[LoopIdiom] Let LIR fold memset pointer / stride SCEV regarding loop guards

[eopXD]

Summary

Expression guraded in loop entry can be folded prior to comparison. This patch
proceeds D107353 and makes LIR able to deal with nested for-loop.

[lebedev.ri]

I think this needs to be somewhere in SCEV itself.

[eopXD]

In D108112#2946374, @lebedev.ri wrote:

I think this needs to be somewhere in SCEV itself.

Do you mean the structure SCEVFolder?

[eopXD]

Fix clang-format warning.

[lebedev.ri]

In D108112#2946376, @eopXD wrote:

In D108112#2946374, @lebedev.ri wrote:

I think this needs to be somewhere in SCEV itself.

Do you mean the structure SCEVFolder?

No, i mean that the ScalarEvolution should perform this canonicalization internally, in ScalarEvolution::getSignExtendExpr()
It already does it, in fact: https://github.com/llvm/llvm-project/blob/0dc6b597db4d8b25f3df96a8f6574be732e18fdd/llvm/lib/Analysis/ScalarEvolution.cpp#L2100-L2103
Though perhaps that is not feasible because we don't know the loop we're after when in ScalarEvolution::getSignExtendExpr().

[eopXD]

No, i mean that the ScalarEvolution should perform this canonicalization internally, in ScalarEvolution::getSignExtendExpr()
It already does it, in fact: https://github.com/llvm/llvm-project/blob/0dc6b597db4d8b25f3df96a8f6574be732e18fdd/llvm/lib/Analysis/ScalarEvolution.cpp#L2100-L2103

I see. Thank you for the elaboration.

Though perhaps that is not feasible because we don't know the loop we're after when in ScalarEvolution::getSignExtendExpr().

I agree.

[eopXD]

Fix some minor type-o.

[bmahjour]

llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp

310

that is -> that are

319

greater of -> greater or

321–325

  if (SE.isLoopEntryGuardedByCond(CurLoop, ICmpInst::ICMP_SGE, Expr,
                                  SE.getZero(Expr->getType())))
    return SE.getZeroExtendExpr(visit(Expr->getOperand()), Expr->getType());
 return Expr;
}

988–989

fold expressions that is -> fold an expression that is

988–989

proceed optimization if equal. -> proceed with optimization, if equal.

993

with respect to -> based on

llvm/test/Transforms/LoopIdiom/memset-runtime.ll

116 ↗

(On Diff #366579)

Given that the stride is o x 4, it's not clear how the loop guard can help make assumptions about sign/zero extension. Can you explain this example in the description section?

[eopXD]

Address comments.
Thank you for pointing out the grammatical errors.

[eopXD]

Update test case CHECK based on change of debug log.

[qianzhen]

LGTM

[fhahn]

Could you clean up the basic block names & co in the test case?

[eopXD]

Update test case with better renaming and clear redundant blocks.

[eopXD]

Thank you @lebedev.ri and @bmahjour for reviewing the patch. I am now going to commit it.

[bmahjour]

I still have some reservations about the SCEVFolder implementation in this patch. It doesn't really fold anything, it just zero extends an expression if it is safe to do so. The assumption being made is that the expression it compares against is unsigned (so a "folding" opportunity is realized). In the LIT test provided the trip count is zero extended by loop simplify because it is originally 32-bit and the code is targeting 64-bit mode. But what if we are compiling in 32-bit mode, or if 'n' and 'm' where 'signed long long'?

[eopXD]

In D108112#2957417, @bmahjour wrote:

But what if we are compiling in 32-bit mode,

This is a valid point. Thank you for pointing this out.
Created D108507 to add test case that is aware of 32-bit mode.

or if 'n' and 'm' where 'signed long long'?

The accepted scenario for LIR is MemsetSizeSCEV == PositivePointerStrideSCEV.
The n, m accepted would always be non-negative. I can't think of scenario when n, m needs to remain as a signed long long. (Please correct me if I'm wrong.

I still have some reservations about the SCEVFolder implementation in this patch. It doesn't really fold anything, it just zero extends an expression if it is safe to do so. The assumption being made is that the expression it compares against is unsigned (so a "folding" opportunity is realized). In the LIT test provided the trip count is zero extended by loop simplify because it is originally 32-bit and the code is targeting 64-bit mode.

