[X86][SSE] Float comparisons can sometimes be safely commuted

[RKSimon]

Summary

For ordered, unordered, equal and not-equal tests, packed float and double comparison instructions can be safely commuted without affecting the results. This patch checks the comparison mode of the (v)cmpps + (v)cmppd instructions and commutes the result if it can.

[qcolombet]

Hi Simon,

LGTM.

Thanks,
-Quentin

llvm/trunk/lib/Target/X86/X86InstrInfo.cpp

case X86::VPCLMULQDQrr:{
     if (NewMI) {
       MachineFunction &MF = *MI->getParent()->getParent();
       MI = MF.CloneMachineInstr(MI);
       NewMI = false;
     }
     MI->getOperand(3).setImm((Src1Hi << 4) | (Src2Hi >> 4));
     return TargetInstrInfo::commuteInstruction(MI, NewMI);
   }
+  case X86::CMPPDrri:
+  case X86::CMPPSrri:
+  case X86::VCMPPDrri:
+  case X86::VCMPPSrri:
+  case X86::VCMPPDYrri:
+  case X86::VCMPPSYrri: {
+    // Float comparison can be safely commuted for
+    // Ordered/Unordered/Equal/NotEqual tests
+    unsigned Imm = MI->getOperand(3).getImm() & 0x7;
+    switch (Imm) {
+    case 0x00: // EQUAL
+    case 0x03: // UNORDERED
+    case 0x04: // NOT EQUAL
+    case 0x07: // ORDERED
+      if (NewMI) {
+        MachineFunction &MF = *MI->getParent()->getParent();
+        MI = MF.CloneMachineInstr(MI);
+        NewMI = false;
+      }
+      return TargetInstrInfo::commuteInstruction(MI, NewMI);
+    default:
+      return nullptr;
+    }
+  }
   case X86::CMOVB16rr:  case X86::CMOVB32rr:  case X86::CMOVB64rr:
   case X86::CMOVAE16rr: case X86::CMOVAE32rr: case X86::CMOVAE64rr:
   case X86::CMOVE16rr:  case X86::CMOVE32rr:  case X86::CMOVE64rr:
   case X86::CMOVNE16rr: case X86::CMOVNE32rr: case X86::CMOVNE64rr:
   case X86::CMOVBE16rr: case X86::CMOVBE32rr: case X86::CMOVBE64rr:
   case X86::CMOVA16rr:  case X86::CMOVA32rr:  case X86::CMOVA64rr:
   case X86::CMOVL16rr:  case X86::CMOVL32rr:  case X86::CMOVL64rr:
   case X86::CMOVGE16rr: case X86::CMOVGE32rr: case X86::CMOVGE64rr:
switch (MI->getOpcode()) {
     case X86::VBLENDPSYrri:
     case X86::VPBLENDDrri:
     case X86::VPBLENDDYrri:
     case X86::VPBLENDWrri:
     case X86::VPBLENDWYrri:
       SrcOpIdx1 = 1;
       SrcOpIdx2 = 2;
       return true;
+    case X86::CMPPDrri:
+    case X86::CMPPSrri:
+    case X86::VCMPPDrri:
+    case X86::VCMPPSrri:
+    case X86::VCMPPDYrri:
+    case X86::VCMPPSYrri: {
+      // Float comparison can be safely commuted for
+      // Ordered/Unordered/Equal/NotEqual tests
+      unsigned Imm = MI->getOperand(3).getImm() & 0x7;
+      switch (Imm) {
+      case 0x00: // EQUAL
+      case 0x03: // UNORDERED
+      case 0x04: // NOT EQUAL
+      case 0x07: // ORDERED
+        SrcOpIdx1 = 1;
+        SrcOpIdx2 = 2;
+        return true;
+      }
+      return false;
+    }
     case X86::VFMADDPDr231r:
     case X86::VFMADDPSr231r:
     case X86::VFMADDSDr231r:
     case X86::VFMADDSSr231r:
     case X86::VFMSUBPDr231r:
     case X86::VFMSUBPSr231r:
     case X86::VFMSUBSDr231r:
     case X86::VFMSUBSSr231r:

llvm/trunk/lib/Target/X86/X86InstrSSE.td

let Defs = [EFLAGS] in {
   }
 } // Defs = [EFLAGS]
 
