https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128284
--- Comment #2 from GCC Commits <cvs-commit at gcc dot gnu.org> --- The master branch has been updated by Kyrylo Tkachov <[email protected]>: https://gcc.gnu.org/g:bb273fcdc0945f3f07a03ba90a39443494563f06 commit r17-6340-gbb273fcdc0945f3f07a03ba90a39443494563f06 Author: Kyrylo Tkachov <[email protected]> Date: Wed Nov 11 05:12:38 2026 -0800 aarch64: Add memory clobber to GCS stack switch patterns [PR128284] The GCSSS1 and GCSSS2 instructions validate and consume the cap token at the new GCS address and write a return token to the old GCS. The aarch64_gcsss1 and aarch64_gcsss2 patterns are plain unspec_volatile with no memory effects, so RTL optimisers treat memory as unchanged across them. For: uint64_t f (uint64_t *p) { uint64_t a = *p; __gcsss (p); return a ^ *p; } cse1 reuses the load before the switch for the load after it and the function folds to returning 0. Add (clobber (mem:BLK (scratch))) to both patterns, as done for other instructions with arbitrary memory side effects. GCSPOPM only reads memory and changes no memory, so it is left unchanged. Before: sys #3, c7, c7, #2, x0 // gcsss1 mov w0, 0 sysl x0, #3, c7, c7, #3 // gcsss2 mov w0, 0 ret After: ldr x2, [x0] sys #3, c7, c7, #2, x0 // gcsss1 mov w1, 0 sysl x1, #3, c7, c7, #3 // gcsss2 ldr x0, [x0] eor x0, x2, x0 ret Bootstrapped and tested on aarch64-none-linux-gnu. gcc/ChangeLog: PR target/128284 * config/aarch64/aarch64.md (aarch64_gcsss1): Add a memory clobber. (aarch64_gcsss2): Likewise. gcc/testsuite/ChangeLog: PR target/128284 * gcc.target/aarch64/pr128284.c: New test. Signed-off-by: Kyrylo Tkachov <[email protected]>
