diff options
| author | ptitSeb <sebastien.chev@gmail.com> | 2022-04-01 10:45:53 +0200 |
|---|---|---|
| committer | ptitSeb <sebastien.chev@gmail.com> | 2022-04-01 10:45:53 +0200 |
| commit | 5cd8176478bfba8b321a25e2ac849af7f2f93c82 (patch) | |
| tree | bb2c08562ce5752c8da1c175e5c519417964e983 /tests | |
| parent | ea0ee7c7aa8bc6f6fa2abcc98936ac53b629e30c (diff) | |
| download | box64-5cd8176478bfba8b321a25e2ac849af7f2f93c82.tar.gz box64-5cd8176478bfba8b321a25e2ac849af7f2f93c82.zip | |
More coverage for test17, and added NaN handling to 66 0F 7C opcode
Diffstat (limited to 'tests')
| -rw-r--r-- | tests/ref17.txt | 54 | ||||
| -rwxr-xr-x | tests/test17 | bin | 131040 -> 147912 bytes | |||
| -rw-r--r-- | tests/test17.c | 55 |
3 files changed, 108 insertions, 1 deletions
diff --git a/tests/ref17.txt b/tests/ref17.txt index 5dc2dad2..533dff62 100644 --- a/tests/ref17.txt +++ b/tests/ref17.txt @@ -262,4 +262,58 @@ punpcklbw(0xff 0x80 0x7f 0x0 0x1 0x2 0x3 0x81 0xfe 0x84 0x72 0x52 0xa5 0x0 0xc0 punpcklwd(0xffff 0x8000 0x7fff 0x0 0x1 0x2 0x3 0x8001 , 0x8000 0x7fff 0xffff 0xffff 0x50 0x9000 0xfffe 0x8001 ) = 0xffff 0x8000 0x8000 0x7fff 0x7fff 0xffff 0x0 0xffff punpckldq(0xffffffff 0x80000000 0x7fffffff 0x0 , 0x1 0x80000000 0x5 0xfffffffe ) = 0xffffffff 0x1 0x80000000 0x80000000 ppacksswb(0xffff 0x8000 0x7fff 0x0 0x1 0x2 0x3 0x8001 , 0x8000 0x7fff 0xffff 0xffff 0x50 0x9000 0xfffe 0x8001 ) = 0x80ff 0x7f 0x201 0x8003 0x7f80 0xffff 0x8050 0x80fe +pcmpgtb(0xff 0x80 0x7f 0x0 0x1 0x2 0x3 0x81 0xfe 0x84 0x72 0x52 0xa5 0x0 0xc0 0x32 , 0x0 0x1 0x5 0x15 0x20 0x80 0xff 0x0 0x8 0x7 0x81 0x6 0xa 0xf 0x10 0x1 ) = 0x0 0x0 0xff 0x0 0x0 0xff 0xff 0x0 0x0 0x0 0xff 0xff 0x0 0x0 0x0 0xff +pcmpgtw(0xffff 0x8000 0x7fff 0x0 0x1 0x2 0x3 0x8001 , 0x8000 0x7fff 0xffff 0xffff 0x50 0x9000 0xfffe 0x8001 ) = 0xffff 0x0 0xffff 0xffff 0x0 0xffff 0xffff 0x0 +pcmpgtd(0xffffffff 0x80000000 0x7fffffff 0x0 , 0x1 0x80000000 0x5 0xfffffffe ) = 0x0 0x0 0xffffffff 0xffffffff +packuswb(0xffff 0x8000 0x7fff 0x0 0x1 0x2 0x3 0x8001 , 0x8000 0x7fff 0xffff 0xffff 0x50 0x9000 0xfffe 0x8001 ) = 0x0 0xff 0x201 0x3 0xff00 0x0 0x50 0x0 +punpckhbw(0xff 0x80 0x7f 0x0 0x1 0x2 0x3 0x81 0xfe 0x84 0x72 0x52 0xa5 0x0 0xc0 0x32 , 0x0 0x1 0x5 0x15 0x20 0x80 0xff 0x0 0x8 0x7 0x81 0x6 0xa 0xf 0x10 0x1 ) = 0xfe 0x8 0x84 0x7 0x72 0x81 0x52 0x6 0xa5 0xa 0x0 0xf 0xc0 0x10 0x32 0x1 +punpckhwd(0xffff 0x8000 0x7fff 0x0 0x1 0x2 0x3 0x8001 , 0x8000 0x7fff 0xffff 0xffff 0x50 0x9000 