Loop Vectorizer - skiping "bitcast" allows to vectorize more loops

[delena]

Summary

Vectorization of memory instruction (Load/Store) is possible when the pointer is coming from GEP. The GEP analysis allows to estimate the profit.
In some cases we have a "bitcast" between GEP and memory instruction.
I added code that skips the "bitcast".

I replaced vectorization-remarks-profitable.ll test. As a result of the optimization, the remarks about non-beneficial vectorization disappeared and loop became vectorizable. I put another loops, that generate the same remarks as the previous.

[delena]

Ping *

[jmolloy]

Hi Elena,

I think this looks reasonable.

Cheers,

James

[delena]

Thanks a lot, James!

[mzolotukhin]

Hi Elena,

The change looks reasonable, one question from me inline.

Thanks,
Michael

../test/Transforms/LoopVectorize/gep_with_bitcast.ll

28–29

Could it be a problem if we cast from, say, double** to i8*? Should we check the type we're pointing to when we do this?

[chapuni]

Seems this caused miscompilation for lib/ExecutionEngine/Interpreter/Execution.cpp
Investigating.

../lib/Transforms/Vectorize/LoopVectorize.cpp

262

if (auto GEP = dyn_cast<GetElementPtrInst>(Ptr)) return GEP;

266

if (auto BC = dyn_cast<BCI>(Ptr)) if (auto GEP = dyn_cast<GEP>(BC->getOperand(0))) return GEP;

[chapuni]

See also; https://llvm.org/bugs/show_bug.cgi?id=25324

[delena]

I will revert the commit tomorrow morning. Today is too late for me, I'm connected from home.

• Elena

[chapuni]

Elena, I have reverted this in r251436.

../lib/Transforms/Vectorize/LoopVectorize.cpp

 /// If the incoming type is void, we return void. If the VF is 1, we return
 /// the scalar type.
 static Type* ToVectorTy(Type *Scalar, unsigned VF) {
   if (Scalar->isVoidTy() || VF == 1)
     return Scalar;
   return VectorType::get(Scalar, VF);
 }
 
+/// A helper function that returns GEP instruction and knows to skip
+/// 'bitcast'.
+static GetElementPtrInst *getGEPInstruction(Value *Ptr) {
+
+  if (isa<GetElementPtrInst>(Ptr))
+    return cast<GetElementPtrInst>(Ptr);
+
+  if (isa<BitCastInst>(Ptr) &&
+      isa<GetElementPtrInst>(cast<BitCastInst>(Ptr)->getOperand(0)))
+    return cast<GetElementPtrInst>(cast<BitCastInst>(Ptr)->getOperand(0));
+  return nullptr;
+}
+
 /// InnerLoopVectorizer vectorizes loops which contain only one basic
 /// block to a specified vectorization factor (VF).
 /// This class performs the widening of scalars into vectors, or multiple
 /// scalars. This class also implements the following features:
 /// * It inserts an epilogue loop for handling loops that don't have iteration
 ///   counts that are known to be a multiple of the vectorization factor.
 /// * It handles the code generation for reduction variables.
 /// * Scalarization (implementation using scalars) of un-vectorizable
int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
 
   // If this value is a pointer induction variable we know it is consecutive.
   PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
   if (Phi && Inductions.count(Phi)) {
     InductionDescriptor II = Inductions[Phi];
     return II.getConsecutiveDirection();
   }
 
-  GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
+  GetElementPtrInst *Gep = getGEPInstruction(Ptr);
   if (!Gep)
     return 0;
 
   unsigned NumOperands = Gep->getNumOperands();
   Value *GpPtr = Gep->getPointerOperand();
   // If this GEP value is a consecutive pointer induction variable and all of
   // the indices are constant then we know it is consecutive. We can
   Phi = dyn_cast<PHINode>(GpPtr);
void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
   bool UniformLoad = LI && Legal->isUniform(Ptr);
   if (!ConsecutiveStride || UniformLoad)
     return scalarizeInstruction(Instr);
 
   Constant *Zero = Builder.getInt32(0);
   VectorParts &Entry = WidenMap.get(Instr);
 
