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: or $a0, $a0, $a2 +; LA32-UAL-NEXT: srli.w $a2, $a1, 8 +; LA32-UAL-NEXT: and $a2, $a2, $a3 +; LA32-UAL-NEXT: srli.w $a4, $a1, 24 +; LA32-UAL-NEXT: or $a2, $a2, $a4 +; LA32-UAL-NEXT: and $a3, $a1, $a3 +; LA32-UAL-NEXT: slli.w $a3, $a3, 8 +; LA32-UAL-NEXT: slli.w $a1, $a1, 24 +; LA32-UAL-NEXT: or $a1, $a1, $a3 +; LA32-UAL-NEXT: or $a1, $a1, $a2 +; LA32-UAL-NEXT: sltu $a2, $a0, $a1 +; LA32-UAL-NEXT: sltu $a0, $a1, $a0 +; LA32-UAL-NEXT: sub.w $a0, $a0, $a2 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_size_4: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 4 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_size_4: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.w $a0, $a0, 0 +; LA64-UAL-NEXT: ld.w $a1, $a1, 0 +; LA64-UAL-NEXT: revb.2w $a0, $a0 +; LA64-UAL-NEXT: addi.w $a0, $a0, 0 +; LA64-UAL-NEXT: revb.2w $a1, $a1 +; LA64-UAL-NEXT: addi.w $a1, $a1, 0 +; LA64-UAL-NEXT: sltu $a2, $a0, $a1 +; LA64-UAL-NEXT: sltu $a0, $a1, $a0 +; LA64-UAL-NEXT: sub.d $a0, $a0, $a2 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_size_4: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 4 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_size_4: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 4 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 4) ret i32 %memcmp } define signext i32 @memcmp_size_5(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: memcmp_size_5: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 5 -; LA32-NEXT: bl memcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_size_5: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a2, $a0, 0 +; LA32-UAL-NEXT: ld.w $a3, $a1, 0 +; LA32-UAL-NEXT: srli.w $a4, $a2, 8 +; LA32-UAL-NEXT: lu12i.w $a5, 15 +; LA32-UAL-NEXT: ori $a5, $a5, 3840 +; LA32-UAL-NEXT: and $a4, $a4, $a5 +; LA32-UAL-NEXT: srli.w $a6, $a2, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: and $a6, $a2, $a5 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a2, $a2, 24 +; LA32-UAL-NEXT: or $a2, $a2, $a6 +; LA32-UAL-NEXT: or $a2, $a2, $a4 +; LA32-UAL-NEXT: srli.w $a4, $a3, 8 +; LA32-UAL-NEXT: and $a4, $a4, $a5 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: and $a5, $a3, $a5 +; LA32-UAL-NEXT: slli.w $a5, $a5, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: or $a3, $a3, $a4 +; LA32-UAL-NEXT: bne $a2, $a3, .LBB28_2 +; LA32-UAL-NEXT: # %bb.1: # %loadbb1 +; LA32-UAL-NEXT: ld.bu $a0, $a0, 4 +; LA32-UAL-NEXT: ld.bu $a1, $a1, 4 +; LA32-UAL-NEXT: sub.w $a0, $a0, $a1 +; LA32-UAL-NEXT: ret +; LA32-UAL-NEXT: .LBB28_2: # %res_block +; LA32-UAL-NEXT: sltu $a0, $a2, $a3 +; LA32-UAL-NEXT: sub.w $a0, $zero, $a0 +; LA32-UAL-NEXT: ori $a0, $a0, 1 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_size_5: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 5 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_size_5: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.bu $a2, $a0, 4 +; LA64-UAL-NEXT: ld.wu $a0, $a0, 0 +; LA64-UAL-NEXT: ld.bu $a3, $a1, 4 +; LA64-UAL-NEXT: ld.wu $a1, $a1, 0 +; LA64-UAL-NEXT: slli.d $a2, $a2, 32 +; LA64-UAL-NEXT: or $a0, $a0, $a2 +; LA64-UAL-NEXT: slli.d $a2, $a3, 32 +; LA64-UAL-NEXT: or $a1, $a1, $a2 +; LA64-UAL-NEXT: revb.d $a0, $a0 +; LA64-UAL-NEXT: revb.d $a1, $a1 +; LA64-UAL-NEXT: sltu $a2, $a0, $a1 +; LA64-UAL-NEXT: sltu $a0, $a1, $a0 +; LA64-UAL-NEXT: sub.d $a0, $a0, $a2 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_size_5: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 5 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_size_5: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 5 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 5) ret i32 %memcmp } define signext i32 @memcmp_size_6(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: memcmp_size_6: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 6 -; LA32-NEXT: bl memcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_size_6: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a3, $a0, 0 +; LA32-UAL-NEXT: ld.w $a4, $a1, 0 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: lu12i.w $a2, 15 +; LA32-UAL-NEXT: ori $a6, $a2, 3840 +; LA32-UAL-NEXT: and $a5, $a5, $a6 +; LA32-UAL-NEXT: srli.w $a7, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a7 +; LA32-UAL-NEXT: and $a7, $a3, $a6 +; LA32-UAL-NEXT: slli.w $a7, $a7, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a7 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a6 +; LA32-UAL-NEXT: srli.w $a7, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a7 +; LA32-UAL-NEXT: and $a6, $a4, $a6 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB29_3 +; LA32-UAL-NEXT: # %bb.1: # %loadbb1 +; LA32-UAL-NEXT: ld.hu $a0, $a0, 4 +; LA32-UAL-NEXT: ld.hu $a1, $a1, 4 +; LA32-UAL-NEXT: srli.w $a3, $a0, 8 +; LA32-UAL-NEXT: slli.w $a0, $a0, 8 +; LA32-UAL-NEXT: or $a0, $a0, $a3 +; LA32-UAL-NEXT: srli.w $a3, $a1, 8 +; LA32-UAL-NEXT: slli.w $a1, $a1, 8 +; LA32-UAL-NEXT: or $a1, $a1, $a3 +; LA32-UAL-NEXT: ori $a2, $a2, 4095 +; LA32-UAL-NEXT: and $a3, $a0, $a2 +; LA32-UAL-NEXT: and $a4, $a1, $a2 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB29_3 +; LA32-UAL-NEXT: # %bb.2: +; LA32-UAL-NEXT: move $a0, $zero +; LA32-UAL-NEXT: ret +; LA32-UAL-NEXT: .LBB29_3: # %res_block +; LA32-UAL-NEXT: sltu $a0, $a3, $a4 +; LA32-UAL-NEXT: sub.w $a0, $zero, $a0 +; LA32-UAL-NEXT: ori $a0, $a0, 1 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_size_6: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 6 