On Thu, 7 Jun 2018, Craig Topper via cfe-commits wrote:
Author: ctopper
Date: Thu Jun 7 10:28:03 2018
New Revision: 334208
URL: http://llvm.org/viewvc/llvm-project?rev=334208&view=rev
Log:
[X86] Add back builtins for _mm_slli_si128/_mm_srli_si128 and similar
intrinsics.
We still lower them to native shuffle IR, but we do it in CGBuiltin.cpp now.
This allows us to check the target feature and ensure the immediate fits in 8
bits.
FWIW, this change broke building libaom:
https://bugs.chromium.org/p/aomedia/issues/detail?id=1945
In libaom, there's a macro construct like this:
#define v256_shr_n_byte(a, n) \
((n) < 16 \
? _mm256_alignr_epi8( \
_mm256_permute2x128_si256(a, a, _MM_SHUFFLE(2, 0, 0, 1)), a, n) \
: ((n) > 16 \
? _mm256_srli_si256( \
_mm256_permute2x128_si256(a, a, _MM_SHUFFLE(2, 0, 0, 1)), \
(n)-16) \
: _mm256_permute2x128_si256(a, a, _MM_SHUFFLE(2, 0, 0, 1))))
Since this commit, the compilation errors out due to the _mm256_srli_si256
with invalid range, even though the toplevel ternary operator won't
actually pick them to be used. Not sure if there's anything to do from the
clang point of view here, I guess it's a tradeoff between having stricter
parameter checks for the intrinsics, vs the convenience of piling them up
in a macro like this in libaom.
// Martin
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