   // Handle consecutive loads/stores.
-  GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
+  GetElementPtrInst *Gep = getGEPInstruction(Ptr);
   if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
     setDebugLocFromInst(Builder, Gep);
     Value *PtrOperand = Gep->getPointerOperand();
     Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
     FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
 
     // Create the new GEP with the new induction variable.
     GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr) {
   // Handle Stores:
   if (SI) {
     assert(!Legal->isUniform(SI->getPointerOperand()) &&
            "We do not allow storing to uniform addresses");
     setDebugLocFromInst(Builder, SI);
     // We don't want to update the value in the map as it might be used in
     // another expression. So don't use a reference type for "StoredVal".
     VectorParts StoredVal = getVectorValue(SI->getValueOperand());
-    
     for (unsigned Part = 0; Part < UF; ++Part) {
       // Calculate the pointer for the specific unroll-part.
       Value *PartPtr =
           Builder.CreateGEP(nullptr, Ptr, Builder.getInt32(Part * VF));
 
       if (Reverse) {
         // If we store to reverse consecutive memory locations, then we need
         // to reverse the order of elements in the stored value.

../test/Transforms/LoopVectorize/X86/vectorization-remarks-profitable.ll

 ; RUN: opt < %s -loop-vectorize -pass-remarks-analysis='loop-vectorize' -mtriple=x86_64-unknown-linux -S 2>&1 | FileCheck %s
 
 ; Verify analysis remarks are generated when interleaving is not beneficial.
-; CHECK: remark: vectorization-remarks-profitable.c:4:14: the cost-model indicates that vectorization is not beneficial
-; CHECK: remark: vectorization-remarks-profitable.c:4:14: the cost-model indicates that interleaving is not beneficial and is explicitly disabled or interleave count is set to 1
-; CHECK: remark: vectorization-remarks-profitable.c:11:14: the cost-model indicates that vectorization is not beneficial
-; CHECK: remark: vectorization-remarks-profitable.c:11:14: the cost-model indicates that interleaving is not beneficial
+; CHECK: remark: vectorization-remarks-profitable.c:5:17: the cost-model indicates that vectorization is not beneficial
+; CHECK: remark: vectorization-remarks-profitable.c:5:17: the cost-model indicates that interleaving is not beneficial and is explicitly disabled or interleave count is set to 1
+; CHECK: remark: vectorization-remarks-profitable.c:12:17: the cost-model indicates that vectorization is not beneficial
+; CHECK: remark: vectorization-remarks-profitable.c:12:17: the cost-model indicates that interleaving is not beneficial
 
 ; First loop.
 ;  #pragma clang loop interleave(disable) unroll(disable)
 ;  for(int i = 0; i < n; i++) {
-;    out[i] = in[i];
+;    out[i] = *in[i];
 ;  }
 
 ; Second loop.
 ;  #pragma clang loop unroll(disable)
 ;  for(int i = 0; i < n; i++) {
-;    out[i] = in[i];
+;    out[i] = *in[i];
 ;  }
 
+; ModuleID = 'vectorization-remarks-profitable.ll'
 target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-apple-macosx10.10.0"
 
-; Function Attrs: nounwind ssp uwtable
-define void @do_not_interleave(float* nocapture %out, float* nocapture readonly %in, i32 %n) #0 {
+; Function Attrs: nounwind uwtable
+define void @do_not_interleave(float** noalias nocapture readonly %in, float* noalias nocapture %out, i32 %size) #0 {
 entry:
-  %cmp.7 = icmp sgt i32 %n, 0, !dbg !3
-  br i1 %cmp.7, label %for.body.preheader, label %for.cond.cleanup, !dbg !8
+  %cmp.4 = icmp eq i32 %size, 0, !dbg !10
+  br i1 %cmp.4, label %for.end, label %for.body.preheader, !dbg !11
 
 for.body.preheader:                               ; preds = %entry
-  br label %for.body, !dbg !9
-
-for.cond.cleanup.loopexit:                        ; preds = %for.body
-  br label %for.cond.cleanup, !dbg !10
-
-for.cond.cleanup:                                 ; preds = %for.cond.cleanup.loopexit, %entry
-  ret void, !dbg !10
+  br label %for.body, !dbg !12
 
