[ScalarEvolutionExpander] Don't drop nowrap flags on addrec expansion

[wsmoses]

When performing expand on a SCEVAddRecNode with nowrap applied (e.g. if deduced from an iv with an increment with nowrap), the generated multiply will not have the nowrap. This PR preserves the nowrap behavior on the generated code.

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: William Moses (wsmoses)

Changes

When performing expand on a SCEVAddRecNode with nowrap applied (e.g. if deduced from an iv with an increment with nowrap), the generated multiply will not have the nowrap. This PR preserves the nowrap behavior on the generated code.

---

Full diff: https://github.com/llvm/llvm-project/pull/78199.diff

1 Files Affected:

• (modified) llvm/lib/Transforms/Utils/ScalarEvolutionExpander.cpp (+2-1)

diff --git a/llvm/lib/Transforms/Utils/ScalarEvolutionExpander.cpp b/llvm/lib/Transforms/Utils/ScalarEvolutionExpander.cpp
index a1d7f0f9ba0f74..e87b430d6bee8f 100644
--- a/llvm/lib/Transforms/Utils/ScalarEvolutionExpander.cpp
+++ b/llvm/lib/Transforms/Utils/ScalarEvolutionExpander.cpp
@@ -1254,7 +1254,8 @@ Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
       expand(SE.getTruncateOrNoop(
         SE.getMulExpr(SE.getUnknown(CanonicalIV),
                       SE.getNoopOrAnyExtend(S->getOperand(1),
-                                            CanonicalIV->getType())),
+                                            CanonicalIV->getType()),
+                                        S->getNoWrapFlags(SCEV::FlagNW)),
         Ty));
 
   // If this is a chain of recurrences, turn it into a closed form, using the

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[nikic]

Tests?

[nikic]

This doesn't make sense. NW is only meaningful/supported on addrecs, while you're applying it to a mul.

[wsmoses]

Tests?

Yeah I wasn't quite sure where to add them, if you had any pointers

[nikic]

I'm not sure where exactly we expand addrecs in canonical mode, so I'd suggest adding an assert into the branch and seeing which tests fail, then modify those.

[nikic]

I'm not sure where exactly we expand addrecs in canonical mode, so I'd suggest adding an assert into the branch and seeing which tests fail, then modify those.

Ah well, it looks like you effectively already did that by trying to use an invalid nowrap flag. List of failing tests:

  LLVM :: CodeGen/AMDGPU/loop-prefetch-data.ll
  LLVM :: CodeGen/PowerPC/hardware-loops-crash.ll
  LLVM :: CodeGen/PowerPC/loop-data-prefetch-inner.ll
  LLVM :: CodeGen/PowerPC/loop-data-prefetch.ll
  LLVM :: CodeGen/PowerPC/loop-hoist-toc-save.ll
  LLVM :: CodeGen/PowerPC/ppc64-get-cache-line-size.ll
  LLVM :: CodeGen/SystemZ/prefetch-02.ll
  LLVM :: CodeGen/SystemZ/prefetch-03.ll
  LLVM :: CodeGen/SystemZ/prefetch-04.ll
  LLVM :: CodeGen/Thumb2/LowOverheadLoops/tail-pred-reduce.ll
  LLVM :: CodeGen/Thumb2/LowOverheadLoops/varying-outer-2d-reduction.ll
  LLVM :: Transforms/IRCE/pr57335.ll
  LLVM :: Transforms/LICM/AliasSetMemSet.ll
  LLVM :: Transforms/LoopDataPrefetch/AArch64/kryo-large-stride.ll
  LLVM :: Transforms/LoopDataPrefetch/AArch64/large-stride.ll
  LLVM :: Transforms/LoopDataPrefetch/AArch64/opt-remark-with-hotness.ll
  LLVM :: Transforms/LoopDataPrefetch/AArch64/opt-remark.ll
  LLVM :: Transforms/LoopDataPrefetch/PowerPC/basic.ll
  LLVM :: Transforms/LoopDataPrefetch/RISCV/basic.ll
  LLVM :: Transforms/LoopIdiom/basic-address-space.ll
  LLVM :: Transforms/LoopIdiom/basic.ll
  LLVM :: Transforms/LoopIdiom/lir-heurs-multi-block-loop.ll
  LLVM :: Transforms/LoopVectorize/multiple-strides-vectorization.ll
  LLVM :: Transforms/LoopVectorize/pr55100-expand-scev-predicate-used.ll
  LLVM :: Transforms/LoopVectorize/runtime-checks-difference.ll
  LLVM :: Transforms/LoopVersioningLICM/loopversioningLICM1.ll
  LLVM :: Transforms/LoopVersioningLICM/loopversioningLICM2.ll
  LLVM :: Transforms/LoopVersioningLICM/metadata.ll

Probably LoopIdiom is the best candidate for a test case.