[X86] Combine fminnum/fmaxnum with non-nan operand to fmin/fmax

[nikic]

Summary

If we have a known non-nan operand, place it in the second operand of fmin/fmax that is returned if either operand is nan.

[spatel]

LGTM - thanks!

llvm/lib/Target/X86/X86ISelLowering.cpp

 
 static SDValue combineFMinNumFMaxNum(SDNode *N, SelectionDAG &DAG,
                                      const X86Subtarget &Subtarget) {
   if (Subtarget.useSoftFloat())
     return SDValue();
 
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
 
-  // TODO: If an operand is already known to be a NaN or not a NaN, this
-  //       should be an optional swap and FMAX/FMIN.
-
   EVT VT = N->getValueType(0);
   if (!((Subtarget.hasSSE1() && VT == MVT::f32) ||
         (Subtarget.hasSSE2() && VT == MVT::f64) ||
         (VT.isVector() && TLI.isTypeLegal(VT))))
     return SDValue();
 
   SDValue Op0 = N->getOperand(0);
   SDValue Op1 = N->getOperand(1);
   SDLoc DL(N);
   auto MinMaxOp = N->getOpcode() == ISD::FMAXNUM ? X86ISD::FMAX : X86ISD::FMIN;
 
   // If we don't have to respect NaN inputs, this is a direct translation to x86
   // min/max instructions.
   if (DAG.getTarget().Options.NoNaNsFPMath || N->getFlags().hasNoNaNs())
     return DAG.getNode(MinMaxOp, DL, VT, Op0, Op1, N->getFlags());
 
+  // If one of the operands is known non-NaN use the native min/max instructions
+  // with the non-NaN input as second operand.
+  if (DAG.isKnownNeverNaN(Op1))
+    return DAG.getNode(MinMaxOp, DL, VT, Op0, Op1, N->getFlags());
+  if (DAG.isKnownNeverNaN(Op0))
+    return DAG.getNode(MinMaxOp, DL, VT, Op1, Op0, N->getFlags());
+
   // If we have to respect NaN inputs, this takes at least 3 instructions.
   // Favor a library call when operating on a scalar and minimizing code size.
   if (!VT.isVector() && DAG.getMachineFunction().getFunction().hasMinSize())
     return SDValue();
 
   EVT SetCCType = TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(),
                                          VT);
 

llvm/test/CodeGen/X86/extract-fp.ll

; CHECK-NEXT:    jmp fmodf # TAILCALL
   %bo = frem <4 x float> <float 1.0, float 2.0, float 3.0, float 42.0>, %x
   %ext = extractelement <4 x float> %bo, i32 1
   ret float %ext
 }
 
 define float @ext_maxnum_v4f32(<4 x float> %x) nounwind {
 ; CHECK-LABEL: ext_maxnum_v4f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
 ; CHECK-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
-; CHECK-NEXT:    movaps %xmm0, %xmm1
-; CHECK-NEXT:    cmpunordss %xmm0, %xmm1
-; CHECK-NEXT:    movaps %xmm1, %xmm3
-; CHECK-NEXT:    andps %xmm2, %xmm3
-; CHECK-NEXT:    maxss %xmm0, %xmm2
-; CHECK-NEXT:    andnps %xmm2, %xmm1
-; CHECK-NEXT:    orps %xmm3, %xmm1
-; CHECK-NEXT:    movaps %xmm1, %xmm0
+; CHECK-NEXT:    maxss {{.*}}(%rip), %xmm0
 ; CHECK-NEXT:    retq
   %v = call <4 x float> @llvm.maxnum.v4f32(<4 x float> %x, <4 x float> <float 0.0, float 1.0, float 2.0, float 3.0>)
   %r = extractelement <4 x float> %v, i32 2
   ret float %r
 }
 
 define double @ext_minnum_v2f64(<2 x double> %x) nounwind {
 ; CHECK-LABEL: ext_minnum_v2f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    movsd {{.*#+}} xmm2 = mem[0],zero
 ; CHECK-NEXT:    unpckhpd {{.*#+}} xmm0 = xmm0[1,1]
-; CHECK-NEXT:    movapd %xmm0, %xmm1
-; CHECK-NEXT:    cmpunordsd %xmm0, %xmm1
-; CHECK-NEXT:    movapd %xmm1, %xmm3
-; CHECK-NEXT:    andpd %xmm2, %xmm3
-; CHECK-NEXT:    minsd %xmm0, %xmm2
-; CHECK-NEXT:    andnpd %xmm2, %xmm1
-; CHECK-NEXT:    orpd %xmm3, %xmm1
-; CHECK-NEXT:    movapd %xmm1, %xmm0
+; CHECK-NEXT:    minsd {{.*}}(%rip), %xmm0
 ; CHECK-NEXT:    retq
   %v = call <2 x double> @llvm.minnum.v2f64(<2 x double> <double 0.0, double 1.0>, <2 x double> %x)
   %r = extractelement <2 x double> %v, i32 1
   ret double %r
 }
 