0xfffe 0x8001 ) = 0x1 0x50 0x2 0x9000 0x3 0xfffe 0x8001 0x8001 +punpckhdq(0xffffffff 0x80000000 0x7fffffff 0x0 , 0x1 0x80000000 0x5 0xfffffffe ) = 0x7fffffff 0x5 0x0 0xfffffffe ppackssdw(0xffffffff 0x80000000 0x7fffffff 0x0 , 0x1 0x80000000 0x5 0xfffffffe ) = 0x8000ffff 0x7fff 0x80000001 0xfffe0005 +punpcklqdq(0xffffffffffffffff 0x8000000000000000 , 0x1 0x8000000000000000 ) = 0xffffffffffffffff 0x1 +punpckhqdq(0xffffffffffffffff 0x8000000000000000 , 0x1 0x8000000000000000 ) = 0x8000000000000000 0x8000000000000000 +pshufd(0xffffffff 0x80000000 0x7fffffff 0x0 0) = 0xffffffff 0xffffffff 0xffffffff 0xffffffff +pshufd(0xffffffff 0x80000000 0x7fffffff 0x0 255) = 0x0 0x0 0x0 0x0 +pshufd(0xffffffff 0x80000000 0x7fffffff 0x0 170) = 0x7fffffff 0x7fffffff 0x7fffffff 0x7fffffff +pshufd(0xffffffff 0x80000000 0x7fffffff 0x0 2) = 0x7fffffff 0xffffffff 0xffffffff 0xffffffff +psrlw(0xffff 0x8000 0x7fff 0x0 0x1 0x2 0x3 0x8001 0) = 0xffff 0x8000 0x7fff 0x0 0x1 0x2 0x3 0x8001 +psrlw(0xffff 0x8000 0x7fff 0x0 0x1 0x2 0x3 0x8001 255) = 0x0 0x0 0x0 0x0 0x0 0x0 0x0 0x0 +psrlw(0xffff 0x8000 0x7fff 0x0 0x1 0x2 0x3 0x8001 170) = 0x0 0x0 0x0 0x0 0x0 0x0 0x0 0x0 +psrlw(0xffff 0x8000 0x7fff 0x0 0x1 0x2 0x3 0x8001 2) = 0x3fff 0x2000 0x1fff 0x0 0x0 0x0 0x0 0x2000 +psrld(0xffffffff 0x80000000 0x7fffffff 0x0 0) = 0xffffffff 0x80000000 0x7fffffff 0x0 +psrld(0xffffffff 0x80000000 0x7fffffff 0x0 255) = 0x0 0x0 0x0 0x0 +psrld(0xffffffff 0x80000000 0x7fffffff 0x0 170) = 0x0 0x0 0x0 0x0 +psrld(0xffffffff 0x80000000 0x7fffffff 0x0 2) = 0x3fffffff 0x20000000 0x1fffffff 0x0 +psrlq(0xffffffffffffffff 0x8000000000000000 0) = 0xffffffffffffffff 0x8000000000000000 +psrlq(0xffffffffffffffff 0x8000000000000000 255) = 0x0 0x0 +psrlq(0xffffffffffffffff 0x8000000000000000 170) = 0x0 0x0 +psrlq(0xffffffffffffffff 0x8000000000000000 2) = 0x3fffffffffffffff 0x2000000000000000 +psraw(0xffff 0x8000 0x7fff 0x0 0x1 0x2 0x3 0x8001 0) = 0xffff 0x8000 0x7fff 0x0 0x1 0x2 0x3 0x8001 +psraw(0xffff 0x8000 0x7fff 0x0 0x1 0x2 0x3 0x8001 255) = 0xffff 0xffff 0x0 0x0 0x0 0x0 0x0 0xffff +psraw(0xffff 0x8000 0x7fff 0x0 0x1 0x2 0x3 0x8001 170) = 0xffff 0xffff 0x0 0x0 0x0 0x0 0x0 0xffff +psraw(0xffff 0x8000 0x7fff 0x0 0x1 0x2 0x3 0x8001 2) = 0xffff 0xe000 0x1fff 0x0 0x0 0x0 0x0 0xe000 +psrad(0xffffffff 0x80000000 0x7fffffff 0x0 0) = 0xffffffff 0x80000000 0x7fffffff 0x0 +psrad(0xffffffff 0x80000000 0x7fffffff 0x0 255) = 0xffffffff 0xffffffff 0x0 0x0 +psrad(0xffffffff 0x80000000 