https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107445
Bug ID: 107445
Summary: Redundant moves for subregs move
Product: gcc
Version: 13.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: juzhe.zhong at rivai dot ai
Target Milestone: ---
Here is the testcase:
https://godbolt.org/z/4bcfx5a86
coalesce:
ld4d {z24.d - z27.d}, p0/z, [x0]
mov z5.d, z24.d
mov z4.d, z25.d
mov z3.d, z26.d
mov z2.d, z27.d
cmp w1, 0
ble .L2
mov w0, 0
ld1d z1.d, p0/z, [x2]
ld1d z0.d, p0/z, [x3]
.L3:
add w0, w0, 1
movprfx z5.d, p0/z, z5.d
mad z5.d, p0/m, z1.d, z0.d
movprfx z4.d, p0/z, z4.d
mad z4.d, p0/m, z1.d, z0.d
movprfx z3.d, p0/z, z3.d
mad z3.d, p0/m, z1.d, z0.d
movprfx z2.d, p0/z, z2.d
mad z2.d, p0/m, z1.d, z0.d
cmp w1, w0
bne .L3
It's obvious that we should be able to directly use z24 - z27 to perform mad
oberation and remove the redundant moves in the prologue.
I tried to implement register coalescing in IRA to fix it, but I am not sure
whether it's a good idea to fix it?
