non-temporal stores: use inline assembly

Author: RalfJung

crates/core_arch/src/x86/avx.rs

@@ -1718,7 +1718,11 @@ pub unsafe fn _mm256_lddqu_si256(mem_addr: *const __m256i) -> __m256i {
 #[cfg_attr(test, assert_instr(vmovntps))] // FIXME vmovntdq
 #[stable(feature = "simd_x86", since = "1.27.0")]
 pub unsafe fn _mm256_stream_si256(mem_addr: *mut __m256i, a: __m256i) {
-    intrinsics::nontemporal_store(mem_addr, a);
+    crate::arch::asm!(
+        "vmovntps [{mem_addr}], {a}",
+        mem_addr = in(reg) mem_addr,
+        a = in(ymm_reg) a,
+    );
 }
 
 /// Moves double-precision values from a 256-bit vector of `[4 x double]`
@@ -1741,7 +1745,11 @@ pub unsafe fn _mm256_stream_si256(mem_addr: *mut __m256i, a: __m256i) {
 #[stable(feature = "simd_x86", since = "1.27.0")]
 #[allow(clippy::cast_ptr_alignment)]
 pub unsafe fn _mm256_stream_pd(mem_addr: *mut f64, a: __m256d) {
-    intrinsics::nontemporal_store(mem_addr as *mut __m256d, a);
+    crate::arch::asm!(
+        "vmovntps [{mem_addr}], {a}",
+        mem_addr = in(reg) mem_addr,
+        a = in(ymm_reg) a,
+    );
 }
 
 /// Moves single-precision floating point values from a 256-bit vector
@@ -1765,7 +1773,11 @@ pub unsafe fn _mm256_stream_pd(mem_addr: *mut f64, a: __m256d) {
 #[stable(feature = "simd_x86", since = "1.27.0")]
 #[allow(clippy::cast_ptr_alignment)]
 pub unsafe fn _mm256_stream_ps(mem_addr: *mut f32, a: __m256) {
-    intrinsics::nontemporal_store(mem_addr as *mut __m256, a);
+    crate::arch::asm!(
+        "vmovntps [{mem_addr}], {a}",
+        mem_addr = in(reg) mem_addr,
+        a = in(ymm_reg) a,
+    );
 }
 
 /// Computes the approximate reciprocal of packed single-precision (32-bit)

crates/core_arch/src/x86/avx512f.rs

@@ -28014,7 +28014,11 @@ pub unsafe fn _mm_mask_testn_epi64_mask(k: __mmask8, a: __m128i, b: __m128i) ->
 #[cfg_attr(test, assert_instr(vmovntps))]
 #[allow(clippy::cast_ptr_alignment)]
 pub unsafe fn _mm512_stream_ps(mem_addr: *mut f32, a: __m512) {
-    intrinsics::nontemporal_store(mem_addr as *mut __m512, a);
+    crate::arch::asm!(
+        "vmovntps [{mem_addr}], {a}",
+        mem_addr = in(reg) mem_addr,
+        a = in(zmm_reg) a,
+    );
 }
 
 /// Store 512-bits (composed of 8 packed double-precision (64-bit) floating-point elements) from a into memory using a non-temporal memory hint. mem_addr must be aligned on a 64-byte boundary or a general-protection exception may be generated.
@@ -28035,7 +28039,11 @@ pub unsafe fn _mm512_stream_ps(mem_addr: *mut f32, a: __m512) {
 #[cfg_attr(test, assert_instr(vmovntps))] //should be vmovntpd
 #[allow(clippy::cast_ptr_alignment)]
 pub unsafe fn _mm512_stream_pd(mem_addr: *mut f64, a: __m512d) {
-    intrinsics::nontemporal_store(mem_addr as *mut __m512d, a);
+    crate::arch::asm!(
+        "vmovntps [{mem_addr}], {a}",
+        mem_addr = in(reg) mem_addr,
+        a = in(zmm_reg) a,
+    );
 }
 
 /// Store 512-bits of integer data from a into memory using a non-temporal memory hint. mem_addr must be aligned on a 64-byte boundary or a general-protection exception may be generated.
@@ -28056,7 +28064,11 @@ pub unsafe fn _mm512_stream_pd(mem_addr: *mut f64, a: __m512d) {
 #[cfg_attr(test, assert_instr(vmovntps))] //should be vmovntdq
 #[allow(clippy::cast_ptr_alignment)]
 pub unsafe fn _mm512_stream_si512(mem_addr: *mut i64, a: __m512i) {
-    intrinsics::nontemporal_store(mem_addr as *mut __m512i, a);
+    crate::arch::asm!(
+        "vmovntps [{mem_addr}], {a}",
+        mem_addr = in(reg) mem_addr,
+        a = in(zmm_reg) a,
+    );
 }
 
