Refactor packed moves to use xmm_mov instead of xmm_rm_r

Refactors previous packed move implementation to use xmm_mov
instead of xmm_rm_r which looks to simplify register accounting
during lowering.

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

@@ -1873,18 +1873,6 @@ pub(crate) fn emit(
                 SseOpcode::Pcmpgtw => (LegacyPrefixes::_66, 0x0F65, 2),
                 SseOpcode::Pcmpgtd => (LegacyPrefixes::_66, 0x0F66, 2),
                 SseOpcode::Pcmpgtq => (LegacyPrefixes::_66, 0x0F3837, 3),
-                SseOpcode::Pmovsxbd => (LegacyPrefixes::_66, 0x0F3821, 3),
-                SseOpcode::Pmovsxbw => (LegacyPrefixes::_66, 0x0F3820, 3),
-                SseOpcode::Pmovsxbq => (LegacyPrefixes::_66, 0x0F3822, 3),
-                SseOpcode::Pmovsxwd => (LegacyPrefixes::_66, 0x0F3823, 3),
-                SseOpcode::Pmovsxwq => (LegacyPrefixes::_66, 0x0F3824, 3),
-                SseOpcode::Pmovsxdq => (LegacyPrefixes::_66, 0x0F3825, 3),
-                SseOpcode::Pmovzxbd => (LegacyPrefixes::_66, 0x0F3831, 3),
-                SseOpcode::Pmovzxbw => (LegacyPrefixes::_66, 0x0F3830, 3),
-                SseOpcode::Pmovzxbq => (LegacyPrefixes::_66, 0x0F3832, 3),
-                SseOpcode::Pmovzxwd => (LegacyPrefixes::_66, 0x0F3833, 3),
-                SseOpcode::Pmovzxwq => (LegacyPrefixes::_66, 0x0F3834, 3),
-                SseOpcode::Pmovzxdq => (LegacyPrefixes::_66, 0x0F3835, 3),
                 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

@@ -10,7 +10,7 @@
 //! RUST_BACKTRACE=1 cargo test --features test-programs/test_programs \
 //!   --features experimental_x64 --all --exclude peepmatic --exclude lightbeam \
 //!   --exclude wasmtime-lightbeam --exclude peepmatic-automata --exclude peepmatic-fuzzing \
-//!  --exclude peepmatic-macro -- isa::x64::inst::emit_tests::test_x64_emit
+//!   --exclude peepmatic-macro --exclude wasmtime-wasi-nn -- isa::x64::inst::emit_tests::test_x64_emit
 
 use super::*;
 use crate::isa::test_utils;
@@ -3201,81 +3201,6 @@ fn test_x64_emit() {
         "cvttps2dq %xmm9, %xmm8",
     ));
 
