https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88833
Bug ID: 88833
Summary: [SVE] Redundant moves for WHILELO-based loops
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: enhancement
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: rsandifo at gcc dot gnu.org
Target Milestone: ---
Compiling this function with -O3 -march=armv8-a+sve:
subroutine foo(x, y, z)
real :: x(100), y(100), z(100)
x = y + z(1)
end subroutine foo
gives:
foo_:
.LFB0:
.cfi_startproc
mov x4, 100
mov x5, x4 // Redundant
mov x3, 0
ptrue p1.s, all
whilelo p0.s, xzr, x4
ld1rw z1.s, p1/z, [x2]
.p2align 3,,7
.L2:
ld1w z0.s, p0/z, [x1, x3, lsl 2]
fadd z0.s, z0.s, z1.s
st1w z0.s, p0, [x0, x3, lsl 2]
incw x3
whilelo p0.s, x3, x5
bne .L2
ret
.cfi_endproc
There's no need for the move here. We should just be able to use x4 for both
WHILELOs.
Although the move itself shouldn't be expensive in context, it suggests that
the RA isn't seeing an accurate picture, which could hurt in more complex
cases.
