[all-commits] [llvm/llvm-project] 5a3a39: [LoopCacheAnalysis] Guard against bit-width overfl...

[Amy Kwan via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 5a3a399c2773bf4d86413e1855f6be2cf35da8af
      https://github.com/llvm/llvm-project/commit/5a3a399c2773bf4d86413e1855f6be2cf35da8af
  Author: Amy Kwan <amy.kwan1 at ibm.com>
  Date:   2026-10-03 (Sat, 03 Oct 2026)

  Changed paths:
    M llvm/lib/Analysis/LoopCacheAnalysis.cpp
    A llvm/test/Analysis/LoopCacheAnalysis/nestedloop-cost.ll

  Log Message:
  -----------
  [LoopCacheAnalysis] Guard against bit-width overflow in computeRefCost (#224427)

In `computeRefCost()`, the non-consecutive access path estimates the
cache cost by multiplying the trip counts across outer loop dimensions.
Each multiplication ends up calling `getExtendedType()` to obtain an
integer type twice as wide as the current one.

For a deeply-nested loop (such as the 20-dimensional array found in this
change), it is possible for the doubling to exceed `MAX_INT_BITS`, which
triggers the following assert in `IntegerType::get()`.
```
Assertion failed: NumBits <= MAX_INT_BITS && "bitwidth too large", file llvm-project/llvm/lib/IR/Type.cpp, line 340, static IntegerType *llvm::IntegerType::get(LLVMContext &, unsigned int)()
```

This change checks whether the current integer width already exceeds
`MAX_INT_BITS / 2`. If this is true, the doubled width would overflow,
so we return `CacheCostTy::getInvalid()` instead of asserting.



To unsubscribe from these emails, change your notification settings at https://github.com/llvm/llvm-project/settings/notifications