[SDAG][x86] check for relaxed math when matching an FP reduction

If the last step in an FP add reduction allows reassociation and doesn't care
about -0.0, then we are free to recognize that computation as a reduction
that may reorder the intermediate steps.

This is requested directly by PR42705:
https://bugs.llvm.org/show_bug.cgi?id=42705
and solves PR42947 (if horizontal math instructions are actually faster than
the alternative):
https://bugs.llvm.org/show_bug.cgi?id=42947

Differential Revision: https://reviews.llvm.org/D66236

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@368995 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index be41e12b78518b51a60b5eeaf8edf4351a6848cf..d8711f82ad15a78636f650e594e3ada0cbb61262 100644 (file)
--- a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -9002,14 +9002,27 @@ SelectionDAG::matchBinOpReduction(SDNode *Extract, ISD::NodeType &BinOp,
       !isNullConstant(Extract->getOperand(1)))
     return SDValue();
 
-  SDValue Op = Extract->getOperand(0);
-
   // Match against one of the candidate binary ops.
+  SDValue Op = Extract->getOperand(0);
   if (llvm::none_of(CandidateBinOps, [Op](ISD::NodeType BinOp) {
         return Op.getOpcode() == unsigned(BinOp);
       }))
     return SDValue();
+
+  // Floating-point reductions may require relaxed constraints on the final step
+  // of the reduction because they may reorder intermediate operations.
   unsigned CandidateBinOp = Op.getOpcode();
+  if (Op.getValueType().isFloatingPoint()) {
+    SDNodeFlags Flags = Op->getFlags();
+    switch (CandidateBinOp) {
+    case ISD::FADD:
+      if (!Flags.hasNoSignedZeros() || !Flags.hasAllowReassociation())
+        return SDValue();
+      break;
+    default:
+      llvm_unreachable("Unhandled FP opcode for binop reduction");
+    }
+  }
 
   // Matching failed - attempt to see if we did enough stages that a partial
   // reduction from a subvector is possible.

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 4779743378ebfa5c5c7f4b94c9ddf61f74f39883..7e36a05e8b126e7a73f5fa4949aa1b02d6fea62c 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -35395,9 +35395,9 @@ static SDValue combineReductionToHorizontal(SDNode *ExtElt, SelectionDAG &DAG,
                                             const X86Subtarget &Subtarget) {
   assert(ExtElt->getOpcode() == ISD::EXTRACT_VECTOR_ELT && "Unexpected caller");
 
-  // TODO: Allow FADD with reduction and/or reassociation and no-signed-zeros.
   ISD::NodeType Opc;
-  SDValue Rdx = DAG.matchBinOpReduction(ExtElt, Opc, {ISD::ADD}, true);
+  SDValue Rdx =
+      DAG.matchBinOpReduction(ExtElt, Opc, {ISD::ADD, ISD::FADD}, true);
   if (!Rdx)
     return SDValue();
 

