Commit Graph

2 Commits

Author SHA1 Message Date
W. J. van der Laan
8ca3005a91 Merge bitcoin/bitcoin#22645: scripts: prevent GCC optimising test symbols in test-symbol-check
5449d44e37982fcd5251fd47873c5f7d34c39fc9 scripts: prevent GCC optimising test symbols in test-symbol-check (fanquake)

Pull request description:

  I noticed in #22381 that when the test-symbol-check target was being built with Clang and run in the CI it would fail due to using a too-new version of `pow` (used [here](d67330d112/contrib/devtools/test-symbol-check.py (L85))). Our CIs use Focal (glibc 2.31) and the version of `pow` was the optimized version introduced in [glibc 2.29](https://lwn.net/Articles/778286/):
  ```bash
  * Optimized generic exp, exp2, log, log2, pow, sinf, cosf, sincosf and tanf.
  ```
  This made sense, except for that if it was failing when built using Clang, why hadn't it also been failing when being built with GCC?

  Turns out GCC is optimizing away that call to `pow` at all optimization levels, including `-O0`, see: https://godbolt.org/z/53MhzMxT7, and this has been the case forever, or at least since GCC 5.x. Clang on the other hand, will only optimize away the `pow` call at `-O1` and `-O2`, not `-O0`: https://godbolt.org/z/Wbnqj3q6c. Thus when this test was built with Clang (we don't pass `-O` so we default to `-O0`) it was failing in the CI environment, because it would actually have a call to the "new" `pow`.

  Avoid this issue by using a symbol that won't be optimized away, or that we are unlikely to ever have versioning issues with.

ACKs for top commit:
  laanwj:
    ACK 5449d44e37982fcd5251fd47873c5f7d34c39fc9

Tree-SHA512: 3a26c5c3a5f2905fd0dd90892470e241ba625c0af3be2629d06d5da3a97534c1d6a55b796bbdd41e2e6a26a8fab7d981b98c45d4238565b0eb7edf3c5da02007
2023-04-16 23:40:59 +03:00
Kittywhiskers Van Gogh
d37f30ce46 merge bitcoin#20476: Add test for ELF symbol-check 2023-01-18 19:02:39 -06:00