 // sse12_cmp_packed - sse 1 & 2 compare packed instructions
 multiclass sse12_cmp_packed<RegisterClass RC, X86MemOperand x86memop,
                             Operand CC, Intrinsic Int, string asm,
                             string asm_alt, Domain d, ImmLeaf immLeaf,
                             OpndItins itins = SSE_ALU_F32P> {
+  let isCommutable = 1 in
   def rri : PIi8<0xC2, MRMSrcReg,
              (outs RC:$dst), (ins RC:$src1, RC:$src2, CC:$cc), asm,
              [(set RC:$dst, (Int RC:$src1, RC:$src2, immLeaf:$cc))],
              itins.rr, d>,
             Sched<[WriteFAdd]>;
   def rmi : PIi8<0xC2, MRMSrcMem,
              (outs RC:$dst), (ins RC:$src1, x86memop:$src2, CC:$cc), asm,
              [(set RC:$dst, (Int RC:$src1, (memop addr:$src2), immLeaf:$cc))],

llvm/trunk/test/CodeGen/X86/commute-fcmp.ll

+; RUN: llc -O3 -mtriple=x86_64-unknown -mcpu=x86-64 -mattr=+sse2 < %s | FileCheck %s --check-prefix=SSE
+; RUN: llc -O3 -mtriple=x86_64-unknown -mcpu=x86-64 -mattr=+avx2 < %s | FileCheck %s --check-prefix=AVX
+
+;
+; Float Comparisons
+; Only equal/not-equal/ordered/unordered can be safely commuted
+;
+
+define <4 x i32> @commute_cmpps_eq(<4 x float>* %a0, <4 x float> %a1) #0 {
+  ;SSE-LABEL: commute_cmpps_eq
+  ;SSE:       cmpeqps (%rdi), %xmm0
+  ;SSE-NEXT:  retq
+
+  ;AVX-LABEL: commute_cmpps_eq
+  ;AVX:       vcmpeqps (%rdi), %xmm0, %xmm0
+  ;AVX-NEXT:  retq
+
+  %1 = load <4 x float>* %a0
+  %2 = fcmp oeq <4 x float> %1, %a1
+  %3 = sext <4 x i1> %2 to <4 x i32>
+  ret <4 x i32> %3
+}
+
+define <4 x i32> @commute_cmpps_ne(<4 x float>* %a0, <4 x float> %a1) #0 {
+  ;SSE-LABEL: commute_cmpps_ne
+  ;SSE:       cmpneqps (%rdi), %xmm0
+  ;SSE-NEXT:  retq
+
+  ;AVX-LABEL: commute_cmpps_ne
+  ;AVX:       vcmpneqps (%rdi), %xmm0, %xmm0
+  ;AVX-NEXT:  retq
+
+  %1 = load <4 x float>* %a0
+  %2 = fcmp une <4 x float> %1, %a1
+  %3 = sext <4 x i1> %2 to <4 x i32>
+  ret <4 x i32> %3
+}
+
+define <4 x i32> @commute_cmpps_ord(<4 x float>* %a0, <4 x float> %a1) #0 {
+  ;SSE-LABEL: commute_cmpps_ord
+  ;SSE:       cmpordps (%rdi), %xmm0
+  ;SSE-NEXT:  retq
+
+  ;AVX-LABEL: commute_cmpps_ord
+  ;AVX:       vcmpordps (%rdi), %xmm0, %xmm0
+  ;AVX-NEXT:  retq
+
+  %1 = load <4 x float>* %a0
+  %2 = fcmp ord <4 x float> %1, %a1
+  %3 = sext <4 x i1> %2 to <4 x i32>
+  ret <4 x i32> %3
+}
+
+define <4 x i32> @commute_cmpps_uno(<4 x float>* %a0, <4 x float> %a1) #0 {
+  ;SSE-LABEL: commute_cmpps_uno
+  ;SSE:       cmpunordps (%rdi), %xmm0
+  ;SSE-NEXT:  retq
+
+  ;AVX-LABEL: commute_cmpps_uno
+  ;AVX:       vcmpunordps (%rdi), %xmm0, %xmm0