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_size_6: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.hu $a2, $a0, 4 +; LA64-UAL-NEXT: ld.wu $a0, $a0, 0 +; LA64-UAL-NEXT: ld.hu $a3, $a1, 4 +; LA64-UAL-NEXT: ld.wu $a1, $a1, 0 +; LA64-UAL-NEXT: slli.d $a2, $a2, 32 +; LA64-UAL-NEXT: or $a0, $a0, $a2 +; LA64-UAL-NEXT: slli.d $a2, $a3, 32 +; LA64-UAL-NEXT: or $a1, $a1, $a2 +; LA64-UAL-NEXT: revb.d $a0, $a0 +; LA64-UAL-NEXT: revb.d $a1, $a1 +; LA64-UAL-NEXT: sltu $a2, $a0, $a1 +; LA64-UAL-NEXT: sltu $a0, $a1, $a0 +; LA64-UAL-NEXT: sub.d $a0, $a0, $a2 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_size_6: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 6 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_size_6: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 6 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 6) ret i32 %memcmp } define signext i32 @memcmp_size_7(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: memcmp_size_7: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 7 -; LA32-NEXT: bl memcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_size_7: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a3, $a0, 0 +; LA32-UAL-NEXT: ld.w $a4, $a1, 0 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: lu12i.w $a2, 15 +; LA32-UAL-NEXT: ori $a2, $a2, 3840 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB30_3 +; LA32-UAL-NEXT: # %bb.1: # %loadbb1 +; LA32-UAL-NEXT: ld.w $a0, $a0, 3 +; LA32-UAL-NEXT: ld.w $a1, $a1, 3 +; LA32-UAL-NEXT: srli.w $a3, $a0, 8 +; LA32-UAL-NEXT: and $a3, $a3, $a2 +; LA32-UAL-NEXT: srli.w $a4, $a0, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a4 +; LA32-UAL-NEXT: and $a4, $a0, $a2 +; LA32-UAL-NEXT: slli.w $a4, $a4, 8 +; LA32-UAL-NEXT: slli.w $a0, $a0, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: or $a3, $a0, $a3 +; LA32-UAL-NEXT: srli.w $a0, $a1, 8 +; LA32-UAL-NEXT: and $a0, $a0, $a2 +; LA32-UAL-NEXT: srli.w $a4, $a1, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: and $a2, $a1, $a2 +; LA32-UAL-NEXT: slli.w $a2, $a2, 8 +; LA32-UAL-NEXT: slli.w $a1, $a1, 24 +; LA32-UAL-NEXT: or $a1, $a1, $a2 +; LA32-UAL-NEXT: or $a4, $a1, $a0 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB30_3 +; LA32-UAL-NEXT: # %bb.2: +; LA32-UAL-NEXT: move $a0, $zero +; LA32-UAL-NEXT: ret +; LA32-UAL-NEXT: .LBB30_3: # %res_block +; LA32-UAL-NEXT: sltu $a0, $a3, $a4 +; LA32-UAL-NEXT: sub.w $a0, $zero, $a0 +; LA32-UAL-NEXT: ori $a0, $a0, 1 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_size_7: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 7 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_size_7: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.w $a2, $a0, 0 +; LA64-UAL-NEXT: ld.w $a3, $a1, 0 +; LA64-UAL-NEXT: revb.2w $a2, $a2 +; LA64-UAL-NEXT: addi.w $a4, $a2, 0 +; LA64-UAL-NEXT: revb.2w $a3, $a3 +; LA64-UAL-NEXT: addi.w $a5, $a3, 0 +; LA64-UAL-NEXT: bne $a4, $a5, .LBB30_3 +; LA64-UAL-NEXT: # %bb.1: # %loadbb1 +; LA64-UAL-NEXT: ld.w $a0, $a0, 3 +; LA64-UAL-NEXT: ld.w $a1, $a1, 3 +; LA64-UAL-NEXT: revb.2w $a2, $a0 +; LA64-UAL-NEXT: addi.w $a0, $a2, 0 +; LA64-UAL-NEXT: revb.2w $a3, $a1 +; LA64-UAL-NEXT: addi.w $a1, $a3, 0 +; LA64-UAL-NEXT: bne $a0, $a1, .LBB30_3 +; LA64-UAL-NEXT: # %bb.2: +; LA64-UAL-NEXT: move $a0, $zero +; LA64-UAL-NEXT: ret +; LA64-UAL-NEXT: .LBB30_3: # %res_block +; LA64-UAL-NEXT: addi.w $a0, $a3, 0 +; LA64-UAL-NEXT: addi.w $a1, $a2, 0 +; LA64-UAL-NEXT: sltu $a0, $a1, $a0 +; LA64-UAL-NEXT: sub.d $a0, $zero, $a0 +; LA64-UAL-NEXT: ori $a0, $a0, 1 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_size_7: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 7 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_size_7: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 7 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 7) ret i32 %memcmp } define signext i32 @memcmp_size_8(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: memcmp_size_8: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 8 -; LA32-NEXT: bl memcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_size_8: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a3, $a0, 0 +; LA32-UAL-NEXT: ld.w $a4, $a1, 0 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: lu12i.w $a2, 15 +; LA32-UAL-NEXT: ori $a2, $a2, 3840 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB31_3 +; LA32-UAL-NEXT: # %bb.1: # %loadbb1 +; LA32-UAL-NEXT: ld.w $a0, $a0, 4 +; LA32-UAL-NEXT: ld.w $a1, $a1, 4 +; LA32-UAL-NEXT: srli.w $a3, $a0, 8 +; LA32-UAL-NEXT: and $a3, $a3, $a2 +; LA32-UAL-NEXT: srli.w $a4, $a0, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a4 +; LA32-UAL-NEXT: and $a4, $a0, $a2 +; LA32-UAL-NEXT: slli.w $a4, $a4, 8 +; LA32-UAL-NEXT: slli.w $a0, $a0, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: or $a3, $a0, $a3 +; LA32-UAL-NEXT: srli.w $a0, $a1, 8 +; LA32-UAL-NEXT: and $a0, $a0, $a2 +; LA32-UAL-NEXT: srli.w $a4, $a1, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: and $a2, $a1, $a2 +; LA32-UAL-NEXT: slli.w $a2, $a2, 8 +; LA32-UAL-NEXT: slli.w $a1, $a1, 24 +; LA32-UAL-NEXT: or $a1, $a1, $a2 +; LA32-UAL-NEXT: or $a4, $a1, $a0 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB31_3 +; LA32-UAL-NEXT: # %bb.2: +; LA32-UAL-NEXT: move $a0, $zero +; LA32-UAL-NEXT: ret +; LA32-UAL-NEXT: .LBB31_3: # %res_block +; LA32-UAL-NEXT: sltu $a0, $a3, $a4 +; LA32-UAL-NEXT: sub.w $a0, $zero, $a0 +; LA32-UAL-NEXT: ori $a0, $a0, 1 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_size_8: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 8 