 for.body:                                         ; preds = %for.body.preheader, %for.body
   %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %for.body.preheader ]
-  %arrayidx = getelementptr inbounds float, float* %in, i64 %indvars.iv, !dbg !9
-  %0 = bitcast float* %arrayidx to i32*, !dbg !9
-  %1 = load i32, i32* %0, align 4, !dbg !9, !tbaa !11
-  %arrayidx2 = getelementptr inbounds float, float* %out, i64 %indvars.iv, !dbg !15
-  %2 = bitcast float* %arrayidx2 to i32*, !dbg !16
-  store i32 %1, i32* %2, align 4, !dbg !16, !tbaa !11
-  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1, !dbg !8
-  %lftr.wideiv = trunc i64 %indvars.iv.next to i32, !dbg !8
-  %exitcond = icmp eq i32 %lftr.wideiv, %n, !dbg !8
-  br i1 %exitcond, label %for.cond.cleanup.loopexit, label %for.body, !dbg !8, !llvm.loop !17
-}
+  %arrayidx = getelementptr inbounds float*, float** %in, i64 %indvars.iv, !dbg !12
+  %0 = bitcast float** %arrayidx to i32**, !dbg !12
+  %1 = load i32*, i32** %0, align 8, !dbg !12
+  %2 = load i32, i32* %1, align 4, !dbg !13
+  %arrayidx2 = getelementptr inbounds float, float* %out, i64 %indvars.iv, !dbg !14
+  %3 = bitcast float* %arrayidx2 to i32*, !dbg !15
+  store i32 %2, i32* %3, align 4, !dbg !15
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1, !dbg !11
+  %lftr.wideiv = trunc i64 %indvars.iv.next to i32, !dbg !11
+  %exitcond = icmp eq i32 %lftr.wideiv, %size, !dbg !11
+  br i1 %exitcond, label %for.end.loopexit, label %for.body, !dbg !11, !llvm.loop !16
 
-; Function Attrs: nounwind ssp uwtable
-define void @interleave_not_profitable(float* nocapture %out, float* nocapture readonly %in, i32 %n) #0 {
-entry:
-  %cmp.7 = icmp sgt i32 %n, 0, !dbg !20
-  br i1 %cmp.7, label %for.body.preheader, label %for.cond.cleanup, !dbg !22
+for.end.loopexit:                                 ; preds = %for.body
+  br label %for.end, !dbg !19
 
-for.body.preheader:                               ; preds = %entry
-  br label %for.body, !dbg !23
+for.end:                                          ; preds = %for.end.loopexit, %entry
+  ret void, !dbg !19
+}
 
-for.cond.cleanup.loopexit:                        ; preds = %for.body
-  br label %for.cond.cleanup, !dbg !24
+; Function Attrs: nounwind uwtable
+define void @interleave_not_profitable(float** noalias nocapture readonly %in, float* noalias nocapture %out, i32 %size) #0 {
+entry:
+  %cmp.4 = icmp eq i32 %size, 0, !dbg !20
+  br i1 %cmp.4, label %for.end, label %for.body, !dbg !21
 
-for.cond.cleanup:                                 ; preds = %for.cond.cleanup.loopexit, %entry
-  ret void, !dbg !24
+for.body:                                         ; preds = %entry, %for.body
+  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
+  %arrayidx = getelementptr inbounds float*, float** %in, i64 %indvars.iv, !dbg !22
+  %0 = bitcast float** %arrayidx to i32**, !dbg !22
+  %1 = load i32*, i32** %0, align 8, !dbg !22
+  %2 = load i32, i32* %1, align 4, !dbg !23
+  %arrayidx2 = getelementptr inbounds float, float* %out, i64 %indvars.iv, !dbg !24
+  %3 = bitcast float* %arrayidx2 to i32*, !dbg !25
+  store i32 %2, i32* %3, align 4, !dbg !25
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1, !dbg !21
+  %lftr.wideiv = trunc i64 %indvars.iv.next to i32, !dbg !21
+  %exitcond = icmp eq i32 %lftr.wideiv, %size, !dbg !21
+  br i1 %exitcond, label %for.end, label %for.body, !dbg !21, !llvm.loop !26
 
