Add comparison operator integer comparisons

Author: Daniel Smith

crates/core_arch/src/x86/avx512f.rs

@@ -220,6 +220,47 @@ pub unsafe fn _mm512_mask_cmpneq_epu64_mask(m: __mmask8, a: __m512i, b: __m512i)
     _mm512_cmpneq_epu64_mask(a, b) & m
 }
 
+/// Compare packed unsigned 64-bit integers in a and b based on the comparison operand specified by op.
+///
+/// [Intel's documentation](https://software.intel.com/sites/landingpage/IntrinsicsGuide/#expand=727,1063,4909,1062,1062,1063&text=_mm512_mask_cmp_epu64)
+#[inline]
+#[target_feature(enable = "avx512f")]
+#[rustc_args_required_const(2)]
+#[cfg_attr(test, assert_instr(vpcmp, op = 0))]
+pub unsafe fn _mm512_cmp_epu64_mask(a: __m512i, b: __m512i, op: _MM_CMPINT_ENUM) -> __mmask8 {
+    let neg_one = -1;
+    macro_rules! call {
+        ($imm3:expr) => {
+            vpcmpuq(a.as_i64x8(), b.as_i64x8(), $imm3, neg_one)
+        };
+    }
+    let r = constify_imm3!(op, call);
+    transmute(r)
+}
+
+/// Compare packed unsigned 64-bit integers in a and b based on the comparison operand specified by op,
+///  using zeromask m (elements are zeroed out when the corresponding mask bit is not set).
+///
+/// [Intel's documentation](https://software.intel.com/sites/landingpage/IntrinsicsGuide/#expand=727,1063,4909,1062,1062,1063&text=_mm512_mask_cmp_epu64)
+#[inline]
+#[target_feature(enable = "avx512f")]
+#[rustc_args_required_const(3)]
+#[cfg_attr(test, assert_instr(vpcmp, op = 0))]
+pub unsafe fn _mm512_mask_cmp_epu64_mask(
+    m: __mmask8,
+    a: __m512i,
+    b: __m512i,
+    op: _MM_CMPINT_ENUM,
+) -> __mmask8 {
+    macro_rules! call {
+        ($imm3:expr) => {
+            vpcmpuq(a.as_i64x8(), b.as_i64x8(), $imm3, m as i8)
+        };
+    }
+    let r = constify_imm3!(op, call);
+    transmute(r)
+}
+
 /// Compare packed signed 64-bit integers in a and b for less-than, and store the results in a mask vector.
 ///
 /// [Intel's documentation](https://software.intel.com/sites/landingpage/IntrinsicsGuide/#expand=727,1063,4909,1062,1062&text=_mm512_cmplt_epi64)
@@ -335,7 +376,7 @@ pub unsafe fn _mm512_cmpneq_epi64_mask(a: __m512i, b: __m512i) -> __mmask8 {
     simd_bitmask::<__m512i, _>(simd_ne(a.as_i64x8(), b.as_i64x8()))
 }
 
-///Compare packed signed 64-bit integers in a and b for inequality, and store the results in a mask vector k
+/// Compare packed signed 64-bit integers in a and b for inequality, and store the results in a mask vector k
 /// using zeromask m (elements are zeroed out when the corresponding mask bit is not set).
 ///
 /// [Intel's documentation](https://software.intel.com/sites/landingpage/IntrinsicsGuide/#expand=727,1063,4909,1062,1062,1063&text=_mm512_mask_cmpneq_epi64)
@@ -346,6 +387,72 @@ pub unsafe fn _mm512_mask_cmpneq_epi64_mask(m: __mmask8, a: __m512i, b: __m512i)
     _mm512_cmpneq_epi64_mask(a, b) & m
 }
 