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_size_8: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.d $a0, $a0, 0 +; LA64-UAL-NEXT: ld.d $a1, $a1, 0 +; LA64-UAL-NEXT: revb.d $a0, $a0 +; LA64-UAL-NEXT: revb.d $a1, $a1 +; LA64-UAL-NEXT: sltu $a2, $a0, $a1 +; LA64-UAL-NEXT: sltu $a0, $a1, $a0 +; LA64-UAL-NEXT: sub.d $a0, $a0, $a2 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_size_8: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 8 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_size_8: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 8 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 8) ret i32 %memcmp } define signext i32 @memcmp_size_15(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: memcmp_size_15: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 15 -; LA32-NEXT: bl memcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_size_15: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a3, $a0, 0 +; LA32-UAL-NEXT: ld.w $a4, $a1, 0 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: lu12i.w $a2, 15 +; LA32-UAL-NEXT: ori $a2, $a2, 3840 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB32_5 +; LA32-UAL-NEXT: # %bb.1: # %loadbb1 +; LA32-UAL-NEXT: ld.w $a3, $a0, 4 +; LA32-UAL-NEXT: ld.w $a4, $a1, 4 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB32_5 +; LA32-UAL-NEXT: # %bb.2: # %loadbb2 +; LA32-UAL-NEXT: ld.w $a3, $a0, 8 +; LA32-UAL-NEXT: ld.w $a4, $a1, 8 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB32_5 +; LA32-UAL-NEXT: # %bb.3: # %loadbb3 +; LA32-UAL-NEXT: ld.w $a0, $a0, 11 +; LA32-UAL-NEXT: ld.w $a1, $a1, 11 +; LA32-UAL-NEXT: srli.w $a3, $a0, 8 +; LA32-UAL-NEXT: and $a3, $a3, $a2 +; LA32-UAL-NEXT: srli.w $a4, $a0, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a4 +; LA32-UAL-NEXT: and $a4, $a0, $a2 +; LA32-UAL-NEXT: slli.w $a4, $a4, 8 +; LA32-UAL-NEXT: slli.w $a0, $a0, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: or $a3, $a0, $a3 +; LA32-UAL-NEXT: srli.w $a0, $a1, 8 +; LA32-UAL-NEXT: and $a0, $a0, $a2 +; LA32-UAL-NEXT: srli.w $a4, $a1, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: and $a2, $a1, $a2 +; LA32-UAL-NEXT: slli.w $a2, $a2, 8 +; LA32-UAL-NEXT: slli.w $a1, $a1, 24 +; LA32-UAL-NEXT: or $a1, $a1, $a2 +; LA32-UAL-NEXT: or $a4, $a1, $a0 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB32_5 +; LA32-UAL-NEXT: # %bb.4: +; LA32-UAL-NEXT: move $a0, $zero +; LA32-UAL-NEXT: ret +; LA32-UAL-NEXT: .LBB32_5: # %res_block +; LA32-UAL-NEXT: sltu $a0, $a3, $a4 +; LA32-UAL-NEXT: sub.w $a0, $zero, $a0 +; LA32-UAL-NEXT: ori $a0, $a0, 1 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_size_15: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 15 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_size_15: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.d $a2, $a0, 0 +; LA64-UAL-NEXT: ld.d $a3, $a1, 0 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB32_3 +; LA64-UAL-NEXT: # %bb.1: # %loadbb1 +; LA64-UAL-NEXT: ld.d $a0, $a0, 7 +; LA64-UAL-NEXT: ld.d $a1, $a1, 7 +; LA64-UAL-NEXT: revb.d $a2, $a0 +; LA64-UAL-NEXT: revb.d $a3, $a1 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB32_3 +; LA64-UAL-NEXT: # %bb.2: +; LA64-UAL-NEXT: move $a0, $zero +; LA64-UAL-NEXT: ret +; LA64-UAL-NEXT: .LBB32_3: # %res_block +; LA64-UAL-NEXT: sltu $a0, $a2, $a3 +; LA64-UAL-NEXT: sub.d $a0, $zero, $a0 +; LA64-UAL-NEXT: ori $a0, $a0, 1 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_size_15: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 15 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_size_15: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 15 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 15) ret i32 %memcmp } define signext i32 @memcmp_size_16(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: memcmp_size_16: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 16 -; LA32-NEXT: bl memcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_size_16: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a3, $a0, 0 +; LA32-UAL-NEXT: ld.w $a4, $a1, 0 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: lu12i.w $a2, 15 +; LA32-UAL-NEXT: ori $a2, $a2, 3840 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB33_5 +; LA32-UAL-NEXT: # %bb.1: # %loadbb1 +; LA32-UAL-NEXT: ld.w $a3, $a0, 4 +; LA32-UAL-NEXT: ld.w $a4, $a1, 4 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB33_5 +; LA32-UAL-NEXT: # %bb.2: # %loadbb2 +; LA32-UAL-NEXT: ld.w $a3, $a0, 8 +; LA32-UAL-NEXT: ld.w $a4, $a1, 8 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB33_5 +; LA32-UAL-NEXT: # %bb.3: # %loadbb3 +; LA32-UAL-NEXT: ld.w $a0, $a0, 12 +; LA32-UAL-NEXT: ld.w $a1, $a1, 12 +; LA32-UAL-NEXT: srli.w $a3, $a0, 8 +; LA32-UAL-NEXT: and $a3, $a3, $a2 +; LA32-UAL-NEXT: srli.w $a4, $a0, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a4 +; LA32-UAL-NEXT: and $a4, $a0, $a2 +; LA32-UAL-NEXT: slli.w $a4, $a4, 8 +; LA32-UAL-NEXT: slli.w $a0, $a0, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: or $a3, $a0, $a3 +; LA32-UAL-NEXT: srli.w $a0, $a1, 8 +; LA32-UAL-NEXT: and $a0, $a0, $a2 +; LA32-UAL-NEXT: srli.w $a4, $a1, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: and $a2, $a1, $a2 +; LA32-UAL-NEXT: slli.w $a2, $a2, 8 +; LA32-UAL-NEXT: slli.w $a1, $a1, 24 +; LA32-UAL-NEXT: or $a1, $a1, $a2 +; LA32-UAL-NEXT: or $a4, $a1, $a0 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB33_5 +; LA32-UAL-NEXT: # %bb.4: +; LA32-UAL-NEXT: move $a0, $zero +; LA32-UAL-NEXT: ret +; LA32-UAL-NEXT: .LBB33_5: # %res_block +; LA32-UAL-NEXT: sltu $a0, $a3, $a4 +; LA32-UAL-NEXT: sub.w $a0, $zero, $a0 +; LA32-UAL-NEXT: ori $a0, $a0, 1 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_size_16: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 16 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_size_16: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.d $a2, $a0, 0 +; LA64-UAL-NEXT: ld.d $a3, $a1, 0 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB33_3 +; LA64-UAL-NEXT: # %bb.1: # %loadbb1 +; LA64-UAL-NEXT: ld.d $a0, $a0, 8 +; LA64-UAL-NEXT: ld.d $a1, $a1, 8 +; LA64-UAL-NEXT: revb.d $a2, $a0 +; LA64-UAL-NEXT: revb.d $a3, $a1 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB33_3 +; LA64-UAL-NEXT: # %bb.2: +; LA64-UAL-NEXT: move $a0, $zero +; LA64-UAL-NEXT: ret +; LA64-UAL-NEXT: .LBB33_3: # %res_block +; LA64-UAL-NEXT: sltu $a0, $a2, $a3 +; LA64-UAL-NEXT: sub.d $a0, $zero, $a0 +; LA64-UAL-NEXT: ori $a0, $a0, 1 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_size_16: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 16 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_size_16: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 16 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 16) ret i32 %memcmp } define signext i32 @memcmp_size_31(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: memcmp_size_31: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 31 -; LA32-NEXT: bl memcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_size_31: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a3, $a0, 0 +; LA32-UAL-NEXT: ld.w $a4, $a1, 0 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: lu12i.w $a2, 15 +; LA32-UAL-NEXT: ori $a2, $a2, 3840 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB34_9 +; LA32-UAL-NEXT: # %bb.1: # %loadbb1 +; LA32-UAL-NEXT: ld.w $a3, $a0, 4 +; LA32-UAL-NEXT: ld.w $a4, $a1, 4 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB34_9 +; LA32-UAL-NEXT: # %bb.2: # %loadbb2 +; LA32-UAL-NEXT: ld.w $a3, $a0, 8 +; LA32-UAL-NEXT: ld.w $a4, $a1, 8 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB34_9 +; LA32-UAL-NEXT: # %bb.3: # %loadbb3 +; LA32-UAL-NEXT: ld.w $a3, $a0, 12 +; LA32-UAL-NEXT: ld.w $a4, $a1, 12 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB34_9 +; LA32-UAL-NEXT: # %bb.4: # %loadbb4 +; LA32-UAL-NEXT: ld.w $a3, $a0, 16 +; LA32-UAL-NEXT: ld.w $a4, $a1, 16 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB34_9 +; LA32-UAL-NEXT: # %bb.5: # %loadbb5 +; LA32-UAL-NEXT: ld.w $a3, $a0, 20 +; LA32-UAL-NEXT: ld.w $a4, $a1, 20 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB34_9 +; LA32-UAL-NEXT: # %bb.6: # %loadbb6 +; LA32-UAL-NEXT: ld.w $a3, $a0, 24 +; LA32-UAL-NEXT: ld.w $a4, $a1, 24 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB34_9 +; LA32-UAL-NEXT: # %bb.7: # %loadbb7 +; LA32-UAL-NEXT: ld.w $a0, $a0, 27 +; LA32-UAL-NEXT: ld.w $a1, $a1, 27 +; LA32-UAL-NEXT: srli.w $a3, $a0, 8 +; LA32-UAL-NEXT: and $a3, $a3, $a2 +; LA32-UAL-NEXT: srli.w $a4, $a0, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a4 +; LA32-UAL-NEXT: and $a4, $a0, $a2 +; LA32-UAL-NEXT: slli.w $a4, $a4, 8 +; LA32-UAL-NEXT: slli.w $a0, $a0, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: or $a3, $a0, $a3 +; LA32-UAL-NEXT: srli.w $a0, $a1, 8 +; LA32-UAL-NEXT: and $a0, $a0, $a2 +; LA32-UAL-NEXT: srli.w $a4, $a1, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: and $a2, $a1, $a2 +; LA32-UAL-NEXT: slli.w $a2, $a2, 8 +; LA32-UAL-NEXT: slli.w $a1, $a1, 24 +; LA32-UAL-NEXT: or $a1, $a1, $a2 +; LA32-UAL-NEXT: or $a4, $a1, $a0 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB34_9 +; LA32-UAL-NEXT: # %bb.8: +; LA32-UAL-NEXT: move $a0, $zero +; LA32-UAL-NEXT: ret +; LA32-UAL-NEXT: .LBB34_9: # %res_block +; LA32-UAL-NEXT: sltu $a0, $a3, $a4 +; LA32-UAL-NEXT: sub.w $a0, $zero, $a0 +; LA32-UAL-NEXT: ori $a0, $a0, 1 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_size_31: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 31 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_size_31: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.d $a2, $a0, 0 +; LA64-UAL-NEXT: ld.d $a3, $a1, 0 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB34_5 +; LA64-UAL-NEXT: # %bb.1: # %loadbb1 +; LA64-UAL-NEXT: ld.d $a2, $a0, 8 +; LA64-UAL-NEXT: ld.d $a3, $a1, 8 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB34_5 +; LA64-UAL-NEXT: # %bb.2: # %loadbb2 +; LA64-UAL-NEXT: ld.d $a2, $a0, 16 +; LA64-UAL-NEXT: ld.d $a3, $a1, 16 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB34_5 +; LA64-UAL-NEXT: # %bb.3: # %loadbb3 +; LA64-UAL-NEXT: ld.d $a0, $a0, 23 +; LA64-UAL-NEXT: ld.d $a1, $a1, 23 +; LA64-UAL-NEXT: revb.d $a2, $a0 +; LA64-UAL-NEXT: revb.d $a3, $a1 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB34_5 +; LA64-UAL-NEXT: # %bb.4: +; LA64-UAL-NEXT: move $a0, $zero +; LA64-UAL-NEXT: ret +; LA64-UAL-NEXT: .LBB34_5: # %res_block +; LA64-UAL-NEXT: sltu $a0, $a2, $a3 +; LA64-UAL-NEXT: sub.d $a0, $zero, $a0 +; LA64-UAL-NEXT: ori $a0, $a0, 1 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_size_31: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 31 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_size_31: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 31 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 31) ret i32 %memcmp } define signext i32 @memcmp_size_32(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: memcmp_size_32: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 32 -; LA32-NEXT: bl memcmp -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_size_32: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a3, $a0, 