https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116136
--- Comment #3 from GCC Commits <cvs-commit at gcc dot gnu.org> --- The master branch has been updated by Jeff Law <l...@gcc.gnu.org>: https://gcc.gnu.org/g:89ed5ab210b5e30c325b92e9e40c50e337be1b44 commit r15-2452-g89ed5ab210b5e30c325b92e9e40c50e337be1b44 Author: Jeff Law <j...@ventanamicro.com> Date: Wed Jul 31 10:15:01 2024 -0600 [PR rtl-optimization/116136] Fix previously latent SUBREG simplification bug This fixes a testsuite regression seen on m68k after some of the recent ext-dce changes. Ultimately Richard S and I have concluded the bug was a latent issue in subreg simplification. Essentially when simplifying something like (set (target:M1) (subreg:M1 (subreg:M2 (reg:M1) 0) 0)) Where M1 > M2. We'd simplify to: (set (target:M1) (reg:M1)) The problem is on a big endian target that's wrong. Consider if M1 is DI and M2 is SI. The original should extract bits 32..63 from the source register and store them into bits 0..31 of the target register. In the simplified form it's just a copy, so bits 0..63 of the source end up bits 0..63 of the target. This shows up as the following regressions on the m68k: > Tests that now fail, but worked before (3 tests): > > gcc: gcc.c-torture/execute/960416-1.c -O2 execution test > gcc: gcc.c-torture/execute/960416-1.c -O2 -flto -fno-use-linker-plugin -flto-partition=none execution test > gcc: gcc.c-torture/execute/960416-1.c -Os execution test The fix is pretty trivial, instead of hardcoding "0" as the byte offset in the test for the simplification, instead we need to use the subreg_lowpart_offset. Anyway, bootstrapped and regression tested on m68k and x86_64 and tested on the other embedded targets as well without regressions. Naturally it fixes the regression noted above. I haven't see other testsuite improvements when I spot checked some of the big endian crosses. PR rtl-optimization/116136 gcc/ * simplify-rtx.cc (simplify_context::simplify_subreg): Check that we're working with the lowpart offset rather than byte 0.
