https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109766
Roger Sayle <roger at nextmovesoftware dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
Last reconfirmed| |2023-05-08
Ever confirmed|0 |1
Status|UNCONFIRMED |NEW
--- Comment #2 from Roger Sayle <roger at nextmovesoftware dot com> ---
I believe the problem is in the cprop_hardreg pass, which undoes reload's
register assignments (to use DImode GPR registers with -Os), by propagating DF
mode values into *pushdi2_rex64, which then get split during the split3 pass
into lea/movq pairs, that are each larger than a DImode push. The work around,
for this test case, is to use -Os -fno-cprop-registers which produces code
that's shorter than -O2.
0000000000000000 <callfunc>:
0: 66 48 0f 7e ca movq %xmm1,%rdx
5: 66 48 0f 7e d1 movq %xmm2,%rcx
a: 66 48 0f 7e de movq %xmm3,%rsi
f: 50 push %rax
10: 66 49 0f 7e e0 movq %xmm4,%r8
15: 66 48 0f 7e c0 movq %xmm0,%rax
1a: 66 49 0f 7e e9 movq %xmm5,%r9
1f: 66 49 0f 7e f2 movq %xmm6,%r10
24: 66 49 0f 7e fb movq %xmm7,%r11
29: 41 53 push %r11
2b: 41 52 push %r10
2d: 41 51 push %r9
2f: 41 50 push %r8
31: 56 push %rsi
32: 51 push %rcx
33: 52 push %rdx
34: 50 push %rax
35: b0 08 mov $0x8,%al
37: e8 00 00 00 00 callq 3c <callfunc+0x3c>
3c: 48 83 c4 48 add $0x48,%rsp
40: c3 retq
Now to figure out if there's a way, using target rtx_costs or pushdi2_rex64's
constraints/predicates, to prevent hardreg cprop performing this substitution.
Plan B might be to investigate reload's choice of DFmode SSE vs DImode GPR, but
this is within one or two bytes of optimal (for four arguments I believe GCC
would produce shorter code than clang).
