https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126747
--- Comment #4 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:eef42c68c7f66835747f82e4532a26d7a0f6bb98 commit r17-3556-geef42c68c7f66835747f82e4532a26d7a0f6bb98 Author: Kyrylo Tkachov <[email protected]> Date: Thu Jul 30 12:20:54 2026 +0200 ifcvt: Do not clobber a live condition-code register [PR126501, PR126747] If-conversion emits its replacement sequence at the end of the test block. Expanding a conditional move there can require a fresh comparison, which writes a condition-code register. end_ifcvt_sequence already rejects a sequence that would destroy the condition code tested by the branch, but cc_in_cond only reports that register when the branch reads it directly. A branch such as AArch64 CBZ tests a general register, so the guard was inert and an unrelated live condition code could be destroyed. For the testcase at -O2 on aarch64, late-combine sinks a cset into the join block, which leaves the flags live across the branch: bb2: cmp w3, 1 // sets cc cset w4, ls cbz w1, .L2 // does not touch cc bb5: cinc w2, w2, ls // reads cc ce2 then if-converted bb3 and bb4 and inserted "cmp w1, 0" ahead of the branch, so cinc read the wrong flags: cmp w3, 1 cset w4, ls cmp w1, 0 // clobbers the live flags csel w2, w2, w3, eq cinc w2, w2, ls // reads cmp w1, 0 rtl.h documents that ports in the "lowered" form, which includes aarch64 before register allocation, may keep the flags live between instructions, so if-conversion has to respect that. Reject a generated sequence that writes a condition-code register while it is live on exit from the test block. Recognise condition-code registers from the mode of each store destination and from the target's fixed flags and condition-code register hooks. The destination mode preserves condition-code semantics that reg_raw_mode can lose when a hard register also holds integers. The hooks cover condition state that is itself represented in an integer mode. DF liveness is already up to date here and the same paths query it for pseudos. noce_convert_multiple_sets validates its sequence itself rather than through end_ifcvt_sequence, and reaches noce_emit_cmove in the same way, so it gets the same check. PR126747 is the same defect reached from a different direction. At -Os the multiplication overflow idiom becomes one .MUL_OVERFLOW, so both arms read a single cset, and late-combine folds it into the second one: bb2: cmp xzr, x0, lsr 32 // sets cc cset w3, ne cbz w0, .L4 // does not touch cc bb5: cinc w0, w0, ne // reads cc ce2 if-converted bb3 the same way and "cmp w0, 0" landed ahead of the cinc, so foo (1, 1) returned 1 instead of 0. Bootstrapped and tested on aarch64-none-linux-gnu. gcc/ChangeLog: PR rtl-optimization/126501 PR rtl-optimization/126747 * ifcvt.cc (noce_cc_reg_set): New struct. (noce_record_cc_reg_set): New function. (noce_clobbers_live_cc_p): New function. (end_ifcvt_sequence): Use it to reject sequences that clobber a condition-code register that is live out of the test block. (noce_convert_multiple_sets): Likewise. gcc/testsuite/ChangeLog: PR rtl-optimization/126501 PR rtl-optimization/126747 * gcc.c-torture/execute/pr126501.c: New test. * gcc.c-torture/execute/pr126747.c: New test. Signed-off-by: Kyrylo Tkachov <[email protected]>
