[MLIR][Utils] Fix the overflow issue in computeSuffixProductImpl for 32-bit system.

[chencha3]

In int64_t r = strides.size() - 2, it may cause overflow on 32-bit system when strides.size() is 1, because strides.size() is a size_t which is a unsigned int (32-bit), leading a wrong computation of 1 - 2 = 4294967295 on unsigned int after casting to int64_t.

[llvmbot]

@llvm/pr-subscribers-mlir

Author: Chao Chen (chencha3)

Changes

In int64_t r = strides.size() - 2, it may cause overflow on 32-bit system when strides.size() is 1. be cause strides.size() is a size_t which is a unsigned int (32-bit), leading 1 - 2 = 4294967295 (on unsigned computing).

---

Full diff: https://github.com/llvm/llvm-project/pull/140567.diff

1 Files Affected:

• (modified) mlir/lib/Dialect/Utils/IndexingUtils.cpp (+1-1)

diff --git a/mlir/lib/Dialect/Utils/IndexingUtils.cpp b/mlir/lib/Dialect/Utils/IndexingUtils.cpp
index d9edabef6693d..8de77e2c3cb08 100644
--- a/mlir/lib/Dialect/Utils/IndexingUtils.cpp
+++ b/mlir/lib/Dialect/Utils/IndexingUtils.cpp
@@ -24,7 +24,7 @@ SmallVector<ExprType> computeSuffixProductImpl(ArrayRef<ExprType> sizes,
   if (sizes.empty())
     return {};
   SmallVector<ExprType> strides(sizes.size(), unit);
-  for (int64_t r = strides.size() - 2; r >= 0; --r)
+  for (int64_t r = static_cast<int64_t>(strides.size()) - 2; r >= 0; --r)
     strides[r] = strides[r + 1] * sizes[r + 1];
   return strides;
 }

[mgorny]

Thanks. I've started the test run.

[mgorny]

Thanks a lot! I can confirm that this fixes both the regression, and a bunch of other test failures we've seen earlier. Also, I'm sorry — this wasn't your fault after all.

[chencha3]

Thanks a lot! I can confirm that this fixes both the regression, and a bunch of other test failures we've seen earlier. Also, I'm sorry — this wasn't your fault after all.

No worries. Happy to contribute my effort.