The shrinkwrap optimization added late in GCC 7 allows each callee-save to
be delayed and done only across blocks which need a particular callee-save.
Although this reduces unnecessary memory traffic on code paths that need
few callee-saves, it typically uses LDR/STR rather than LDP/STP. The
number of LDP/STP instructions is reduced by ~7%. This means more memory
accesses and increased codesize, ~1.0% on average.
To improve this, if a particular callee-save must be saved/restored, also
add the adjacent callee-save to allow use of LDP/STP. This significantly
reduces codesize (for example gcc_r, povray_r, parest_r, xalancbmk_r are
1% smaller). This is a simple fix which can be backported. A more advanced
approach would scan blocks for pairs of callee-saves, but that requires a
rewrite of all the callee-save code which is too late at this stage.
An example epilog in a shrinkwrapped function before:
ldp x21, x22, [sp,#16]
ldr x23, [sp,#32]
ldr x24, [sp,#40]
ldp x25, x26, [sp,#48]
ldr x27, [sp,#64]
ldr x28, [sp,#72]
ldr x30, [sp,#80]
ldr d8, [sp,#88]
ldp x19, x20, [sp],#96
ret
And after this patch:
ldr d8, [sp,#88]
ldp x21, x22, [sp,#16]
ldp x23, x24, [sp,#32]
ldp x25, x26, [sp,#48]
ldp x27, x28, [sp,#64]
ldr x30, [sp,#80]
ldp x19, x20, [sp],#96
ret
Passes bootstrap, OK for commit (and backport to GCC7)?
ChangeLog:
2018-01-05 Wilco Dijkstra <wdijk...@arm.com>
* config/aarch64/aarch64.c (aarch64_components_for_bb):
Increase LDP/STP opportunities by adding adjacent callee-saves.
--
diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c
index
9735fc18402dd8fe2fa4022eef4c0522814a0552..da21032b19413d0361b8d30b51a31124eaaa31a1
100644
--- a/gcc/config/aarch64/aarch64.c
+++ b/gcc/config/aarch64/aarch64.c
@@ -3503,7 +3503,22 @@ aarch64_components_for_bb (basic_block bb)
&& (bitmap_bit_p (in, regno)
|| bitmap_bit_p (gen, regno)
|| bitmap_bit_p (kill, regno)))
- bitmap_set_bit (components, regno);
+ {
+ unsigned regno2, offset, offset2;
+ bitmap_set_bit (components, regno);
+
+ /* If there is a callee-save at an adjacent offset, add it too
+ to increase the use of LDP/STP. */
+ offset = cfun->machine->frame.reg_offset[regno];
+ regno2 = ((offset & 8) == 0) ? regno + 1 : regno - 1;
+
+ if (regno2 <= LAST_SAVED_REGNUM)
+ {
+ offset2 = cfun->machine->frame.reg_offset[regno2];
+ if ((offset & ~8) == (offset2 & ~8))
+ bitmap_set_bit (components, regno2);
+ }
+ }
return components;
}
