https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113779
--- Comment #10 from Oleg Endo <olegendo at gcc dot gnu.org> ---
I've been trying to resurrect some previous work for addressing mode selection
optimization (RTL pass).
This is what current GCC17 produces on SH for the example in the original
description:
_f:
cmp/pl r6
bf .L11
.align 2
.L9:
mov.l @r4,r1
dt r6
mov.l r1,@r5
mov.l @(4,r4),r1
mov.l r1,@(4,r5)
mov.l @(8,r4),r1
mov.l r1,@(8,r5)
mov.l @(12,r4),r1
mov.l r1,@(12,r5)
mov.l @(16,r4),r1
mov.l r1,@(16,r5)
mov.l @(20,r4),r1
mov.l r1,@(20,r5)
mov.l @(24,r4),r1
mov.l r1,@(24,r5)
mov.l @(28,r4),r1
mov.l r1,@(28,r5)
mov.l @(32,r4),r1
mov.l r1,@(32,r5)
mov.l @(36,r4),r1
mov.l r1,@(36,r5)
mov.l @(40,r4),r1
mov.l r1,@(40,r5)
mov.l @(44,r4),r1
mov.l r1,@(44,r5)
mov.l @(48,r4),r1
mov.l r1,@(48,r5)
mov.l @(52,r4),r1
mov.l r1,@(52,r5)
mov.l @(56,r4),r1
mov.l r1,@(56,r5)
mov.l @(60,r4),r1
add #64,r4
mov.l r1,@(60,r5)
bf/s .L9
add #64,r5
.L11:
rts
nop
With the additional optimization pass we get proper post-inc utilization
On SH4 we usually don't have post-inc stores, only post-inc loads.
_f:
cmp/pl r6
bf .L11
.align 2
.L9:
mov.l @r4+,r1
dt r6
mov.l r1,@r5
mov.l @r4+,r1
mov.l r1,@(4,r5)
mov.l @r4+,r1
mov.l r1,@(8,r5)
mov.l @r4+,r1
mov.l r1,@(12,r5)
mov.l @r4+,r1
mov.l r1,@(16,r5)
mov.l @r4+,r1
mov.l r1,@(20,r5)
mov.l @r4+,r1
mov.l r1,@(24,r5)
mov.l @r4+,r1
mov.l r1,@(28,r5)
mov.l @r4+,r1
mov.l r1,@(32,r5)
mov.l @r4+,r1
mov.l r1,@(36,r5)
mov.l @r4+,r1
mov.l r1,@(40,r5)
mov.l @r4+,r1
mov.l r1,@(44,r5)
mov.l @r4+,r1
mov.l r1,@(48,r5)
mov.l @r4+,r1
mov.l r1,@(52,r5)
mov.l @r4+,r1
mov.l r1,@(56,r5)
mov.l @r4+,r1
mov.l r1,@(60,r5)
bf/s .L9
add #64,r5
.L11:
rts
nop
On SH2A there's also a post-inc store insn, so that produces:
_f:
cmp/pl r6
bf .L10
.align 2
.L8:
mov.l @r4+,r0
dt r6
mov.l r0,@r5+
mov.l @r4+,r0
mov.l r0,@r5+
mov.l @r4+,r0
mov.l r0,@r5+
mov.l @r4+,r0
mov.l r0,@r5+
mov.l @r4+,r0
mov.l r0,@r5+
mov.l @r4+,r0
mov.l r0,@r5+
mov.l @r4+,r0
mov.l r0,@r5+
mov.l @r4+,r0
mov.l r0,@r5+
mov.l @r4+,r0
mov.l r0,@r5+
mov.l @r4+,r0
mov.l r0,@r5+
mov.l @r4+,r0
mov.l r0,@r5+
mov.l @r4+,r0
mov.l r0,@r5+
mov.l @r4+,r0
mov.l r0,@r5+
mov.l @r4+,r0
mov.l r0,@r5+
mov.l @r4+,r0
mov.l r0,@r5+
mov.l @r4+,r0
bf.s .L8
mov.l r0,@r5+
.L10:
rts/n
and the case from comment #8 is fixed, too:
_test_8:
cmp/pl r6
bf .L16
.align 2
.L14:
mov.b @r5+,r1
dt r6
mov.b r1,@r4
bf/s .L14
add #1,r4
.L16:
rts
nop
However, this is still very early and very WIP-ish. But early results look
promising, I think.