diff --git a/test/CodeGen/X86/haddsub.ll b/test/CodeGen/X86/haddsub.ll
index 78270a5cfe4f60a1d39fd2d6a1dd9fd7075420d8..597066108bfcae7c0598d284a77899c0ce3375b1 100644 (file)
--- a/test/CodeGen/X86/haddsub.ll
+++ b/test/CodeGen/X86/haddsub.ll
@@ -1645,10 +1645,8 @@ define float @fadd_reduce_v8f32(float %a0, <8 x float> %a1) {
 ;
 ; SSE3-FAST-LABEL: fadd_reduce_v8f32:
 ; SSE3-FAST:       # %bb.0:
-; SSE3-FAST-NEXT:    addps %xmm2, %xmm1
-; SSE3-FAST-NEXT:    movaps %xmm1, %xmm2
-; SSE3-FAST-NEXT:    unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm1[1]
-; SSE3-FAST-NEXT:    addps %xmm1, %xmm2
+; SSE3-FAST-NEXT:    haddps %xmm1, %xmm2
+; SSE3-FAST-NEXT:    haddps %xmm2, %xmm2
 ; SSE3-FAST-NEXT:    haddps %xmm2, %xmm2
 ; SSE3-FAST-NEXT:    addss %xmm2, %xmm0
 ; SSE3-FAST-NEXT:    retq
@@ -1668,9 +1666,8 @@ define float @fadd_reduce_v8f32(float %a0, <8 x float> %a1) {
 ; AVX-FAST-LABEL: fadd_reduce_v8f32:
 ; AVX-FAST:       # %bb.0:
 ; AVX-FAST-NEXT:    vextractf128 $1, %ymm1, %xmm2
-; AVX-FAST-NEXT:    vaddps %xmm2, %xmm1, %xmm1
-; AVX-FAST-NEXT:    vpermilpd {{.*#+}} xmm2 = xmm1[1,0]
-; AVX-FAST-NEXT:    vaddps %xmm2, %xmm1, %xmm1
+; AVX-FAST-NEXT:    vhaddps %xmm1, %xmm2, %xmm1
+; AVX-FAST-NEXT:    vhaddps %xmm1, %xmm1, %xmm1
 ; AVX-FAST-NEXT:    vhaddps %xmm1, %xmm1, %xmm1
 ; AVX-FAST-NEXT:    vaddss %xmm1, %xmm0, %xmm0
 ; AVX-FAST-NEXT:    vzeroupper
@@ -1691,9 +1688,9 @@ define double @fadd_reduce_v4f64(double %a0, <4 x double> %a1) {
 ;
 ; SSE3-FAST-LABEL: fadd_reduce_v4f64:
 ; SSE3-FAST:       # %bb.0:
-; SSE3-FAST-NEXT:    addpd %xmm2, %xmm1
-; SSE3-FAST-NEXT:    haddpd %xmm1, %xmm1
-; SSE3-FAST-NEXT:    addsd %xmm1, %xmm0
+; SSE3-FAST-NEXT:    haddpd %xmm1, %xmm2
+; SSE3-FAST-NEXT:    haddpd %xmm2, %xmm2
+; SSE3-FAST-NEXT:    addsd %xmm2, %xmm0
 ; SSE3-FAST-NEXT:    retq
 ;
 ; AVX-SLOW-LABEL: fadd_reduce_v4f64:
@@ -1709,7 +1706,7 @@ define double @fadd_reduce_v4f64(double %a0, <4 x double> %a1) {
 ; AVX-FAST-LABEL: fadd_reduce_v4f64:
 ; AVX-FAST:       # %bb.0:
 ; AVX-FAST-NEXT:    vextractf128 $1, %ymm1, %xmm2
-; AVX-FAST-NEXT:    vaddpd %xmm2, %xmm1, %xmm1
+; AVX-FAST-NEXT:    vhaddpd %xmm1, %xmm2, %xmm1
 ; AVX-FAST-NEXT:    vhaddpd %xmm1, %xmm1, %xmm1
 ; AVX-FAST-NEXT:    vaddsd %xmm1, %xmm0, %xmm0
 ; AVX-FAST-NEXT:    vzeroupper
@@ -2017,8 +2014,7 @@ define float @partial_reduction_fadd_v8f32(<8 x float> %x) {
 ;
 ; AVX-FAST-LABEL: partial_reduction_fadd_v8f32:
 ; AVX-FAST:       # %bb.0:
-; AVX-FAST-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
-; AVX-FAST-NEXT:    vaddps %xmm1, %xmm0, %xmm0
+; AVX-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
 ; AVX-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
 ; AVX-FAST-NEXT:    vzeroupper
 ; AVX-FAST-NEXT:    retq
@@ -2030,6 +2026,9 @@ define float @partial_reduction_fadd_v8f32(<8 x float> %x) {
   ret float %r
 }
 
+; Negative test - only the flags on the final math op in the
+; sequence determine whether we can transform to horizontal ops.
+
 define float @partial_reduction_fadd_v8f32_wrong_flags(<8 x float> %x) {
 ; SSE3-SLOW-LABEL: partial_reduction_fadd_v8f32_wrong_flags:
 ; SSE3-SLOW:       # %bb.0:
@@ -2105,8 +2104,7 @@ define float @partial_reduction_fadd_v16f32(<16 x float> %x) {
 ;
 ; AVX-FAST-LABEL: partial_reduction_fadd_v16f32:
 ; AVX-FAST:       # %bb.0:
-; AVX-FAST-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
-; AVX-FAST-NEXT:    vaddps %xmm1, %xmm0, %xmm0
+; AVX-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
 ; AVX-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
 ; AVX-FAST-NEXT:    vzeroupper
 ; AVX-FAST-NEXT:    retq

