[AArch64][GlobalISel] Fold G_ADD into the cset for G_ICMP

[paquette]

Summary

When we have a G_ADD which is fed by a G_ICMP on one side, we can fold it into the cset for the G_ICMP.

e.g. Given

%cmp = G_ICMP ... %x, %y
%add = G_ADD %cmp, %z

We would normally emit a cmp, cset, and add.

However, %add is either %z or %z + 1. So, we can just use %z as the source of the cset rather than wzr, saving an instruction.

This would probably be cleaner in AArch64PostLegalizerLowering, but we'd need to change the way we represent G_ICMP to do that, I think. For now, it's easiest to implement in selection.

This is a 0.1% code size improvement on CTMark/pairlocalalign at -Os.

Example: https://godbolt.org/z/7KdrP8

[aemerson]

llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp

2181

What if the type is s64? emitCSetForICMP seems to only work for s32.

[paquette]

llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp

2181

I don't think that should happen, because G_ICMP can only output a s32. So, any G_ADD fed by it should take in + produce a s32.

That being said, it's probably better to early exit if we don't have a s32 anyway.

(And maybe emitCSetForICMP should be changed to a general purpose emitCSet function.)

[paquette]

Check the G_ADD type before performing the optimization, and add a vector test.

llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp

private:
   /// Helper function for selecting G_FCONSTANT. If the G_FCONSTANT can be
   /// materialized using a FMOV instruction, then update MI and return it.
   /// Otherwise, do nothing and return a nullptr.
   MachineInstr *emitFMovForFConstant(MachineInstr &MI,
                                      MachineRegisterInfo &MRI) const;
 
   /// Emit a CSet for an integer compare.
   ///
-  /// \p DefReg is expected to be a 32-bit scalar register.
+  /// \p DefReg and \p SrcReg are expected to be 32-bit scalar registers.
   MachineInstr *emitCSetForICMP(Register DefReg, unsigned Pred,
-                                MachineIRBuilder &MIRBuilder) const;
+                                MachineIRBuilder &MIRBuilder,
+                                Register SrcReg = AArch64::WZR) const;
   /// Emit a CSet for a FP compare.
   ///
   /// \p Dst is expected to be a 32-bit scalar register.
   MachineInstr *emitCSetForFCmp(Register Dst, CmpInst::Predicate Pred,
                                 MachineIRBuilder &MIRBuilder) const;
 
   /// Emit the overflow op for \p Opcode.
   ///
if (Ty.getSizeInBits() == 64) {
       I.getOperand(1).ChangeToRegister(AArch64::WZR, false);
       RBI.constrainGenericRegister(DefReg, AArch64::GPR32RegClass, MRI);
     } else
       return false;
 
     I.setDesc(TII.get(TargetOpcode::COPY));
     return true;
   }
+
+  case TargetOpcode::G_ADD: {
+    // Check if this is being fed by a G_ICMP on either side.
+    //
+    // (cmp pred, x, y) + z
+    //
+    // In the above case, when the cmp is true, we increment z by 1. So, we can
+    // fold the add into the cset for the cmp by using cinc.
+    //
+    // FIXME: This would probably be a lot nicer in PostLegalizerLowering.
+    Register X = I.getOperand(1).getReg();
+    Register CmpReg = I.getOperand(2).getReg();
+    MachineInstr *Cmp = getOpcodeDef(TargetOpcode::G_ICMP, CmpReg, MRI);
+    if (!Cmp) {
+      std::swap(X, CmpReg);
+      Cmp = getOpcodeDef(TargetOpcode::G_ICMP, CmpReg, MRI);
+      if (!Cmp)
+        return false;
+    }
+    MachineIRBuilder MIRBuilder(I);
+    auto Pred =
+        static_cast<CmpInst::Predicate>(Cmp->getOperand(1).getPredicate());
+    emitIntegerCompare(Cmp->getOperand(2), Cmp->getOperand(3),
+                       Cmp->getOperand(1), MIRBuilder);
+    emitCSetForICMP(I.getOperand(0).getReg(), Pred, MIRBuilder, X);
+    I.eraseFromParent();
+    return true;
+  }
   default:
     return false;
   }
 }
 
 bool AArch64InstructionSelector::select(MachineInstr &I) {
   assert(I.getParent() && "Instruction should be in a basic block!");
   assert(I.getParent()->getParent() && "Instruction should be in a function!");
MachineInstr *AArch64InstructionSelector::emitFMovForFConstant(
   unsigned MovOpc = DefSize == 32 ? AArch64::FMOVSi : AArch64::FMOVDi;
   I.setDesc(TII.get(MovOpc));
   constrainSelectedInstRegOperands(I, TII, TRI, RBI);
   return &I;
 }
 
