<div dir="ltr">Looks like this caused <a href="https://llvm.org/bugs/show_bug.cgi?id=26652">https://llvm.org/bugs/show_bug.cgi?id=26652</a></div><div class="gmail_extra"><br><div class="gmail_quote">On Tue, Feb 16, 2016 at 10:37 PM, Cong Hou via llvm-commits <span dir="ltr"><<a href="mailto:llvm-commits@lists.llvm.org" target="_blank">llvm-commits@lists.llvm.org</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">Author: conghou<br>
Date: Wed Feb 17 00:37:04 2016<br>
New Revision: 261070<br>
<br>
URL: <a href="http://llvm.org/viewvc/llvm-project?rev=261070&view=rev" rel="noreferrer" target="_blank">http://llvm.org/viewvc/llvm-project?rev=261070&view=rev</a><br>
Log:<br>
Detecte vector reduction operations just before instruction selection.<br>
<br>
This patch detects vector reductions before instruction selection. Vector<br>
reductions are vectorized reduction operations, and for such operations we have<br>
freedom to reorganize the elements of the result as long as the reduction of them<br>
stay unchanged. This will enable some reduction pattern recognition during<br>
instruction combine such as SAD/dot-product on X86. A flag is added to<br>
SDNodeFlags to mark those vector reduction nodes to be checked during instruction<br>
combine.<br>
<br>
To detect those vector reductions, we search def-use chains starting from the<br>
given instruction, and check if all uses fall into two categories:<br>
<br>
1. Reduction with another vector.<br>
2. Reduction on all elements.<br>
<br>
in which 2 is detected by recognizing the pattern that the loop vectorizer<br>
generates to reduce all elements in the vector outside of the loop, which<br>
includes several ShuffleVector and one ExtractElement instructions.<br>
<br>
<br>
Differential revision: <a href="http://reviews.llvm.org/D15250" rel="noreferrer" target="_blank">http://reviews.llvm.org/D15250</a><br>
<br>
<br>
Added:<br>
llvm/trunk/test/CodeGen/Generic/vector-redux.ll<br>
Modified:<br>
llvm/trunk/include/llvm/CodeGen/SelectionDAGNodes.h<br>
llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp<br>
<br>
Modified: llvm/trunk/include/llvm/CodeGen/SelectionDAGNodes.h<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/include/llvm/CodeGen/SelectionDAGNodes.h?rev=261070&r1=261069&r2=261070&view=diff" rel="noreferrer" target="_blank">http://llvm.org/viewvc/llvm-project/llvm/trunk/include/llvm/CodeGen/SelectionDAGNodes.h?rev=261070&r1=261069&r2=261070&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/include/llvm/CodeGen/SelectionDAGNodes.h (original)<br>
+++ llvm/trunk/include/llvm/CodeGen/SelectionDAGNodes.h Wed Feb 17 00:37:04 2016<br>
@@ -328,6 +328,7 @@ private:<br>
bool NoInfs : 1;<br>
bool NoSignedZeros : 1;<br>
bool AllowReciprocal : 1;<br>
+ bool VectorReduction : 1;<br>
<br>
public:<br>
/// Default constructor turns off all optimization flags.<br>
@@ -340,6 +341,7 @@ public:<br>
NoInfs = false;<br>
NoSignedZeros = false;<br>
AllowReciprocal = false;<br>
+ VectorReduction = false;<br>
}<br>
<br>
// These are mutators for each flag.<br>
@@ -351,6 +353,7 @@ public:<br>
void setNoInfs(bool b) { NoInfs = b; }<br>
void setNoSignedZeros(bool b) { NoSignedZeros = b; }<br>
void setAllowReciprocal(bool b) { AllowReciprocal = b; }<br>
+ void setVectorReduction(bool b) { VectorReduction = b; }<br>
<br>
// These are accessors for each flag.<br>
bool hasNoUnsignedWrap() const { return NoUnsignedWrap; }<br>
@@ -361,6 +364,7 @@ public:<br>
bool hasNoInfs() const { return NoInfs; }<br>
bool hasNoSignedZeros() const { return NoSignedZeros; }<br>
bool hasAllowReciprocal() const { return AllowReciprocal; }<br>
+ bool hasVectorReduction() const { return VectorReduction; }<br>
<br>
/// Return a raw encoding of the flags.<br>
/// This function should only be used to add data to the NodeID value.<br>