0x7fffffff 0x0 170) = 0xffffffff 0xffffffff 0x0 0x0 +psrad(0xffffffff 0x80000000 0x7fffffff 0x0 2) = 0xffffffff 0xe0000000 0x1fffffff 0x0 +psllw(0xffff 0x8000 0x7fff 0x0 0x1 0x2 0x3 0x8001 0) = 0xffff 0x8000 0x7fff 0x0 0x1 0x2 0x3 0x8001 +psllw(0xffff 0x8000 0x7fff 0x0 0x1 0x2 0x3 0x8001 255) = 0x0 0x0 0x0 0x0 0x0 0x0 0x0 0x0 +psllw(0xffff 0x8000 0x7fff 0x0 0x1 0x2 0x3 0x8001 170) = 0x0 0x0 0x0 0x0 0x0 0x0 0x0 0x0 +psllw(0xffff 0x8000 0x7fff 0x0 0x1 0x2 0x3 0x8001 2) = 0xfffc 0x0 0xfffc 0x0 0x4 0x8 0xc 0x4 +pslld(0xffffffff 0x80000000 0x7fffffff 0x0 0) = 0xffffffff 0x80000000 0x7fffffff 0x0 +pslld(0xffffffff 0x80000000 0x7fffffff 0x0 255) = 0x0 0x0 0x0 0x0 +pslld(0xffffffff 0x80000000 0x7fffffff 0x0 170) = 0x0 0x0 0x0 0x0 +pslld(0xffffffff 0x80000000 0x7fffffff 0x0 2) = 0xfffffffc 0x0 0xfffffffc 0x0 +psllq(0xffffffffffffffff 0x8000000000000000 0) = 0xffffffffffffffff 0x8000000000000000 +psllq(0xffffffffffffffff 0x8000000000000000 255) = 0x0 0x0 +psllq(0xffffffffffffffff 0x8000000000000000 170) = 0x0 0x0 +psllq(0xffffffffffffffff 0x8000000000000000 2) = 0xfffffffffffffffc 0x0 +pcmpeqb(0xff 0x80 0x7f 0x0 0x1 0x2 0x3 0x81 0xfe 0x84 0x72 0x52 0xa5 0x0 0xc0 0x32 , 0x0 0x1 0x5 0x15 0x20 0x80 0xff 0x0 0x8 0x7 0x81 0x6 0xa 0xf 0x10 0x1 ) = 0x0 0x0 0x0 0x0 0x0 0x0 0x0 0x0 0x0 0x0 0x0 0x0 0x0 0x0 0x0 0x0 +pcmpeqw(0xffff 0x8000 0x7fff 0x0 0x1 0x2 0x3 0x8001 , 0x8000 0x7fff 0xffff 0xffff 0x50 0x9000 0xfffe 0x8001 ) = 0x0 0x0 0x0 0x0 0x0 0x0 0x0 0xffff +pcmpeqd(0xffffffff 0x80000000 0x7fffffff 0x0 , 0x1 0x80000000 0x5 0xfffffffe ) = 0x0 0xffffffff 0x0 0x0 +haddpd(1 2 , 0 -2 ) = 3 -2 +haddpd(0 -2 , inf -inf ) = -2 0xfff8000000000000 +haddpd(1 2 , 0x7ff8000000000000 -0 ) = 3 0x7ff8000000000000 +haddpd(0 -2 , 0x7ff8000000000000 -0 ) = -2 0x7ff8000000000000 +haddpd(inf -inf , 0x7ff8000000000000 -0 ) = 0xfff8000000000000 0x7ff8000000000000 +haddpd(0x7ff8000000000000 -0 , 0x7ff8000000000000 -0 ) = 0x7ff8000000000000 0x7ff8000000000000 diff --git a/tests/test17 b/tests/test17 index d4bca8cc..aa5cd04f 100755 --- a/tests/test17 +++ b/tests/test17 Binary files differdiff --git a/tests/test17.c b/tests/test17.c index 883abf66..aa0d80a5 100644 --- a/tests/test17.c +++ b/tests/test17.c @@ -252,6 +252,10 @@ printf(N " %g, %g => %g\n", b, a, *(float*)&r); a128.mm = _mm_##A##_epi##N(a128_##N.mm); \ printf("%s(", #C); print_##N(a128_##N); \ printf(") = "); print_##N(a128); printf("\n"); + #define GO1C(A, N, C, A1, I) \ + a128.mm = _mm_##A##_epi##N(A1.mm, I); \ + printf("%s(", #C); print_##N(A1); \ + printf("%d) = ", I); print_##N(a128); printf("\n"); #define GO2(A, N, C, A1, A2) \ a128.mm = _mm_##A##_epi##N(A1.mm, A2.mm); \ printf("%s(", #C); print_##N(A1); \ @@ -355,9 +359,58 @@ printf(N " %g, %g => %g\n", b, a, *(float*)&r); GO2(unpacklo, 16, punpcklwd, a128_16, b128_16) GO2(unpacklo, 32, punpckldq, a128_32, b128_32) GO2(packs, 16, ppacksswb, a128_16, b128_16) - + GO2(cmpgt, 8, pcmpgtb, a128_8, b128_8) + GO2(cmpgt, 16, pcmpgtw, a128_16, b128_16) + GO2(cmpgt, 32, pcmpgtd, a128_32, b128_32) + GO2(packus, 16, packuswb, a128_16, b128_16) + GO2(unpackhi, 8, punpckhbw, a128_8, b128_8) + GO2(unpackhi, 16, punpckhwd, a128_16, b128_16) + GO2(unpackhi, 32, punpckhdq, a128_32, b128_32) GO2(packs, 32, ppackssdw, a128_32, b128_32) + GO2(unpacklo, 64, punpcklqdq, a128_64, b128_64) + GO2(unpackhi, 64, punpckhqdq, a128_64, b128_64) + GO1C(shuffle, 32, pshufd, a128_32, 0) + GO1C(shuffle, 32, pshufd, a128_32, 0xff) + GO1C(shuffle, 32, pshufd, a128_32, 0xaa) + GO1C(shuffle, 32, pshufd, a128_32, 2) + GO1C(srli, 16, psrlw, a128_16, 0) + GO1C(srli, 16, psrlw, a128_16, 0xff) + GO1C(srli, 16, psrlw, a128_16, 0xaa) + GO1C(srli, 16, psrlw, a128_16, 2) + GO1C(srli, 32, psrld, a128_32, 0) + GO1C(srli, 32, psrld, a128_32, 0xff) + GO1C(srli, 32, psrld, a128_32, 0xaa) + GO1C(srli, 32, psrld, a128_32, 2) + GO1C(srli, 64, psrlq, a128_64, 0) + GO1C(srli, 64, psrlq, a128_64, 0xff) + GO1C(srli, 64, psrlq, a128_64, 0xaa) + GO1C(srli, 64, psrlq, a128_64, 2) + GO1C(srai, 16, psraw, a128_16, 0) + GO1C(srai, 16, psraw, a128_16, 0xff) + GO1C(srai, 16, psraw, a128_16, 0xaa) + GO1C(srai, 16, psraw, a128_16, 2) + GO1C(srai, 32, psrad, a128_32, 0) + GO1C(srai, 32, psrad, a128_32, 0xff) + GO1C(srai, 32, psrad, a128_32, 0xaa) + GO1C(srai, 32, psrad, a128_32, 2) + GO1C(slli, 16, psllw, a128_16, 0) + GO1C(slli, 16, psllw, a128_16, 0xff) + GO1C(slli, 16, psllw, a128_16, 0xaa) + GO1C(slli, 16, psllw, a128_16, 2) + GO1C(slli, 32, pslld, a128_32, 0) + GO1C(slli, 32, pslld, a128_32, 0xff) + GO1C(slli, 32, pslld, a128_32, 0xaa) + GO1C(slli, 32, pslld, a128_32, 2) + GO1C(slli, 64, psllq, a128_64, 0) + GO1C(slli, 64, psllq, a128_64, 0xff) + GO1C(slli, 64, psllq, a128_64, 0xaa) + GO1C(slli, 64, psllq, a128_64, 2) + GO2(cmpeq, 8, pcmpeqb, a128_8, b128_8) + GO2(cmpeq, 16, pcmpeqw, a128_16, b128_16) + GO2(cmpeq, 32, pcmpeqd, a128_32, b128_32) + MULITGO2pd(hadd, haddpd) return 0; } + |