-for.body:                                         ; preds = %for.body.preheader, %for.body
-  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %for.body.preheader ]
-  %arrayidx = getelementptr inbounds float, float* %in, i64 %indvars.iv, !dbg !23
-  %0 = bitcast float* %arrayidx to i32*, !dbg !23
-  %1 = load i32, i32* %0, align 4, !dbg !23, !tbaa !11
-  %arrayidx2 = getelementptr inbounds float, float* %out, i64 %indvars.iv, !dbg !25
-  %2 = bitcast float* %arrayidx2 to i32*, !dbg !26
-  store i32 %1, i32* %2, align 4, !dbg !26, !tbaa !11
-  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1, !dbg !22
-  %lftr.wideiv = trunc i64 %indvars.iv.next to i32, !dbg !22
-  %exitcond = icmp eq i32 %lftr.wideiv, %n, !dbg !22
-  br i1 %exitcond, label %for.cond.cleanup.loopexit, label %for.body, !dbg !22, !llvm.loop !27
+for.end:                                          ; preds = %for.body, %entry
+  ret void, !dbg !27
 }
 
-attributes #0 = { nounwind }
+attributes #0 = { nounwind uwtable "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+mmx,+sse,+sse2" "unsafe-fp-math"="false" "use-soft-float"="false" }
 
-!llvm.module.flags = !{!0, !1}
-!llvm.ident = !{!2}
+!llvm.dbg.cu = !{!0}
+!llvm.module.flags = !{!7, !8}
+!llvm.ident = !{!9}
+
+!0 = distinct !DICompileUnit(language: DW_LANG_C99, file: !1, producer: "clang version 3.8.0 (trunk 250016)", isOptimized: false, runtimeVersion: 0, emissionKind: 2, enums: !2, subprograms: !3)
+!1 = !DIFile(filename: "vectorization-remarks-profitable.c", directory: "")
+!2 = !{}
+!3 = !{!4, !6}
+!4 = distinct !DISubprogram(name: "do_not_interleave", scope: !1, file: !1, line: 1, type: !5, isLocal: false, isDefinition: true, scopeLine: 1, flags: DIFlagPrototyped, isOptimized: false, function: void (float**, float*, i32)* @do_not_interleave, variables: !2)
+!5 = !DISubroutineType(types: !2)
+!6 = distinct !DISubprogram(name: "interleave_not_profitable", scope: !1, file: !1, line: 8, type: !5, isLocal: false, isDefinition: true, scopeLine: 8, flags: DIFlagPrototyped, isOptimized: false, function: void (float**, float*, i32)* @interleave_not_profitable, variables: !2)
+!7 = !{i32 2, !"Dwarf Version", i32 4}
+!8 = !{i32 2, !"Debug Info Version", i32 3}
+!9 = !{!"clang version 3.8.0 (trunk 250016)"}
+!10 = !DILocation(line: 4, column: 23, scope: !4)
+!11 = !DILocation(line: 4, column: 3, scope: !4)
+!12 = !DILocation(line: 5, column: 17, scope: !4)
+!13 = !DILocation(line: 5, column: 16, scope: !4)
+!14 = !DILocation(line: 5, column: 7, scope: !4)
+!15 = !DILocation(line: 5, column: 14, scope: !4)
+!16 = distinct !{!16, !17, !18}
+!17 = !{!"llvm.loop.interleave.count", i32 1}
+!18 = !{!"llvm.loop.unroll.disable"}
+!19 = !DILocation(line: 6, column: 1, scope: !4)
+!20 = !DILocation(line: 11, column: 23, scope: !6)
+!21 = !DILocation(line: 11, column: 3, scope: !6)
+!22 = !DILocation(line: 12, column: 17, scope: !6)
+!23 = !DILocation(line: 12, column: 16, scope: !6)
+!24 = !DILocation(line: 12, column: 7, scope: !6)
+!25 = !DILocation(line: 12, column: 14, scope: !6)
+!26 = distinct !{!26, !18}
+!27 = !DILocation(line: 13, column: 1, scope: !6)
 
