Fix the stream intrinsics

They should use a platform-specific address management.

Author: sayantn

crates/core_arch/src/x86/avx.rs

@@ -1738,8 +1738,8 @@ pub unsafe fn _mm256_lddqu_si256(mem_addr: *const __m256i) -> __m256i {
 #[stable(feature = "simd_x86", since = "1.27.0")]
 pub unsafe fn _mm256_stream_si256(mem_addr: *mut __m256i, a: __m256i) {
     crate::arch::asm!(
-        "vmovntdq [{mem_addr}], {a}",
-        mem_addr = in(reg) mem_addr,
+        vps!("vmovntdq", ",{a}"),
+        p = in(reg) mem_addr,
         a = in(ymm_reg) a,
         options(nostack, preserves_flags),
     );
@@ -1766,8 +1766,8 @@ pub unsafe fn _mm256_stream_si256(mem_addr: *mut __m256i, a: __m256i) {
 #[allow(clippy::cast_ptr_alignment)]
 pub unsafe fn _mm256_stream_pd(mem_addr: *mut f64, a: __m256d) {
     crate::arch::asm!(
-        "vmovntpd [{mem_addr}], {a}",
-        mem_addr = in(reg) mem_addr,
+        vps!("vmovntpd", ",{a}"),
+        p = in(reg) mem_addr,
         a = in(ymm_reg) a,
         options(nostack, preserves_flags),
     );
@@ -1795,8 +1795,8 @@ pub unsafe fn _mm256_stream_pd(mem_addr: *mut f64, a: __m256d) {
 #[allow(clippy::cast_ptr_alignment)]
 pub unsafe fn _mm256_stream_ps(mem_addr: *mut f32, a: __m256) {
     crate::arch::asm!(
-        "vmovntps [{mem_addr}], {a}",
-        mem_addr = in(reg) mem_addr,
+        vps!("vmovntps", ",{a}"),
+        p = in(reg) mem_addr,
         a = in(ymm_reg) a,
         options(nostack, preserves_flags),
     );

crates/core_arch/src/x86/avx2.rs

@@ -3149,9 +3149,9 @@ pub unsafe fn _mm256_srlv_epi64(a: __m256i, count: __m256i) -> __m256i {
 pub unsafe fn _mm256_stream_load_si256(mem_addr: *const __m256i) -> __m256i {
     let dst: __m256i;
     crate::arch::asm!(
-        "vmovntdqa {a}, [{mem_addr}]",
+        vpl!("vmovntdqa {a}"),
         a = out(ymm_reg) dst,
-        mem_addr = in(reg) mem_addr,
+        p = in(reg) mem_addr,
         options(pure, readonly, nostack, preserves_flags),
     );
     dst

crates/core_arch/src/x86/avx512bw.rs

@@ -8,8 +8,6 @@ use crate::{
 #[cfg(test)]
 use stdarch_test::assert_instr;
 
-use super::avx512f::{vpl, vps};
-
 /// Compute the absolute value of packed signed 16-bit integers in a, and store the unsigned results in dst.
 ///
 /// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm512_abs_epi16&expand=30)

crates/core_arch/src/x86/avx512f.rs

@@ -5,37 +5,6 @@ use crate::{
     mem, ptr,
 };
 
-// x86-32 wants to use a 32-bit address size, but asm! defaults to using the full
-// register name (e.g. rax). We have to explicitly override the placeholder to
-// use the 32-bit register name in that case.
-
-#[cfg(target_pointer_width = "32")]
-macro_rules! vpl {
-    ($inst:expr) => {
-        concat!($inst, ", [{p:e}]")
-    };
-}
-#[cfg(target_pointer_width = "64")]
-macro_rules! vpl {
-    ($inst:expr) => {
-        concat!($inst, ", [{p}]")
-    };
-}
-#[cfg(target_pointer_width = "32")]
-macro_rules! vps {
-    ($inst1:expr, $inst2:expr) => {
-        concat!($inst1, " [{p:e}]", $inst2)
-    };
-}
-#[cfg(target_pointer_width = "64")]
-macro_rules! vps {
-    ($inst1:expr, $inst2:expr) => {
-        concat!($inst1, " [{p}]", $inst2)
-    };
-}
-
-pub(crate) use {vpl, vps};
-
 #[cfg(test)]
 use stdarch_test::assert_instr;
 
