From 1b2fe71ecf6bd647a244fc64c231e31dfe1faaa2 Mon Sep 17 00:00:00 2001
From: Qiu Chaofan <qiucofan@cn.ibm.com>
Date: Tue, 20 Oct 2020 10:13:01 +0800
Subject: [PATCH] [DAGCombiner] Tighten reasscociation of visitFMA

From LangRef, FMF contract should not enable reassociating to form
arbitrary contractions. So it should not help rearrange nodes like
(fma (fmul x, c1), c2, y) into (fma x, c1*c2, y).

Reviewed By: spatel

Differential Revision: https://reviews.llvm.org/D89527
---
 llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp |  5 +--
 llvm/test/CodeGen/PowerPC/fma-combine.ll      | 32 +++++++++++++++++++
 llvm/test/CodeGen/X86/fma-scalar-combine.ll   | 16 ++++++++++
 3 files changed, 51 insertions(+), 2 deletions(-)

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index b9b82931e410..f4cf77ba8bc0 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -13217,10 +13217,11 @@ SDValue DAGCombiner::visitFMA(SDNode *N) {
   EVT VT = N->getValueType(0);
   SDLoc DL(N);
   const TargetOptions &Options = DAG.getTarget().Options;
+  // FMA nodes have flags that propagate to the created nodes.
   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) &&
diff --git a/llvm/test/CodeGen/PowerPC/fma-combine.ll b/llvm/test/CodeGen/PowerPC/fma-combine.ll
index 217d520f8918..c07cb1cc12e5 100644
--- a/llvm/test/CodeGen/PowerPC/fma-combine.ll
+++ b/llvm/test/CodeGen/PowerPC/fma-combine.ll
@@ -313,5 +313,37 @@ entry:
   ret <2 x double> %0
 }
 
+define double @fma_combine_const(double %a, double %b) {
+; CHECK-FAST-LABEL: fma_combine_const:
+; CHECK-FAST:       # %bb.0: # %entry
+; CHECK-FAST-NEXT:    addis 3, 2, .LCPI9_0@toc@ha
+; CHECK-FAST-NEXT:    lfd 0, .LCPI9_0@toc@l(3)
+; CHECK-FAST-NEXT:    xsmaddadp 2, 1, 0
+; CHECK-FAST-NEXT:    fmr 1, 2
+; CHECK-FAST-NEXT:    blr
+;
+; CHECK-FAST-NOVSX-LABEL: fma_combine_const:
+; CHECK-FAST-NOVSX:       # %bb.0: # %entry
+; CHECK-FAST-NOVSX-NEXT:    addis 3, 2, .LCPI9_0@toc@ha
+; CHECK-FAST-NOVSX-NEXT:    lfd 0, .LCPI9_0@toc@l(3)
+; CHECK-FAST-NOVSX-NEXT:    fmadd 1, 1, 0, 2
+; CHECK-FAST-NOVSX-NEXT:    blr
+;
+; CHECK-LABEL: fma_combine_const:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    addis 3, 2, .LCPI9_0@toc@ha
+; CHECK-NEXT:    lfd 0, .LCPI9_0@toc@l(3)
+; CHECK-NEXT:    addis 3, 2, .LCPI9_1@toc@ha
+; CHECK-NEXT:    lfd 3, .LCPI9_1@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
+  %1 = call contract double @llvm.fma.f64(double %0, double 2.1, double %b)
+  ret double %1
+}
+
 declare double @llvm.fma.f64(double, double, double) nounwind readnone
 declare <2 x double> @llvm.fma.v2f64(<2 x double>, <2 x double>, <2 x double>) nounwind readnone
diff --git a/llvm/test/CodeGen/X86/fma-scalar-combine.ll b/llvm/test/CodeGen/X86/fma-scalar-combine.ll
index aa7374747cf8..1804828df293 100644
--- a/llvm/test/CodeGen/X86/fma-scalar-combine.ll
+++ b/llvm/test/CodeGen/X86/fma-scalar-combine.ll
@@ -542,3 +542,19 @@ entry:
   %8 = insertelement <2 x double> %c, double %7, i64 0
   ret <2 x double> %8
 }
+
+; Don't fold into (fmul x, c1+c2) if reassoc not set
+define float @fma_const_fmul(float %x) {
+; CHECK-LABEL: fma_const_fmul:
+; CHECK:       # %bb.0:
+; 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
+  %add1 = fadd contract float %mul1, %mul2
+  ret float %add1
+}