The following testcase shows two separate issues caused by the cselib
changes.
One is that through the cselib sp tracking improvements on
... r12 = rsp; rsp -= 8; push cst1; push cst2; push cst3; call
rsp += 32; rsp -= 8; push cst4; push cst5; push cst6; call
rsp += 32; rsp -= 8; push cst7; push cst8; push cst9; call
rsp += 32
reload_cse_simplify_set decides to optimize the rsp += 32 insns
into rsp = r12 because cselib figures that the r12 register holds the right
value. From the pure cost perspective that seems like a win and on its own
at least for -Os that would be beneficial, except that there are those
rsp -= 8 stack adjustments after it, where rsp += 32; rsp -= 8; is optimized
into rsp += 24; by the csa pass, but rsp = r12; rsp -= 8 can't. Dunno
what to do about this part, the PR has a hack in a comment.
Anyway, the following patch fixes the other part, which isn't a missed
optimization, but a wrong-code issue. The problem is that the pushes of
constant are on x86 represented through PRE_MODIFY and while
move2add_note_store has some code to handle {PRE,POST}_{INC,DEC} without
REG_INC note, it doesn't handle {PRE,POST}_MODIFY (that would be enough
to fix this testcase). But additionally it looks misplaced, because
move2add_note_store is only called on the rtxes that are stored into,
while RTX_AUTOINC can happen not just in those, but anywhere else in the
instruction (e.g. pop insn can have autoinc in the SET_SRC MEM).
REG_INC note seems to be required for any autoinc except for stack pointer
autoinc which doesn't have those notes, so this patch just handles
the sp autoinc after the REG_INC note handling loop.
2020-04-08 Jakub Jelinek <jakub@redhat.com>
PR rtl-optimization/94516
* postreload.c: Include rtl-iter.h.
(reload_cse_move2add): Handle SP autoinc here by FOR_EACH_SUBRTX_VAR
looking for all MEMs with RTX_AUTOINC operand.
(move2add_note_store): Remove {PRE,POST}_{INC,DEC} handling.
* gcc.dg/torture/pr94516.c: New test.