0 +; LA32-UAL-NEXT: ld.w $a4, $a1, 0 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: lu12i.w $a2, 15 +; LA32-UAL-NEXT: ori $a2, $a2, 3840 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB35_9 +; LA32-UAL-NEXT: # %bb.1: # %loadbb1 +; LA32-UAL-NEXT: ld.w $a3, $a0, 4 +; LA32-UAL-NEXT: ld.w $a4, $a1, 4 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB35_9 +; LA32-UAL-NEXT: # %bb.2: # %loadbb2 +; LA32-UAL-NEXT: ld.w $a3, $a0, 8 +; LA32-UAL-NEXT: ld.w $a4, $a1, 8 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB35_9 +; LA32-UAL-NEXT: # %bb.3: # %loadbb3 +; LA32-UAL-NEXT: ld.w $a3, $a0, 12 +; LA32-UAL-NEXT: ld.w $a4, $a1, 12 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB35_9 +; LA32-UAL-NEXT: # %bb.4: # %loadbb4 +; LA32-UAL-NEXT: ld.w $a3, $a0, 16 +; LA32-UAL-NEXT: ld.w $a4, $a1, 16 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB35_9 +; LA32-UAL-NEXT: # %bb.5: # %loadbb5 +; LA32-UAL-NEXT: ld.w $a3, $a0, 20 +; LA32-UAL-NEXT: ld.w $a4, $a1, 20 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB35_9 +; LA32-UAL-NEXT: # %bb.6: # %loadbb6 +; LA32-UAL-NEXT: ld.w $a3, $a0, 24 +; LA32-UAL-NEXT: ld.w $a4, $a1, 24 +; LA32-UAL-NEXT: srli.w $a5, $a3, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a3, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a3, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a3, $a3, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a6 +; LA32-UAL-NEXT: or $a3, $a3, $a5 +; LA32-UAL-NEXT: srli.w $a5, $a4, 8 +; LA32-UAL-NEXT: and $a5, $a5, $a2 +; LA32-UAL-NEXT: srli.w $a6, $a4, 24 +; LA32-UAL-NEXT: or $a5, $a5, $a6 +; LA32-UAL-NEXT: and $a6, $a4, $a2 +; LA32-UAL-NEXT: slli.w $a6, $a6, 8 +; LA32-UAL-NEXT: slli.w $a4, $a4, 24 +; LA32-UAL-NEXT: or $a4, $a4, $a6 +; LA32-UAL-NEXT: or $a4, $a4, $a5 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB35_9 +; LA32-UAL-NEXT: # %bb.7: # %loadbb7 +; LA32-UAL-NEXT: ld.w $a0, $a0, 28 +; LA32-UAL-NEXT: ld.w $a1, $a1, 28 +; LA32-UAL-NEXT: srli.w $a3, $a0, 8 +; LA32-UAL-NEXT: and $a3, $a3, $a2 +; LA32-UAL-NEXT: srli.w $a4, $a0, 24 +; LA32-UAL-NEXT: or $a3, $a3, $a4 +; LA32-UAL-NEXT: and $a4, $a0, $a2 +; LA32-UAL-NEXT: slli.w $a4, $a4, 8 +; LA32-UAL-NEXT: slli.w $a0, $a0, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: or $a3, $a0, $a3 +; LA32-UAL-NEXT: srli.w $a0, $a1, 8 +; LA32-UAL-NEXT: and $a0, $a0, $a2 +; LA32-UAL-NEXT: srli.w $a4, $a1, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: and $a2, $a1, $a2 +; LA32-UAL-NEXT: slli.w $a2, $a2, 8 +; LA32-UAL-NEXT: slli.w $a1, $a1, 24 +; LA32-UAL-NEXT: or $a1, $a1, $a2 +; LA32-UAL-NEXT: or $a4, $a1, $a0 +; LA32-UAL-NEXT: bne $a3, $a4, .LBB35_9 +; LA32-UAL-NEXT: # %bb.8: +; LA32-UAL-NEXT: move $a0, $zero +; LA32-UAL-NEXT: ret +; LA32-UAL-NEXT: .LBB35_9: # %res_block +; LA32-UAL-NEXT: sltu $a0, $a3, $a4 +; LA32-UAL-NEXT: sub.w $a0, $zero, $a0 +; LA32-UAL-NEXT: ori $a0, $a0, 1 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_size_32: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 32 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_size_32: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.d $a2, $a0, 0 +; LA64-UAL-NEXT: ld.d $a3, $a1, 0 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB35_5 +; LA64-UAL-NEXT: # %bb.1: # %loadbb1 +; LA64-UAL-NEXT: ld.d $a2, $a0, 8 +; LA64-UAL-NEXT: ld.d $a3, $a1, 8 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB35_5 +; LA64-UAL-NEXT: # %bb.2: # %loadbb2 +; LA64-UAL-NEXT: ld.d $a2, $a0, 16 +; LA64-UAL-NEXT: ld.d $a3, $a1, 16 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB35_5 +; LA64-UAL-NEXT: # %bb.3: # %loadbb3 +; LA64-UAL-NEXT: ld.d $a0, $a0, 24 +; LA64-UAL-NEXT: ld.d $a1, $a1, 24 +; LA64-UAL-NEXT: revb.d $a2, $a0 +; LA64-UAL-NEXT: revb.d $a3, $a1 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB35_5 +; LA64-UAL-NEXT: # %bb.4: +; LA64-UAL-NEXT: move $a0, $zero +; LA64-UAL-NEXT: ret +; LA64-UAL-NEXT: .LBB35_5: # %res_block +; LA64-UAL-NEXT: sltu $a0, $a2, $a3 +; LA64-UAL-NEXT: sub.d $a0, $zero, $a0 +; LA64-UAL-NEXT: ori $a0, $a0, 1 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_size_32: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 32 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_size_32: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 32 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 32) ret i32 %memcmp @@ -957,16 +2537,74 @@ define signext i32 @memcmp_size_63(ptr %s1, ptr %s2) nounwind { ; LA32-NEXT: addi.w $sp, $sp, 16 ; LA32-NEXT: ret ; -; LA64-LABEL: memcmp_size_63: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 63 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_size_63: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.d $a2, $a0, 0 +; LA64-UAL-NEXT: ld.d $a3, $a1, 0 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB36_9 +; LA64-UAL-NEXT: # %bb.1: # %loadbb1 +; LA64-UAL-NEXT: ld.d $a2, $a0, 8 +; LA64-UAL-NEXT: ld.d $a3, $a1, 8 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB36_9 +; LA64-UAL-NEXT: # %bb.2: # %loadbb2 +; LA64-UAL-NEXT: ld.d $a2, $a0, 16 +; LA64-UAL-NEXT: ld.d $a3, $a1, 16 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB36_9 +; LA64-UAL-NEXT: # %bb.3: # %loadbb3 +; LA64-UAL-NEXT: ld.d $a2, $a0, 24 +; LA64-UAL-NEXT: ld.d $a3, $a1, 24 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB36_9 +; LA64-UAL-NEXT: # %bb.4: # %loadbb4 +; LA64-UAL-NEXT: ld.d $a2, $a0, 32 +; LA64-UAL-NEXT: ld.d $a3, $a1, 32 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB36_9 +; LA64-UAL-NEXT: # %bb.5: # %loadbb5 +; LA64-UAL-NEXT: ld.d $a2, $a0, 40 +; LA64-UAL-NEXT: ld.d $a3, $a1, 40 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB36_9 +; LA64-UAL-NEXT: # %bb.6: # %loadbb6 +; LA64-UAL-NEXT: ld.d $a2, $a0, 48 +; LA64-UAL-NEXT: ld.d $a3, $a1, 48 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB36_9 +; LA64-UAL-NEXT: # %bb.7: # %loadbb7 +; LA64-UAL-NEXT: ld.d $a0, $a0, 55 +; LA64-UAL-NEXT: ld.d $a1, $a1, 55 +; LA64-UAL-NEXT: revb.d $a2, $a0 +; LA64-UAL-NEXT: revb.d $a3, $a1 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB36_9 +; LA64-UAL-NEXT: # %bb.8: +; LA64-UAL-NEXT: move $a0, $zero +; LA64-UAL-NEXT: ret +; LA64-UAL-NEXT: .LBB36_9: # %res_block +; LA64-UAL-NEXT: sltu $a0, $a2, $a3 +; LA64-UAL-NEXT: sub.d $a0, $zero, $a0 +; LA64-UAL-NEXT: ori $a0, $a0, 1 +; LA64-UAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_size_63: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 63 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 63) ret i32 %memcmp @@ -983,16 +2621,74 @@ define signext i32 @memcmp_size_64(ptr %s1, ptr %s2) nounwind { ; LA32-NEXT: addi.w $sp, $sp, 16 ; LA32-NEXT: ret ; -; LA64-LABEL: memcmp_size_64: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 64 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_size_64: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.d $a2, $a0, 0 +; LA64-UAL-NEXT: ld.d $a3, $a1, 0 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB37_9 +; LA64-UAL-NEXT: # %bb.1: # %loadbb1 +; LA64-UAL-NEXT: ld.d $a2, $a0, 8 +; LA64-UAL-NEXT: ld.d $a3, $a1, 8 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB37_9 +; LA64-UAL-NEXT: # %bb.2: # %loadbb2 +; LA64-UAL-NEXT: ld.d $a2, $a0, 16 +; LA64-UAL-NEXT: ld.d $a3, $a1, 16 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB37_9 +; LA64-UAL-NEXT: # %bb.3: # %loadbb3 +; LA64-UAL-NEXT: ld.d $a2, $a0, 24 +; LA64-UAL-NEXT: ld.d $a3, $a1, 24 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB37_9 +; LA64-UAL-NEXT: # %bb.4: # %loadbb4 +; LA64-UAL-NEXT: ld.d $a2, $a0, 32 +; LA64-UAL-NEXT: ld.d $a3, $a1, 32 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB37_9 +; LA64-UAL-NEXT: # %bb.5: # %loadbb5 +; LA64-UAL-NEXT: ld.d $a2, $a0, 40 +; LA64-UAL-NEXT: ld.d $a3, $a1, 40 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB37_9 +; LA64-UAL-NEXT: # %bb.6: # %loadbb6 +; LA64-UAL-NEXT: ld.d $a2, $a0, 48 +; LA64-UAL-NEXT: ld.d $a3, $a1, 48 +; LA64-UAL-NEXT: revb.d $a2, $a2 +; LA64-UAL-NEXT: revb.d $a3, $a3 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB37_9 +; LA64-UAL-NEXT: # %bb.7: # %loadbb7 +; LA64-UAL-NEXT: ld.d $a0, $a0, 56 +; LA64-UAL-NEXT: ld.d $a1, $a1, 56 +; LA64-UAL-NEXT: revb.d $a2, $a0 +; LA64-UAL-NEXT: revb.d $a3, $a1 +; LA64-UAL-NEXT: bne $a2, $a3, .LBB37_9 +; LA64-UAL-NEXT: # %bb.8: +; LA64-UAL-NEXT: move $a0, $zero +; LA64-UAL-NEXT: ret +; LA64-UAL-NEXT: .LBB37_9: # %res_block +; LA64-UAL-NEXT: sltu $a0, $a2, $a3 +; LA64-UAL-NEXT: sub.d $a0, $zero, $a0 +; LA64-UAL-NEXT: ori $a0, $a0, 1 +; LA64-UAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_size_64: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 64 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 64) ret i32 %memcmp @@ -1075,28 +2771,60 @@ entry: } define i1 @memcmp_eq_zero(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: memcmp_eq_zero: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 16 -; LA32-NEXT: bl memcmp -; LA32-NEXT: sltui $a0, $a0, 1 -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_eq_zero: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a2, $a0, 0 +; LA32-UAL-NEXT: ld.w $a3, $a1, 0 +; LA32-UAL-NEXT: ld.w $a4, $a0, 4 +; LA32-UAL-NEXT: ld.w $a5, $a1, 4 +; LA32-UAL-NEXT: ld.w $a6, $a0, 8 +; LA32-UAL-NEXT: ld.w $a7, $a1, 8 +; LA32-UAL-NEXT: ld.w $a0, $a0, 12 +; LA32-UAL-NEXT: ld.w $a1, $a1, 12 +; LA32-UAL-NEXT: xor $a2, $a2, $a3 +; LA32-UAL-NEXT: xor $a3, $a4, $a5 +; LA32-UAL-NEXT: xor $a4, $a6, $a7 +; LA32-UAL-NEXT: xor $a0, $a0, $a1 +; LA32-UAL-NEXT: or $a1, $a2, $a3 +; LA32-UAL-NEXT: or $a0, $a4, $a0 +; LA32-UAL-NEXT: or $a0, $a1, $a0 +; LA32-UAL-NEXT: sltui $a0, $a0, 1 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_eq_zero: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 16 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: sltui $a0, $a0, 1 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_eq_zero: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.d $a2, $a0, 0 +; LA64-UAL-NEXT: ld.d $a3, $a1, 0 +; LA64-UAL-NEXT: ld.d $a0, $a0, 8 +; LA64-UAL-NEXT: ld.d $a1, $a1, 8 +; LA64-UAL-NEXT: xor $a2, $a2, $a3 +; LA64-UAL-NEXT: xor $a0, $a0, $a1 +; LA64-UAL-NEXT: or $a0, $a2, $a0 +; LA64-UAL-NEXT: sltui $a0, $a0, 1 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_eq_zero: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 16 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: sltui $a0, $a0, 1 +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_eq_zero: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 16 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: sltui $a0, $a0, 1 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 16) %ret = icmp eq i32 %memcmp, 0 @@ -1104,28 +2832,66 @@ entry: } define i1 @memcmp_lt_zero(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: memcmp_lt_zero: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 4 -; LA32-NEXT: bl memcmp -; LA32-NEXT: srli.w $a0, $a0, 31 -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_lt_zero: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a0, $a0, 0 +; LA32-UAL-NEXT: ld.w $a1, $a1, 0 +; LA32-UAL-NEXT: srli.w $a2, $a0, 8 +; LA32-UAL-NEXT: lu12i.w $a3, 15 +; LA32-UAL-NEXT: ori $a3, $a3, 3840 +; LA32-UAL-NEXT: and $a2, $a2, $a3 +; LA32-UAL-NEXT: srli.w $a4, $a0, 24 +; LA32-UAL-NEXT: or $a2, $a2, $a4 +; LA32-UAL-NEXT: and $a4, $a0, $a3 +; LA32-UAL-NEXT: slli.w $a4, $a4, 8 +; LA32-UAL-NEXT: slli.w $a0, $a0, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: or $a0, $a0, $a2 +; LA32-UAL-NEXT: srli.w $a2, $a1, 8 +; LA32-UAL-NEXT: and $a2, $a2, $a3 +; LA32-UAL-NEXT: srli.w $a4, $a1, 24 +; LA32-UAL-NEXT: or $a2, $a2, $a4 +; LA32-UAL-NEXT: and $a3, $a1, $a3 +; LA32-UAL-NEXT: slli.w $a3, $a3, 8 +; LA32-UAL-NEXT: slli.w $a1, $a1, 24 +; LA32-UAL-NEXT: or $a1, $a1, $a3 +; LA32-UAL-NEXT: or $a1, $a1, $a2 +; LA32-UAL-NEXT: sltu $a0, $a0, $a1 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_lt_zero: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 4 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: slti $a0, $a0, 0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_lt_zero: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.w $a0, $a0, 0 +; LA64-UAL-NEXT: ld.w $a1, $a1, 0 +; LA64-UAL-NEXT: revb.2w $a0, $a0 +; LA64-UAL-NEXT: addi.w $a0, $a0, 0 +; LA64-UAL-NEXT: revb.2w $a1, $a1 +; LA64-UAL-NEXT: addi.w $a1, $a1, 0 +; LA64-UAL-NEXT: sltu $a0, $a0, $a1 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_lt_zero: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 4 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: srli.w $a0, $a0, 31 +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_lt_zero: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 4 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: slti $a0, $a0, 0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 4) %ret = icmp slt i32 %memcmp, 0 @@ -1133,28 +2899,66 @@ entry: } define i1 @memcmp_gt_zero(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: memcmp_gt_zero: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 4 -; LA32-NEXT: bl memcmp -; LA32-NEXT: slt $a0, $zero, $a0 -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_gt_zero: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a0, $a0, 0 +; LA32-UAL-NEXT: ld.w $a1, $a1, 0 +; LA32-UAL-NEXT: srli.w $a2, $a0, 8 +; LA32-UAL-NEXT: lu12i.w $a3, 15 +; LA32-UAL-NEXT: ori $a3, $a3, 3840 +; LA32-UAL-NEXT: and $a2, $a2, $a3 +; LA32-UAL-NEXT: srli.w $a4, $a0, 24 +; LA32-UAL-NEXT: or $a2, $a2, $a4 +; LA32-UAL-NEXT: and $a4, $a0, $a3 +; LA32-UAL-NEXT: slli.w $a4, $a4, 8 +; LA32-UAL-NEXT: slli.w $a0, $a0, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: or $a0, $a0, $a2 +; LA32-UAL-NEXT: srli.w $a2, $a1, 8 +; LA32-UAL-NEXT: and $a2, $a2, $a3 +; LA32-UAL-NEXT: srli.w $a4, $a1, 24 +; LA32-UAL-NEXT: or $a2, $a2, $a4 +; LA32-UAL-NEXT: and $a3, $a1, $a3 +; LA32-UAL-NEXT: slli.w $a3, $a3, 8 +; LA32-UAL-NEXT: slli.w $a1, $a1, 24 +; LA32-UAL-NEXT: or $a1, $a1, $a3 +; LA32-UAL-NEXT: or $a1, $a1, $a2 +; LA32-UAL-NEXT: sltu $a0, $a1, $a0 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_gt_zero: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 4 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: slt $a0, $zero, $a0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_gt_zero: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.w $a0, $a0, 0 +; LA64-UAL-NEXT: ld.w $a1, $a1, 0 +; LA64-UAL-NEXT: revb.2w $a0, $a0 +; LA64-UAL-NEXT: addi.w $a0, $a0, 0 +; LA64-UAL-NEXT: revb.2w $a1, $a1 +; LA64-UAL-NEXT: addi.w $a1, $a1, 0 +; LA64-UAL-NEXT: sltu $a0, $a1, $a0 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_gt_zero: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 4 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: slt $a0, $zero, $a0 +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_gt_zero: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 4 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: slt $a0, $zero, $a0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 4) %ret = icmp sgt i32 %memcmp, 0 @@ -1162,28 +2966,68 @@ entry: } define i1 @memcmp_le_zero(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: memcmp_le_zero: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 4 -; LA32-NEXT: bl memcmp -; LA32-NEXT: slti $a0, $a0, 1 -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_le_zero: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a0, $a0, 0 +; LA32-UAL-NEXT: ld.w $a1, $a1, 0 +; LA32-UAL-NEXT: srli.w $a2, $a0, 8 +; LA32-UAL-NEXT: lu12i.w $a3, 15 +; LA32-UAL-NEXT: ori $a3, $a3, 3840 +; LA32-UAL-NEXT: and $a2, $a2, $a3 +; LA32-UAL-NEXT: srli.w $a4, $a0, 24 +; LA32-UAL-NEXT: or $a2, $a2, $a4 +; LA32-UAL-NEXT: and $a4, $a0, $a3 +; LA32-UAL-NEXT: slli.w $a4, $a4, 8 +; LA32-UAL-NEXT: slli.w $a0, $a0, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: or $a0, $a0, $a2 +; LA32-UAL-NEXT: srli.w $a2, $a1, 8 +; LA32-UAL-NEXT: and $a2, $a2, $a3 +; LA32-UAL-NEXT: srli.w $a4, $a1, 24 +; LA32-UAL-NEXT: or $a2, $a2, $a4 +; LA32-UAL-NEXT: and $a3, $a1, $a3 +; LA32-UAL-NEXT: slli.w $a3, $a3, 8 +; LA32-UAL-NEXT: slli.w $a1, $a1, 24 +; LA32-UAL-NEXT: or $a1, $a1, $a3 +; LA32-UAL-NEXT: or $a1, $a1, $a2 +; LA32-UAL-NEXT: sltu $a0, $a1, $a0 +; LA32-UAL-NEXT: xori $a0, $a0, 1 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_le_zero: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 4 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: slti $a0, $a0, 1 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_le_zero: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.w $a0, $a0, 0 +; LA64-UAL-NEXT: ld.w $a1, $a1, 0 +; LA64-UAL-NEXT: revb.2w $a0, $a0 +; LA64-UAL-NEXT: addi.w $a0, $a0, 0 +; LA64-UAL-NEXT: revb.2w $a1, $a1 +; LA64-UAL-NEXT: addi.w $a1, $a1, 0 +; LA64-UAL-NEXT: sltu $a0, $a1, $a0 +; LA64-UAL-NEXT: xori $a0, $a0, 1 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_le_zero: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 4 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: slti $a0, $a0, 1 +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_le_zero: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 4 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: slti $a0, $a0, 1 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 4) %ret = icmp slt i32 %memcmp, 1 @@ -1191,37 +3035,72 @@ entry: } define i1 @memcmp_ge_zero(ptr %s1, ptr %s2) nounwind { -; LA32-LABEL: memcmp_ge_zero: -; LA32: # %bb.0: # %entry -; LA32-NEXT: addi.w $sp, $sp, -16 -; LA32-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill -; LA32-NEXT: ori $a2, $zero, 4 -; LA32-NEXT: bl memcmp -; LA32-NEXT: addi.w $a1, $zero, -1 -; LA32-NEXT: slt $a0, $a1, $a0 -; LA32-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload -; LA32-NEXT: addi.w $sp, $sp, 16 -; LA32-NEXT: ret +; LA32-UAL-LABEL: memcmp_ge_zero: +; LA32-UAL: # %bb.0: # %entry +; LA32-UAL-NEXT: ld.w $a0, $a0, 0 +; LA32-UAL-NEXT: ld.w $a1, $a1, 0 +; LA32-UAL-NEXT: srli.w $a2, $a0, 8 +; LA32-UAL-NEXT: lu12i.w $a3, 15 +; LA32-UAL-NEXT: ori $a3, $a3, 3840 +; LA32-UAL-NEXT: and $a2, $a2, $a3 +; LA32-UAL-NEXT: srli.w $a4, $a0, 24 +; LA32-UAL-NEXT: or $a2, $a2, $a4 +; LA32-UAL-NEXT: and $a4, $a0, $a3 +; LA32-UAL-NEXT: slli.w $a4, $a4, 8 +; LA32-UAL-NEXT: slli.w $a0, $a0, 24 +; LA32-UAL-NEXT: or $a0, $a0, $a4 +; LA32-UAL-NEXT: or $a0, $a0, $a2 +; LA32-UAL-NEXT: srli.w $a2, $a1, 8 +; LA32-UAL-NEXT: and $a2, $a2, $a3 +; LA32-UAL-NEXT: srli.w $a4, $a1, 24 +; LA32-UAL-NEXT: or $a2, $a2, $a4 +; LA32-UAL-NEXT: and $a3, $a1, $a3 +; LA32-UAL-NEXT: slli.w $a3, $a3, 8 +; LA32-UAL-NEXT: slli.w $a1, $a1, 24 +; LA32-UAL-NEXT: or $a1, $a1, $a3 +; LA32-UAL-NEXT: or $a1, $a1, $a2 +; LA32-UAL-NEXT: sltu $a0, $a0, $a1 +; LA32-UAL-NEXT: xori $a0, $a0, 1 +; LA32-UAL-NEXT: ret ; -; LA64-LABEL: memcmp_ge_zero: -; LA64: # %bb.0: # %entry -; LA64-NEXT: addi.d $sp, $sp, -16 -; LA64-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; LA64-NEXT: ori $a2, $zero, 4 -; LA64-NEXT: pcaddu18i $ra, %call36(memcmp) -; LA64-NEXT: jirl $ra, $ra, 0 -; LA64-NEXT: addi.w $a1, $zero, -1 -; LA64-NEXT: slt $a0, $a1, $a0 -; LA64-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; LA64-NEXT: addi.d $sp, $sp, 16 -; LA64-NEXT: ret +; LA64-UAL-LABEL: memcmp_ge_zero: +; LA64-UAL: # %bb.0: # %entry +; LA64-UAL-NEXT: ld.w $a0, $a0, 0 +; LA64-UAL-NEXT: ld.w $a1, $a1, 0 +; LA64-UAL-NEXT: revb.2w $a0, $a0 +; LA64-UAL-NEXT: addi.w $a0, $a0, 0 +; LA64-UAL-NEXT: revb.2w $a1, $a1 +; LA64-UAL-NEXT: addi.w $a1, $a1, 0 +; LA64-UAL-NEXT: sltu $a0, $a0, $a1 +; LA64-UAL-NEXT: xori $a0, $a0, 1 +; LA64-UAL-NEXT: ret +; +; LA32-NUAL-LABEL: memcmp_ge_zero: +; LA32-NUAL: # %bb.0: # %entry +; LA32-NUAL-NEXT: addi.w $sp, $sp, -16 +; LA32-NUAL-NEXT: st.w $ra, $sp, 12 # 4-byte Folded Spill +; LA32-NUAL-NEXT: ori $a2, $zero, 4 +; LA32-NUAL-NEXT: bl memcmp +; LA32-NUAL-NEXT: addi.w $a1, $zero, -1 +; LA32-NUAL-NEXT: slt $a0, $a1, $a0 +; LA32-NUAL-NEXT: ld.w $ra, $sp, 12 # 4-byte Folded Reload +; LA32-NUAL-NEXT: addi.w $sp, $sp, 16 +; LA32-NUAL-NEXT: ret +; +; LA64-NUAL-LABEL: memcmp_ge_zero: +; LA64-NUAL: # %bb.0: # %entry +; LA64-NUAL-NEXT: addi.d $sp, $sp, -16 +; LA64-NUAL-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill +; LA64-NUAL-NEXT: ori $a2, $zero, 4 +; LA64-NUAL-NEXT: pcaddu18i $ra, %call36(memcmp) +; LA64-NUAL-NEXT: jirl $ra, $ra, 0 +; LA64-NUAL-NEXT: addi.w $a1, $zero, -1 +; LA64-NUAL-NEXT: slt $a0, $a1, $a0 +; LA64-NUAL-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload +; LA64-NUAL-NEXT: addi.d $sp, $sp, 16 +; LA64-NUAL-NEXT: ret entry: %memcmp = call signext i32 @memcmp(ptr %s1, ptr %s2, iGRLen 4) %ret = icmp sgt i32 %memcmp, -1 ret i1 %ret } -;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line: -; LA32-NUAL: {{.*}} -; LA32-UAL: {{.*}} -; LA64-NUAL: {{.*}} -; LA64-UAL: {{.*}} diff --git a/llvm/test/CodeGen/LoongArch/memcmp.ll b/llvm/test/CodeGen/LoongArch/memcmp.ll index c4aaf9a75a852..c3811c0357793 100644 --- a/llvm/test/CodeGen/LoongArch/memcmp.ll +++ b/llvm/test/CodeGen/LoongArch/memcmp.ll @@ -7,15 +7,24 @@ define signext i32 @test1(ptr %buffer1, ptr %buffer2) { ; CHECK-LABEL: test1: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: addi.d $sp, $sp, -16 -; CHECK-NEXT: .cfi_def_cfa_offset 16 -; CHECK-NEXT: st.d $ra, $sp, 8 # 8-byte Folded Spill -; CHECK-NEXT: .cfi_offset 1, -8 -; CHECK-NEXT: ori $a2, $zero, 16 -; CHECK-NEXT: pcaddu18i $ra, %call36(memcmp) -; CHECK-NEXT: jirl $ra, $ra, 0 -; CHECK-NEXT: ld.d $ra, $sp, 8 # 8-byte Folded Reload -; CHECK-NEXT: addi.d $sp, $sp, 16 +; CHECK-NEXT: ld.d $a2, $a0, 0 +; CHECK-NEXT: ld.d $a3, $a1, 0 +; CHECK-NEXT: revb.d $a2, $a2 +; CHECK-NEXT: revb.d $a3, $a3 +; CHECK-NEXT: bne $a2, $a3, .LBB0_3 +; CHECK-NEXT: # %bb.1: # %loadbb1 +; CHECK-NEXT: ld.d $a0, $a0, 8 +; CHECK-NEXT: ld.d $a1, $a1, 8 +; CHECK-NEXT: revb.d $a2, $a0 +; CHECK-NEXT: revb.d $a3, $a1 +; CHECK-NEXT: bne $a2, $a3, .LBB0_3 +; CHECK-NEXT: # %bb.2: +; CHECK-NEXT: move $a0, $zero +; CHECK-NEXT: ret +; CHECK-NEXT: .LBB0_3: # %res_block +; CHECK-NEXT: sltu $a0, $a2, $a3 +; CHECK-NEXT: sub.d $a0, $zero, $a0 +; CHECK-NEXT: ori $a0, $a0, 1 ; CHECK-NEXT: ret entry: %call = call signext i32 @memcmp(ptr %buffer1, ptr %buffer2, i64 16)