<br>
Modified: llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp?rev=261070&r1=261069&r2=261070&view=diff" rel="noreferrer" target="_blank">http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp?rev=261070&r1=261069&r2=261070&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp (original)<br>
+++ llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp Wed Feb 17 00:37:04 2016<br>
@@ -2308,6 +2308,125 @@ void SelectionDAGBuilder::visitFSub(cons<br>
visitBinary(I, ISD::FSUB);<br>
}<br>
<br>
+/// Checks if the given instruction performs a vector reduction, in which case<br>
+/// we have the freedom to alter the elements in the result as long as the<br>
+/// reduction of them stays unchanged.<br>
+static bool isVectorReductionOp(const User *I) {<br>
+ const Instruction *Inst = dyn_cast<Instruction>(I);<br>
+ if (!Inst || !Inst->getType()->isVectorTy())<br>
+ return false;<br>
+<br>
+ auto OpCode = Inst->getOpcode();<br>
+ switch (OpCode) {<br>
+ case Instruction::Add:<br>
+ case Instruction::Mul:<br>
+ case Instruction::And:<br>
+ case Instruction::Or:<br>
+ case Instruction::Xor:<br>
+ break;<br>
+ case Instruction::FAdd:<br>
+ case Instruction::FMul:<br>
+ if (const FPMathOperator *FPOp = dyn_cast<const FPMathOperator>(Inst))<br>
+ if (FPOp->getFastMathFlags().unsafeAlgebra())<br>
+ break;<br>
+ // Fall through.<br>
+ default:<br>
+ return false;<br>
+ }<br>
+<br>
+ unsigned ElemNum = Inst->getType()->getVectorNumElements();<br>
+ unsigned ElemNumToReduce = ElemNum;<br>
+<br>
+ // Do DFS search on the def-use chain from the given instruction. We only<br>
+ // allow four kinds of operations during the search until we reach the<br>
+ // instruction that extracts the first element from the vector:<br>
+ //<br>
+ // 1. The reduction operation of the same opcode as the given instruction.<br>
+ //<br>
+ // 2. PHI node.<br>
+ //<br>
+ // 3. ShuffleVector instruction together with a reduction operation that<br>
+ // does a partial reduction.<br>
+ //<br>
+ // 4. ExtractElement that extracts the first element from the vector, and we<br>
+ // stop searching the def-use chain here.<br>
+ //<br>
+ // 3 & 4 above perform a reduction on all elements of the vector. We push defs<br>
+ // from 1-3 to the stack to continue the DFS. The given instruction is not<br>
+ // a reduction operation if we meet any other instructions other than those<br>
+ // listed above.<br>
+<br>
+ SmallVector<const User *, 16> UsersToVisit{Inst};<br>
+ SmallPtrSet<const User *, 16> Visited;<br>
+ bool ReduxExtracted = false;<br>
+<br>
+ while (!UsersToVisit.empty()) {<br>
+ auto User = UsersToVisit.back();<br>
+ UsersToVisit.pop_back();<br>
+ if (!Visited.insert(User).second)<br>
+ continue;<br>
+<br>
+ for (const auto &U : User->users()) {<br>
+ auto Inst = dyn_cast<Instruction>(U);<br>
+ if (!Inst)<br>
+ return false;<br>
+<br>
+ if (Inst->getOpcode() == OpCode || isa<PHINode>(U)) {<br>
+ if (const FPMathOperator *FPOp = dyn_cast<const FPMathOperator>(Inst))<br>
+ if (!isa<PHINode>(FPOp) && !FPOp->getFastMathFlags().unsafeAlgebra())<br>
+ return false;<br>
+ UsersToVisit.push_back(U);<br>
+ } else if (const ShuffleVectorInst *ShufInst =<br>
+ dyn_cast<ShuffleVectorInst>(U)) {<br>
+ // Detect the following pattern: A ShuffleVector instruction together<br>
+ // with a reduction that do partial reduction on the first and second<br>
+ // ElemNumToReduce / 2 elements, and store the result in<br>
+ // ElemNumToReduce / 2 elements in another vector.<br>
+<br>
+ if (ElemNumToReduce == 1)<br>
+ return false;<br>
+ if (!isa<UndefValue>(U->getOperand(1)))<br>
+ return false;<br>
+ for (unsigned i = 0; i < ElemNumToReduce / 2; ++i)<br>
+ if (ShufInst->getMaskValue(i) != int(i + ElemNumToReduce / 2))<br>
+ return false;<br>
