[X86][FMA] Optimize FNEG(FMA) Patterns

X86 needs to use its own FMA opcodes, preventing the standard FNEG(FMA) pattern table recognition method used by other platforms. This patch adds support for lowering FNEG(FMA(X,Y,Z)) into a single suitably negated FMA instruction.

Fix for PR24364

Differential Revision: http://reviews.llvm.org/D14906

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@254016 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 59e16d0a09432ace4887481c30ebaab13185c9a7..0bf5ee68d1341e5174dc71316e58f465cf73edc6 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -1773,6 +1773,7 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
   setTargetDAGCombine(ISD::ADD);
   setTargetDAGCombine(ISD::FADD);
   setTargetDAGCombine(ISD::FSUB);
+  setTargetDAGCombine(ISD::FNEG);
   setTargetDAGCombine(ISD::FMA);
   setTargetDAGCombine(ISD::SUB);
   setTargetDAGCombine(ISD::LOAD);
@@ -26148,6 +26149,33 @@ static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
   return SDValue();
 }
 
+/// Do target-specific dag combines on floating point negations.
+static SDValue PerformFNEGCombine(SDNode *N, SelectionDAG &DAG,
+                                  const X86Subtarget *Subtarget) {
+  EVT VT = N->getValueType(0);
+  SDValue Arg = N->getOperand(0);
+
+  // If we're negating a FMA node, then we can adjust the
+  // instruction to include the extra negation.
+  if (Arg.hasOneUse()) {
+    switch (Arg.getOpcode()) {
+      case X86ISD::FMADD:
+        return DAG.getNode(X86ISD::FNMSUB, SDLoc(N), VT, Arg.getOperand(0),
+                           Arg.getOperand(1), Arg.getOperand(2));
+      case X86ISD::FMSUB:
+        return DAG.getNode(X86ISD::FNMADD, SDLoc(N), VT, Arg.getOperand(0),
+                           Arg.getOperand(1), Arg.getOperand(2));
+      case X86ISD::FNMADD:
+        return DAG.getNode(X86ISD::FMSUB, SDLoc(N), VT, Arg.getOperand(0),
+                           Arg.getOperand(1), Arg.getOperand(2));
+      case X86ISD::FNMSUB:
+        return DAG.getNode(X86ISD::FMADD, SDLoc(N), VT, Arg.getOperand(0),
+                           Arg.getOperand(1), Arg.getOperand(2));
+    }
+  }
+  return SDValue();
+}
+
 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG,
                                  const X86Subtarget *Subtarget) {
@@ -27042,6 +27070,7 @@ SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
   case ISD::UINT_TO_FP:     return PerformUINT_TO_FPCombine(N, DAG, Subtarget);
   case ISD::FADD:           return PerformFADDCombine(N, DAG, Subtarget);
   case ISD::FSUB:           return PerformFSUBCombine(N, DAG, Subtarget);
+  case ISD::FNEG:           return PerformFNEGCombine(N, DAG, Subtarget);
   case ISD::TRUNCATE:       return PerformTRUNCATECombine(N, DAG, Subtarget);
   case X86ISD::FXOR:
   case X86ISD::FOR:         return PerformFORCombine(N, DAG, Subtarget);

diff --git a/test/CodeGen/X86/fma_patterns.ll b/test/CodeGen/X86/fma_patterns.ll
index c8667d7b8ad5ceb418b1ed8f639c561dc9e621b2..9c8b12d63faddf67a8c04f85e6e5967ae5d0df9f 100644 (file)
--- a/test/CodeGen/X86/fma_patterns.ll
+++ b/test/CodeGen/X86/fma_patterns.ll
@@ -568,6 +568,74 @@ define <4 x double> @test_v4f64_interp(<4 x double> %x, <4 x double> %y, <4 x do
   ret <4 x double> %r
 }
 
