https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109940
--- Comment #5 from CVS Commits <cvs-commit at gcc dot gnu.org> --- The trunk branch has been updated by Richard Sandiford <rsand...@gcc.gnu.org>: https://gcc.gnu.org/g:ee2a8b373a88bae4c533aa68bed56bf01afea0e2 commit r14-1157-gee2a8b373a88bae4c533aa68bed56bf01afea0e2 Author: Richard Sandiford <richard.sandif...@arm.com> Date: Wed May 24 09:53:12 2023 +0100 early-remat: Resync with new DF postorders [PR109940] When I wrote early-remat, the DF_FORWARD block order was a postorder of a reverse/backward walk (i.e. of the inverted cfg), rather than a reverse postorder of a forward walk. A postorder of a backward walk lacked the important property that dominators come before the blocks they dominate; instead it ensures that postdominators come after the blocks that they postdominate. The DF_BACKWARD block order was similarly a postorder of a forward walk. Since early-remat wanted a standard postorder and reverse postorder with normal dominator properties, it used the DF_BACKWARD order instead of the DF_FORWARD order. g:53dddbfeb213ac4ec39f fixed the DF orders so that DF_FORWARD was an RPO of a forward walk and so that DF_BACKWARD was an RPO of a backward walk. This meant that iterating backwards over the DF_BACKWARD order had the exact problem that the original DF_FORWARD order had, triggering a flurry of ICEs for SVE. This fixes the build with SVE enabled. It also fixes an ICE in g++.target/aarch64/sve/pr99766.C with normal builds. I've included the test from the PR as well, for extra coverage. gcc/ PR rtl-optimization/109940 * early-remat.cc (postorder_index): Rename to... (rpo_index): ...this. (compare_candidates): Sort by decreasing rpo_index rather than increasing postorder_index. (early_remat::sort_candidates): Calculate the forward RPO from DF_FORWARD. (early_remat::local_phase): Follow forward RPO using DF_FORWARD, rather than DF_BACKWARD in reverse. gcc/testsuite/ * gcc.dg/torture/pr109940.c: New test.
