https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99668
Bug ID: 99668
Summary: Converting argument _Complex double to double vector
causes STLF stall
Product: gcc
Version: 11.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: middle-end
Assignee: unassigned at gcc dot gnu.org
Reporter: rguenth at gcc dot gnu.org
Target Milestone: ---
Consider
_Complex double a, b;
typedef double v2df __attribute__((vector_size(16)));
v2df foo (_Complex double x)
{
return *(v2df *)&x;
}
which on GIMPLE is represented as the simple
v2df foo (complex double x)
{
v2df _2;
_2 = VIEW_CONVERT_EXPR<v2df>(x_3(D));
return _2;
}
but since on x86_64 _Complex double are passed in %xmm0 and %xmm1 we end up
with spilling x to the stack:
foo:
.LFB0:
.cfi_startproc
movsd %xmm0, -24(%rsp)
movsd %xmm1, -16(%rsp)
movupd -24(%rsp), %xmm0
ret
Ideally GIMPLE would know that a V_C_E isn't a good representation here
but improving RTL expansion should be possible as well. We expand
x_3(D) to (concat:DC ...) but the V_C_E expansion path misses special-casing
of this. The MEM_REF expansion code can deal with this so eventually
splitting out parts of that support is possible.
In the end we want a simple
movlhps %xmm1, %xmm0
