[PATCH] D89527: [DAGCombiner] Tighten reasscociation of visitFMA

[Qiu Chaofan via Phabricator via llvm-commits]

qiucf updated this revision to Diff 298934.
qiucf marked an inline comment as done.
qiucf added a comment.

Add another test about preventing combined into `(fmul x, c1+c2)`. (test was not pre-committed, diff is for review)


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D89527/new/

https://reviews.llvm.org/D89527

Files:
  llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
  llvm/test/CodeGen/PowerPC/fma-combine.ll
  llvm/test/CodeGen/X86/fma-scalar-combine.ll


Index: llvm/test/CodeGen/X86/fma-scalar-combine.ll
===================================================================
--- llvm/test/CodeGen/X86/fma-scalar-combine.ll
+++ llvm/test/CodeGen/X86/fma-scalar-combine.ll
@@ -547,8 +547,11 @@
 define float @fma_const_fmul(float %x) {
 ; CHECK-LABEL: fma_const_fmul:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vmulss {{.*}}(%rip), %xmm0, %xmm0 # EVEX TO VEX Compression encoding: [0xc5,0xfa,0x59,0x05,A,A,A,A]
+; CHECK-NEXT:    vmulss {{.*}}(%rip), %xmm0, %xmm1 # EVEX TO VEX Compression encoding: [0xc5,0xfa,0x59,0x0d,A,A,A,A]
 ; CHECK-NEXT:    # fixup A - offset: 4, value: {{\.LCPI.*}}-4, kind: reloc_riprel_4byte
+; CHECK-NEXT:    vfmadd132ss {{.*}}(%rip), %xmm1, %xmm0 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x71,0x99,0x05,A,A,A,A]
+; CHECK-NEXT:    # fixup A - offset: 5, value: {{\.LCPI.*}}-4, kind: reloc_riprel_4byte
+; CHECK-NEXT:    # xmm0 = (xmm0 * mem) + xmm1
 ; CHECK-NEXT:    retq # encoding: [0xc3]
   %mul1 = fmul contract float %x, 10.0
   %mul2 = fmul contract float %x, 11.0
Index: llvm/test/CodeGen/PowerPC/fma-combine.ll
===================================================================
--- llvm/test/CodeGen/PowerPC/fma-combine.ll
+++ llvm/test/CodeGen/PowerPC/fma-combine.ll
@@ -333,8 +333,11 @@
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    addis 3, 2, .LCPI9_0 at toc@ha
 ; CHECK-NEXT:    lfd 0, .LCPI9_0 at toc@l(3)
-; CHECK-NEXT:    xsmaddadp 2, 1, 0
+; CHECK-NEXT:    addis 3, 2, .LCPI9_1 at toc@ha
+; CHECK-NEXT:    lfd 3, .LCPI9_1 at toc@l(3)
+; CHECK-NEXT:    xsmuldp 0, 1, 0
 ; CHECK-NEXT:    fmr 1, 2
+; CHECK-NEXT:    xsmaddadp 1, 0, 3
 ; CHECK-NEXT:    blr
 entry:
   %0 = fmul double %a, 1.1
Index: llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
===================================================================
--- llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -13212,7 +13212,8 @@
   SelectionDAG::FlagInserter FlagsInserter(DAG, N);
 
   // FMA nodes have flags that propagate to the created nodes.
-  bool UnsafeFPMath = Options.UnsafeFPMath || isContractable(N);
+  bool UnsafeFPMath =
+      Options.UnsafeFPMath || N->getFlags().hasAllowReassociation();
 
   // Constant fold FMA.
   if (isa<ConstantFPSDNode>(N0) &&


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D89527.298934.patch
Type: text/x-patch
Size: 2262 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20201019/f945e9a7/attachment.bin>