[PATCH] aarch64: Improve size optimisation heuristic for setmem expansion

[Kyrylo Tkachov via Gcc-patches]

Hi all,

This patch adjusts the setmem expansion in the backend to track the number of 
ops it generates
for the DUP + STR/STP inline sequences. This way we can compare the 
size/complexity of the sequence
against alternatives, notably just returning "false" and thus just emitting a 
call to memset.

The simple heuristic change here is that if we were going to emit more than 4 
operations then
we shouldn't bother and just call memset. The number 4 is chosen because in the 
worst case for memset
we need to emit 4 instructions: 3 to move the arguments into the right 
registers and 1 for the call.

The speed optimisation decisions are not affected, though I do want to extend 
these expansions in a later
patch and I'd like to reuse this ops counting logic there. In any case this 
patch should make sense on its own.

For the code:
void __attribute__((__noinline__))
set127byte (int64_t *src, int c)
{
  __builtin_memset (src, c, 127);
}

void __attribute__((__noinline__))
set128byte (int64_t *src, int c)
{
  __builtin_memset (src, c, 128);
}

when optimising for size we now get just an immediate move + a call to memset 
(2 instructions) where before we'd have generated:
set127byte(long*, int):
        dup     v0.16b, w1
        str     q0, [x0, 96]
        stp     q0, q0, [x0]
        stp     q0, q0, [x0, 32]
        stp     q0, q0, [x0, 64]
        str     q0, [x0, 111]
        ret
set128byte(long*, int):
        dup     v0.16b, w1
        stp     q0, q0, [x0]
        stp     q0, q0, [x0, 32]
        stp     q0, q0, [x0, 64]
        stp     q0, q0, [x0, 96]
        ret

which is clearly undesirable for -Os.

I've adjusted the recently-added gcc.target/aarch64/memset-strict-align-1.c 
testcase to use a bigger struct
and switch to speed optimisation as with this patch we'll just call memset 
rather than expanding inline.
That is the right decision for size optimisation (the resulting code is indeed 
shorter).
With -O2 and the new struct size we still try the SIMD expansion and still 
trigger the path that the testcase is supposed to exercise.

Bootstrapped and tested on aarch64-none-linux-gnu.
Pushing to trunk.

2021-09-29  Kyrylo Tkachov  <kyrylo.tkac...@arm.com>

        * config/aarch64/aarch64.c (aarch64_expand_setmem): Count number of
        emitted operations and adjust heuristic for code size.
    
2021-09-29  Kyrylo Tkachov  <kyrylo.tkac...@arm.com>

        * gcc.target/aarch64/memset-corner-cases-2.c: New test.
        * gcc.target/aarch64/memset-strict-align-1.c: Adjust.

setmem-s.patch
Description: setmem-s.patch