Add i64x2 integer multiplication using AVX512DQ

cranelift/codegen/meta/src/isa/x86/encodings.rs

@@ -1612,6 +1612,7 @@ fn define_simd(
     let x86_ptest = x86.by_name("x86_ptest");
 
     // Shorthands for recipes.
+    let rec_evex_reg_vvvv_rm_128 = r.template("evex_reg_vvvv_rm_128");
     let rec_f_ib = r.template("f_ib");
     let rec_fa = r.template("fa");
     let rec_fa_ib = r.template("fa_ib");
@@ -1647,6 +1648,7 @@ fn define_simd(
     let use_ssse3_simd = settings.predicate_by_name("use_ssse3_simd");
     let use_sse41_simd = settings.predicate_by_name("use_sse41_simd");
     let use_sse42_simd = settings.predicate_by_name("use_sse42_simd");
+    let use_avx512dq_simd = settings.predicate_by_name("use_avx512dq_simd");
 
     // SIMD vector size: eventually multiple vector sizes may be supported but for now only
     // SSE-sized vectors are available.
@@ -1927,6 +1929,16 @@ fn define_simd(
         e.enc_32_64_maybe_isap(imul, rec_fa.opcodes(opcodes), *isap);
     }
 
+    // SIMD integer multiplication for I64x2 using a AVX512.
+    {
+        let imul = imul.bind(vector(I64, sse_vector_size));
+        e.enc_32_64_maybe_isap(
+            imul,
+            rec_evex_reg_vvvv_rm_128.opcodes(&PMULLQ).w(),
+            Some(use_avx512dq_simd), // TODO need an OR predicate to join with AVX512VL
+        );
+    }
+
     // SIMD integer average with rounding.
     for (ty, opcodes) in &[(I8, &PAVGB[..]), (I16, &PAVGW[..])] {
         let avgr = avg_round.bind(vector(*ty, sse_vector_size));

cranelift/codegen/meta/src/isa/x86/opcodes.rs

@@ -421,6 +421,10 @@ pub static PMULLW: [u8; 3] = [0x66, 0x0f, 0xd5];
 /// bits of each product in xmm1 (SSE4.1).
 pub static PMULLD: [u8; 4] = [0x66, 0x0f, 0x38, 0x40];
 
+/// Multiply the packed quadword signed integers in xmm2 and xmm3/m128 and store the low 64
+/// bits of each product in xmm1 (AVX512VL/DQ). Requires an EVEX encoding.
+pub static PMULLQ: [u8; 4] = [0x66, 0x0f, 0x38, 0x40];
+
 /// Pop top of stack into r{16,32,64}; increment stack pointer.
 pub static POP_REG: [u8; 1] = [0x58];
 

filetests/isa/x86/simd-avx512-arithmetic-binemit.clif

@@ -0,0 +1,17 @@
+test binemit
+set enable_simd
+target x86_64 skylake has_avx512dq=true
+
+function %imul_i64x2() {
+block0:
+    [-, %xmm1]    v0 = vconst.i64x2 [1 2]
+    [-, %xmm2]    v1 = vconst.i64x2 [2 2]
+    [-, %xmm19]   v2 = imul v0, v1 ; bin: 62 e2 f5 08 40 da
+    ; 62, mandatory EVEX prefix
+    ; e2 = 1110 0010, R, X, B are unset (inverted) while R' is set (MSB in %xmm19); mm is set to 0F38
+    ; f5 = 1111 0101, W is set (64-bit op), vvvv set to 1 (inverted), bit 2 always set, pp set to 01
+    ; 08 = 0000 1000, everything, LL' indicates 128-bit, V' is unset (inverted, %xmm1 has MSB of 0)
+    ; 40, opcode (correct)
+    ; da = 1100 1010, ModR/M byte using 0b011 from %xmm19 in reg and 0b010 from %xmm2 in r/m
+    return
+}