I agree with you that SCEVFolder has trivial functionality. It only to turn sext to zext. This patch only fixes the test case added, which is very limited.
On the other hand, I don't think I have deep understanding to benchmarks. So I am not sure if I have covered most of the "practical memset idioms" that can be optimized here. Do you have other "fold loop-guard into SCEV-expression` cases that you wish this patch to have? I would be happy to resolve it.

[eopXD]

gentle ping @bmahjour

[bmahjour]

In D108112#2958737, @eopXD wrote:

In D108112#2957417, @bmahjour wrote:

or if 'n' and 'm' where 'signed long long'?

The accepted scenario for LIR is MemsetSizeSCEV == PositivePointerStrideSCEV.
The n, m accepted would always be non-negative. I can't think of scenario when n, m needs to remain as a signed long long. (Please correct me if I'm wrong.

Since we don't normalize loops we may have reverse loops where the upper bound is a negative signed value. Another example is if IVs start with negative values:

void foo(int n, int m, float Arr[n][m]) {
  for (int i = -100; i < n; i++)
    for (int j = -200; j < m; j++)
      Arr[i+100][j+200] = 1.0;
}

I still have some reservations about the SCEVFolder implementation in this patch. It doesn't really fold anything, it just zero extends an expression if it is safe to do so. The assumption being made is that the expression it compares against is unsigned (so a "folding" opportunity is realized). In the LIT test provided the trip count is zero extended by loop simplify because it is originally 32-bit and the code is targeting 64-bit mode.

I agree with you that SCEVFolder has trivial functionality. It only to turn sext to zext. This patch only fixes the test case added, which is very limited.

The term "fold" doesn't seem to be the right terminology. I think "rewrite" is more appropriate. For example, SCEVFolder may be called SCEVSignToZeroExtentionRewriter (or something similar).

On the other hand, I don't think I have deep understanding to benchmarks. So I am not sure if I have covered most of the "practical memset idioms" that can be optimized here. Do you have other "fold loop-guard into SCEV-expression` cases that you wish this patch to have? I would be happy to resolve it.

I don't have any specific benchmarks in mind.

[eopXD]

Rebase to latest main.

[eopXD]

Update testcase due to rebase.

[eopXD]

Rename SCEVFolder to SCEVSignToZeroExtentionRewriter.
Add example with negative start.

[eopXD]

Hi @bmahjour,

Sorry for such delay.

I think the example you raised don't cause a problem to the current implementation.
The comparison is between "pointer stride" and "memset size", which all need to be non-negative.
The start/end of the induction variable don't matter because if the origin stride is negative it will be turned positive before comparison.

I have addressed your comment on renaming the SCEV folding structure and added test case for negative start loops like your example.
(Your example is loop strided store, I modified it to memset in the examples)

[bmahjour]

llvm/test/Transforms/LoopIdiom/memset-runtime.ll

110–140 ↗

(On Diff #367706)

Why is this test removed?

[eopXD]

llvm/test/Transforms/LoopIdiom/memset-runtime.ll

110–140 ↗

(On Diff #367706)

The test file is split into 2 files, memset-runtime-32bit.ll and memset-runtime-64bit.ll to test on 32-bit mode and 64-bit mode in D108507. So memset-runtime.ll is gone.

[bmahjour]

LGTM

llvm/test/Transforms/LoopIdiom/memset-runtime.ll

110–140 ↗

(On Diff #367706)

If you could put the comment, describing what these tests are meant for (ie lines 120-125), at the beginning of memset-runtime-32bit.ll and/or memset-runtime-64bit.ll it would be helpful.

[eopXD]

Rebase.

[eopXD]

Last rebase was with the wrong version, rebase again.

[eopXD]

Address @bmahjour 's comment, add comment to test case this patch improves.

llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp

public:
   void getAnalysisUsage(AnalysisUsage &AU) const override {
     AU.addRequired<TargetLibraryInfoWrapperPass>();
     AU.addRequired<TargetTransformInfoWrapperPass>();
     AU.addPreserved<MemorySSAWrapperPass>();
     getLoopAnalysisUsage(AU);
   }
 };
 
