https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91560
Bug ID: 91560
Summary: Try harder for AVX non-AVX2 cross-lane permutations
Product: gcc
Version: 9.1.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: jakub at gcc dot gnu.org
Target Milestone: ---
On:
typedef float __v8sf __attribute__((vector_size (32)));
typedef double __v4df __attribute__((vector_size (32)));
typedef int __v8si __attribute__((vector_size (32)));
typedef long long __v4di __attribute__((vector_size (32)));
#ifdef __clang__
#define S(x, y, t, ...) __builtin_shufflevector (x, y, __VA_ARGS__)
#else
#define S(x, y, t, ...) __builtin_shuffle (x, y, (t) { __VA_ARGS__ })
#endif
__v8sf f1 (__v8sf x, __v8sf y) { return S (x, y, __v8si, 0, 8, 9, 10, 11, 12,
13, 14 ); }
__v8sf f2 (__v8sf x, __v8sf y) { return S (x, y, __v8si, 0, 1, 8, 9, 10, 11,
12, 13 ); }
__v8sf f3 (__v8sf x, __v8sf y) { return S (x, y, __v8si, 0, 1, 2, 3, 8, 9, 10,
11 ); }
__v8sf f4 (__v8sf x, __v8sf y) { return S (x, y, __v8si, 7, 7, 7, 7, 7, 7, 7, 7
); }
__v4df f5 (__v4df x, __v4df y) { return S (x, y, __v4di, 0, 4, 5, 6 ); }
__v4df f6 (__v4df x, __v4df y) { return S (x, y, __v4di, 0, 1, 4, 5 ); }
__v4df f7 (__v4df x, __v4df y) { return S (x, y, __v4di, 3, 3, 3, 3 ); }
LLVM generates for -O2 -mavx -mno-avx2 shorter code for f1 and f2 (but worse
code for f5), GCC simply gives up for f1/f2 and then expands the shuffle as
lots of BIT_FIELD_REF extractions plus vector creation.
Wonder if ix86_expand_vec_perm_const_1 for the if (TARGET_AVX && !TARGET_AVX2)
and 32-byte vectors shouldn't try harder (though, with rightly estimated
costs).
The above permutations are what is used for OpenMP scans in scan-13.c (and
variant thereof with double instead of float) and reason why we don't vectorize
using 32-byte vectors.
