Merge pull request #3526 from laytan/target-features

Improve target features support
This commit is contained in:
gingerBill
2024-05-05 21:50:57 +01:00
committed by GitHub
19 changed files with 1067 additions and 205 deletions
+2 -2
View File
@@ -36,7 +36,7 @@ _mm_lddqu_si128 :: #force_inline proc "c" (mem_addr: ^__m128i) -> __m128i {
_mm_movedup_pd :: #force_inline proc "c" (a: __m128d) -> __m128d {
return simd.shuffle(a, a, 0, 0)
}
@(require_results, enable_target_feature="sse3")
@(require_results, enable_target_feature="sse2,sse3")
_mm_loaddup_pd :: #force_inline proc "c" (mem_addr: [^]f64) -> __m128d {
return _mm_load1_pd(mem_addr)
}
@@ -65,4 +65,4 @@ foreign _ {
hsubps :: proc(a, b: __m128) -> __m128 ---
@(link_name = "llvm.x86.sse3.ldu.dq")
lddqu :: proc(mem_addr: rawptr) -> i8x16 ---
}
}
+2 -2
View File
@@ -268,7 +268,7 @@ _mm_testnzc_si128 :: #force_inline proc "c" (a: __m128i, mask: __m128i) -> i32 {
_mm_test_all_zeros :: #force_inline proc "c" (a: __m128i, mask: __m128i) -> i32 {
return _mm_testz_si128(a, mask)
}
@(require_results, enable_target_feature="sse4.1")
@(require_results, enable_target_feature="sse2,sse4.1")
_mm_test_all_ones :: #force_inline proc "c" (a: __m128i) -> i32 {
return _mm_testc_si128(a, _mm_cmpeq_epi32(a, a))
}
@@ -349,4 +349,4 @@ foreign _ {
ptestc :: proc(a, mask: i64x2) -> i32 ---
@(link_name = "llvm.x86.sse41.ptestnzc")
ptestnzc :: proc(a, mask: i64x2) -> i32 ---
}
}
+31 -13
View File
@@ -5,31 +5,49 @@ package sync
import "base:intrinsics"
import "core:time"
// NOTE: because `core:sync` is in the dependency chain of a lot of the core packages (mostly through `core:mem`)
// without actually calling into it much, I opted for a runtime panic instead of a compile error here.
_futex_wait :: proc "contextless" (f: ^Futex, expected: u32) -> bool {
s := intrinsics.wasm_memory_atomic_wait32((^u32)(f), expected, -1)
return s != 0
when !intrinsics.has_target_feature("atomics") {
_panic("usage of `core:sync` requires the `-target-feature:\"atomics\"` or a `-microarch` that supports it")
} else {
s := intrinsics.wasm_memory_atomic_wait32((^u32)(f), expected, -1)
return s != 0
}
}
_futex_wait_with_timeout :: proc "contextless" (f: ^Futex, expected: u32, duration: time.Duration) -> bool {
s := intrinsics.wasm_memory_atomic_wait32((^u32)(f), expected, i64(duration))
return s != 0
when !intrinsics.has_target_feature("atomics") {
_panic("usage of `core:sync` requires the `-target-feature:\"atomics\"` or a `-microarch` that supports it")
} else {
s := intrinsics.wasm_memory_atomic_wait32((^u32)(f), expected, i64(duration))
return s != 0
}
}
_futex_signal :: proc "contextless" (f: ^Futex) {
loop: for {
s := intrinsics.wasm_memory_atomic_notify32((^u32)(f), 1)
if s >= 1 {
return
when !intrinsics.has_target_feature("atomics") {
_panic("usage of `core:sync` requires the `-target-feature:\"atomics\"` or a `-microarch` that supports it")
} else {
loop: for {
s := intrinsics.wasm_memory_atomic_notify32((^u32)(f), 1)
if s >= 1 {
return
}
}
}
}
_futex_broadcast :: proc "contextless" (f: ^Futex) {
loop: for {
s := intrinsics.wasm_memory_atomic_notify32((^u32)(f), ~u32(0))
if s >= 0 {
return
when !intrinsics.has_target_feature("atomics") {
_panic("usage of `core:sync` requires the `-target-feature:\"atomics\"` or a `-microarch` that supports it")
} else {
loop: for {
s := intrinsics.wasm_memory_atomic_notify32((^u32)(f), ~u32(0))
if s >= 0 {
return
}
}
}
}