+// The Folder will fold expressions that are guarded by the loop entry.
+class SCEVSignToZeroExtentionRewriter
+    : public SCEVRewriteVisitor<SCEVSignToZeroExtentionRewriter> {
+public:
+  ScalarEvolution &SE;
+  const Loop *CurLoop;
+  SCEVSignToZeroExtentionRewriter(ScalarEvolution &SE, const Loop *CurLoop)
+      : SCEVRewriteVisitor(SE), SE(SE), CurLoop(CurLoop) {}
+
+  const SCEV *visitSignExtendExpr(const SCEVSignExtendExpr *Expr) {
+    // If expression is guarded by CurLoop to be greater or equal to zero
+    // then convert sext to zext. Otherwise return the original expression.
+    if (SE.isLoopEntryGuardedByCond(CurLoop, ICmpInst::ICMP_SGE, Expr,
+                                    SE.getZero(Expr->getType())))
+      return SE.getZeroExtendExpr(visit(Expr->getOperand()), Expr->getType());
+    return Expr;
+  }
+};
+
 } // end anonymous namespace
 
 char LoopIdiomRecognizeLegacyPass::ID = 0;
 
 PreservedAnalyses LoopIdiomRecognizePass::run(Loop &L, LoopAnalysisManager &AM,
                                               LoopStandardAnalysisResults &AR,
                                               LPMUpdater &) {
   if (DisableLIRP::All)
if (IsConstantSize) {
     IsNegStride = PointerStrideSCEV->isNonConstantNegative();
     const SCEV *PositiveStrideSCEV =
         IsNegStride ? SE->getNegativeSCEV(PointerStrideSCEV)
                     : PointerStrideSCEV;
     LLVM_DEBUG(dbgs() << "  MemsetSizeSCEV: " << *MemsetSizeSCEV << "\n"
                       << "  PositiveStrideSCEV: " << *PositiveStrideSCEV
                       << "\n");
 
     if (PositiveStrideSCEV != MemsetSizeSCEV) {
-      // TODO: folding can be done to the SCEVs
-      // The folding is to fold expressions that is covered by the loop guard
-      // at loop entry. After the folding, compare again and proceed
-      // optimization if equal.
+      // The folding is to fold an expression that is covered by the loop guard
+      // at loop entry. After the folding, compare again and proceed with
+      // optimization, if equal.
+      SCEVSignToZeroExtentionRewriter Folder(*SE, CurLoop);
+      const SCEV *FoldedPositiveStride = Folder.visit(PositiveStrideSCEV);
+      const SCEV *FoldedMemsetSize = Folder.visit(MemsetSizeSCEV);
+
+      LLVM_DEBUG(dbgs() << "  Try to fold SCEV based on loop guard\n"
+                        << "    FoldedMemsetSize: " << *FoldedMemsetSize << "\n"
+                        << "    FoldedPositiveStride: " << *FoldedPositiveStride
+                        << "\n");
+
+      if (FoldedPositiveStride != FoldedMemsetSize) {
         LLVM_DEBUG(dbgs() << "  SCEV don't match, abort\n");
         return false;
       }
     }
+  }
 
   // Verify that the memset value is loop invariant.  If not, we can't promote
   // the memset.
   Value *SplatValue = MSI->getValue();
   if (!SplatValue || !CurLoop->isLoopInvariant(SplatValue))
     return false;
 
   SmallPtrSet<Instruction *, 1> MSIs;

llvm/test/Transforms/LoopIdiom/memset-runtime-32bit.ll

 
 for.end17.loopexit:                               ; preds = %for.cond2.for.end_crit_edge.us
   br label %for.end17
 
 for.end17:                                        ; preds = %for.end17.loopexit, %for.body.lr.ph, %entry
   ret void
 }
 