-    // ========================================================
-    // XMM_RM_R: Packed Move
-
-    insns.push((
-        Inst::xmm_rm_r(SseOpcode::Pmovsxbd, RegMem::reg(xmm6), w_xmm8),
-        "66440F3821C6",
-        "pmovsxbd %xmm6, %xmm8",
-    ));
-
-    insns.push((
-        Inst::xmm_rm_r(SseOpcode::Pmovsxbw, RegMem::reg(xmm9), w_xmm10),
-        "66450F3820D1",
-        "pmovsxbw %xmm9, %xmm10",
-    ));
-
-    insns.push((
-        Inst::xmm_rm_r(SseOpcode::Pmovsxbq, RegMem::reg(xmm1), w_xmm1),
-        "660F3822C9",
-        "pmovsxbq %xmm1, %xmm1",
-    ));
-
-    insns.push((
-        Inst::xmm_rm_r(SseOpcode::Pmovsxwd, RegMem::reg(xmm13), w_xmm10),
-        "66450F3823D5",
-        "pmovsxwd %xmm13, %xmm10",
-    ));
-
-    insns.push((
-        Inst::xmm_rm_r(SseOpcode::Pmovsxwq, RegMem::reg(xmm12), w_xmm12),
-        "66450F3824E4",
-        "pmovsxwq %xmm12, %xmm12",
-    ));
-
-    insns.push((
-        Inst::xmm_rm_r(SseOpcode::Pmovsxdq, RegMem::reg(xmm10), w_xmm8),
-        "66450F3825C2",
-        "pmovsxdq %xmm10, %xmm8",
-    ));
-
-    insns.push((
-        Inst::xmm_rm_r(SseOpcode::Pmovzxbd, RegMem::reg(xmm5), w_xmm6),
-        "660F3831F5",
-        "pmovzxbd %xmm5, %xmm6",
-    ));
-
-    insns.push((
-        Inst::xmm_rm_r(SseOpcode::Pmovzxbw, RegMem::reg(xmm5), w_xmm13),
-        "66440F3830ED",
-        "pmovzxbw %xmm5, %xmm13",
-    ));
-
-    insns.push((
-        Inst::xmm_rm_r(SseOpcode::Pmovzxbq, RegMem::reg(xmm10), w_xmm11),
-        "66450F3832DA",
-        "pmovzxbq %xmm10, %xmm11",
-    ));
-
-    insns.push((
-        Inst::xmm_rm_r(SseOpcode::Pmovzxwd, RegMem::reg(xmm2), w_xmm10),
-        "66440F3833D2",
-        "pmovzxwd %xmm2, %xmm10",
-    ));
-
-    insns.push((
-        Inst::xmm_rm_r(SseOpcode::Pmovzxwq, RegMem::reg(xmm7), w_xmm4),
-        "660F3834E7",
-        "pmovzxwq %xmm7, %xmm4",
-    ));
-
-    insns.push((
-        Inst::xmm_rm_r(SseOpcode::Pmovzxdq, RegMem::reg(xmm3), w_xmm4),
-        "660F3835E3",
-        "pmovzxdq %xmm3, %xmm4",
-    ));
-
     // XMM_Mov_R_M: float stores
     insns.push((
         Inst::xmm_mov_r_m(SseOpcode::Movss, xmm15, Amode::imm_reg(128, r12)),
@@ -3288,6 +3213,81 @@ fn test_x64_emit() {
         "movsd   %xmm1, 0(%rsi)",
     ));
 
+    // ========================================================
+    // XMM_MOV: Packed Move
+
+    insns.push((
+        Inst::xmm_mov(SseOpcode::Pmovsxbd, RegMem::reg(xmm6), w_xmm8),
+        "66440F3821C6",
+        "pmovsxbd %xmm6, %xmm8",
+    ));
+
+    insns.push((
+        Inst::xmm_mov(SseOpcode::Pmovsxbw, RegMem::reg(xmm9), w_xmm10),
+        "66450F3820D1",
+        "pmovsxbw %xmm9, %xmm10",
+    ));
+
+    insns.push((
+        Inst::xmm_mov(SseOpcode::Pmovsxbq, RegMem::reg(xmm1), w_xmm1),
+        "660F3822C9",
+        "pmovsxbq %xmm1, %xmm1",
+    ));
+
+    insns.push((
+        Inst::xmm_mov(SseOpcode::Pmovsxwd, RegMem::reg(xmm13), w_xmm10),
+        "66450F3823D5",
+        "pmovsxwd %xmm13, %xmm10",
+    ));
+
+    insns.push((
+        Inst::xmm_mov(SseOpcode::Pmovsxwq, RegMem::reg(xmm12), w_xmm12),
+        "66450F3824E4",
+        "pmovsxwq %xmm12, %xmm12",
+    ));
+
+    insns.push((
+        Inst::xmm_mov(SseOpcode::Pmovsxdq, RegMem::reg(xmm10), w_xmm8),
+        "66450F3825C2",
+        "pmovsxdq %xmm10, %xmm8",
+    ));
+
+    insns.push((
+        Inst::xmm_mov(SseOpcode::Pmovzxbd, RegMem::reg(xmm5), w_xmm6),
+        "660F3831F5",
+        "pmovzxbd %xmm5, %xmm6",
+    ));
+
+    insns.push((
+        Inst::xmm_mov(SseOpcode::Pmovzxbw, RegMem::reg(xmm5), w_xmm13),
+        "66440F3830ED",
+        "pmovzxbw %xmm5, %xmm13",
+    ));
+
+    insns.push((
+        Inst::xmm_mov(SseOpcode::Pmovzxbq, RegMem::reg(xmm10), w_xmm11),
+        "66450F3832DA",
+        "pmovzxbq %xmm10, %xmm11",
+    ));
+
+    insns.push((
+        Inst::xmm_mov(SseOpcode::Pmovzxwd, RegMem::reg(xmm2), w_xmm10),
+        "66440F3833D2",
+        "pmovzxwd %xmm2, %xmm10",
+    ));
+
+    insns.push((
+        Inst::xmm_mov(SseOpcode::Pmovzxwq, RegMem::reg(xmm7), w_xmm4),
+        "660F3834E7",
+        "pmovzxwq %xmm7, %xmm4",
+    ));
+
+    insns.push((
+        Inst::xmm_mov(SseOpcode::Pmovzxdq, RegMem::reg(xmm3), w_xmm4),
+        "660F3835E3",
+        "pmovzxdq %xmm3, %xmm4",
+    ));
+
     // XmmUnary: moves and unary float ops
     insns.push((
         Inst::xmm_unary_rm_r(SseOpcode::Movss, RegMem::reg(xmm13), w_xmm2),

