The following testcase is miscompiled, because the AVX2 patterns don't
describe correctly what the insn does. E.g. vphaddd with %ymm* operands
(the second pattern) instruction as per:
https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm256_hadd_epi32&expand=2941
does { a0+a1, a2+a3, b0+b1, b2+b3, a4+a5, a6+a7, b4+b5, b6+b7 }
but our RTL pattern did
{ a0+a1, a2+a3, a4+a5, a6+a7, b0+b1, b2+b3, b4+b5, b6+b7 }
where the first and last 64 bits are the same and two middle 64 bits
swapped.
https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm256_hadd_epi16&expand=2939
similarly, insn does:
{ a0+a1, a2+a3, a4+a5, a6+a7, b0+b1, b2+b3, b4+b5, b6+b7,
a8+a9, a10+a11, a12+a13, a14+a15, b8+b9, b10+b11, b12+b13, b14+b15 }
but RTL pattern did
{ a0+a1, a2+a3, a4+a5, a6+a7, a8+a9, a10+a11, a12+a13, a14+a15,
b0+b1, b2+b3, b4+b5, b6+b7, b8+b9, b10+b11, b12+b13, b14+b15 }
again, first and last 64 bits are the same and the two middle 64 bits
swapped.
2020-04-03 Jakub Jelinek <jakub@redhat.com>
PR target/94460
* config/i386/sse.md (avx2_ph<plusminus_mnemonic>wv16hi3,
avx2_ph<plusminus_mnemonic>dv8si3): Fix up RTL pattern to do
second half of first lane from first lane of second operand and
first half of second lane from second lane of first operand.
* gcc.target/i386/avx2-pr94460.c: New test.