+; void NegStart(int n, int m, int *ar) {
+;   for (int i = -100; i < n; i++) {
+;     int *arr = ar + (i + 100) * m;
+;     memset(arr, 0, m * sizeof(int));
+;   }
+; }
+define void @NegStart(i32 %n, i32 %m, i32* %ar) {
+; CHECK-LABEL: @NegStart(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[AR1:%.*]] = bitcast i32* [[AR:%.*]] to i8*
+; CHECK-NEXT:    [[CMP1:%.*]] = icmp slt i32 -100, [[N:%.*]]
+; CHECK-NEXT:    br i1 [[CMP1]], label [[FOR_BODY_LR_PH:%.*]], label [[FOR_END:%.*]]
+; CHECK:       for.body.lr.ph:
+; CHECK-NEXT:    [[MUL1:%.*]] = mul i32 [[M:%.*]], 4
+; CHECK-NEXT:    [[TMP0:%.*]] = add i32 [[N]], 100
+; CHECK-NEXT:    [[TMP1:%.*]] = mul i32 [[M]], [[TMP0]]
+; CHECK-NEXT:    [[TMP2:%.*]] = shl i32 [[TMP1]], 2
+; CHECK-NEXT:    call void @llvm.memset.p0i8.i32(i8* align 4 [[AR1]], i8 0, i32 [[TMP2]], i1 false)
+; CHECK-NEXT:    br label [[FOR_END]]
+; CHECK:       for.end:
+; CHECK-NEXT:    ret void
+;
+entry:
+  %cmp1 = icmp slt i32 -100, %n
+  br i1 %cmp1, label %for.body.lr.ph, label %for.end
+
+for.body.lr.ph:                                   ; preds = %entry
+  %mul1 = mul i32 %m, 4
+  br label %for.body
+
+for.body:                                         ; preds = %for.body.lr.ph, %for.body
+  %i.02 = phi i32 [ -100, %for.body.lr.ph ], [ %inc, %for.body ]
+  %add = add nsw i32 %i.02, 100
+  %mul = mul nsw i32 %add, %m
+  %add.ptr = getelementptr inbounds i32, i32* %ar, i32 %mul
+  %0 = bitcast i32* %add.ptr to i8*
+  call void @llvm.memset.p0i8.i32(i8* align 4 %0, i8 0, i32 %mul1, i1 false)
+  %inc = add nsw i32 %i.02, 1
+  %exitcond = icmp ne i32 %inc, %n
+  br i1 %exitcond, label %for.body, label %for.end.loopexit
+
+for.end.loopexit:                                 ; preds = %for.body
+  br label %for.end
+
+for.end:                                          ; preds = %for.end.loopexit, %entry
+  ret void
+}
+
 declare void @llvm.memset.p0i8.i32(i8* nocapture writeonly, i8, i32, i1 immarg)

llvm/test/Transforms/LoopIdiom/memset-runtime-64bit.ll

for.body:                                         ; preds = %for.body.lr.ph, %for.body
   call void @llvm.memset.p0i8.i64(i8* align 4 %0, i8 0, i64 %mul3, i1 false)
   %dec = add nsw i64 %i.02, -1
   %cmp = icmp sge i64 %dec, 0
   br i1 %cmp, label %for.body, label %for.end
 