+  ;AVX-NEXT:  retq
+
+  %1 = load <4 x float>* %a0
+  %2 = fcmp uno <4 x float> %1, %a1
+  %3 = sext <4 x i1> %2 to <4 x i32>
+  ret <4 x i32> %3
+}
+
+define <4 x i32> @commute_cmpps_lt(<4 x float>* %a0, <4 x float> %a1) #0 {
+  ;SSE-LABEL: commute_cmpps_lt
+  ;SSE:       movaps (%rdi), %xmm1
+  ;SSE-NEXT:  cmpltps %xmm0, %xmm1
+  ;SSE-NEXT:  movaps %xmm1, %xmm0
+  ;SSE-NEXT:  retq
+
+  ;AVX-LABEL: commute_cmpps_lt
+  ;AVX:       vmovaps (%rdi), %xmm1
+  ;AVX-NEXT:  vcmpltps %xmm0, %xmm1, %xmm0
+  ;AVX-NEXT:  retq
+
+  %1 = load <4 x float>* %a0
+  %2 = fcmp olt <4 x float> %1, %a1
+  %3 = sext <4 x i1> %2 to <4 x i32>
+  ret <4 x i32> %3
+}
+
+define <4 x i32> @commute_cmpps_le(<4 x float>* %a0, <4 x float> %a1) #0 {
+  ;SSE-LABEL: commute_cmpps_le
+  ;SSE:       movaps (%rdi), %xmm1
+  ;SSE-NEXT:  cmpleps %xmm0, %xmm1
+  ;SSE-NEXT:  movaps %xmm1, %xmm0
+  ;SSE-NEXT:  retq
+
+  ;AVX-LABEL: commute_cmpps_le
+  ;AVX:       vmovaps (%rdi), %xmm1
+  ;AVX-NEXT:  vcmpleps %xmm0, %xmm1, %xmm0
+  ;AVX-NEXT:  retq
+
+  %1 = load <4 x float>* %a0
+  %2 = fcmp ole <4 x float> %1, %a1
+  %3 = sext <4 x i1> %2 to <4 x i32>
+  ret <4 x i32> %3
+}
+
+define <8 x i32> @commute_cmpps_eq_ymm(<8 x float>* %a0, <8 x float> %a1) #0 {
+  ;AVX-LABEL: commute_cmpps_eq_ymm
+  ;AVX:       vcmpeqps (%rdi), %ymm0, %ymm0
+  ;AVX-NEXT:  retq
+
+  %1 = load <8 x float>* %a0
+  %2 = fcmp oeq <8 x float> %1, %a1
+  %3 = sext <8 x i1> %2 to <8 x i32>
+  ret <8 x i32> %3
+}
+
+define <8 x i32> @commute_cmpps_ne_ymm(<8 x float>* %a0, <8 x float> %a1) #0 {
+  ;AVX-LABEL: commute_cmpps_ne_ymm
+  ;AVX:       vcmpneqps (%rdi), %ymm0, %ymm0
+  ;AVX-NEXT:  retq
+
+  %1 = load <8 x float>* %a0
+  %2 = fcmp une <8 x float> %1, %a1
+  %3 = sext <8 x i1> %2 to <8 x i32>
+  ret <8 x i32> %3
+}
+
+define <8 x i32> @commute_cmpps_ord_ymm(<8 x float>* %a0, <8 x float> %a1) #0 {
+  ;AVX-LABEL: commute_cmpps_ord_ymm
+  ;AVX:       vcmpordps (%rdi), %ymm0, %ymm0
+  ;AVX-NEXT:  retq
+
+  %1 = load <8 x float>* %a0
+  %2 = fcmp ord <8 x float> %1, %a1
+  %3 = sext <8 x i1> %2 to <8 x i32>
+  ret <8 x i32> %3
+}
+
+define <8 x i32> @commute_cmpps_uno_ymm(<8 x float>* %a0, <8 x float> %a1) #0 {
+  ;AVX-LABEL: commute_cmpps_uno_ymm
+  ;AVX:       vcmpunordps (%rdi), %ymm0, %ymm0
+  ;AVX-NEXT:  retq
+
+  %1 = load <8 x float>* %a0
+  %2 = fcmp uno <8 x float> %1, %a1
+  %3 = sext <8 x i1> %2 to <8 x i32>
+  ret <8 x i32> %3
+}
+
+define <8 x i32> @commute_cmpps_lt_ymm(<8 x float>* %a0, <8 x float> %a1) #0 {
+  ;AVX-LABEL: commute_cmpps_lt_ymm
+  ;AVX:       vmovaps (%rdi), %ymm1
+  ;AVX-NEXT:  vcmpltps %ymm0, %ymm1, %ymm0