 MachineInstr *
 AArch64InstructionSelector::emitCSetForICMP(Register DefReg, unsigned Pred,
-                                     MachineIRBuilder &MIRBuilder) const {
+                                            MachineIRBuilder &MIRBuilder,
+                                            Register SrcReg) const {
   // CSINC increments the result when the predicate is false. Invert it.
   const AArch64CC::CondCode InvCC = changeICMPPredToAArch64CC(
       CmpInst::getInversePredicate((CmpInst::Predicate)Pred));
-  auto I =
-      MIRBuilder
-    .buildInstr(AArch64::CSINCWr, {DefReg}, {Register(AArch64::WZR), Register(AArch64::WZR)})
+  auto I = MIRBuilder.buildInstr(AArch64::CSINCWr, {DefReg}, {SrcReg, SrcReg})
                .addImm(InvCC);
   constrainSelectedInstRegOperands(*I, TII, TRI, RBI);
   return &*I;
 }
 
 std::pair<MachineInstr *, AArch64CC::CondCode>
 AArch64InstructionSelector::emitOverflowOp(unsigned Opcode, Register Dst,
                                            MachineOperand &LHS,
                                            MachineOperand &RHS,

llvm/test/CodeGen/AArch64/GlobalISel/select-cmp.mir

bb.0:
     %ext:gpr(s64) = G_ZEXT %reg0(s32)
     %cst:gpr(s64) = G_CONSTANT i64 5
     %shift:gpr(s64) = G_SHL %ext, %cst(s64)
     %cmp:gpr(s32) = G_ICMP intpred(ugt), %reg1(s64), %shift
     $w0 = COPY %cmp(s32)
     RET_ReallyLR implicit $w0
 
 ...
+---
+name:            cmp_add_rhs
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo: {}
+body:             |
+  bb.0:
+    liveins: $w0, $w1, $w2
+
+    ; The CSINC should use the add's RHS.
+
+    ; CHECK-LABEL: name: cmp_add_rhs
+    ; CHECK: liveins: $w0, $w1, $w2
+    ; CHECK: %cmp_lhs:gpr32 = COPY $w0
+    ; CHECK: %cmp_rhs:gpr32 = COPY $w1
+    ; CHECK: %add_rhs:gpr32 = COPY $w2
+    ; CHECK: [[SUBSWrr:%[0-9]+]]:gpr32 = SUBSWrr %cmp_lhs, %cmp_rhs, implicit-def $nzcv
+    ; CHECK: %add:gpr32 = CSINCWr %add_rhs, %add_rhs, 1, implicit $nzcv
+    ; CHECK: $w0 = COPY %add
+    ; CHECK: RET_ReallyLR implicit $w0
+    %cmp_lhs:gpr(s32) = COPY $w0
+    %cmp_rhs:gpr(s32) = COPY $w1
+    %add_rhs:gpr(s32) = COPY $w2
+    %cmp:gpr(s32) = G_ICMP intpred(eq), %cmp_lhs(s32), %cmp_rhs
+    %add:gpr(s32) = G_ADD %cmp, %add_rhs
+    $w0 = COPY %add(s32)
+    RET_ReallyLR implicit $w0
+
+...
+---
+name:            cmp_add_lhs
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo: {}
+body:             |
+  bb.0:
+    liveins: $w0, $w1, $w2
+
+    ; The CSINC should use the add's LHS.
+
+    ; CHECK-LABEL: name: cmp_add_lhs
+    ; CHECK: liveins: $w0, $w1, $w2
+    ; CHECK: %cmp_lhs:gpr32 = COPY $w0
+    ; CHECK: %cmp_rhs:gpr32 = COPY $w1
+    ; CHECK: %add_lhs:gpr32 = COPY $w2
+    ; CHECK: [[SUBSWrr:%[0-9]+]]:gpr32 = SUBSWrr %cmp_lhs, %cmp_rhs, implicit-def $nzcv
+    ; CHECK: %add:gpr32 = CSINCWr %add_lhs, %add_lhs, 1, implicit $nzcv
+    ; CHECK: $w0 = COPY %add
+    ; CHECK: RET_ReallyLR implicit $w0
+    %cmp_lhs:gpr(s32) = COPY $w0
+    %cmp_rhs:gpr(s32) = COPY $w1
+    %add_lhs:gpr(s32) = COPY $w2
+    %cmp:gpr(s32) = G_ICMP intpred(eq), %cmp_lhs(s32), %cmp_rhs
+    %add:gpr(s32) = G_ADD %add_lhs, %cmp
+    $w0 = COPY %add(s32)
+    RET_ReallyLR implicit $w0