@@ -28013,8 +27982,8 @@ pub unsafe fn _mm_mask_testn_epi64_mask(k: __mmask8, a: __m128i, b: __m128i) ->
 #[allow(clippy::cast_ptr_alignment)]
 pub unsafe fn _mm512_stream_ps(mem_addr: *mut f32, a: __m512) {
     crate::arch::asm!(
-        "vmovntps [{mem_addr}], {a}",
-        mem_addr = in(reg) mem_addr,
+        vps!("vmovntps", ",{a}"),
+        p = in(reg) mem_addr,
         a = in(zmm_reg) a,
         options(nostack, preserves_flags),
     );
@@ -28039,8 +28008,8 @@ pub unsafe fn _mm512_stream_ps(mem_addr: *mut f32, a: __m512) {
 #[allow(clippy::cast_ptr_alignment)]
 pub unsafe fn _mm512_stream_pd(mem_addr: *mut f64, a: __m512d) {
     crate::arch::asm!(
-        "vmovntpd [{mem_addr}], {a}",
-        mem_addr = in(reg) mem_addr,
+        vps!("vmovntpd", ",{a}"),
+        p = in(reg) mem_addr,
         a = in(zmm_reg) a,
         options(nostack, preserves_flags),
     );
@@ -28065,13 +28034,32 @@ pub unsafe fn _mm512_stream_pd(mem_addr: *mut f64, a: __m512d) {
 #[allow(clippy::cast_ptr_alignment)]
 pub unsafe fn _mm512_stream_si512(mem_addr: *mut i32, a: __m512i) {
     crate::arch::asm!(
-        "vmovntdq [{mem_addr}], {a}",
-        mem_addr = in(reg) mem_addr,
+        vps!("vmovntdq", ",{a}"),
+        p = in(reg) mem_addr,
         a = in(zmm_reg) a,
         options(nostack, preserves_flags),
     );
 }
 
+/// Load 512-bits of integer data from memory into dst using a non-temporal memory hint. mem_addr
+/// must be aligned on a 64-byte boundary or a general-protection exception may be generated. To
+/// minimize caching, the data is flagged as non-temporal (unlikely to be used again soon)
+///
+/// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm512_stream_load_si256)
+#[inline]
+#[target_feature(enable = "avx512f")]
+#[unstable(feature = "stdarch_x86_avx512", issue = "111137")]
+pub unsafe fn _mm512_stream_load_si512(mem_addr: *const __m512i) -> __m512i {
+    let dst: __m512i;
+    crate::arch::asm!(
+        vpl!("vmovntdqa {a}"),
+        a = out(zmm_reg) dst,
+        p = in(reg) mem_addr,
+        options(pure, readonly, nostack, preserves_flags),
+    );
+    dst
+}
+
 /// Sets packed 32-bit integers in `dst` with the supplied values.
 ///
 /// [Intel's documentation]( https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm512_set_ps&expand=4931)
@@ -54510,6 +54498,13 @@ mod tests {
         }
     }
 
+    #[simd_test(enable = "avx512f")]
+    unsafe fn test_mm512_stream_load_si512() {
+        let a = _mm512_set_epi64(1, 2, 3, 4, 5, 6, 7, 8);
+        let r = _mm512_stream_load_si512(core::ptr::addr_of!(a) as *const _);
+        assert_eq_m512i(a, r);
+    }
+
     #[simd_test(enable = "avx512f")]
     unsafe fn test_mm512_reduce_add_epi32() {
         let a = _mm512_set1_epi32(1);

crates/core_arch/src/x86/macros.rs

@@ -57,3 +57,33 @@ macro_rules! assert_approx_eq {
         );
     }};
 }
+
+// x86-32 wants to use a 32-bit address size, but asm! defaults to using the full
+// register name (e.g. rax). We have to explicitly override the placeholder to
+// use the 32-bit register name in that case.
+
+#[cfg(target_pointer_width = "32")]
+macro_rules! vpl {
+    ($inst:expr) => {
+        concat!($inst, ", [{p:e}]")
+    };
+}
+#[cfg(target_pointer_width = "64")]
+macro_rules! vpl {
+    ($inst:expr) => {
+        concat!($inst, ", [{p}]")
+    };
+}
+
+#[cfg(target_pointer_width = "32")]
+macro_rules! vps {
+    ($inst1:expr, $inst2:expr) => {
+        concat!($inst1, " [{p:e}]", $inst2)
+    };
+}
+#[cfg(target_pointer_width = "64")]
+macro_rules! vps {
+    ($inst1:expr, $inst2:expr) => {
+        concat!($inst1, " [{p}]", $inst2)
+    };
+}

