[RISCV] Add fmin/fmax scalar instructions to isAssociativeAndCommutative

[HsiangKai]

Summary

Follow-up patch of D140530.

We can add FMIN, FMAX to isAssociativeAndCommutative to
increase instruction-level parallelism by the existing MachineCombiner
pass.

[craig.topper]

llvm/lib/Target/RISCV/RISCVInstrInfo.cpp

1367

I don't think FMIN/FMAX require Reassoc or FmNsz

[HsiangKai]

No need to check RmReassoc and FmNsz for fmin/fmax.

[asi-sc]

llvm/test/CodeGen/RISCV/machine-combiner.ll

743

I believe we can remove nsz and reassoc flags here as well.

[HsiangKai]

llvm/test/CodeGen/RISCV/machine-combiner.ll

743

Thank you. I copied the test cases and forgot to remove them.

[HsiangKai]

Remove unnecessary fast-math flags in the test cases.

[craig.topper]

LGTM

[HsiangKai]

Rebase.

[asi-sc]

LGTM too.

llvm/lib/Target/RISCV/RISCVInstrInfo.cpp

bool RISCVInstrInfo::isAssociativeAndCommutative(const MachineInstr &Inst,
   unsigned Opc = Inst.getOpcode();
   if (Invert) {
     auto InverseOpcode = getInverseOpcode(Opc);
     if (!InverseOpcode)
       return false;
     Opc = *InverseOpcode;
   }
 
   if (isFADD(Opc) || isFMUL(Opc))
     return Inst.getFlag(MachineInstr::MIFlag::FmReassoc) &&
            Inst.getFlag(MachineInstr::MIFlag::FmNsz);
 
   switch (Opc) {
   default:
     return false;
   case RISCV::ADD:
   case RISCV::ADDW:
   case RISCV::AND:
   case RISCV::OR:
   case RISCV::XOR:
   case RISCV::MUL:
   case RISCV::MULW:
+  case RISCV::FMIN_H:
+  case RISCV::FMIN_S:
+  case RISCV::FMIN_D:
+  case RISCV::FMAX_H:
+  case RISCV::FMAX_S:
+  case RISCV::FMAX_D:
     return true;
   }
 
   return false;
 }
 
 std::optional<unsigned>
 RISCVInstrInfo::getInverseOpcode(unsigned Opcode) const {

llvm/test/CodeGen/RISCV/machine-combiner.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc -mtriple=riscv64 -mattr=+d -verify-machineinstrs -mcpu=sifive-u74 \
+; RUN: llc -mtriple=riscv64 -mattr=+d,+zfh -verify-machineinstrs -mcpu=sifive-u74 \
 ; RUN: -O1 -riscv-enable-machine-combiner=true < %s | \
 ; RUN: FileCheck %s
 
 define double @test_reassoc_fadd1(double %a0, double %a1, double %a2, double %a3) {
 ; CHECK-LABEL: test_reassoc_fadd1:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    fadd.d ft0, fa0, fa1
 ; CHECK-NEXT:    fadd.d ft1, fa2, fa3

 ; CHECK-NEXT:    mul a1, a2, a3
 ; CHECK-NEXT:    mul a0, a0, a1
 ; CHECK-NEXT:    ret
   %t0 = mul i64 %a0, %a1
   %t1 = mul i64 %t0, %a2
   %t2 = mul i64 %t1, %a3
   ret i64 %t2
 }
+
+define half @test_fmin_f16(half %a0, half %a1, half %a2, half %a3) {
+; CHECK-LABEL: test_fmin_f16:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    fmin.h ft0, fa0, fa1
+; CHECK-NEXT:    fmin.h ft1, fa2, fa3
+; CHECK-NEXT:    fmin.h fa0, ft0, ft1
+; CHECK-NEXT:    ret
+  %t0 = call half @llvm.minnum.f16(half %a0, half %a1)
+  %t1 = call half @llvm.minnum.f16(half %t0, half %a2)
+  %t2 = call half @llvm.minnum.f16(half %t1, half %a3)
+  ret half %t2
+}
+
+define float @test_fmin_f32(float %a0, float %a1, float %a2, float %a3) {
+; CHECK-LABEL: test_fmin_f32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    fmin.s ft0, fa0, fa1
+; CHECK-NEXT:    fmin.s ft1, fa2, fa3
+; CHECK-NEXT:    fmin.s fa0, ft0, ft1
+; CHECK-NEXT:    ret
+  %t0 = call float @llvm.minnum.f32(float %a0, float %a1)
+  %t1 = call float @llvm.minnum.f32(float %t0, float %a2)
+  %t2 = call float @llvm.minnum.f32(float %t1, float %a3)
+  ret float %t2
+}
+
+define double @test_fmin_f64(double %a0, double %a1, double %a2, double %a3) {
+; CHECK-LABEL: test_fmin_f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    fmin.d ft0, fa0, fa1
+; CHECK-NEXT:    fmin.d ft1, fa2, fa3
+; CHECK-NEXT:    fmin.d fa0, ft0, ft1
+; CHECK-NEXT:    ret
+  %t0 = call double @llvm.minnum.f64(double %a0, double %a1)
+  %t1 = call double @llvm.minnum.f64(double %t0, double %a2)
+  %t2 = call double @llvm.minnum.f64(double %t1, double %a3)
+  ret double %t2
+}
+
+define half @test_fmax_f16(half %a0, half %a1, half %a2, half %a3) {
+; CHECK-LABEL: test_fmax_f16:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    fmax.h ft0, fa0, fa1
+; CHECK-NEXT:    fmax.h ft1, fa2, fa3
+; CHECK-NEXT:    fmax.h fa0, ft0, ft1
+; CHECK-NEXT:    ret
+  %t0 = call half @llvm.maxnum.f16(half %a0, half %a1)
+  %t1 = call half @llvm.maxnum.f16(half %t0, half %a2)
+  %t2 = call half @llvm.maxnum.f16(half %t1, half %a3)
+  ret half %t2
+}
+
+define float @test_fmax_f32(float %a0, float %a1, float %a2, float %a3) {
+; CHECK-LABEL: test_fmax_f32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    fmax.s ft0, fa0, fa1
+; CHECK-NEXT:    fmax.s ft1, fa2, fa3
+; CHECK-NEXT:    fmax.s fa0, ft0, ft1
+; CHECK-NEXT:    ret
+  %t0 = call float @llvm.maxnum.f32(float %a0, float %a1)
+  %t1 = call float @llvm.maxnum.f32(float %t0, float %a2)
+  %t2 = call float @llvm.maxnum.f32(float %t1, float %a3)
+  ret float %t2
+}
+
+define double @test_fmax_f64(double %a0, double %a1, double %a2, double %a3) {
+; CHECK-LABEL: test_fmax_f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    fmax.d ft0, fa0, fa1
+; CHECK-NEXT:    fmax.d ft1, fa2, fa3
+; CHECK-NEXT:    fmax.d fa0, ft0, ft1
+; CHECK-NEXT:    ret
+  %t0 = call double @llvm.maxnum.f64(double %a0, double %a1)
+  %t1 = call double @llvm.maxnum.f64(double %t0, double %a2)
+  %t2 = call double @llvm.maxnum.f64(double %t1, double %a3)
+  ret double %t2
+}
+
+declare half @llvm.minnum.f16(half, half)
+declare float @llvm.minnum.f32(float, float)
+declare double @llvm.minnum.f64(double, double)
+declare half @llvm.maxnum.f16(half, half)
+declare float @llvm.maxnum.f32(float, float)
+declare double @llvm.maxnum.f64(double, double)