Difference between gcc and clang tsan on combined read/write

[Anders Waldenborg]

Hi,

Have noticed a difference between what clang and gcc emits when istrumenting something that is an inplace operation such as "*anintpointer |= 1".

gcc emits both __tsan_read4 and __tsan_write4  calls while clang only emits a call to __tsan_write4

full example: https://godbolt.org/z/TrM7YzMnd

Is this an intentional optimization that is done in the clang/llvm implementation (and thus a missed optimization in gcc)? Or is this an inaccuracy in the clang/llvm implementation?

(we have our own runtime that currently just borrows the tsan instrumentation to trace memory accesses for other reasons. This is why we have notice this)

[Dmitry Vyukov]

Hi Anders,

This is an intentional optimization in clang and a missed optimization in gcc.