https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116179
--- Comment #13 from GCC Commits <cvs-commit at gcc dot gnu.org> --- The trunk branch has been updated by Andrew Pinski <pins...@gcc.gnu.org>: https://gcc.gnu.org/g:01cca42f65f25d709264fe277d25d9db3d68c437 commit r15-2712-g01cca42f65f25d709264fe277d25d9db3d68c437 Author: Andrew Pinski <quic_apin...@quicinc.com> Date: Fri Aug 2 10:04:40 2024 -0700 IRA: Ignore debug insns for uses in split_live_ranges_for_shrink_wrap. [PR116179] Late_combine exposed this latent bug in split_live_ranges_for_shrink_wrap. What it did was copy-prop regno 151 from regno 119 from: ``` (insn 2 264 3 2 (set (reg/f:DI 119 [ thisD.3697 ]) (reg:DI 151)) "/app/example.cpp":19:13 70 {*movdi_aarch64} (expr_list:REG_DEAD (reg:DI 151) (nil))) ``` into these insns: ``` (debug_insn 147 146 148 5 (var_location:DI thisD.3727 (reg/f:DI 119 [ thisD.3697 ])) "/app/example.cpp":21:5 -1 (nil)) .... (insn 167 166 168 7 (set (reg:DI 1 x1) (reg/f:DI 119 [ thisD.3697 ])) "/app/example.cpp":14:21 70 {*movdi_aarch64} (nil)) ``` Both are valid things to do. The problem is split_live_ranges_for_shrink_wrap looks at the uses of reg 151 and with and without debugging reg 151 have a different usage in different BBs. The function is trying to find a splitting point for reg 151 and they are different. In the end this causes register allocation difference. The fix is for split_live_ranges_for_shrink_wrap to ignore uses that were in debug insns. Bootstrappped and tested on x86_64-linux-gnu with no regressions. PR rtl-optimization/116179 gcc/ChangeLog: * ira.cc (split_live_ranges_for_shrink_wrap): For the uses loop, only look at non-debug insns. gcc/testsuite/ChangeLog: * g++.dg/torture/pr116179-1.C: New test. Signed-off-by: Andrew Pinski <quic_apin...@quicinc.com>