 ;define double @ext_maximum_v4f64(<2 x double> %x) nounwind {
 ;  %v = call <2 x double> @llvm.maximum.v2f64(<2 x double> %x, <2 x double> <double 42.0, double 43.0>)

llvm/test/CodeGen/X86/fmaxnum.ll

 ; AVX-NEXT:    retq
   %r = tail call <2 x double> @llvm.maxnum.v2f64(<2 x double> %a, <2 x double> %b)
   ret <2 x double> %r
 }
 
 define float @test_maxnum_const_op1(float %x) {
 ; SSE-LABEL: test_maxnum_const_op1:
 ; SSE:       # %bb.0:
-; SSE-NEXT:    movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
-; SSE-NEXT:    movaps %xmm0, %xmm1
-; SSE-NEXT:    cmpunordss %xmm0, %xmm1
-; SSE-NEXT:    movaps %xmm1, %xmm3
-; SSE-NEXT:    andps %xmm2, %xmm3
-; SSE-NEXT:    maxss %xmm0, %xmm2
-; SSE-NEXT:    andnps %xmm2, %xmm1
-; SSE-NEXT:    orps %xmm3, %xmm1
-; SSE-NEXT:    movaps %xmm1, %xmm0
+; SSE-NEXT:    maxss {{.*}}(%rip), %xmm0
 ; SSE-NEXT:    retq
 ;
-; AVX1-LABEL: test_maxnum_const_op1:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero
-; AVX1-NEXT:    vmaxss %xmm0, %xmm1, %xmm2
-; AVX1-NEXT:    vcmpunordss %xmm0, %xmm0, %xmm0
-; AVX1-NEXT:    vblendvps %xmm0, %xmm1, %xmm2, %xmm0
-; AVX1-NEXT:    retq
-;
-; AVX512-LABEL: test_maxnum_const_op1:
-; AVX512:       # %bb.0:
-; AVX512-NEXT:    vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero
-; AVX512-NEXT:    vmaxss %xmm0, %xmm2, %xmm1
-; AVX512-NEXT:    vcmpunordss %xmm0, %xmm0, %k1
-; AVX512-NEXT:    vmovss %xmm2, %xmm0, %xmm1 {%k1}
-; AVX512-NEXT:    vmovaps %xmm1, %xmm0
-; AVX512-NEXT:    retq
+; AVX-LABEL: test_maxnum_const_op1:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vmaxss {{.*}}(%rip), %xmm0, %xmm0
+; AVX-NEXT:    retq
   %r = call float @llvm.maxnum.f32(float 1.0, float %x)
   ret float %r
 }
 
 define float @test_maxnum_const_op2(float %x) {
 ; SSE-LABEL: test_maxnum_const_op2:
 ; SSE:       # %bb.0:
-; SSE-NEXT:    movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
-; SSE-NEXT:    movaps %xmm0, %xmm1
-; SSE-NEXT:    cmpunordss %xmm0, %xmm1
-; SSE-NEXT:    movaps %xmm1, %xmm3
-; SSE-NEXT:    andps %xmm2, %xmm3
-; SSE-NEXT:    maxss %xmm0, %xmm2
-; SSE-NEXT:    andnps %xmm2, %xmm1
-; SSE-NEXT:    orps %xmm3, %xmm1
-; SSE-NEXT:    movaps %xmm1, %xmm0
+; SSE-NEXT:    maxss {{.*}}(%rip), %xmm0
 ; SSE-NEXT:    retq
 ;
-; AVX1-LABEL: test_maxnum_const_op2:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero
-; AVX1-NEXT:    vmaxss %xmm0, %xmm1, %xmm2
-; AVX1-NEXT:    vcmpunordss %xmm0, %xmm0, %xmm0
-; AVX1-NEXT:    vblendvps %xmm0, %xmm1, %xmm2, %xmm0
-; AVX1-NEXT:    retq
-;
-; AVX512-LABEL: test_maxnum_const_op2:
-; AVX512:       # %bb.0:
-; AVX512-NEXT:    vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero
-; AVX512-NEXT:    vmaxss %xmm0, %xmm2, %xmm1
-; AVX512-NEXT:    vcmpunordss %xmm0, %xmm0, %k1
-; AVX512-NEXT:    vmovss %xmm2, %xmm0, %xmm1 {%k1}
-; AVX512-NEXT:    vmovaps %xmm1, %xmm0
-; AVX512-NEXT:    retq
+; AVX-LABEL: test_maxnum_const_op2:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vmaxss {{.*}}(%rip), %xmm0, %xmm0
+; AVX-NEXT:    retq
   %r = call float @llvm.maxnum.f32(float %x, float 1.0)
   ret float %r
 }
 
