Infer REX prefix for SIMD load instruction

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

@@ -1835,7 +1835,10 @@ fn define_simd(
 
         // Load
         let bound_load = load.bind(vector(ty, sse_vector_size)).bind(Any);
-        e.enc_32_64(bound_load.clone(), rec_fld.opcodes(&MOVUPS_LOAD));
+        e.enc_32_64(
+            bound_load.clone(),
+            rec_fld.opcodes(&MOVUPS_LOAD).infer_rex(),
+        );
         e.enc_32_64(bound_load.clone(), rec_fldDisp8.opcodes(&MOVUPS_LOAD));
         e.enc_32_64(bound_load, rec_fldDisp32.opcodes(&MOVUPS_LOAD));
 

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

@@ -2002,7 +2002,7 @@ pub(crate) fn define<'shared>(
         );
 
         // XX /r float load with no offset.
-        recipes.add_template_recipe(
+        recipes.add_template_inferred(
             EncodingRecipeBuilder::new("fld", &formats.load, 1)
                 .operands_in(vec![gpr])
                 .operands_out(vec![fpr])
@@ -2026,6 +2026,7 @@ pub(crate) fn define<'shared>(
                         }
                     "#,
                 ),
+            "size_plus_maybe_sib_or_offset_for_inreg_0_plus_rex_prefix_for_inreg0_outreg0",
         );
 
         let has_small_offset =

cranelift/codegen/src/isa/x86/enc_tables.rs

@@ -139,6 +139,23 @@ fn size_plus_maybe_sib_or_offset_inreg1_plus_rex_prefix_for_inreg0_inreg1(
         + if needs_rex { 1 } else { 0 }
 }
 
+/// Calculates the size while inferring if the first input register (inreg0) and first output
+/// register (outreg0) require a dynamic REX and if the first input register (inreg0) requires a
+/// SIB or offset.
+fn size_plus_maybe_sib_or_offset_for_inreg_0_plus_rex_prefix_for_inreg0_outreg0(
+    sizing: &RecipeSizing,
+    enc: Encoding,
+    inst: Inst,
+    divert: &RegDiversions,
+    func: &Function,
+) -> u8 {
+    let needs_rex = (EncodingBits::from(enc.bits()).rex_w() != 0)
+        || test_input(0, inst, divert, func, is_extended_reg)
+        || test_result(0, inst, divert, func, is_extended_reg);
+    size_plus_maybe_sib_or_offset_for_inreg_0(sizing, enc, inst, divert, func)
+        + if needs_rex { 1 } else { 0 }
+}
+
 /// Infers whether a dynamic REX prefix will be emitted, for use with one input reg.
 ///
 /// A REX prefix is known to be emitted if either:

cranelift/filetests/filetests/isa/x86/binary64.clif

@@ -1683,7 +1683,10 @@ block0:
 function %V128() {
 block0:
     [-,%r10]            v3 = iconst.i64 0x2102_0304_f1f2_f3f4  ; bin: 49 ba 21020304f1f2f3f4
-    [-, %xmm9]          v4 = vconst.i32x4 [0 1 2 3]  ; bin: 44 0f 10 0d 00000005 PCRelRodata4(23)
+    [-, %xmm9]          v4 = vconst.i32x4 [0 1 2 3]  ; bin: 44 0f 10 0d 0000000f PCRelRodata4(33)
                         store v4, v3  ; bin: heap_oob 45 0f 11 0a
+
+    [-, %r11]           v5 = iconst.i64 0x1234
+    [-, %xmm2]          v6 = load.i32x4 v5  ; bin: heap_oob 41 0f 10 13
                         return
 }