 /// Sets packed 32-bit integers in `dst` with the supplied values.

crates/core_arch/src/x86/mod.rs

@@ -2,7 +2,7 @@
 
 #[allow(unused_imports)]
 use crate::marker::Sized;
-use crate::{intrinsics, mem::transmute};
+use crate::mem::transmute;
 
 #[macro_use]
 mod macros;

crates/core_arch/src/x86/sse.rs

@@ -2002,7 +2002,11 @@ extern "C" {
 #[stable(feature = "simd_x86", since = "1.27.0")]
 #[allow(clippy::cast_ptr_alignment)]
 pub unsafe fn _mm_stream_ps(mem_addr: *mut f32, a: __m128) {
-    intrinsics::nontemporal_store(mem_addr as *mut __m128, a);
+    crate::arch::asm!(
+        "movntps [{mem_addr}], {a}",
+        mem_addr = in(reg) mem_addr,
+        a = in(xmm_reg) a,
+    );
 }
 
 #[cfg(test)]

crates/core_arch/src/x86/sse2.rs

@@ -1327,11 +1327,15 @@ pub unsafe fn _mm_storel_epi64(mem_addr: *mut __m128i, a: __m128i) {
 ///
 /// See [`_mm_sfence`] for details.
 #[inline]
-#[target_feature(enable = "sse2")]
+#[target_feature(enable = "sse,sse2")]
 #[cfg_attr(test, assert_instr(movntps))] // FIXME movntdq
 #[stable(feature = "simd_x86", since = "1.27.0")]
 pub unsafe fn _mm_stream_si128(mem_addr: *mut __m128i, a: __m128i) {
-    intrinsics::nontemporal_store(mem_addr, a);
+    crate::arch::asm!(
+        "movntps [{mem_addr}], {a}",
+        mem_addr = in(reg) mem_addr,
+        a = in(xmm_reg) a,
+    );
 }
 
 /// Stores a 32-bit integer value in the specified memory location.
@@ -1353,7 +1357,11 @@ pub unsafe fn _mm_stream_si128(mem_addr: *mut __m128i, a: __m128i) {
 #[cfg_attr(test, assert_instr(movnti))]
 #[stable(feature = "simd_x86", since = "1.27.0")]
 pub unsafe fn _mm_stream_si32(mem_addr: *mut i32, a: i32) {
-    intrinsics::nontemporal_store(mem_addr, a);
+    crate::arch::asm!(
+        "movnti [{mem_addr}], {a:e}", // `:e` for 32bit value
+        mem_addr = in(reg) mem_addr,
+        a = in(reg) a,
+    );
 }
 
 /// Returns a vector where the low element is extracted from `a` and its upper
@@ -2543,12 +2551,16 @@ pub unsafe fn _mm_loadl_pd(a: __m128d, mem_addr: *const f64) -> __m128d {
 ///
 /// See [`_mm_sfence`] for details.
 #[inline]
-#[target_feature(enable = "sse2")]
+#[target_feature(enable = "sse,sse2")]
 #[cfg_attr(test, assert_instr(movntps))] // FIXME movntpd
 #[stable(feature = "simd_x86", since = "1.27.0")]
 #[allow(clippy::cast_ptr_alignment)]
 pub unsafe fn _mm_stream_pd(mem_addr: *mut f64, a: __m128d) {
-    intrinsics::nontemporal_store(mem_addr as *mut __m128d, a);
+    crate::arch::asm!(
+        "movntps [{mem_addr}], {a}",
+        mem_addr = in(reg) mem_addr,
+        a = in(xmm_reg) a,
+    );
 }
 
 /// Stores the lower 64 bits of a 128-bit vector of `[2 x double]` to a

crates/core_arch/src/x86_64/sse2.rs

@@ -1,9 +1,6 @@
 //! `x86_64`'s Streaming SIMD Extensions 2 (SSE2)
 
-use crate::{
-    core_arch::x86::*,
-    intrinsics::{self, simd::*},
-};
+use crate::{core_arch::x86::*, intrinsics::simd::*};
 
 #[cfg(test)]
 use stdarch_test::assert_instr;
@@ -81,7 +78,11 @@ pub unsafe fn _mm_cvttsd_si64x(a: __m128d) -> i64 {
 #[cfg_attr(test, assert_instr(movnti))]
 #[stable(feature = "simd_x86", since = "1.27.0")]
 pub unsafe fn _mm_stream_si64(mem_addr: *mut i64, a: i64) {
-    intrinsics::nontemporal_store(mem_addr, a);
+    crate::arch::asm!(
+        "movnti [{mem_addr}], {a}",
+        mem_addr = in(reg) mem_addr,
+        a = in(reg) a,
+    );
 }
 
 /// Returns a vector whose lowest element is `a` and all higher elements are