| 1 |
/* Test the `vmaxnmf32' ARM Neon intrinsic. */ |
| 2 |
|
| 3 |
/* { dg-do run } */ |
| 4 |
/* { dg-require-effective-target arm_v8_neon_hw } */ |
| 5 |
/* { dg-options "-save-temps -O3 -march=armv8-a" } */ |
| 6 |
/* { dg-add-options arm_v8_neon } */ |
| 7 |
|
| 8 |
#include "arm_neon.h" |
| 9 |
|
| 10 |
extern void abort (); |
| 11 |
|
| 12 |
void __attribute__ ((noinline)) |
| 13 |
test_vmaxnm_f32__regular_input1 () |
| 14 |
{ |
| 15 |
float32_t a1[] = {1,2}; |
| 16 |
float32_t b1[] = {3,4}; |
| 17 |
float32x2_t a = vld1_f32 (a1); |
| 18 |
float32x2_t b = vld1_f32 (b1); |
| 19 |
float32x2_t c = vmaxnm_f32 (a, b); |
| 20 |
float32_t actual[2]; |
| 21 |
vst1_f32 (actual, c); |
| 22 |
|
| 23 |
for (int i = 0; i < 2; ++i) |
| 24 |
if (actual[i] != b1[i]) |
| 25 |
abort (); |
| 26 |
} |
| 27 |
|
| 28 |
void __attribute__ ((noinline)) |
| 29 |
test_vmaxnm_f32__regular_input2 () |
| 30 |
{ |
| 31 |
float32_t a1[] = {3,2}; |
| 32 |
float32_t b1[] = {1,4}; |
| 33 |
float32_t e[] = {3,4}; |
| 34 |
float32x2_t a = vld1_f32 (a1); |
| 35 |
float32x2_t b = vld1_f32 (b1); |
| 36 |
float32x2_t c = vmaxnm_f32 (a, b); |
| 37 |
float32_t actual[2]; |
| 38 |
vst1_f32 (actual, c); |
| 39 |
|
| 40 |
for (int i = 0; i < 2; ++i) |
| 41 |
if (actual[i] != e[i]) |
| 42 |
abort (); |
| 43 |
} |
| 44 |
|
| 45 |
void __attribute__ ((noinline)) |
| 46 |
test_vmaxnm_f32__quiet_NaN_one_arg () |
| 47 |
{ |
| 48 |
/* When given a quiet NaN, vmaxnm returns the other operand. |
| 49 |
In this test case we have NaNs in only one operand. */ |
| 50 |
float32_t n = __builtin_nanf (""); |
| 51 |
float32_t a1[] = {1,2}; |
| 52 |
float32_t b1[] = {n,n}; |
| 53 |
float32_t e[] = {1,2}; |
| 54 |
float32x2_t a = vld1_f32 (a1); |
| 55 |
float32x2_t b = vld1_f32 (b1); |
| 56 |
float32x2_t c = vmaxnm_f32 (a, b); |
| 57 |
float32_t actual[2]; |
| 58 |
vst1_f32 (actual, c); |
| 59 |
|
| 60 |
for (int i = 0; i < 2; ++i) |
| 61 |
if (actual[i] != e[i]) |
| 62 |
abort (); |
| 63 |
} |
| 64 |
|
| 65 |
void __attribute__ ((noinline)) |
| 66 |
test_vmaxnm_f32__quiet_NaN_both_args () |
| 67 |
{ |
| 68 |
/* When given a quiet NaN, vmaxnm returns the other operand. |
| 69 |
In this test case we have NaNs in both operands. */ |
| 70 |
float32_t n = __builtin_nanf (""); |
| 71 |
float32_t a1[] = {n,2}; |
| 72 |
float32_t b1[] = {1,n}; |
| 73 |
float32_t e[] = {1,2}; |
| 74 |
float32x2_t a = vld1_f32 (a1); |
| 75 |
float32x2_t b = vld1_f32 (b1); |
| 76 |
float32x2_t c = vmaxnm_f32 (a, b); |
| 77 |
float32_t actual[2]; |
| 78 |
vst1_f32 (actual, c); |
| 79 |
|
| 80 |
for (int i = 0; i < 2; ++i) |
| 81 |
if (actual[i] != e[i]) |
| 82 |
abort (); |
