[SIMD][x86_64] Add encoding for PMADDWD (#2530)

* [SIMD][x86_64] Add encoding for PMADDWD

* also for "experimental_x64"

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

@@ -1691,6 +1691,7 @@ fn define_simd(
     let usub_sat = shared.by_name("usub_sat");
     let vconst = shared.by_name("vconst");
     let vselect = shared.by_name("vselect");
+    let widening_pairwise_dot_product_s = shared.by_name("widening_pairwise_dot_product_s");
     let x86_cvtt2si = x86.by_name("x86_cvtt2si");
     let x86_insertps = x86.by_name("x86_insertps");
     let x86_fmax = x86.by_name("x86_fmax");
@@ -2213,6 +2214,9 @@ fn define_simd(
     // SIMD multiplication with lane expansion.
     e.enc_both_inferred(x86_pmuludq, rec_fa.opcodes(&PMULUDQ));
 
+    // SIMD multiplication and add adjacent pairs, from SSE2.
+    e.enc_both_inferred(widening_pairwise_dot_product_s, rec_fa.opcodes(&PMADDWD));
+
     // SIMD integer multiplication for I64x2 using a AVX512.
     {
         e.enc_32_64_maybe_isap(

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

@@ -508,6 +508,9 @@ pub static VPMULLQ: [u8; 4] = [0x66, 0x0f, 0x38, 0x40];
 /// in xmm2/m128, and store the quadword results in xmm1 (SSE2).
 pub static PMULUDQ: [u8; 3] = [0x66, 0x0f, 0xf4];
 
+/// Multiply the packed word integers, add adjacent doubleword results.
+pub static PMADDWD: [u8; 3] = [0x66, 0x0f, 0xf5];
+
 /// Pop top of stack into r{16,32,64}; increment stack pointer.
 pub static POP_REG: [u8; 1] = [0x58];
 

cranelift/codegen/src/isa/x64/inst/args.rs

@@ -498,6 +498,7 @@ pub enum SseOpcode {
     Pinsrb,
     Pinsrw,
     Pinsrd,
+    Pmaddwd,
     Pmaxsb,
     Pmaxsw,
     Pmaxsd,
@@ -661,6 +662,7 @@ impl SseOpcode {
             | SseOpcode::Pcmpgtd
             | SseOpcode::Pextrw
             | SseOpcode::Pinsrw
+            | SseOpcode::Pmaddwd
             | SseOpcode::Pmaxsw
             | SseOpcode::Pmaxub
             | SseOpcode::Pminsw
@@ -842,6 +844,7 @@ impl fmt::Debug for SseOpcode {
             SseOpcode::Pinsrb => "pinsrb",
             SseOpcode::Pinsrw => "pinsrw",
             SseOpcode::Pinsrd => "pinsrd",
+            SseOpcode::Pmaddwd => "pmaddwd",
             SseOpcode::Pmaxsb => "pmaxsb",
             SseOpcode::Pmaxsw => "pmaxsw",
             SseOpcode::Pmaxsd => "pmaxsd",

cranelift/codegen/src/isa/x64/inst/emit.rs

@@ -1873,6 +1873,7 @@ pub(crate) fn emit(
                 SseOpcode::Pcmpgtw => (LegacyPrefixes::_66, 0x0F65, 2),
                 SseOpcode::Pcmpgtd => (LegacyPrefixes::_66, 0x0F66, 2),
                 SseOpcode::Pcmpgtq => (LegacyPrefixes::_66, 0x0F3837, 3),
+                SseOpcode::Pmaddwd => (LegacyPrefixes::_66, 0x0FF5, 2),
                 SseOpcode::Pmaxsb => (LegacyPrefixes::_66, 0x0F383C, 3),
                 SseOpcode::Pmaxsw => (LegacyPrefixes::_66, 0x0FEE, 2),
                 SseOpcode::Pmaxsd => (LegacyPrefixes::_66, 0x0F383D, 3),

cranelift/codegen/src/isa/x64/inst/emit_tests.rs

@@ -3067,6 +3067,12 @@ fn test_x64_emit() {
         "pmuludq %xmm8, %xmm9",
     ));
 
+    insns.push((
+        Inst::xmm_rm_r(SseOpcode::Pmaddwd, RegMem::reg(xmm8), w_xmm1),
+        "66410FF5C8",
+        "pmaddwd %xmm8, %xmm1",
+    ));
+
     insns.push((
         Inst::xmm_rm_r(SseOpcode::Pmaxsb, RegMem::reg(xmm15), w_xmm6),
         "66410F383CF7",

cranelift/codegen/src/isa/x64/lower.rs

@@ -2235,6 +2235,24 @@ fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
             }
         }
 
+        Opcode::WideningPairwiseDotProductS => {
+            let lhs = put_input_in_reg(ctx, inputs[0]);
+            let rhs = input_to_reg_mem(ctx, inputs[1]);
+            let dst = get_output_reg(ctx, outputs[0]);
+            let ty = ty.unwrap();
+
+            ctx.emit(Inst::gen_move(dst, lhs, ty));
+
+            if ty == types::I32X4 {
+                ctx.emit(Inst::xmm_rm_r(SseOpcode::Pmaddwd, rhs, dst));
+            } else {
+                panic!(
+                    "Opcode::WideningPairwiseDotProductS: unsupported laneage: {:?}",
+                    ty
+                );
+            }
+        }
+
         Opcode::Fadd | Opcode::Fsub | Opcode::Fmul | Opcode::Fdiv => {
             let lhs = put_input_in_reg(ctx, inputs[0]);
             let rhs = input_to_reg_mem(ctx, inputs[1]);

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

@@ -108,3 +108,9 @@ block0(v0: i64x2 [%xmm3], v1: i64x2 [%xmm5]):
 [-, %xmm3]    v2 = x86_pmuludq v0, v1      ; bin: 66 0f f4 dd
     return v2
 }
+
+function %pmaddwd(i16x8, i16x8) -> i32x4 {
+block0(v0: i16x8 [%xmm8], v1: i16x8 [%xmm9]):
+[-, %xmm8]    v2 = widening_pairwise_dot_product_s v0, v1      ; bin: 66 45 0f f5 c1
+    return v2
+}