[AArch64] Predicated SVE comparison folds
[official-gcc.git] / gcc / testsuite / gcc.target / aarch64 / sve / vcond_6.c
bloba59f08d553a5d191d941717b5005b48e0f138907
1 /* { dg-do compile } */
2 /* { dg-options "-O2 -ftree-vectorize" } */
4 #include <stdint.h>
6 #define b_and(A, B) ((A) & (B))
7 #define b_orr(A, B) ((A) | (B))
8 #define b_eor(A, B) ((A) ^ (B))
9 #define b_nand(A, B) (!((A) & (B)))
10 #define b_nor(A, B) (!((A) | (B)))
11 #define b_bic(A, B) ((A) & !(B))
12 #define b_orn(A, B) ((A) | !(B))
14 #define LOOP(TYPE, BINOP) \
15 void __attribute__ ((noinline, noclone)) \
16 test_##TYPE##_##BINOP (TYPE *restrict dest, TYPE *restrict src, \
17 TYPE *restrict a, TYPE *restrict b, \
18 TYPE *restrict c, TYPE *restrict d, \
19 TYPE fallback, int count) \
20 { \
21 for (int i = 0; i < count; ++i) \
22 { \
23 TYPE srcv = src[i]; \
24 dest[i] = (BINOP (__builtin_isunordered (a[i], b[i]), \
25 __builtin_isunordered (c[i], d[i])) \
26 ? srcv : fallback); \
27 } \
30 #define TEST_BINOP(T, BINOP) \
31 T (_Float16, BINOP) \
32 T (float, BINOP) \
33 T (double, BINOP)
35 #define TEST_ALL(T) \
36 TEST_BINOP (T, b_and) \
37 TEST_BINOP (T, b_orr) \
38 TEST_BINOP (T, b_eor) \
39 TEST_BINOP (T, b_nand) \
40 TEST_BINOP (T, b_nor) \
41 TEST_BINOP (T, b_bic) \
42 TEST_BINOP (T, b_orn)
44 TEST_ALL (LOOP)
46 /* ??? We predicate one of the comparisons on the result of the other,
47 but whether that's a win or a loss will depend on the schedule. */
48 /* { dg-final { scan-assembler-not {\tand\t} } } */
49 /* { dg-final { scan-assembler-times {\torr\tp[0-9]+\.b, p[0-9]+/z, p[0-9]+\.b, p[0-9]+\.b} 3 } } */
50 /* { dg-final { scan-assembler-times {\teor\tp[0-9]+\.b, p[0-9]+/z, p[0-9]+\.b, p[0-9]+\.b} 3 } } */
51 /* { dg-final { scan-assembler-times {\tnand\tp[0-9]+\.b, p[0-9]+/z, p[0-9]+\.b, p[0-9]+\.b} 3 } } */
52 /* { dg-final { scan-assembler-times {\tnor\tp[0-9]+\.b, p[0-9]+/z, p[0-9]+\.b, p[0-9]+\.b} 3 } } */
53 /* Currently we predicate one of the comparisons on the result of the other
54 and then use NOT, but the original BIC sequence is better. It's a fairly
55 niche failure though. We'd handle most other types of comparison by
56 using the inverse operation instead of a separate NOT. */
57 /* { dg-final { scan-assembler-times {\tbic\tp[0-9]+\.b, p[0-9]+/z, p[0-9]+\.b, p[0-9]+\.b} 3 { xfail *-*-* } } */
58 /* { dg-final { scan-assembler-times {\torn\tp[0-9]+\.b, p[0-9]+/z, p[0-9]+\.b, p[0-9]+\.b} 3 } } */