| 83 |
} |
| 84 |
|
| 85 |
void __attribute__ ((noinline)) |
| 86 |
test_vmaxnm_f32__zero_both_args () |
| 87 |
{ |
| 88 |
/* For 0 and -0, vmaxnm returns 0. Since 0 == -0, check sign bit. */ |
| 89 |
float32_t a1[] = {0.0, 0.0}; |
| 90 |
float32_t b1[] = {-0.0, -0.0}; |
| 91 |
float32_t e[] = {0.0, 0.0}; |
| 92 |
|
| 93 |
float32x2_t a = vld1_f32 (a1); |
| 94 |
float32x2_t b = vld1_f32 (b1); |
| 95 |
float32x2_t c = vmaxnm_f32 (a, b); |
| 96 |
|
| 97 |
float32_t actual1[2]; |
| 98 |
vst1_f32 (actual1, c); |
| 99 |
|
| 100 |
for (int i = 0; i < 2; ++i) |
| 101 |
if (actual1[i] != e[i] || __builtin_signbit (actual1[i]) != 0) |
| 102 |
abort (); |
| 103 |
} |
| 104 |
|
| 105 |
void __attribute__ ((noinline)) |
| 106 |
test_vmaxnm_f32__inf_both_args () |
| 107 |
{ |
| 108 |
/* The max of inf and inf is inf. The max of -inf and -inf is -inf. */ |
| 109 |
float32_t inf = __builtin_huge_valf (); |
| 110 |
float32_t a1[] = {inf, -inf}; |
| 111 |
float32_t b1[] = {inf, -inf}; |
| 112 |
float32_t e[] = {inf, -inf}; |
| 113 |
|
| 114 |
float32x2_t a = vld1_f32 (a1); |
| 115 |
float32x2_t b = vld1_f32 (b1); |
| 116 |
float32x2_t c = vmaxnm_f32 (a, b); |
| 117 |
|
| 118 |
float32_t actual1[2]; |
| 119 |
vst1_f32 (actual1, c); |
| 120 |
|
| 121 |
for (int i = 0; i < 2; ++i) |
| 122 |
if (actual1[i] != e[i]) |
| 123 |
abort (); |
| 124 |
} |
| 125 |
|
| 126 |
void __attribute__ ((noinline)) |
| 127 |
test_vmaxnm_f32__two_quiet_NaNs_both_args () |
| 128 |
{ |
| 129 |
/* When given 2 NaNs, return a NaN. Since a NaN is not equal to anything, |
| 130 |
not even another NaN, use __builtin_isnan () to check. */ |
| 131 |
float32_t n = __builtin_nanf (""); |
| 132 |
float32_t a1[] = {n,n}; |
| 133 |
float32_t b1[] = {n,n}; |
| 134 |
float32_t e[] = {n,n}; |
| 135 |
float32x2_t a = vld1_f32 (a1); |
| 136 |
float32x2_t b = vld1_f32 (b1); |
| 137 |
float32x2_t c = vmaxnm_f32 (a, b); |
| 138 |
float32_t actual[2]; |
| 139 |
vst1_f32 (actual, c); |
| 140 |
|
| 141 |
for (int i = 0; i < 2; ++i) |
| 142 |
if (!__builtin_isnan (actual[i])) |
| 143 |
abort (); |
| 144 |
} |
| 145 |
|
| 146 |
int |
| 147 |
main () |
| 148 |
{ |
| 149 |
test_vmaxnm_f32__regular_input1 (); |
| 150 |
test_vmaxnm_f32__regular_input2 (); |
| 151 |
test_vmaxnm_f32__quiet_NaN_one_arg (); |
| 152 |
test_vmaxnm_f32__quiet_NaN_both_args (); |
| 153 |
test_vmaxnm_f32__zero_both_args (); |
| 154 |
test_vmaxnm_f32__inf_both_args (); |
| 155 |
test_vmaxnm_f32__two_quiet_NaNs_both_args (); |
| 156 |
return 0; |
| 157 |
} |
| 158 |
|
| 159 |
/* { dg-final { scan-assembler-times "vmaxnm\.f32\t\[dD\]\[0-9\]+, ?\[dD\]\[0-9\]+, ?\[dD\]\[0-9\]+\n" 7 } } */ |