| Message ID | 20190223012206.30498-1-richard.henderson@linaro.org |
|---|---|
| Headers | show |
| Series | target/arm: Implement ARMv8.5-FRINT | expand |
Hi Richard, On Sat, Feb 23, 2019 at 2:23 AM Richard Henderson <richard.henderson@linaro.org> wrote: > > Based-on: the ARMv8.2-FHM patch set, although I don't know that > this is an actual dependency; it's just the tree I started with. > > There is not yet support for this extension within FVP, so I've > self-tested it against my own understanding of what is supposed > to go on with the file below. I ran some tests and found no issue. Your understanding looks correct :-) Tested-by: Laurent Desnogues <laurent.desnogues@gmail.com> Thanks, Laurent > > r~ > > > Richard Henderson (2): > target/arm: Restructure handle_fp_1src_{single,double} > target/arm: Implement ARMv8.5-FRINT > > target/arm/cpu.h | 5 ++ > target/arm/helper.h | 5 ++ > target/arm/cpu64.c | 1 + > target/arm/translate-a64.c | 161 ++++++++++++++++++++++++++----------- > target/arm/vfp_helper.c | 96 ++++++++++++++++++++++ > 5 files changed, 222 insertions(+), 46 deletions(-) > > -- > > #include <stdio.h> > #include <limits.h> > > asm(".arch armv8.5-a"); > > static const float f32_i32_z[][2] = { > { 0, 0 }, > { -0.0, -0.0 }, > { 0.9, 0.0 }, > { 1.0, 1.0 }, > { 1.1, 1.0 }, > { 1.9, 1.0 }, > { 0x1.fffffep30, 0x1.fffffep30 }, /* 2^31 - epsilon */ > { -0x1p31, -0x1.0p31 }, /* -2^31 = INT32_MIN */ > { 0x1.0p31, -0x1.0p31 }, /* overflow */ > { 0x1.0p32, -0x1.0p31 }, /* overflow */ > { -0x1.0p32, -0x1.0p31 }, /* overflow */ > { __builtin_inf(), -0x1.0p31 }, /* overflow */ > }; > > static const double f64_i32_z[][2] = { > { 0, 0 }, > { -0.0, -0.0 }, > { 0.9, 0.0 }, > { 1.0, 1.0 }, > { 1.1, 1.0 }, > { 1.9, 1.0 }, > { 2147483647, 2147483647 }, /* 2^31 - 1 = INT32_MAX */ > { -0x1p31, -0x1.0p31 }, /* -2^31 = INT32_MIN */ > { 0x1.0p31, -0x1.0p31 }, /* overflow */ > { 0x1.0p32, -0x1.0p31 }, /* overflow */ > { -0x1.0p32, -0x1.0p31 }, /* overflow */ > { __builtin_inf(), -0x1.0p31 }, /* overflow */ > }; > > static const float f32_i64_z[][2] = { > { 0, 0 }, > { -0.0, -0.0 }, > { 0.9, 0.0 }, > { 1.0, 1.0 }, > { 1.1, 1.0 }, > { 1.9, 1.0 }, > { 0x1.fffffep30, 0x1.fffffep30 }, /* 2^31 - epsilon */ > { -0x1p31, -0x1.0p31 }, /* -2^31 = INT32_MIN */ > { 0x1.0p31, 0x1.0p31 }, > { 0x1.0p32, 0x1.0p32 }, > { -0x1.0p32, -0x1.0p32 }, > { 0x1.0p62, 0x1.0p62 }, > { 0x1.fffffep62, 0x1.fffffep62 }, /* 2^63 - epsilon */ > { -0x1.0p63, -0x1.0p63 }, /* -2^63 = INT64_MIN */ > { 0x1.0p63, -0x1.0p63 }, /* overflow */ > { 0x1.0p64, -0x1.0p63 }, /* overflow */ > { -0x1.0p64, -0x1.0p63 }, /* overflow */ > { __builtin_inf(), -0x1.0p63 }, /* overflow */ > }; > > static const double f64_i64_z[][2] = { > { 