diff --git a/test/CodeGen/X86/vector-reduce-fadd-fast.ll b/test/CodeGen/X86/vector-reduce-fadd-fast.ll
index acc02a2d84be02f576fc4e22ad76cd5aeabdaf7f..c94f584dfdb0f2ecb937c487feb42441b2c3d897 100644 (file)
--- a/test/CodeGen/X86/vector-reduce-fadd-fast.ll
+++ b/test/CodeGen/X86/vector-reduce-fadd-fast.ll
@@ -90,8 +90,7 @@ define float @test_v4f32(float %a0, <4 x float> %a1) {
 ;
 ; AVX1-FAST-LABEL: test_v4f32:
 ; AVX1-FAST:       # %bb.0:
-; AVX1-FAST-NEXT:    vpermilpd {{.*#+}} xmm2 = xmm1[1,0]
-; AVX1-FAST-NEXT:    vaddps %xmm2, %xmm1, %xmm1
+; AVX1-FAST-NEXT:    vhaddps %xmm1, %xmm1, %xmm1
 ; AVX1-FAST-NEXT:    vhaddps %xmm1, %xmm1, %xmm1
 ; AVX1-FAST-NEXT:    vaddss %xmm1, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    retq
@@ -156,9 +155,8 @@ define float @test_v8f32(float %a0, <8 x float> %a1) {
 ; AVX1-FAST-LABEL: test_v8f32:
 ; AVX1-FAST:       # %bb.0:
 ; AVX1-FAST-NEXT:    vextractf128 $1, %ymm1, %xmm2
-; AVX1-FAST-NEXT:    vaddps %xmm2, %xmm1, %xmm1
-; AVX1-FAST-NEXT:    vpermilpd {{.*#+}} xmm2 = xmm1[1,0]
-; AVX1-FAST-NEXT:    vaddps %xmm2, %xmm1, %xmm1
+; AVX1-FAST-NEXT:    vhaddps %xmm1, %xmm2, %xmm1
+; AVX1-FAST-NEXT:    vhaddps %xmm1, %xmm1, %xmm1
 ; AVX1-FAST-NEXT:    vhaddps %xmm1, %xmm1, %xmm1
 ; AVX1-FAST-NEXT:    vaddss %xmm1, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    vzeroupper
@@ -350,8 +348,7 @@ define float @test_v4f32_zero(<4 x float> %a0) {
 ;
 ; AVX1-FAST-LABEL: test_v4f32_zero:
 ; AVX1-FAST:       # %bb.0:
-; AVX1-FAST-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
-; AVX1-FAST-NEXT:    vaddps %xmm1, %xmm0, %xmm0
+; AVX1-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    retq
 ;
@@ -411,9 +408,8 @@ define float @test_v8f32_zero(<8 x float> %a0) {
 ; AVX1-FAST-LABEL: test_v8f32_zero:
 ; AVX1-FAST:       # %bb.0:
 ; AVX1-FAST-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; AVX1-FAST-NEXT:    vaddps %xmm1, %xmm0, %xmm0
-; AVX1-FAST-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
-; AVX1-FAST-NEXT:    vaddps %xmm1, %xmm0, %xmm0
+; AVX1-FAST-NEXT:    vhaddps %xmm0, %xmm1, %xmm0
+; AVX1-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    vzeroupper
 ; AVX1-FAST-NEXT:    retq
@@ -597,8 +593,7 @@ define float @test_v4f32_undef(<4 x float> %a0) {
 ;
 ; AVX1-FAST-LABEL: test_v4f32_undef:
 ; AVX1-FAST:       # %bb.0:
-; AVX1-FAST-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
-; AVX1-FAST-NEXT:    vaddps %xmm1, %xmm0, %xmm0
+; AVX1-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    retq
 ;
@@ -658,9 +653,8 @@ define float @test_v8f32_undef(<8 x float> %a0) {
 ; AVX1-FAST-LABEL: test_v8f32_undef:
 ; AVX1-FAST:       # %bb.0:
 ; AVX1-FAST-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; AVX1-FAST-NEXT:    vaddps %xmm1, %xmm0, %xmm0
-; AVX1-FAST-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm0[1,0]
-; AVX1-FAST-NEXT:    vaddps %xmm1, %xmm0, %xmm0
+; AVX1-FAST-NEXT:    vhaddps %xmm0, %xmm1, %xmm0
+; AVX1-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    vzeroupper
 ; AVX1-FAST-NEXT:    retq
@@ -834,7 +828,7 @@ define double @test_v4f64(double %a0, <4 x double> %a1) {
 ; AVX1-FAST-LABEL: test_v4f64:
 ; AVX1-FAST:       # %bb.0:
 ; AVX1-FAST-NEXT:    vextractf128 $1, %ymm1, %xmm2
-; AVX1-FAST-NEXT:    vaddpd %xmm2, %xmm1, %xmm1
+; AVX1-FAST-NEXT:    vhaddpd %xmm1, %xmm2, %xmm1
 ; AVX1-FAST-NEXT:    vhaddpd %xmm1, %xmm1, %xmm1
 ; AVX1-FAST-NEXT:    vaddsd %xmm1, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    vzeroupper
@@ -1053,7 +1047,7 @@ define double @test_v4f64_zero(<4 x double> %a0) {
 ; AVX1-FAST-LABEL: test_v4f64_zero:
 ; AVX1-FAST:       # %bb.0:
 ; AVX1-FAST-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; AVX1-FAST-NEXT:    vaddpd %xmm1, %xmm0, %xmm0
+; AVX1-FAST-NEXT:    vhaddpd %xmm0, %xmm1, %xmm0
 ; AVX1-FAST-NEXT:    vhaddpd %xmm0, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    vzeroupper
 ; AVX1-FAST-NEXT:    retq
@@ -1260,7 +1254,7 @@ define double @test_v4f64_undef(<4 x double> %a0) {
 ; AVX1-FAST-LABEL: test_v4f64_undef:
 ; AVX1-FAST:       # %bb.0:
 ; AVX1-FAST-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; AVX1-FAST-NEXT:    vaddpd %xmm1, %xmm0, %xmm0
+; AVX1-FAST-NEXT:    vhaddpd %xmm0, %xmm1, %xmm0
 ; AVX1-FAST-NEXT:    vhaddpd %xmm0, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    vzeroupper
 ; AVX1-FAST-NEXT:    retq