+  ;AVX-NEXT:  retq
+
+  %1 = load <8 x float>* %a0
+  %2 = fcmp olt <8 x float> %1, %a1
+  %3 = sext <8 x i1> %2 to <8 x i32>
+  ret <8 x i32> %3
+}
+
+define <8 x i32> @commute_cmpps_le_ymm(<8 x float>* %a0, <8 x float> %a1) #0 {
+  ;AVX-LABEL: commute_cmpps_le_ymm
+  ;AVX:       vmovaps (%rdi), %ymm1
+  ;AVX-NEXT:  vcmpleps %ymm0, %ymm1, %ymm0
+  ;AVX-NEXT:  retq
+
+  %1 = load <8 x float>* %a0
+  %2 = fcmp ole <8 x float> %1, %a1
+  %3 = sext <8 x i1> %2 to <8 x i32>
+  ret <8 x i32> %3
+}
+
+;
+; Double Comparisons
+; Only equal/not-equal/ordered/unordered can be safely commuted
+;
+
+define <2 x i64> @commute_cmppd_eq(<2 x double>* %a0, <2 x double> %a1) #0 {
+  ;SSE-LABEL: commute_cmppd_eq
+  ;SSE:       cmpeqpd (%rdi), %xmm0
+  ;SSE-NEXT:  retq
+
+  ;AVX-LABEL: commute_cmppd_eq
+  ;AVX:       vcmpeqpd (%rdi), %xmm0, %xmm0
+  ;AVX-NEXT:  retq
+
+  %1 = load <2 x double>* %a0
+  %2 = fcmp oeq <2 x double> %1, %a1
+  %3 = sext <2 x i1> %2 to <2 x i64>
+  ret <2 x i64> %3
+}
+
+define <2 x i64> @commute_cmppd_ne(<2 x double>* %a0, <2 x double> %a1) #0 {
+  ;SSE-LABEL: commute_cmppd_ne
+  ;SSE:       cmpneqpd (%rdi), %xmm0
+  ;SSE-NEXT:  retq
+
+  ;AVX-LABEL: commute_cmppd_ne
+  ;AVX:       vcmpneqpd (%rdi), %xmm0, %xmm0
+  ;AVX-NEXT:  retq
+
+  %1 = load <2 x double>* %a0
+  %2 = fcmp une <2 x double> %1, %a1
+  %3 = sext <2 x i1> %2 to <2 x i64>
+  ret <2 x i64> %3
+}
+
+define <2 x i64> @commute_cmppd_ord(<2 x double>* %a0, <2 x double> %a1) #0 {
+  ;SSE-LABEL: commute_cmppd_ord
+  ;SSE:       cmpordpd (%rdi), %xmm0
+  ;SSE-NEXT:  retq
+
+  ;AVX-LABEL: commute_cmppd_ord
+  ;AVX:       vcmpordpd (%rdi), %xmm0, %xmm0
+  ;AVX-NEXT:  retq
+
+  %1 = load <2 x double>* %a0
+  %2 = fcmp ord <2 x double> %1, %a1
+  %3 = sext <2 x i1> %2 to <2 x i64>
+  ret <2 x i64> %3
+}
+
+define <2 x i64> @commute_cmppd_uno(<2 x double>* %a0, <2 x double> %a1) #0 {
+  ;SSE-LABEL: commute_cmppd_uno
+  ;SSE:       cmpunordpd (%rdi), %xmm0
+  ;SSE-NEXT:  retq
+
+  ;AVX-LABEL: commute_cmppd_uno
+  ;AVX:       vcmpunordpd (%rdi), %xmm0, %xmm0
+  ;AVX-NEXT:  retq
+
+  %1 = load <2 x double>* %a0
+  %2 = fcmp uno <2 x double> %1, %a1
+  %3 = sext <2 x i1> %2 to <2 x i64>
+  ret <2 x i64> %3
+}
+
+define <2 x i64> @commute_cmppd_lt(<2 x double>* %a0, <2 x double> %a1) #0 {
+  ;SSE-LABEL: commute_cmppd_lt
+  ;SSE:       movapd (%rdi), %xmm1
+  ;SSE-NEXT:  cmpltpd %xmm0, %xmm1
+  ;SSE-NEXT:  movapd %xmm1, %xmm0
+  ;SSE-NEXT:  retq
+
+  ;AVX-LABEL: commute_cmppd_lt
+  ;AVX:       vmovapd (%rdi), %xmm1
+  ;AVX-NEXT:  vcmpltpd %xmm0, %xmm1, %xmm0
+  ;AVX-NEXT:  retq
+