 for.end:                                          ; preds = %for.body, %entry
   ret void
 }
+; This case requires SCEVFolder in LoopIdiomRecognize.cpp to fold SCEV prior to comparison.
+; For the inner-loop, SCEVFolder is not needed, however the promoted memset size would be based
+; on the trip count of inner-loop (which is an unsigned integer).
+; Then in the outer loop, the pointer stride SCEV for memset needs to be converted based on the
+; loop guard for it to equal to the memset size SCEV. The loop guard guaranteeds that m >= 0
+; inside the loop, so m can be converted from sext to zext, making the two SCEV-s equal.
 ; void NestedFor32(int *ar, int n, int m, int o)
 ; {
 ;   int i, j;
 ;   for (i=0; i<n; ++i) {
 ;     for (j=0; j<m; j++) {
 ;       int *arr = ar + i * m * o + j * o;
 ;       memset(arr, 0, o * sizeof(int));
 ;     }
 ;   }
 ; }
 define void @NestedFor32(i32* %ar, i32 %n, i32 %m, i32 %o) {
 ; CHECK-LABEL: @NestedFor32(
 ; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[AR2:%.*]] = bitcast i32* [[AR:%.*]] to i8*
 ; CHECK-NEXT:    [[CMP3:%.*]] = icmp slt i32 0, [[N:%.*]]
 ; CHECK-NEXT:    br i1 [[CMP3]], label [[FOR_BODY_LR_PH:%.*]], label [[FOR_END11:%.*]]
 ; CHECK:       for.body.lr.ph:
 ; CHECK-NEXT:    [[CMP21:%.*]] = icmp slt i32 0, [[M:%.*]]
 ; CHECK-NEXT:    [[CONV:%.*]] = sext i32 [[O:%.*]] to i64
 ; CHECK-NEXT:    [[MUL8:%.*]] = mul i64 [[CONV]], 4
 ; CHECK-NEXT:    br i1 [[CMP21]], label [[FOR_BODY_US_PREHEADER:%.*]], label [[FOR_END11]]
 ; CHECK:       for.body.us.preheader:
 ; CHECK-NEXT:    [[TMP0:%.*]] = sext i32 [[O]] to i64
 ; CHECK-NEXT:    [[TMP1:%.*]] = sext i32 [[M]] to i64
 ; CHECK-NEXT:    [[WIDE_TRIP_COUNT10:%.*]] = zext i32 [[N]] to i64
 ; CHECK-NEXT:    [[TMP2:%.*]] = mul i64 [[TMP0]], [[TMP1]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = zext i32 [[M]] to i64
 ; CHECK-NEXT:    [[TMP4:%.*]] = mul i64 [[TMP0]], [[TMP3]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = shl i64 [[TMP4]], 2
-; CHECK-NEXT:    br label [[FOR_BODY_US:%.*]]
-; CHECK:       for.body.us:
-; CHECK-NEXT:    [[INDVARS_IV6:%.*]] = phi i64 [ 0, [[FOR_BODY_US_PREHEADER]] ], [ [[INDVARS_IV_NEXT7:%.*]], [[FOR_BODY_US]] ]
-; CHECK-NEXT:    [[TMP6:%.*]] = mul i64 [[TMP2]], [[INDVARS_IV6]]
-; CHECK-NEXT:    [[SCEVGEP:%.*]] = getelementptr i32, i32* [[AR:%.*]], i64 [[TMP6]]
-; CHECK-NEXT:    [[SCEVGEP1:%.*]] = bitcast i32* [[SCEVGEP]] to i8*
-; CHECK-NEXT:    [[WIDE_TRIP_COUNT:%.*]] = zext i32 [[M]] to i64
-; CHECK-NEXT:    call void @llvm.memset.p0i8.i64(i8* align 4 [[SCEVGEP1]], i8 0, i64 [[TMP5]], i1 false)
-; CHECK-NEXT:    [[INDVARS_IV_NEXT7]] = add nuw nsw i64 [[INDVARS_IV6]], 1
-; CHECK-NEXT:    [[EXITCOND11:%.*]] = icmp ne i64 [[INDVARS_IV_NEXT7]], [[WIDE_TRIP_COUNT10]]
-; CHECK-NEXT:    br i1 [[EXITCOND11]], label [[FOR_BODY_US]], label [[FOR_END11]]
+; CHECK-NEXT:    [[TMP6:%.*]] = mul i64 [[TMP4]], [[WIDE_TRIP_COUNT10]]
+; CHECK-NEXT:    [[TMP7:%.*]] = shl i64 [[TMP6]], 2
+; CHECK-NEXT:    call void @llvm.memset.p0i8.i64(i8* align 4 [[AR2]], i8 0, i64 [[TMP7]], i1 false)
+; CHECK-NEXT:    br label [[FOR_END11]]
 ; CHECK:       for.end11:
 ; CHECK-NEXT:    ret void
 ;
 entry:
   %cmp3 = icmp slt i32 0, %n
   br i1 %cmp3, label %for.body.lr.ph, label %for.end11
 
 for.body.lr.ph:                                   ; preds = %entry

 
 for.end11.loopexit:                               ; preds = %for.cond1.for.end_crit_edge.us
   br label %for.end11
 
 for.end11:                                        ; preds = %for.end11.loopexit, %for.body.lr.ph, %entry
   ret void
 }
 