+ for (unsigned i = ElemNumToReduce / 2; i < ElemNum; ++i)<br>
+ if (ShufInst->getMaskValue(i) != -1)<br>
+ return false;<br>
+<br>
+ // There is only one user of this ShuffleVector instruction, which must<br>
+ // be a reduction operation.<br>
+ if (!U->hasOneUse())<br>
+ return false;<br>
+<br>
+ auto U2 = dyn_cast<Instruction>(*U->user_begin());<br>
+ if (!U2 || U2->getOpcode() != OpCode)<br>
+ return false;<br>
+<br>
+ // Check operands of the reduction operation.<br>
+ if ((U2->getOperand(0) == U->getOperand(0) && U2->getOperand(1) == U) ||<br>
+ (U2->getOperand(1) == U->getOperand(0) && U2->getOperand(0) == U)) {<br>
+ UsersToVisit.push_back(U2);<br>
+ ElemNumToReduce /= 2;<br>
+ } else<br>
+ return false;<br>
+ } else if (isa<ExtractElementInst>(U)) {<br>
+ // At this moment we should have reduced all elements in the vector.<br>
+ if (ElemNumToReduce != 1)<br>
+ return false;<br>
+<br>
+ const ConstantInt *Val = dyn_cast<ConstantInt>(U->getOperand(1));<br>
+ if (!Val || Val->getZExtValue() != 0)<br>
+ return false;<br>
+<br>
+ ReduxExtracted = true;<br>
+ } else<br>
+ return false;<br>
+ }<br>
+ }<br>
+ return ReduxExtracted;<br>
+}<br>
+<br>
void SelectionDAGBuilder::visitBinary(const User &I, unsigned OpCode) {<br>
SDValue Op1 = getValue(I.getOperand(0));<br>
SDValue Op2 = getValue(I.getOperand(1));<br>
@@ -2315,6 +2434,7 @@ void SelectionDAGBuilder::visitBinary(co<br>
bool nuw = false;<br>
bool nsw = false;<br>
bool exact = false;<br>
+ bool vec_redux = false;<br>
FastMathFlags FMF;<br>
<br>
if (const OverflowingBinaryOperator *OFBinOp =<br>
@@ -2328,10 +2448,16 @@ void SelectionDAGBuilder::visitBinary(co<br>
if (const FPMathOperator *FPOp = dyn_cast<const FPMathOperator>(&I))<br>
FMF = FPOp->getFastMathFlags();<br>
<br>
+ if (isVectorReductionOp(&I)) {<br>
+ vec_redux = true;<br>
+ DEBUG(dbgs() << "Detected a reduction operation:" << I << "\n");<br>
+ }<br>
+<br>
SDNodeFlags Flags;<br>
Flags.setExact(exact);<br>
Flags.setNoSignedWrap(nsw);<br>
Flags.setNoUnsignedWrap(nuw);<br>
+ Flags.setVectorReduction(vec_redux);<br>
if (EnableFMFInDAG) {<br>
Flags.setAllowReciprocal(FMF.allowReciprocal());<br>
Flags.setNoInfs(FMF.noInfs());<br>
<br>
Added: llvm/trunk/test/CodeGen/Generic/vector-redux.ll<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/Generic/vector-redux.ll?rev=261070&view=auto" rel="noreferrer" target="_blank">http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/Generic/vector-redux.ll?rev=261070&view=auto</a><br>
==============================================================================<br>
--- llvm/trunk/test/CodeGen/Generic/vector-redux.ll (added)<br>
+++ llvm/trunk/test/CodeGen/Generic/vector-redux.ll Wed Feb 17 00:37:04 2016<br>
@@ -0,0 +1,85 @@<br>
+; RUN: llc < %s -debug-only=isel -o /dev/null 2>&1 | FileCheck %s<br>
+; REQUIRES: asserts<br>
+<br>
+@a = global [1024 x i32] zeroinitializer, align 16<br>
+<br>
+define float @reduce_add_float(float* nocapture readonly %a) {<br>
+; CHECK-LABEL: reduce_add_float<br>
+; CHECK: Detected a reduction operation: {{.*}} fadd fast<br>
+; CHECK: Detected a reduction operation: {{.*}} fadd fast<br>
+; CHECK: Detected a reduction operation: {{.*}} fadd fast<br>
+; CHECK: Detected a reduction operation: {{.*}} fadd fast<br>
+; CHECK: Detected a reduction operation: {{.*}} fadd fast<br>
+; CHECK: Detected a reduction operation: {{.*}} fadd fast<br>
+; CHECK: Detected a reduction operation: {{.*}} fadd fast<br>
+; CHECK: Detected a reduction operation: {{.*}} fadd fast<br>
+; CHECK: Detected a reduction operation: {{.*}} fadd fast<br>
+; CHECK: Detected a reduction operation: {{.*}} fadd fast<br>
+; CHECK: Detected a reduction operation: {{.*}} fadd fast<br>
+;<br>
+entry:<br>
+ br label %vector.body<br>
+<br>
+vector.body:<br>
+ %index = phi i64 [ 0, %entry ], [ %index.next.4, %vector.body ]<br>
