https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127053
--- Comment #3 from GCC Commits <cvs-commit at gcc dot gnu.org> --- The trunk branch has been updated by Thomas Schwinge <[email protected]>: https://gcc.gnu.org/g:bfdd3d3d8b6877435898c89393177bbcbd713309 commit r17-4667-gbfdd3d3d8b6877435898c89393177bbcbd713309 Author: Thomas Schwinge <[email protected]> Date: Thu Sep 24 12:14:16 2026 +0200 cselim: cond_if_else_store_replacement_limited support skipping on store one either side: Adjust 'libgomp.graphite/force-parallel-8.c' [PR127053] With commit r17-4469-ge35b8dad030bc4164fdaa81dc0a5bfca8c1bbfc3 "cselim: cond_if_else_store_replacement_limited support skipping on store one either side [PR127053]", we regressed: PASS: libgomp.graphite/force-parallel-8.c (test for excess errors) PASS: libgomp.graphite/force-parallel-8.c execution test [-PASS:-]{+FAIL:+} libgomp.graphite/force-parallel-8.c scan-tree-dump-times graphite "5 loops carried no dependency" 1 PASS: libgomp.graphite/force-parallel-8.c scan-tree-dump-times optimized "loopfn.0" 4 PASS: libgomp.graphite/force-parallel-8.c scan-tree-dump-times optimized "loopfn.1" 4 libgomp.graphite/force-parallel-8.c: pattern found 0 times Indeed, instead of "5" it is now "3 loops carried no dependency". Per 'diff' of '-fdump-tree-all-details-folding' for commit r17-4469-ge35b8dad030bc4164fdaa81dc0a5bfca8c1bbfc3 "cselim: cond_if_else_store_replacement_limited support skipping on store one either side [PR127053]" reverted vs. current sources, I see in 'force-parallel-8.c.050t.phiopt1': [...] ;; Function foo (foo, funcdef_no=0, decl_uid=2216, cgraph_uid=3, symbol_order=2) +factoring out stores: + then: +# .MEM_30 = VDEF <.MEM_19> +x[i_12][j_14] = _3; + else: +# .MEM_29 = VDEF <.MEM_19> +x[i_12][j_14] = _7; + +to use phi: +cstore_33 = PHI <_3(11), _7(12)> + +# .MEM_18 = VDEF <.MEM_27> +x[i_12][j_14] = cstore_33; [...] ..., and subsequent simplification (as far as I can quickly tell) of the IR. Restore the previous state via '-fno-expensive-optimizations', to continue testing what was tested before. PR tree-optimization/127053 libgomp/ * testsuite/libgomp.graphite/force-parallel-8.c: 'dg-additional-options -fno-expensive-optimizations'.
