[DAGCombiner] exclude x*2.0 from normal negation profitability rules

This is the codegen part of fixing:
https://bugs.llvm.org/show_bug.cgi?id=32939

Even with the optimal/canonical IR that is ideally created by D65954,
we would reverse that transform in DAGCombiner and end up with the same
asm on AArch64 or x86.

I see 2 options for trying to correct this:

  1. Limit isNegatibleForFree() by special-casing the fmul pattern (this patch).
  2. Avoid creating (fmul X, 2.0) in the 1st place by adding a special-case
     transform to SelectionDAG::getNode() and/or SelectionDAGBuilder::visitFMul()
     that matches the transform done by DAGCombiner.

This seems like the less intrusive patch, but if there's some other reason to
prefer 1 option over the other, we can change to the other option.

Differential Revision: https://reviews.llvm.org/D66016

llvm-svn: 368490

3 files changed, 13 insertions, 10 deletions

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 63431d5b7ec..d9a77899ca3 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -868,6 +868,11 @@ static char isNegatibleForFree(SDValue Op, bool LegalOperations,
                                     Options, ForCodeSize, Depth + 1))
       return V;
 
+    // Ignore X * 2.0 because that is expected to be canonicalized to X + X.
+    if (auto *C = isConstOrConstSplatFP(Op.getOperand(1)))
+      if (C->isExactlyValue(2.0) && Op.getOpcode() == ISD::FMUL)
+        return 0;
+
     return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
                               ForCodeSize, Depth + 1);
 
diff --git a/llvm/test/CodeGen/AArch64/fadd-combines.ll b/llvm/test/CodeGen/AArch64/fadd-combines.ll
index 7e34d2ea9e0..640c1a10046 100644
--- a/llvm/test/CodeGen/AArch64/fadd-combines.ll
+++ b/llvm/test/CodeGen/AArch64/fadd-combines.ll
@@ -169,10 +169,9 @@ define float @fadd_const_multiuse_attr(float %x) {
 define double @fmul2_negated(double %a, double %b, double %c) {
 ; CHECK-LABEL: fmul2_negated:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    fmov d3, #-2.00000000
-; CHECK-NEXT:    fmul d1, d1, d3
+; CHECK-NEXT:    fadd d1, d1, d1
 ; CHECK-NEXT:    fmul d1, d1, d2
-; CHECK-NEXT:    fadd d0, d0, d1
+; CHECK-NEXT:    fsub d0, d0, d1
 ; CHECK-NEXT:    ret
   %mul = fmul double %b, 2.0
   %mul1 = fmul double %mul, %c
@@ -183,10 +182,9 @@ define double @fmul2_negated(double %a, double %b, double %c) {
 define <2 x double> @fmul2_negated_vec(<2 x double> %a, <2 x double> %b, <2 x double> %c) {
 ; CHECK-LABEL: fmul2_negated_vec:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    fmov v3.2d, #-2.00000000
-; CHECK-NEXT:    fmul v1.2d, v1.2d, v3.2d
+; CHECK-NEXT:    fadd v1.2d, v1.2d, v1.2d
 ; CHECK-NEXT:    fmul v1.2d, v1.2d, v2.2d
-; CHECK-NEXT:    fadd v0.2d, v0.2d, v1.2d
+; CHECK-NEXT:    fsub v0.2d, v0.2d, v1.2d
 ; CHECK-NEXT:    ret
   %mul = fmul <2 x double> %b, <double 2.0, double 2.0>
   %mul1 = fmul <2 x double> %mul, %c
diff --git a/llvm/test/CodeGen/X86/fadd-combines.ll b/llvm/test/CodeGen/X86/fadd-combines.ll
index 07e203f3a71..048e5c5d325 100644
--- a/llvm/test/CodeGen/X86/fadd-combines.ll
+++ b/llvm/test/CodeGen/X86/fadd-combines.ll
@@ -252,9 +252,9 @@ define float @fadd_const_multiuse_attr(float %x) #0 {
 define double @fmul2_negated(double %a, double %b, double %c) {
 ; CHECK-LABEL: fmul2_negated:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    mulsd {{.*}}(%rip), %xmm1
+; CHECK-NEXT:    addsd %xmm1, %xmm1
 ; CHECK-NEXT:    mulsd %xmm2, %xmm1
-; CHECK-NEXT:    addsd %xmm1, %xmm0
+; CHECK-NEXT:    subsd %xmm1, %xmm0
 ; CHECK-NEXT:    retq
   %mul = fmul double %b, 2.0
   %mul1 = fmul double %mul, %c
@@ -265,9 +265,9 @@ define double @fmul2_negated(double %a, double %b, double %c) {
 define <2 x double> @fmul2_negated_vec(<2 x double> %a, <2 x double> %b, <2 x double> %c) {
 ; CHECK-LABEL: fmul2_negated_vec:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    mulpd {{.*}}(%rip), %xmm1
+; CHECK-NEXT:    addpd %xmm1, %xmm1
 ; CHECK-NEXT:    mulpd %xmm2, %xmm1
-; CHECK-NEXT:    addpd %xmm1, %xmm0
+; CHECK-NEXT:    subpd %xmm1, %xmm0
 ; CHECK-NEXT:    retq
   %mul = fmul <2 x double> %b, <double 2.0, double 2.0>
   %mul1 = fmul <2 x double> %mul, %c