+ %vec.phi = phi <4 x float> [ zeroinitializer, %entry ], [ %28, %vector.body ]<br>
+ %vec.phi9 = phi <4 x float> [ zeroinitializer, %entry ], [ %29, %vector.body ]<br>
+ %0 = getelementptr inbounds float, float* %a, i64 %index<br>
+ %1 = bitcast float* %0 to <4 x float>*<br>
+ %wide.load = load <4 x float>, <4 x float>* %1, align 4<br>
+ %2 = getelementptr float, float* %0, i64 4<br>
+ %3 = bitcast float* %2 to <4 x float>*<br>
+ %wide.load10 = load <4 x float>, <4 x float>* %3, align 4<br>
+ %4 = fadd fast <4 x float> %wide.load, %vec.phi<br>
+ %5 = fadd fast <4 x float> %wide.load10, %vec.phi9<br>
+ %index.next = add nuw nsw i64 %index, 8<br>
+ %6 = getelementptr inbounds float, float* %a, i64 %index.next<br>
+ %7 = bitcast float* %6 to <4 x float>*<br>
+ %wide.load.1 = load <4 x float>, <4 x float>* %7, align 4<br>
+ %8 = getelementptr float, float* %6, i64 4<br>
+ %9 = bitcast float* %8 to <4 x float>*<br>
+ %wide.load10.1 = load <4 x float>, <4 x float>* %9, align 4<br>
+ %10 = fadd fast <4 x float> %wide.load.1, %4<br>
+ %11 = fadd fast <4 x float> %wide.load10.1, %5<br>
+ %index.next.1 = add nsw i64 %index, 16<br>
+ %12 = getelementptr inbounds float, float* %a, i64 %index.next.1<br>
+ %13 = bitcast float* %12 to <4 x float>*<br>
+ %wide.load.2 = load <4 x float>, <4 x float>* %13, align 4<br>
+ %14 = getelementptr float, float* %12, i64 4<br>
+ %15 = bitcast float* %14 to <4 x float>*<br>
+ %wide.load10.2 = load <4 x float>, <4 x float>* %15, align 4<br>
+ %16 = fadd fast <4 x float> %wide.load.2, %10<br>
+ %17 = fadd fast <4 x float> %wide.load10.2, %11<br>
+ %index.next.2 = add nsw i64 %index, 24<br>
+ %18 = getelementptr inbounds float, float* %a, i64 %index.next.2<br>
+ %19 = bitcast float* %18 to <4 x float>*<br>
+ %wide.load.3 = load <4 x float>, <4 x float>* %19, align 4<br>
+ %20 = getelementptr float, float* %18, i64 4<br>
+ %21 = bitcast float* %20 to <4 x float>*<br>
+ %wide.load10.3 = load <4 x float>, <4 x float>* %21, align 4<br>
+ %22 = fadd fast <4 x float> %wide.load.3, %16<br>
+ %23 = fadd fast <4 x float> %wide.load10.3, %17<br>
+ %index.next.3 = add nsw i64 %index, 32<br>
+ %24 = getelementptr inbounds float, float* %a, i64 %index.next.3<br>
+ %25 = bitcast float* %24 to <4 x float>*<br>
+ %wide.load.4 = load <4 x float>, <4 x float>* %25, align 4<br>
+ %26 = getelementptr float, float* %24, i64 4<br>
+ %27 = bitcast float* %26 to <4 x float>*<br>
+ %wide.load10.4 = load <4 x float>, <4 x float>* %27, align 4<br>
+ %28 = fadd fast <4 x float> %wide.load.4, %22<br>
+ %29 = fadd fast <4 x float> %wide.load10.4, %23<br>
+ %index.next.4 = add nsw i64 %index, 40<br>
+ %30 = icmp eq i64 %index.next.4, 1000<br>
+ br i1 %30, label %middle.block, label %vector.body<br>
+<br>
+middle.block:<br>
+ %.lcssa15 = phi <4 x float> [ %29, %vector.body ]<br>
+ %.lcssa = phi <4 x float> [ %28, %vector.body ]<br>
+ %bin.rdx = fadd fast <4 x float> %.lcssa15, %.lcssa<br>
+ %rdx.shuf = shufflevector <4 x float> %bin.rdx, <4 x float> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef><br>
+ %bin.rdx11 = fadd fast <4 x float> %bin.rdx, %rdx.shuf<br>
+ %rdx.shuf12 = shufflevector <4 x float> %bin.rdx11, <4 x float> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef><br>
+ %bin.rdx13 = fadd fast <4 x float> %bin.rdx11, %rdx.shuf12<br>
+ %31 = extractelement <4 x float> %bin.rdx13, i32 0<br>
+ ret float %31<br>
+}<br>
<br>
<br>
_______________________________________________<br>
llvm-commits mailing list<br>
<a href="mailto:llvm-commits@lists.llvm.org">llvm-commits@lists.llvm.org</a><br>
<a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits" rel="noreferrer" target="_blank">http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits</a><br>
</blockquote></div><br></div>