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

@@ -2910,12 +2910,10 @@ fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
                 match op {
                     Opcode::SwidenLow => match (input_ty, output_ty) {
                         (types::I8X16, types::I16X8) => {
-                            ctx.emit(Inst::gen_move(dst, src, output_ty));
+                            ctx.emit(Inst::xmm_mov(SseOpcode::Pmovsxbw, RegMem::reg(src), dst));
-                            ctx.emit(Inst::xmm_rm_r(SseOpcode::Pmovsxbw, RegMem::from(dst), dst));
                         }
                         (types::I16X8, types::I32X4) => {
-                            ctx.emit(Inst::gen_move(dst, src, output_ty));
+                            ctx.emit(Inst::xmm_mov(SseOpcode::Pmovsxwd, RegMem::reg(src), dst));
-                            ctx.emit(Inst::xmm_rm_r(SseOpcode::Pmovsxwd, RegMem::from(dst), dst));
                         }
                         _ => unreachable!(),
                     },
@@ -2929,7 +2927,7 @@ fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
                                 8,
                                 false,
                             ));
-                            ctx.emit(Inst::xmm_rm_r(SseOpcode::Pmovsxbw, RegMem::from(dst), dst));
+                            ctx.emit(Inst::xmm_mov(SseOpcode::Pmovsxbw, RegMem::from(dst), dst));
                         }
                         (types::I16X8, types::I32X4) => {
                             ctx.emit(Inst::gen_move(dst, src, output_ty));
@@ -2940,18 +2938,16 @@ fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
                                 8,
                                 false,
                             ));
-                            ctx.emit(Inst::xmm_rm_r(SseOpcode::Pmovsxwd, RegMem::from(dst), dst));
+                            ctx.emit(Inst::xmm_mov(SseOpcode::Pmovsxwd, RegMem::from(dst), dst));
                         }
                         _ => unreachable!(),
                     },
                     Opcode::UwidenLow => match (input_ty, output_ty) {
                         (types::I8X16, types::I16X8) => {
-                            ctx.emit(Inst::gen_move(dst, src, output_ty));
+                            ctx.emit(Inst::xmm_mov(SseOpcode::Pmovzxbw, RegMem::reg(src), dst));
-                            ctx.emit(Inst::xmm_rm_r(SseOpcode::Pmovzxbw, RegMem::from(dst), dst));
                         }
                         (types::I16X8, types::I32X4) => {
-                            ctx.emit(Inst::gen_move(dst, src, output_ty));
+                            ctx.emit(Inst::xmm_mov(SseOpcode::Pmovzxwd, RegMem::reg(src), dst));
-                            ctx.emit(Inst::xmm_rm_r(SseOpcode::Pmovzxwd, RegMem::from(dst), dst));
                         }
                         _ => unreachable!(),
                     },
@@ -2965,7 +2961,7 @@ fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
                                 8,
                                 false,
                             ));
-                            ctx.emit(Inst::xmm_rm_r(SseOpcode::Pmovzxbw, RegMem::from(dst), dst));
+                            ctx.emit(Inst::xmm_mov(SseOpcode::Pmovzxbw, RegMem::from(dst), dst));
                         }
                         (types::I16X8, types::I32X4) => {
                             ctx.emit(Inst::gen_move(dst, src, output_ty));
@@ -2976,7 +2972,7 @@ fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
                                 8,
                                 false,
                             ));
-                            ctx.emit(Inst::xmm_rm_r(SseOpcode::Pmovzxwd, RegMem::from(dst), dst));
+                            ctx.emit(Inst::xmm_mov(SseOpcode::Pmovzxwd, RegMem::from(dst), dst));
                         }
                         _ => unreachable!(),
                     },