 attributes #0 = { "no-nans-fp-math"="true" }
 

llvm/test/CodeGen/X86/fminnum.ll

 ; AVX-NEXT:    retq
   %r = tail call <4 x float> @llvm.minnum.v4f32(<4 x float> %a, <4 x float> %b)
   ret <4 x float> %r
 }
 
 define float @test_minnum_const_op1(float %x) {
 ; SSE-LABEL: test_minnum_const_op1:
 ; SSE:       # %bb.0:
-; SSE-NEXT:    movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
-; SSE-NEXT:    movaps %xmm0, %xmm1
-; SSE-NEXT:    cmpunordss %xmm0, %xmm1
-; SSE-NEXT:    movaps %xmm1, %xmm3
-; SSE-NEXT:    andps %xmm2, %xmm3
-; SSE-NEXT:    minss %xmm0, %xmm2
-; SSE-NEXT:    andnps %xmm2, %xmm1
-; SSE-NEXT:    orps %xmm3, %xmm1
-; SSE-NEXT:    movaps %xmm1, %xmm0
+; SSE-NEXT:    minss {{.*}}(%rip), %xmm0
 ; SSE-NEXT:    retq
 ;
-; AVX1-LABEL: test_minnum_const_op1:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero
-; AVX1-NEXT:    vminss %xmm0, %xmm1, %xmm2
-; AVX1-NEXT:    vcmpunordss %xmm0, %xmm0, %xmm0
-; AVX1-NEXT:    vblendvps %xmm0, %xmm1, %xmm2, %xmm0
-; AVX1-NEXT:    retq
-;
-; AVX512-LABEL: test_minnum_const_op1:
-; AVX512:       # %bb.0:
-; AVX512-NEXT:    vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero
-; AVX512-NEXT:    vminss %xmm0, %xmm2, %xmm1
-; AVX512-NEXT:    vcmpunordss %xmm0, %xmm0, %k1
-; AVX512-NEXT:    vmovss %xmm2, %xmm0, %xmm1 {%k1}
-; AVX512-NEXT:    vmovaps %xmm1, %xmm0
-; AVX512-NEXT:    retq
+; AVX-LABEL: test_minnum_const_op1:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vminss {{.*}}(%rip), %xmm0, %xmm0
+; AVX-NEXT:    retq
   %r = call float @llvm.minnum.f32(float 1.0, float %x)
   ret float %r
 }
 
 define float @test_minnum_const_op2(float %x) {
 ; SSE-LABEL: test_minnum_const_op2:
 ; SSE:       # %bb.0:
-; SSE-NEXT:    movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
-; SSE-NEXT:    movaps %xmm0, %xmm1
-; SSE-NEXT:    cmpunordss %xmm0, %xmm1
-; SSE-NEXT:    movaps %xmm1, %xmm3
-; SSE-NEXT:    andps %xmm2, %xmm3
-; SSE-NEXT:    minss %xmm0, %xmm2
-; SSE-NEXT:    andnps %xmm2, %xmm1
-; SSE-NEXT:    orps %xmm3, %xmm1
-; SSE-NEXT:    movaps %xmm1, %xmm0
+; SSE-NEXT:    minss {{.*}}(%rip), %xmm0
 ; SSE-NEXT:    retq
 ;
-; AVX1-LABEL: test_minnum_const_op2:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero
-; AVX1-NEXT:    vminss %xmm0, %xmm1, %xmm2
-; AVX1-NEXT:    vcmpunordss %xmm0, %xmm0, %xmm0
-; AVX1-NEXT:    vblendvps %xmm0, %xmm1, %xmm2, %xmm0
-; AVX1-NEXT:    retq
-;
-; AVX512-LABEL: test_minnum_const_op2:
-; AVX512:       # %bb.0:
-; AVX512-NEXT:    vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero
-; AVX512-NEXT:    vminss %xmm0, %xmm2, %xmm1
-; AVX512-NEXT:    vcmpunordss %xmm0, %xmm0, %k1
-; AVX512-NEXT:    vmovss %xmm2, %xmm0, %xmm1 {%k1}
-; AVX512-NEXT:    vmovaps %xmm1, %xmm0
-; AVX512-NEXT:    retq
+; AVX-LABEL: test_minnum_const_op2:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vminss {{.*}}(%rip), %xmm0, %xmm0
+; AVX-NEXT:    retq
   %r = call float @llvm.minnum.f32(float %x, float 1.0)
   ret float %r
 }
 
 attributes #0 = { "no-nans-fp-math"="true" }