+  %1 = load <2 x double>* %a0
+  %2 = fcmp olt <2 x double> %1, %a1
+  %3 = sext <2 x i1> %2 to <2 x i64>
+  ret <2 x i64> %3
+}
+
+define <2 x i64> @commute_cmppd_le(<2 x double>* %a0, <2 x double> %a1) #0 {
+  ;SSE-LABEL: commute_cmppd_le
+  ;SSE:       movapd (%rdi), %xmm1
+  ;SSE-NEXT:  cmplepd %xmm0, %xmm1
+  ;SSE-NEXT:  movapd %xmm1, %xmm0
+  ;SSE-NEXT:  retq
+
+  ;AVX-LABEL: commute_cmppd_le
+  ;AVX:       vmovapd (%rdi), %xmm1
+  ;AVX-NEXT:  vcmplepd %xmm0, %xmm1, %xmm0
+  ;AVX-NEXT:  retq
+
+  %1 = load <2 x double>* %a0
+  %2 = fcmp ole <2 x double> %1, %a1
+  %3 = sext <2 x i1> %2 to <2 x i64>
+  ret <2 x i64> %3
+}
+
+define <4 x i64> @commute_cmppd_eq_ymmm(<4 x double>* %a0, <4 x double> %a1) #0 {
+  ;AVX-LABEL: commute_cmppd_eq
+  ;AVX:       vcmpeqpd (%rdi), %ymm0, %ymm0
+  ;AVX-NEXT:  retq
+
+  %1 = load <4 x double>* %a0
+  %2 = fcmp oeq <4 x double> %1, %a1
+  %3 = sext <4 x i1> %2 to <4 x i64>
+  ret <4 x i64> %3
+}
+
+define <4 x i64> @commute_cmppd_ne_ymmm(<4 x double>* %a0, <4 x double> %a1) #0 {
+  ;AVX-LABEL: commute_cmppd_ne
+  ;AVX:       vcmpneqpd (%rdi), %ymm0, %ymm0
+  ;AVX-NEXT:  retq
+
+  %1 = load <4 x double>* %a0
+  %2 = fcmp une <4 x double> %1, %a1
+  %3 = sext <4 x i1> %2 to <4 x i64>
+  ret <4 x i64> %3
+}
+
+define <4 x i64> @commute_cmppd_ord_ymmm(<4 x double>* %a0, <4 x double> %a1) #0 {
+  ;AVX-LABEL: commute_cmppd_ord
+  ;AVX:       vcmpordpd (%rdi), %ymm0, %ymm0
+  ;AVX-NEXT:  retq
+
+  %1 = load <4 x double>* %a0
+  %2 = fcmp ord <4 x double> %1, %a1
+  %3 = sext <4 x i1> %2 to <4 x i64>
+  ret <4 x i64> %3
+}
+
+define <4 x i64> @commute_cmppd_uno_ymmm(<4 x double>* %a0, <4 x double> %a1) #0 {
+  ;AVX-LABEL: commute_cmppd_uno
+  ;AVX:       vcmpunordpd (%rdi), %ymm0, %ymm0
+  ;AVX-NEXT:  retq
+
+  %1 = load <4 x double>* %a0
+  %2 = fcmp uno <4 x double> %1, %a1
+  %3 = sext <4 x i1> %2 to <4 x i64>
+  ret <4 x i64> %3
+}
+
+define <4 x i64> @commute_cmppd_lt_ymmm(<4 x double>* %a0, <4 x double> %a1) #0 {
+  ;AVX-LABEL: commute_cmppd_lt
+  ;AVX:       vmovapd (%rdi), %ymm1
+  ;AVX-NEXT:  vcmpltpd %ymm0, %ymm1, %ymm0
+  ;AVX-NEXT:  retq
+
+  %1 = load <4 x double>* %a0
+  %2 = fcmp olt <4 x double> %1, %a1
+  %3 = sext <4 x i1> %2 to <4 x i64>
+  ret <4 x i64> %3
+}
+
+define <4 x i64> @commute_cmppd_le_ymmm(<4 x double>* %a0, <4 x double> %a1) #0 {
+  ;AVX-LABEL: commute_cmppd_le
+  ;AVX:       vmovapd (%rdi), %ymm1
+  ;AVX-NEXT:  vcmplepd %ymm0, %ymm1, %ymm0
+  ;AVX-NEXT:  retq
+
+  %1 = load <4 x double>* %a0
+  %2 = fcmp ole <4 x double> %1, %a1
+  %3 = sext <4 x i1> %2 to <4 x i64>
+  ret <4 x i64> %3
+}