A GCC 15 regression turned out to be a bug in Newlib related to undefined behavior that just started to trigger in some cases.
As it is now fixed, it makes IMHO sense to mention that Newlib commit in GCC's install documentation for AMD GPUs. Comments, suggestions to the attached patch? Tobias PS: Current HTML version is athttps://gcc.gnu.org/install/specific.html#amdgcn-x-amdhsa
install.texi: gcn - suggest to use Newlib with simd math fix [PR119325] Suggest a Newlib with a fix for the SIMD math issue. Newlib commit: https://sourceware.org/git/?p=newlib-cygwin.git;a=commitdiff;h=2ef1a37e7 Additionally, for generic support in ROCm, it is expected that 6.4 will added the support; the current version is 6.3.3 and it does not support it; bump >6.3.2 to >6.3.3 in install.texi to avoid doubts. gcc/ChangeLog: PR middle-end/119325 * doc/install.texi (gcn): Change ROCm > 6.3.2 to >6.3.3 for generic support; mention Newlib commit that fixes a SIMD math issue. diff --git a/gcc/doc/install.texi b/gcc/doc/install.texi index 308529669b1..c1280fd9787 100644 --- a/gcc/doc/install.texi +++ b/gcc/doc/install.texi @@ -4056,11 +4056,12 @@ ISA targets @code{gfx9-generic}, @code{gfx10-3-generic}, and @code{gfx11-generic} reduce the number of required multilibs but note that @code{gfx9-generic} does not include @code{gfx908} or @code{gfx90a}, that linking specific ISA code with generic code is currently not supported, -and that only a future ROCm release (newer than 6.3.2) will be able to execute +and that only a future ROCm release (newer than 6.3.3) will be able to execute generic code. Use Newlib (4.3.0 or newer; 4.4.0 contains some improvements and 4.5.0 fixes -the device console output for GFX10 and GFX11 devices). +the device console output for GFX10 and GFX11 devices; post-4.5.0 +commit 2ef1a37e7 [Mar 25, 2025] fixes a SIMD math issue). To run the binaries, install the HSA Runtime from the @uref{https://rocm.docs.amd.com/,,ROCm Platform}, and use
