https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109944
Bug ID: 109944
Summary: vector CTOR with byte elements and SSE2 has STLF fail
Product: gcc
Version: 13.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: rguenth at gcc dot gnu.org
Target Milestone: ---
I've experimented with CTORs from smaller elements and byte handling with
plain SSE2 is quite bad (for word we have pinsrw).
void foo(char *a, char *m, char d, char e)
{
char b = *m;
char c = m[2];
a[0] = b;
a[1] = c;
a[2] = d;
a[3] = e;
a[4] = b;
a[5] = c;
a[6] = b;
a[7] = c;
a[8] = b;
a[9] = c;
a[10] = b;
a[11] = c;
a[12] = b;
a[13] = c;
a[14] = b;
a[15] = c;
}
generates
movzbl 2(%rsi), %r8d
movl %edx, %r9d
movzbl (%rsi), %edx
movzbl %cl, %ecx
movzbl %r9b, %r9d
movq %r8, %rax
salq $8, %rax
orq %rdx, %rax
salq $8, %rax
orq %r8, %rax
salq $8, %rax
orq %rdx, %rax
salq $8, %rax
orq %rax, %rcx
orq %r8, %rax
salq $8, %rcx
salq $8, %rax
orq %r9, %rcx
orq %rdx, %rax
salq $8, %rcx
salq $8, %rax
orq %r8, %rcx
orq %r8, %rax
salq $8, %rcx
salq $8, %rax
orq %rdx, %rcx
orq %rdx, %rax
movq %rcx, -24(%rsp)
movq %rax, -16(%rsp)
movdqa -24(%rsp), %xmm0
movups %xmm0, (%rdi)
while we can handle a splat from QImode via
movzbl (%rsi), %eax
movd %eax, %xmm0
punpcklbw %xmm0, %xmm0
punpcklwd %xmm0, %xmm0
pshufd $0, %xmm0, %xmm0
movups %xmm0, (%rdi)
I think we can go and for a generic V16QImode CTOR and SSE2 create two
V8HImode vectors using pinsrw, for the first from zero-extended QImode
values of the even elements and for the second from zero-extended and
left-shifted values of the odd elements and then IOR the two vectors.
Alternatively the above needs to be pessimized better in the cost model.
Btw, for HImode elements I see we do
movzwl (%rsi), %eax
movd %eax, %xmm0
movdqa %xmm0, %xmm1
movdqa %xmm0, %xmm2
pinsrw $1, 4(%rsi), %xmm1
...
not sure why we don't do
pxor %xmm1, %xmm1
pinsrw $0, (%rsi), %xmm1
and thus avoid the round-trip through the GPR for the initial element?