+; (fneg (fma x, y, z)) -> (fma x, -y, -z)
+
+define <4 x float> @test_v4f32_fneg_fmadd(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) #0 {
+; CHECK_FMA-LABEL: test_v4f32_fneg_fmadd:
+; CHECK_FMA:       # BB#0:
+; CHECK_FMA-NEXT:    vfnmsub213ps %xmm2, %xmm1, %xmm0
+; CHECK_FMA-NEXT:    retq
+;
+; CHECK_FMA4-LABEL: test_v4f32_fneg_fmadd:
+; CHECK_FMA4:       # BB#0:
+; CHECK_FMA4-NEXT:    vfnmsubps %xmm2, %xmm1, %xmm0, %xmm0
+; CHECK_FMA4-NEXT:    retq
+  %mul = fmul <4 x float> %a0, %a1
+  %add = fadd <4 x float> %mul, %a2
+  %neg = fsub <4 x float> <float -0.0, float -0.0, float -0.0, float -0.0>, %add
+  ret <4 x float> %neg
+}
+
+define <4 x double> @test_v4f64_fneg_fmsub(<4 x double> %a0, <4 x double> %a1, <4 x double> %a2) #0 {
+; CHECK_FMA-LABEL: test_v4f64_fneg_fmsub:
+; CHECK_FMA:       # BB#0:
+; CHECK_FMA-NEXT:    vfnmadd213pd %ymm2, %ymm1, %ymm0
+; CHECK_FMA-NEXT:    retq
+;
+; CHECK_FMA4-LABEL: test_v4f64_fneg_fmsub:
+; CHECK_FMA4:       # BB#0:
+; CHECK_FMA4-NEXT:    vfnmaddpd %ymm2, %ymm1, %ymm0, %ymm0
+; CHECK_FMA4-NEXT:    retq
+  %mul = fmul <4 x double> %a0, %a1
+  %sub = fsub <4 x double> %mul, %a2
+  %neg = fsub <4 x double> <double -0.0, double -0.0, double -0.0, double -0.0>, %sub
+  ret <4 x double> %neg
+}
+
+define <4 x float> @test_v4f32_fneg_fnmadd(<4 x float> %a0, <4 x float> %a1, <4 x float> %a2) #0 {
+; CHECK_FMA-LABEL: test_v4f32_fneg_fnmadd:
+; CHECK_FMA:       # BB#0:
+; CHECK_FMA-NEXT:    vfmsub213ps %xmm2, %xmm1, %xmm0
+; CHECK_FMA-NEXT:    retq
+;
+; CHECK_FMA4-LABEL: test_v4f32_fneg_fnmadd:
+; CHECK_FMA4:       # BB#0:
+; CHECK_FMA4-NEXT:    vfmsubps %xmm2, %xmm1, %xmm0, %xmm0
+; CHECK_FMA4-NEXT:    retq
+  %mul = fmul <4 x float> %a0, %a1
+  %neg0 = fsub <4 x float> <float -0.0, float -0.0, float -0.0, float -0.0>, %mul
+  %add = fadd <4 x float> %neg0, %a2
+  %neg1 = fsub <4 x float> <float -0.0, float -0.0, float -0.0, float -0.0>, %add
+  ret <4 x float> %neg1
+}
+
+define <4 x double> @test_v4f64_fneg_fnmsub(<4 x double> %a0, <4 x double> %a1, <4 x double> %a2) #0 {
+; CHECK_FMA-LABEL: test_v4f64_fneg_fnmsub:
+; CHECK_FMA:       # BB#0:
+; CHECK_FMA-NEXT:    vfmadd213pd %ymm2, %ymm1, %ymm0
+; CHECK_FMA-NEXT:    retq
+;
+; CHECK_FMA4-LABEL: test_v4f64_fneg_fnmsub:
+; CHECK_FMA4:       # BB#0:
+; CHECK_FMA4-NEXT:    vfmaddpd %ymm2, %ymm1, %ymm0, %ymm0
+; CHECK_FMA4-NEXT:    retq
+  %mul = fmul <4 x double> %a0, %a1
+  %neg0 = fsub <4 x double> <double -0.0, double -0.0, double -0.0, double -0.0>, %mul
+  %sub = fsub <4 x double> %neg0, %a2
+  %neg1 = fsub <4 x double> <double -0.0, double -0.0, double -0.0, double -0.0>, %sub
+  ret <4 x double> %neg1
+}
+
 ; (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2)
 
 define <4 x float> @test_v4f32_fma_x_c1_fmul_x_c2(<4 x float> %x) #0 {