+/// Compare packed signed 64-bit integers in a and b based on the comparison operand specified by op.
+///
+/// [Intel's documentation](https://software.intel.com/sites/landingpage/IntrinsicsGuide/#expand=727,1063,4909,1062,1062,1063&text=_mm512_mask_cmp_epi64)
+#[inline]
+#[target_feature(enable = "avx512f")]
+#[rustc_args_required_const(2)]
+#[cfg_attr(test, assert_instr(vpcmp, op = 0))]
+pub unsafe fn _mm512_cmp_epi64_mask(a: __m512i, b: __m512i, op: _MM_CMPINT_ENUM) -> __mmask8 {
+    let neg_one = -1;
+    macro_rules! call {
+        ($imm3:expr) => {
+            vpcmpq(a.as_i64x8(), b.as_i64x8(), $imm3, neg_one)
+        };
+    }
+    let r = constify_imm3!(op, call);
+    transmute(r)
+}
+
+/// Compare packed signed 64-bit integers in a and b based on the comparison operand specified by op,
+///  using zeromask m (elements are zeroed out when the corresponding mask bit is not set).
+///
+/// [Intel's documentation](https://software.intel.com/sites/landingpage/IntrinsicsGuide/#expand=727,1063,4909,1062,1062,1063&text=_mm512_mask_cmp_epi64)
+#[inline]
+#[target_feature(enable = "avx512f")]
+#[rustc_args_required_const(3)]
+#[cfg_attr(test, assert_instr(vpcmp, op = 0))]
+pub unsafe fn _mm512_mask_cmp_epi64_mask(
+    m: __mmask8,
+    a: __m512i,
+    b: __m512i,
+    op: _MM_CMPINT_ENUM,
+) -> __mmask8 {
+    macro_rules! call {
+        ($imm3:expr) => {
+            vpcmpq(a.as_i64x8(), b.as_i64x8(), $imm3, m as i8)
+        };
+    }
+    let r = constify_imm3!(op, call);
+    transmute(r)
+}
+
+/// Equal
+pub const _MM_CMPINT_EQ: _MM_CMPINT_ENUM = 0x00;
+/// Less-than
+pub const _MM_CMPINT_LT: _MM_CMPINT_ENUM = 0x01;
+/// Less-than-or-equal
+pub const _MM_CMPINT_LE: _MM_CMPINT_ENUM = 0x02;
+/// False
+pub const _MM_CMPINT_FALSE: _MM_CMPINT_ENUM = 0x03;
+/// Not-equal
+pub const _MM_CMPINT_NE: _MM_CMPINT_ENUM = 0x04;
+/// Not less-than
+pub const _MM_CMPINT_NLT: _MM_CMPINT_ENUM = 0x05;
+/// Not less-than-or-equal
+pub const _MM_CMPINT_NLE: _MM_CMPINT_ENUM = 0x06;
+/// True
+pub const _MM_CMPINT_TRUE: _MM_CMPINT_ENUM = 0x07;
+
+#[allow(improper_ctypes)]
+extern "C" {
+    #[link_name = "llvm.x86.avx512.mask.ucmp.q.512"]
+    fn vpcmpuq(a: i64x8, b: i64x8, op: i32, m: i8) -> i8;
+    #[link_name = "llvm.x86.avx512.mask.cmp.q.512"]
+    fn vpcmpq(a: i64x8, b: i64x8, op: i32, m: i8) -> i8;
+}
+
 #[cfg(test)]
 mod tests {
     use std;

crates/core_arch/src/x86/mod.rs

@@ -350,6 +350,9 @@ pub type __mmask16 = u16;
 #[allow(non_camel_case_types)]
 pub type __mmask8 = u8;
 
+/// The `_MM_CMPINT_ENUM` type used to specify comparison operations in AVX-512 intrinsics.
+pub type _MM_CMPINT_ENUM = i32;
+
 #[cfg(test)]
 mod test;
 #[cfg(test)]

