Hi Steve,
> This patch checks for SIMD functions and saves the extra registers when
> needed. It does not change the caller behavour, so with just this patch
> there may be values saved by both the caller and callee. This is not
> efficient, but it is correct code.
I tried a few simple test cases. It seems calls to non-vector functions don't
mark
the callee-saves as needing to be saved/restored:
void g(void);
void __attribute__ ((aarch64_vector_pcs))
f1 (void)
{
g();
g();
}
f1:
str x30, [sp, -16]!
bl g
ldr x30, [sp], 16
b g
Here I would expect q8-q23 to be preserved and no tailcall to g() since it is
not a vector
function. This is important for correctness since f1 must preserve q8-q23.
// compile with -O2 -ffixed-d1 -ffixed-d2 -ffixed-d3 -ffixed-d4 -ffixed-d5
-ffixed-d6 -ffixed-d7
float __attribute__ ((aarch64_vector_pcs))
f2 (float *p)
{
float t0 = p[1];
float t1 = p[3];
float t2 = p[5];
return t0 - t1 * (t1 + t0) + (t2 * t0);
}
f2:
stp d16, d17, [sp, -48]!
ldr s17, [x0, 4]
ldr s18, [x0, 12]
ldr s0, [x0, 20]
fadd s16, s17, s18
fmsub s16, s16, s18, s17
fmadd s0, s17, s0, s16
ldp d16, d17, [sp], 48
ret
This uses s16-s18 when it should prefer to use s24-s31 first. Also it needs to
save q16-q18,
not only d16 and d17.
Btw the -ffixed-d* is useful to block the register allocator from using certain
registers.
Wilco
