PR tree-optimization/98335: Improvements to DSE's compute_trims.
This patch is the main middle-end piece of a fix for PR tree-opt/98335,
which is a code-quality regression affecting mainline. The issue occurs
in DSE's (dead store elimination's) compute_trims function that determines
where a store to memory can be trimmed. In the testcase given in the
PR, this function notices that the first byte of a DImode store is dead,
and replaces the 8-byte store at (aligned) offset zero, with a 7-byte store
at (unaligned) offset one. Most architectures can store a power-of-two
bytes (up to a maximum) in single instruction, so writing 7 bytes requires
more instructions than writing 8 bytes. This patch follows Jakub Jelinek's
suggestion in comment 5, that compute_trims needs improved heuristics.
On x86_64-pc-linux-gnu with -O2 the new test case in the PR goes from:
movl $0, -24(%rsp)
movabsq $
72057594037927935, %rdx
movl $0, -21(%rsp)
andq -24(%rsp), %rdx
movq %rdx, %rax
salq $8, %rax
movb c(%rip), %al
ret
to
xorl %eax, %eax
movb c(%rip), %al
ret
2022-03-11 Roger Sayle <roger@nextmovesoftware.com>
Richard Biener <rguenther@suse.de>
gcc/ChangeLog
PR tree-optimization/98335
* builtins.cc (get_object_alignment_2): Export.
* builtins.h (get_object_alignment_2): Likewise.
* tree-ssa-alias.cc (ao_ref_alignment): New.
* tree-ssa-alias.h (ao_ref_alignment): Declare.
* tree-ssa-dse.cc (compute_trims): Improve logic deciding whether
to align head/tail, writing more bytes but using fewer store insns.
gcc/testsuite/ChangeLog
PR tree-optimization/98335
* g++.dg/pr98335.C: New test case.
* gcc.dg/pr86010.c: New test case.
* gcc.dg/pr86010-2.c: New test case.