+; void NegStart(int n, int m, int *ar) {
+;   for (int i = -100; i < n; i++) {
+;     int *arr = ar + (i + 100) * m;
+;     memset(arr, 0, m * sizeof(int));
+;   }
+; }
+define void @NegStart(i32 %n, i32 %m, i32* %ar) {
+; CHECK-LABEL: @NegStart(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[AR1:%.*]] = bitcast i32* [[AR:%.*]] to i8*
+; CHECK-NEXT:    [[CMP1:%.*]] = icmp slt i32 -100, [[N:%.*]]
+; CHECK-NEXT:    br i1 [[CMP1]], label [[FOR_BODY_LR_PH:%.*]], label [[FOR_END:%.*]]
+; CHECK:       for.body.lr.ph:
+; CHECK-NEXT:    [[CONV:%.*]] = sext i32 [[M:%.*]] to i64
+; CHECK-NEXT:    [[MUL1:%.*]] = mul i64 [[CONV]], 4
+; CHECK-NEXT:    [[TMP0:%.*]] = sext i32 [[M]] to i64
+; CHECK-NEXT:    [[WIDE_TRIP_COUNT:%.*]] = sext i32 [[N]] to i64
+; CHECK-NEXT:    [[TMP1:%.*]] = add nsw i64 [[WIDE_TRIP_COUNT]], 100
+; CHECK-NEXT:    [[TMP2:%.*]] = mul i64 [[TMP1]], [[TMP0]]
+; CHECK-NEXT:    [[TMP3:%.*]] = shl i64 [[TMP2]], 2
+; CHECK-NEXT:    call void @llvm.memset.p0i8.i64(i8* align 4 [[AR1]], i8 0, i64 [[TMP3]], i1 false)
+; CHECK-NEXT:    br label [[FOR_END]]
+; CHECK:       for.end:
+; CHECK-NEXT:    ret void
+;
+entry:
+  %cmp1 = icmp slt i32 -100, %n
+  br i1 %cmp1, label %for.body.lr.ph, label %for.end
+
+for.body.lr.ph:                                   ; preds = %entry
+  %conv = sext i32 %m to i64
+  %mul1 = mul i64 %conv, 4
+  %0 = sext i32 %m to i64
+  %wide.trip.count = sext i32 %n to i64
+  br label %for.body
+
+for.body:                                         ; preds = %for.body.lr.ph, %for.body
+  %indvars.iv = phi i64 [ -100, %for.body.lr.ph ], [ %indvars.iv.next, %for.body ]
+  %1 = add nsw i64 %indvars.iv, 100
+  %2 = mul nsw i64 %1, %0
+  %add.ptr = getelementptr inbounds i32, i32* %ar, i64 %2
+  %3 = bitcast i32* %add.ptr to i8*
+  call void @llvm.memset.p0i8.i64(i8* align 4 %3, i8 0, i64 %mul1, i1 false)
+  %indvars.iv.next = add nsw i64 %indvars.iv, 1
+  %exitcond = icmp ne i64 %indvars.iv.next, %wide.trip.count
+  br i1 %exitcond, label %for.body, label %for.end.loopexit
+
+for.end.loopexit:                                 ; preds = %for.body
+  br label %for.end
+
+for.end:                                          ; preds = %for.end.loopexit, %entry
+  ret void
+}
+
 declare void @llvm.memset.p0i8.i64(i8* nocapture writeonly, i8, i64, i1 immarg)

llvm/test/Transforms/LoopIdiom/memset-runtime-debug.ll

 ; Check on debug outputs...
 ; CHECK: loop-idiom Scanning: F[MemsetSize_LoopVariant] Countable Loop %for.body
 ; CHECK-NEXT: memset size is non-constant
 ; CHECK-NEXT: memset size is not a loop-invariant, abort
 ; CHECK: loop-idiom Scanning: F[MemsetSize_Stride_Mismatch] Countable Loop %for.body
 ; CHECK-NEXT: memset size is non-constant
 ; CHECK-NEXT: MemsetSizeSCEV: (4 * (sext i32 %m to i64))<nsw>
 ; CHECK-NEXT: PositiveStrideSCEV: (4 + (4 * (sext i32 %m to i64))<nsw>)<nsw>
+; CHECK-NEXT: Try to fold SCEV based on loop guard
+; CHECK-NEXT: FoldedMemsetSize: (4 * (sext i32 %m to i64))<nsw>
+; CHECK-NEXT: FoldedPositiveStride: (4 + (4 * (sext i32 %m to i64))<nsw>)<nsw>
 ; CHECK-NEXT: SCEV don't match, abort
 ; CHECK: loop-idiom Scanning: F[NonZeroAddressSpace] Countable Loop %for.cond1.preheader
 ; CHECK-NEXT: memset size is non-constant
 ; CHECK-NEXT: pointer is not in address space zero, abort
 ; CHECK: loop-idiom Scanning: F[NonAffinePointer] Countable Loop %for.body
 ; CHECK-NEXT: Pointer is not affine, abort
 
 define void @MemsetSize_LoopVariant(i32* %ar, i32 %n, i32 %m) {