0, 0 }, > { -0.0, -0.0 }, > { 0.9, 0.0 }, > { 1.0, 1.0 }, > { 1.1, 1.0 }, > { 1.9, 1.0 }, > { 2147483647, 2147483647 }, /* 2^31 - 1 = INT32_MAX */ > { -0x1p31, -0x1.0p31 }, /* -2^31 = INT32_MIN */ > { 0x1.0p31, 0x1.0p31 }, > { 0x1.0p32, 0x1.0p32 }, > { -0x1.0p32, -0x1.0p32 }, > { 0x1.0p62, 0x1.0p62 }, > { 0x1.fffffffffffffp62, 0x1.fffffffffffffp62 }, /* 2^63 - epsilon */ > { -0x1.0p63, -0x1.0p63 }, /* -2^63 = INT64_MIN */ > { 0x1.0p63, -0x1.0p63 }, /* overflow */ > { 0x1.0p64, -0x1.0p63 }, /* overflow */ > { -0x1.0p64, -0x1.0p63 }, /* overflow */ > { __builtin_inf(), -0x1.0p63 }, /* overflow */ > }; > > int main() > { > int i; > > for (i = 0; i < sizeof(f32_i32_z)/sizeof(f32_i32_z[0]); i++) { > float x; > asm("frint32z %s0,%s1" : "=w"(x) : "w"(f32_i32_z[i][0])); > if (x != f32_i32_z[i][1]) { > printf("%-2d: frint(%a) -> %a != %a\n", > i, f32_i32_z[i][0], x, f32_i32_z[i][1]); > } > } > > for (i = 0; i < sizeof(f32_i64_z)/sizeof(f32_i64_z[0]); i++) { > float x; > asm("frint64z %s0,%s1" : "=w"(x) : "w"(f32_i64_z[i][0])); > if (x != f32_i64_z[i][1]) { > printf("%-2d: frint(%a) -> %a != %a\n", > i, f32_i64_z[i][0], x, f32_i64_z[i][1]); > } > } > > for (i = 0; i < sizeof(f64_i32_z)/sizeof(f64_i32_z[0]); i++) { > double x; > asm("frint32z %d0,%d1" : "=w"(x) : "w"(f64_i32_z[i][0])); > if (x != f64_i32_z[i][1]) { > printf("%-2d: frint(%a) -> %a != %a\n", > i, f64_i32_z[i][0], x, f64_i32_z[i][1]); > } > } > > for (i = 0; i < sizeof(f64_i64_z)/sizeof(f64_i64_z[0]); i++) { > double x; > asm("frint64z %d0,%d1" : "=w"(x) : "w"(f64_i64_z[i][0])); > if (x != f64_i64_z[i][1]) { > printf("%-2d: frint(%a) -> %a != %a\n", > i, f64_i64_z[i][0], x, f64_i64_z[i][1]); > } > } > > return 0; > } >
Based-on: the ARMv8.2-FHM patch set, although I don't know that this is an actual dependency; it's just the tree I started with. There is not yet support for this extension within FVP, so I've self-tested it against my own understanding of what is supposed to go on with the file below. r~ Richard Henderson (2): target/arm: Restructure handle_fp_1src_{single,double} target/arm: Implement ARMv8.5-FRINT target/arm/cpu.h | 5 ++ target/arm/helper.h | 5 ++ target/arm/cpu64.c | 1 + target/arm/translate-a64.c | 161 ++++++++++++++++++++++++++----------- target/arm/vfp_helper.c | 96 ++++++++++++++++++++++ 5 files changed, 222 insertions(+), 46 deletions(-) -- #include <stdio.h> #include <limits.h> asm(".arch armv8.5-a"); static const float f32_i32_z[][2] = { { 0, 0 }, { -0.0, -0.0 }, { 0.9, 0.0 }, { 1.0, 1.0 }, { 1.1, 1.0 }, { 1.9, 1.0 }, { 0x1.fffffep30, 0x1.fffffep30 }, /* 2^31 - epsilon */ { -0x1p31, -0x1.0p31 }, /* -2^31 = INT32_MIN */ { 0x1.0p31, -0x1.0p31 }, /* overflow */ { 0x1.0p32, -0x1.0p31 }, /* overflow */ { -0x1.0p32, -0x1.0p31 }, /* overflow */ { __builtin_inf(), -0x1.0p31 }, /* overflow */ }; static const double f64_i32_z[][2] = { { 0, 0 }, { -0.0, -0.0 }, { 0.9, 0.0 }, { 1.0, 1.0 }, { 1.1, 1.0 }, { 1.9, 1.0 }, { 2147483647, 2147483647 }, /* 2^31 - 1 = INT32_MAX */ { -0x1p31, -0x1.0p31 }, /* -2^31 = INT32_MIN */ { 0x1.0p31, -0x1.0p31 }, /* overflow */ { 0x1.0p32, -0x1.0p31 }, /* overflow */ { -0x1.0p32, -0x1.0p31 }, /* overflow */ { __builtin_inf(), -0x1.0p31 }, /* overflow */ }; static const float f32_i64_z[][2] = { { 0, 0 }, { -0.0, -0.0 }, { 0.9, 0.0 }, { 1.0, 1.0 }, { 1.1, 1.0 }, { 1.9, 1.0 }, { 0x1.fffffep30, 0x1.fffffep30 }, /* 2^31 - epsilon */ { -0x1p31, -0x1.0p31 }, /* -2^31 = INT32_MIN */ { 0x1.0p31, 0x1.0p31 }, { 0x1.0p32, 0x1.0p32 }, { -0x1.0p32, -0x1.0p32 }, { 0x1.0p62, 0x1.0p62 }, { 0x1.fffffep62, 0x1.fffffep62 }, /* 2^63 - epsilon */ { -0x1.0p63, -0x1.0p63 }, /* -2^63 = INT64_MIN */ { 0x1.0p63, -0x1.0p63 }, /* overflow */ { 0x1.0p64, -0x1.0p63 }, /* overflow */ { -0x1.0p64, -0x1.0p63 }, /* overflow */ { __builtin_inf(), -0x1.0p63 }, /* overflow */ }; static const double f64_i64_z[][2] = { { 0, 0 }, { -0.0, -0.0 }, { 0.9, 0.0 }, { 1.0, 1.0 }, { 1.1, 1.0 }, { 1.9, 1.0 }, { 2147483647, 2147483647 }, /* 2^31 - 1 = INT32_MAX */ { -0x1p31, -0x1.0p31 }, /* -2^31 = INT32_MIN */ { 0x1.0p31, 0x1.0p31 }, { 0x1.0p32, 0x1.0p32 }, { -0x1.0p32, -0x1.0p32 }, { 0x1.0p62, 0x1.0p62 }, { 0x1.fffffffffffffp62, 0x1.fffffffffffffp62 }, /* 2^63 - epsilon */ { -0x1.0p63, -0x1.0p63 }, /* -2^63 = INT64_MIN */ { 0x1.0p63, -0x1.0p63 }, /* overflow */ { 0x1.0p64, -0x1.0p63 }, /* overflow */ { -0x1.0p64, -0x1.0p63 }, /* overflow */ { __builtin_inf(), -0x1.0p63 }, /* overflow */ }; int main() { int i; for (i = 0; i < sizeof(f32_i32_z)/sizeof(f32_i32_z[0]); i++) { float x; asm("frint32z %s0,%s1" : "=w"(x) : "w"(f32_i32_z[i][0])); if (x != f32_i32_z[i][1]) { printf("%-2d: frint(%a) -> %a != %a\n", i, f32_i32_z[i][0], x, f32_i32_z[i][1]); } } for (i = 0; i < sizeof(f32_i64_z)/sizeof(f32_i64_z[0]); i++) { float x; asm("frint64z %s0,%s1" : "=w"(x) : "w"(f32_i64_z[i][0])); if (x != f32_i64_z[i][1]) { printf("%-2d: frint(%a) -> %a != %a\n", i, f32_i64_z[i][0], x, f32_i64_z[i][1]); } } for (i = 0; i < sizeof(f64_i32_z)/sizeof(f64_i32_z[0]); i++) { double x; asm("frint32z %d0,%d1" : "=w"(x) : "w"(f64_i32_z[i][0])); if (x != f64_i32_z[i][1]) { printf("%-2d: frint(%a) -> %a != %a\n", i, f64_i32_z[i][0], x, f64_i32_z[i][1]); } } for (i = 0; i < sizeof(f64_i64_z)/sizeof(f64_i64_z[0]); i++) { double x; asm("frint64z %d0,%d1" : "=w"(x) : "w"(f64_i64_z[i][0])); if (x != f64_i64_z[i][1]) { printf("%-2d: frint(%a) -> %a != %a\n", i, f64_i64_z[i][0], x, f64_i64_z[i][1]); } } return 0; }