crates/core_arch/src/x86/sse.rs

@@ -1992,8 +1992,8 @@ extern "C" {
 #[allow(clippy::cast_ptr_alignment)]
 pub unsafe fn _mm_stream_ps(mem_addr: *mut f32, a: __m128) {
     crate::arch::asm!(
-        "movntps [{mem_addr}], {a}",
-        mem_addr = in(reg) mem_addr,
+        vps!("movntps", ",{a}"),
+        p = in(reg) mem_addr,
         a = in(xmm_reg) a,
         options(nostack, preserves_flags),
     );

crates/core_arch/src/x86/sse2.rs

@@ -1312,8 +1312,8 @@ pub unsafe fn _mm_storel_epi64(mem_addr: *mut __m128i, a: __m128i) {
 #[stable(feature = "simd_x86", since = "1.27.0")]
 pub unsafe fn _mm_stream_si128(mem_addr: *mut __m128i, a: __m128i) {
     crate::arch::asm!(
-        "movntdq [{mem_addr}], {a}",
-        mem_addr = in(reg) mem_addr,
+        vps!("movntdq",  ",{a}"),
+        p = in(reg) mem_addr,
         a = in(xmm_reg) a,
         options(nostack, preserves_flags),
     );
@@ -1339,8 +1339,8 @@ pub unsafe fn _mm_stream_si128(mem_addr: *mut __m128i, a: __m128i) {
 #[stable(feature = "simd_x86", since = "1.27.0")]
 pub unsafe fn _mm_stream_si32(mem_addr: *mut i32, a: i32) {
     crate::arch::asm!(
-        "movnti [{mem_addr}], {a:e}", // `:e` for 32bit value
-        mem_addr = in(reg) mem_addr,
+        vps!("movnti", ",{a:e}"), // `:e` for 32bit value
+        p = in(reg) mem_addr,
         a = in(reg) a,
         options(nostack, preserves_flags),
     );
@@ -2542,8 +2542,8 @@ pub unsafe fn _mm_loadl_pd(a: __m128d, mem_addr: *const f64) -> __m128d {
 #[allow(clippy::cast_ptr_alignment)]
 pub unsafe fn _mm_stream_pd(mem_addr: *mut f64, a: __m128d) {
     crate::arch::asm!(
-        "movntpd [{mem_addr}], {a}",
-        mem_addr = in(reg) mem_addr,
+        vps!("movntpd", ",{a}"),
+        p = in(reg) mem_addr,
         a = in(xmm_reg) a,
         options(nostack, preserves_flags),
     );

crates/core_arch/src/x86/sse41.rs

@@ -1154,9 +1154,9 @@ pub unsafe fn _mm_test_mix_ones_zeros(a: __m128i, mask: __m128i) -> i32 {
 pub unsafe fn _mm_stream_load_si128(mem_addr: *const __m128i) -> __m128i {
     let dst: __m128i;
     crate::arch::asm!(
-        "movntdqa {a}, [{mem_addr}]",
+        vpl!("movntdqa {a}"),
         a = out(xmm_reg) dst,
-        mem_addr = in(reg) mem_addr,
+        p = in(reg) mem_addr,
         options(pure, readonly, nostack, preserves_flags),
     );
     dst

crates/core_arch/src/x86_64/sse2.rs

@@ -79,8 +79,8 @@ pub unsafe fn _mm_cvttsd_si64x(a: __m128d) -> i64 {
 #[stable(feature = "simd_x86", since = "1.27.0")]
 pub unsafe fn _mm_stream_si64(mem_addr: *mut i64, a: i64) {
     crate::arch::asm!(
-        "movnti [{mem_addr}], {a}",
-        mem_addr = in(reg) mem_addr,
+        "movnti [{p}], {a}",
+        p = in(reg) mem_addr,
         a = in(reg) a,
         options(nostack, preserves_flags),
     );