crates/core_arch/src/x86_64/avx512f.rs

@@ -153,6 +153,23 @@ mod tests {
         assert_eq!(r, 0b00110010);
     }
 
+    #[simd_test(enable = "avx512f")]
+    unsafe fn test_mm512_cmp_epu64_mask() {
+        let a = _mm512_set_epi64(0, 1, -1, u64::MAX as i64, i64::MAX, i64::MIN, 100, -100);
+        let b = _mm512_set1_epi64(-1);
+        let m = _mm512_cmp_epu64_mask(a, b, _MM_CMPINT_LT);
+        assert_eq!(m, 0b11001111);
+    }
+
+    #[simd_test(enable = "avx512f")]
+    unsafe fn test_mm512_mask_cmp_epu64_mask() {
+        let a = _mm512_set_epi64(0, 1, -1, u64::MAX as i64, i64::MAX, i64::MIN, 100, -100);
+        let b = _mm512_set1_epi64(-1);
+        let mask = 0b01111010;
+        let r = _mm512_mask_cmp_epu64_mask(mask, a, b, _MM_CMPINT_LT);
+        assert_eq!(r, 0b01001010);
+    }
+
     #[simd_test(enable = "avx512f")]
     unsafe fn test_mm512_cmplt_epi64_mask() {
         let a = _mm512_set_epi64(0, 1, -1, 13, i64::MAX, i64::MIN, 100, -100);
@@ -257,6 +274,23 @@ mod tests {
         assert_eq!(r, 0b00110010)
     }
 
+    #[simd_test(enable = "avx512f")]
+    unsafe fn test_mm512_cmp_epi64_mask() {
+        let a = _mm512_set_epi64(0, 1, -1, 13, i64::MAX, i64::MIN, 100, -100);
+        let b = _mm512_set1_epi64(-1);
+        let m = _mm512_cmp_epi64_mask(a, b, _MM_CMPINT_LT);
+        assert_eq!(m, 0b00000101);
+    }
+
+    #[simd_test(enable = "avx512f")]
+    unsafe fn test_mm512_mask_cmp_epi64_mask() {
+        let a = _mm512_set_epi64(0, 1, -1, 13, i64::MAX, i64::MIN, 100, -100);
+        let b = _mm512_set1_epi64(-1);
+        let mask = 0b01100110;
+        let r = _mm512_mask_cmp_epi64_mask(mask, a, b, _MM_CMPINT_LT);
+        assert_eq!(r, 0b00000100);
+    }
+
     #[simd_test(enable = "avx512f")]
     unsafe fn test_mm512_set_epi64() {
         let r = _mm512_setr_epi64(0, 1, 2, 3, 4, 5, 6, 7);

crates/stdarch-verify/src/lib.rs

@@ -147,6 +147,7 @@ fn to_type(t: &syn::Type) -> proc_macro2::TokenStream {
             "__m512i" => quote! { &M512I },
             "__mmask8" => quote! { &MMASK8 },
             "__mmask16" => quote! { &MMASK16 },
+            "_MM_CMPINT_ENUM" => quote! { &MM_CMPINT_ENUM },
             "__m64" => quote! { &M64 },
             "bool" => quote! { &BOOL },
             "f32" => quote! { &F32 },

crates/stdarch-verify/tests/x86-intel.rs

@@ -55,6 +55,7 @@ static M512I: Type = Type::M512I;
 static M512D: Type = Type::M512D;
 static MMASK8: Type = Type::MMASK8;
 static MMASK16: Type = Type::MMASK16;
+static MM_CMPINT_ENUM: Type = Type::MM_CMPINT_ENUM;
 
 static TUPLE: Type = Type::Tuple;
 static CPUID: Type = Type::CpuidResult;
@@ -79,6 +80,7 @@ enum Type {
     M512I,
     MMASK8,
     MMASK16,
+    MM_CMPINT_ENUM,
     Tuple,
     CpuidResult,
     Never,
@@ -657,6 +659,7 @@ fn equate(t: &Type, intel: &str, intrinsic: &str, is_const: bool) -> Result<(),
 
         (&Type::MMASK8, "__mmask8") => {}
         (&Type::MMASK16, "__mmask16") => {}
+        (&Type::MM_CMPINT_ENUM, "const _MM_CMPINT_ENUM") => require_const()?,
 
         // This is a macro (?) in C which seems to mutate its arguments, but
         // that means that we're taking pointers to arguments in rust