This patch (for when the compiler returns to stage 1) improves support
for vector equality and inequality of V1TImode vectors, and V2DImode
vectors with sse2 but not sse4. Consider the three functions below:
typedef unsigned int uv4si __attribute__ ((__vector_size__ (16)));
typedef unsigned long long uv2di __attribute__ ((__vector_size__ (16)));
typedef unsigned __int128 uv1ti __attribute__ ((__vector_size__ (16)));
uv4si eq_v4si(uv4si x, uv4si y) { return x == y; }
uv2di eq_v2di(uv2di x, uv2di y) { return x == y; }
uv1ti eq_v1ti(uv1ti x, uv1ti y) { return x == y; }
These all perform vector comparisons of 128bit SSE2 registers, generating
the result as a vector, where ~0 (all 1 bits) represents true and a zero
represents false. eq_v4si is trivially implemented by x86_64's pcmpeqd
instruction. This patch improves the other two cases:
For v2di, gcc -O2 currently generates:
movq %xmm0, %rdx
movq %xmm1, %rax
movdqa %xmm0, %xmm2
cmpq %rax, %rdx
movhlps %xmm2, %xmm3
movhlps %xmm1, %xmm4
sete %al
movq %xmm3, %rdx
movzbl %al, %eax
negq %rax
movq %rax, %xmm0
movq %xmm4, %rax
cmpq %rax, %rdx
sete %al
movzbl %al, %eax
negq %rax
movq %rax, %xmm5
punpcklqdq %xmm5, %xmm0
ret
but with this patch we now generate:
pcmpeqd %xmm0, %xmm1
pshufd $177, %xmm1, %xmm0
pand %xmm1, %xmm0
ret
where the results of a V4SI comparison are shuffled and bit-wise ANDed
to produce the desired result. There's no change in the code generated
for "-O2 -msse4" where the compiler generates a single "pcmpeqq" insn.
For V1TI mode, the results are equally dramatic, where the current -O2
output looks like:
movaps %xmm0, -40(%rsp)
movq -40(%rsp), %rax
movq -32(%rsp), %rdx
movaps %xmm1, -24(%rsp)
movq -24(%rsp), %rcx
movq -16(%rsp), %rsi
xorq %rcx, %rax
xorq %rsi, %rdx
orq %rdx, %rax
sete %al
xorl %edx, %edx
movzbl %al, %eax
negq %rax
adcq $0, %rdx
movq %rax, %xmm2
negq %rdx
movq %rdx, -40(%rsp)
movhps -40(%rsp), %xmm2
movdqa %xmm2, %xmm0
ret
with this patch we now generate:
pcmpeqd %xmm0, %xmm1
pshufd $177, %xmm1, %xmm0
pand %xmm1, %xmm0
pshufd $78, %xmm0, %xmm1
pand %xmm1, %xmm0
ret
performing a V2DI comparison, followed by a shuffle and pand, and with
-O2 -msse4 take advantages of SSE4.1's pcmpeqq:
pcmpeqq %xmm0, %xmm1
pshufd $78, %xmm1, %xmm0
pand %xmm1, %xmm0
ret
This patch has been tested on x86_64-pc-linux-gnu with make bootstrap
and make -k check, both with and without --target_board=unix{-m32}, with
no new failures. Is this OK for when we return to stage 1?
2022-04-20 Roger Sayle <ro...@nextmovesoftware.com>
gcc/ChangeLog
* config/i386/sse.md (vec_cmpeqv2div2di): Enable for TARGET_SSE2.
For !TARGET_SSE4_1, expand as a V4SI vector comparison, followed
by a pshufd and pand.
(vec_cmpeqv1tiv1ti): New define_expand implementing V1TImode
vector equality as a V2DImode vector comparison (see above),
followed by a pshufd and pand.
gcc/testsuite/ChangeLog
* gcc.target/i386/sse2-v1ti-veq.c: New test case.
* gcc.target/i386/sse2-v1ti-vne.c: New test case.
Roger
--