-!0 = !{i32 2, !"Debug Info Version", i32 3}
-!1 = !{i32 1, !"PIC Level", i32 2}
-!2 = !{!"clang version 3.7.0"}
-!3 = !DILocation(line: 3, column: 20, scope: !4)
-!4 = distinct !DISubprogram(name: "do_not_interleave", scope: !5, file: !5, line: 1, type: !6, isLocal: false, isDefinition: true, scopeLine: 1, flags: DIFlagPrototyped, isOptimized: true, function: void (float*, float*, i32)* @do_not_interleave, variables: !7)
-!5 = !DIFile(filename: "vectorization-remarks-profitable.c", directory: "")
-!6 = !DISubroutineType(types: !7)
-!7 = !{}
-!8 = !DILocation(line: 3, column: 3, scope: !4)
-!9 = !DILocation(line: 4, column: 14, scope: !4)
-!10 = !DILocation(line: 6, column: 1, scope: !4)
-!11 = !{!12, !12, i64 0}
-!12 = !{!"float", !13, i64 0}
-!13 = !{!"omnipotent char", !14, i64 0}
-!14 = !{!"Simple C/C++ TBAA"}
-!15 = !DILocation(line: 4, column: 5, scope: !4)
-!16 = !DILocation(line: 4, column: 12, scope: !4)
-!17 = distinct !{!17, !18, !19}
-!18 = !{!"llvm.loop.interleave.count", i32 1}
-!19 = !{!"llvm.loop.unroll.disable"}
-!20 = !DILocation(line: 10, column: 20, scope: !21)
-!21 = distinct !DISubprogram(name: "interleave_not_profitable", scope: !5, file: !5, line: 8, type: !6, isLocal: false, isDefinition: true, scopeLine: 8, flags: DIFlagPrototyped, isOptimized: true, function: void (float*, float*, i32)* @interleave_not_profitable, variables: !7)
-!22 = !DILocation(line: 10, column: 3, scope: !21)
-!23 = !DILocation(line: 11, column: 14, scope: !21)
-!24 = !DILocation(line: 13, column: 1, scope: !21)
-!25 = !DILocation(line: 11, column: 5, scope: !21)
-!26 = !DILocation(line: 11, column: 12, scope: !21)
-!27 = distinct !{!27, !19}

../test/Transforms/LoopVectorize/gep_with_bitcast.ll

+; RUN: opt -S -loop-vectorize -instcombine -force-vector-width=4  < %s | FileCheck %s
+
+target datalayout = "e-m:e-i64:64-i128:128-n32:64-S128"
+
+; Vectorization of loop with bitcast between GEP and load
+; Simplified source code:
+;void foo (double** __restrict__  in, bool * __restrict__ res) {
+;
+;  for (int i = 0; i < 4096; ++i)
+;    res[i] = ((unsigned long long)in[i] == 0);
+;}
+
+; CHECK-LABEL: @foo
+; CHECK: vector.body
+; CHECK:  %0 = getelementptr inbounds double*, double** %in, i64 %index
+; CHECK:  %1 = bitcast double** %0 to <4 x i64>*
+; CHECK:  %wide.load = load <4 x i64>, <4 x i64>* %1, align 8
+; CHECK:  %2 = icmp eq <4 x i64> %wide.load, zeroinitializer
+; CHECK:  br i1
+
+define void @foo(double** noalias nocapture readonly %in, double** noalias nocapture readnone %out, i8* noalias nocapture %res) #0 {
+entry:
+  br label %for.body
+
+for.body:
+  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
+  %arrayidx = getelementptr inbounds double*, double** %in, i64 %indvars.iv
+  %tmp53 = bitcast double** %arrayidx to i64*
+  %tmp54 = load i64, i64* %tmp53, align 8
+  %cmp1 = icmp eq i64 %tmp54, 0
+  %arrayidx3 = getelementptr inbounds i8, i8* %res, i64 %indvars.iv
+  %frombool = zext i1 %cmp1 to i8
+  store i8 %frombool, i8* %arrayidx3, align 1
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+  %exitcond = icmp eq i64 %indvars.iv.next, 4096
+  br i1 %exitcond, label %for.end, label %for.body
+
+for.end:
+  ret void
+}
 No newline at end of file