diff --git a/llvm/include/llvm/CodeGen/GlobalISel/CombinerHelper.h b/llvm/include/llvm/CodeGen/GlobalISel/CombinerHelper.h --- a/llvm/include/llvm/CodeGen/GlobalISel/CombinerHelper.h +++ b/llvm/include/llvm/CodeGen/GlobalISel/CombinerHelper.h @@ -784,6 +784,9 @@ /// Match shifts greater or equal to the bitwidth of the operation. bool matchShiftsTooBig(MachineInstr &MI); + /// Match constant LHS ops that should be commuted. + bool matchCommuteConstantToRHS(MachineInstr &MI); + private: /// Given a non-indexed load or store instruction \p MI, find an offset that /// can be usefully and legally folded into it as a post-indexing operation. diff --git a/llvm/include/llvm/Target/GlobalISel/Combine.td b/llvm/include/llvm/Target/GlobalISel/Combine.td --- a/llvm/include/llvm/Target/GlobalISel/Combine.td +++ b/llvm/include/llvm/Target/GlobalISel/Combine.td @@ -389,9 +389,8 @@ // TODO: handle compares (currently not marked as isCommutable) def commute_constant_to_rhs : GICombineRule< (defs root:$root), - (match (wip_match_opcode G_ADD, G_MUL, G_AND, G_OR, G_XOR):$root, [{ - return getIConstantVRegVal(${root}->getOperand(1).getReg(), MRI).has_value(); - }]), + (match (wip_match_opcode G_ADD, G_MUL, G_AND, G_OR, G_XOR):$root, + [{ return Helper.matchCommuteConstantToRHS(*${root}); }]), (apply [{ Observer.changingInstr(*${root}); Register LHSReg = ${root}->getOperand(1).getReg(); diff --git a/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp b/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp --- a/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp +++ b/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp @@ -6013,6 +6013,25 @@ return matchUnaryPredicate(MRI, ShiftReg, IsShiftTooBig); } +bool CombinerHelper::matchCommuteConstantToRHS(MachineInstr &MI) { + Register LHS = MI.getOperand(1).getReg(); + Register RHS = MI.getOperand(2).getReg(); + auto *LHSDef = MRI.getVRegDef(LHS); + if (getIConstantVRegVal(LHS, MRI).has_value()) + return true; + + // LHS may be a G_CONSTANT_FOLD_BARRIER. If so we commute + // as long as we don't already have a constant on the RHS. + if (LHSDef->getOpcode() != TargetOpcode::G_CONSTANT_FOLD_BARRIER) + return false; + auto *RHSDef = MRI.getVRegDef(RHS); + if (RHSDef->getOpcode() != TargetOpcode::G_CONSTANT_FOLD_BARRIER && + !getIConstantVRegVal(RHS, MRI).has_value()) + return true; + + return false; +} + bool CombinerHelper::tryCombine(MachineInstr &MI) { if (tryCombineCopy(MI)) return true; diff --git a/llvm/test/CodeGen/AArch64/GlobalISel/combine-const-fold-barrier-rhs.mir b/llvm/test/CodeGen/AArch64/GlobalISel/combine-const-fold-barrier-rhs.mir new file mode 100644 --- /dev/null +++ b/llvm/test/CodeGen/AArch64/GlobalISel/combine-const-fold-barrier-rhs.mir @@ -0,0 +1,82 @@ +# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 2 +# RUN: llc -mtriple aarch64 -run-pass=aarch64-prelegalizer-combiner -verify-machineinstrs %s -o - | FileCheck %s +--- +name: cfb_lhs +alignment: 4 +tracksRegLiveness: true +body: | + bb.1: + liveins: $w0 + + ; CHECK-LABEL: name: cfb_lhs + ; CHECK: liveins: $w0 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:_(s32) = COPY $w0 + ; CHECK-NEXT: %cst:_(s32) = G_CONSTANT i32 1 + ; CHECK-NEXT: %cfb:_(s32) = G_CONSTANT_FOLD_BARRIER %cst + ; CHECK-NEXT: %add:_(s32) = G_ADD [[COPY]], %cfb + ; CHECK-NEXT: $w0 = COPY %add(s32) + ; CHECK-NEXT: RET_ReallyLR + %0:_(s32) = COPY $w0 + %cst:_(s32) = G_CONSTANT i32 1 + %cfb:_(s32) = G_CONSTANT_FOLD_BARRIER %cst + %cst2:_(s32) = G_CONSTANT i32 2 + %add:_(s32) = G_ADD %cfb, %0 + $w0 = COPY %add + RET_ReallyLR + +... +--- +name: cfb_lhs_cfb_already_rhs +alignment: 4 +tracksRegLiveness: true +body: | + bb.1: + liveins: $w0 + + ; CHECK-LABEL: name: cfb_lhs_cfb_already_rhs + ; CHECK: liveins: $w0 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: %cst:_(s32) = G_CONSTANT i32 1 + ; CHECK-NEXT: %cfb:_(s32) = G_CONSTANT_FOLD_BARRIER %cst + ; CHECK-NEXT: %cst2:_(s32) = G_CONSTANT i32 2 + ; CHECK-NEXT: %cfb2:_(s32) = G_CONSTANT_FOLD_BARRIER %cst2 + ; CHECK-NEXT: %add:_(s32) = G_ADD %cfb, %cfb2 + ; CHECK-NEXT: $w0 = COPY %add(s32) + ; CHECK-NEXT: RET_ReallyLR + %0:_(s32) = COPY $w0 + %cst:_(s32) = G_CONSTANT i32 1 + %cfb:_(s32) = G_CONSTANT_FOLD_BARRIER %cst + %cst2:_(s32) = G_CONSTANT i32 2 + %cfb2:_(s32) = G_CONSTANT_FOLD_BARRIER %cst2 + %add:_(s32) = G_ADD %cfb, %cfb2 + $w0 = COPY %add + RET_ReallyLR + +... +--- +name: cfb_lhs_cst_on_rhs +alignment: 4 +tracksRegLiveness: true +body: | + bb.1: + liveins: $w0 + + ; CHECK-LABEL: name: cfb_lhs_cst_on_rhs + ; CHECK: liveins: $w0 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: %cst:_(s32) = G_CONSTANT i32 1 + ; CHECK-NEXT: %cfb:_(s32) = G_CONSTANT_FOLD_BARRIER %cst + ; CHECK-NEXT: %cst2:_(s32) = G_CONSTANT i32 2 + ; CHECK-NEXT: %add:_(s32) = G_ADD %cfb, %cst2 + ; CHECK-NEXT: $w0 = COPY %add(s32) + ; CHECK-NEXT: RET_ReallyLR + %0:_(s32) = COPY $w0 + %cst:_(s32) = G_CONSTANT i32 1 + %cfb:_(s32) = G_CONSTANT_FOLD_BARRIER %cst + %cst2:_(s32) = G_CONSTANT i32 2 + %add:_(s32) = G_ADD %cfb, %cst2 + $w0 = COPY %add + RET_ReallyLR + +...