PR target/11877: Use xor to write zero to memory with -Os
The following patch attempts to resolve PR target/11877 (without
triggering PR/23102). On x86_64, writing an SImode or DImode zero
to memory uses an instruction encoding that is larger than first
clearing a register (using xor) then writing that to memory. Hence,
after reload, the peephole2 pass can determine if there's a suitable
free register, and if so, use that to shrink the code size with -Os.
To improve code size, and avoid inserting a large number of xor
instructions (PR target/23102), this patch makes use of peephole2's
efficient pattern matching to use a single temporary for a run of
consecutive writes. In theory, one could do better still with a
new target-specific pass, gated on -Os, to shrink these instructions
(like stv), but that's probably overkill for the little remaining
space savings.
Evaluating this patch on the CSiBE benchmark (v2.1.1) results in a
0.26% code size improvement (
3715273 bytes down to
3705477) on x86_64
with -Os [saving 1 byte every 400]. 549 of 894 tests improve, two
tests grow larger. Analysis of these 2 pathological cases reveals
that although peephole2's match_scratch prefers to use a call-clobbered
register (to avoid requiring a new stack frame), very rarely this
interacts with GCC's shrink wrapping optimization, which may previously
have avoided saving/restoring a call clobbered register, such as %eax,
in the calling function.
2021-06-21 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
PR target/11877
* config/i386/i386.md: New define_peephole2s to shrink writing
1, 2 or 4 consecutive zeros to memory when optimizing for size.
gcc/testsuite/ChangeLog
PR target/11877
* gcc.target/i386/pr11877.c: New test case.