Add x86 complex encodings for SIMD load-extend instructions

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

@@ -1607,8 +1607,11 @@ fn define_simd(
     let sadd_sat = shared.by_name("sadd_sat");
     let scalar_to_vector = shared.by_name("scalar_to_vector");
     let sload8x8 = shared.by_name("sload8x8");
+    let sload8x8_complex = shared.by_name("sload8x8_complex");
     let sload16x4 = shared.by_name("sload16x4");
+    let sload16x4_complex = shared.by_name("sload16x4_complex");
     let sload32x2 = shared.by_name("sload32x2");
+    let sload32x2_complex = shared.by_name("sload32x2_complex");
     let spill = shared.by_name("spill");
     let sqrt = shared.by_name("sqrt");
     let sshr_imm = shared.by_name("sshr_imm");
@@ -1617,8 +1620,11 @@ fn define_simd(
     let store_complex = shared.by_name("store_complex");
     let uadd_sat = shared.by_name("uadd_sat");
     let uload8x8 = shared.by_name("uload8x8");
+    let uload8x8_complex = shared.by_name("uload8x8_complex");
     let uload16x4 = shared.by_name("uload16x4");
+    let uload16x4_complex = shared.by_name("uload16x4_complex");
     let uload32x2 = shared.by_name("uload32x2");
+    let uload32x2_complex = shared.by_name("uload32x2_complex");
     let ushr_imm = shared.by_name("ushr_imm");
     let usub_sat = shared.by_name("usub_sat");
     let vconst = shared.by_name("vconst");
@@ -1983,6 +1989,35 @@ fn define_simd(
         }
     }
 
+    // SIMD load extend (complex addressing)
+    let is_load_complex_length_two =
+        InstructionPredicate::new_length_equals(&*formats.load_complex, 2);
+    for (inst, opcodes) in &[
+        (uload8x8_complex, &PMOVZXBW),
+        (uload16x4_complex, &PMOVZXWD),
+        (uload32x2_complex, &PMOVZXDQ),
+        (sload8x8_complex, &PMOVSXBW),
+        (sload16x4_complex, &PMOVSXWD),
+        (sload32x2_complex, &PMOVSXDQ),
+    ] {
+        for recipe in &[
+            rec_fldWithIndex,
+            rec_fldWithIndexDisp8,
+            rec_fldWithIndexDisp32,
+        ] {
+            let template = recipe.opcodes(*opcodes);
+            let predicate = |encoding: EncodingBuilder| {
+                encoding
+                    .isa_predicate(use_sse41_simd)
+                    .inst_predicate(is_load_complex_length_two.clone())
+            };
+            e.enc32_func(inst.clone(), template.clone(), predicate);
+            // No infer_rex calculator for these recipes; place REX version first as in enc_x86_64.
+            e.enc64_func(inst.clone(), template.rex(), predicate);
+            e.enc64_func(inst.clone(), template, predicate);
+        }
+    }
+
     // SIMD integer addition
     for (ty, opcodes) in &[(I8, &PADDB), (I16, &PADDW), (I32, &PADDD), (I64, &PADDQ)] {
         let iadd = iadd.bind(vector(*ty, sse_vector_size));

cranelift/codegen/meta/src/shared/instructions.rs

@@ -1172,6 +1172,20 @@ pub(crate) fn define(
         .can_load(true),
     );
 
+    ig.push(
+        Inst::new(
+            "uload8x8_complex",
+            r#"
+        Load an 8x8 vector (64 bits) from memory at ``sum(args) + Offset`` and zero-extend into an 
+        i16x8 vector.
+        "#,
+            &formats.load_complex,
+        )
+        .operands_in(vec![MemFlags, args, Offset])
+        .operands_out(vec![a])
+        .can_load(true),
+    );
+
     ig.push(
         Inst::new(
             "sload8x8",
@@ -1186,6 +1200,20 @@ pub(crate) fn define(
         .can_load(true),
     );
 
+    ig.push(
+        Inst::new(
+            "sload8x8_complex",
+            r#"
+        Load an 8x8 vector (64 bits) from memory at ``sum(args) + Offset`` and sign-extend into an 
+        i16x8 vector.
+        "#,
+            &formats.load_complex,
+        )
+        .operands_in(vec![MemFlags, args, Offset])
+        .operands_out(vec![a])
+        .can_load(true),
+    );
+
     let I32x4 = &TypeVar::new(
         "I32x4",
         "A SIMD vector with exactly 4 lanes of 32-bit values",
@@ -1201,7 +1229,7 @@ pub(crate) fn define(
         Inst::new(
             "uload16x4",
             r#"
-        Load an 16x4 vector (64 bits) from memory at ``p + Offset`` and zero-extend into an i32x4 
+        Load a 16x4 vector (64 bits) from memory at ``p + Offset`` and zero-extend into an i32x4 
         vector.
         "#,
             &formats.load,
@@ -1211,6 +1239,20 @@ pub(crate) fn define(
         .can_load(true),
     );
 
+    ig.push(
+        Inst::new(
+            "uload16x4_complex",
+            r#"
+        Load a 16x4 vector (64 bits) from memory at ``sum(args) + Offset`` and zero-extend into an 
+        i32x4 vector.
+        "#,
+            &formats.load_complex,
+        )
+        .operands_in(vec![MemFlags, args, Offset])
+        .operands_out(vec![a])
+        .can_load(true),
+    );
+
     ig.push(
         Inst::new(
             "sload16x4",
@@ -1225,6 +1267,20 @@ pub(crate) fn define(
         .can_load(true),
     );
 
+    ig.push(
+        Inst::new(
+            "sload16x4_complex",
+            r#"
+        Load a 16x4 vector (64 bits) from memory at ``sum(args) + Offset`` and sign-extend into an 
+        i32x4 vector.
+        "#,
+            &formats.load_complex,
+        )
+        .operands_in(vec![MemFlags, args, Offset])
+        .operands_out(vec![a])
+        .can_load(true),
+    );
+
     let I64x2 = &TypeVar::new(
         "I64x2",
         "A SIMD vector with exactly 2 lanes of 64-bit values",
@@ -1250,6 +1306,20 @@ pub(crate) fn define(
         .can_load(true),
     );
 
+    ig.push(
+        Inst::new(
+            "uload32x2_complex",
+            r#"
+        Load a 32x2 vector (64 bits) from memory at ``sum(args) + Offset`` and zero-extend into an 
+        i64x2 vector.
+        "#,
+            &formats.load_complex,
+        )
+        .operands_in(vec![MemFlags, args, Offset])
+        .operands_out(vec![a])
+        .can_load(true),
+    );
+
     ig.push(
         Inst::new(
             "sload32x2",
@@ -1264,6 +1334,20 @@ pub(crate) fn define(
         .can_load(true),
     );
 
+    ig.push(
+        Inst::new(
+            "sload32x2_complex",
+            r#"
+        Load a 32x2 vector (64 bits) from memory at ``sum(args) + Offset`` and sign-extend into an 
+        i64x2 vector.
+        "#,
+            &formats.load_complex,
+        )
+        .operands_in(vec![MemFlags, args, Offset])
+        .operands_out(vec![a])
+        .can_load(true),
+    );
+
     let x = &Operand::new("x", Mem).with_doc("Value to be stored");
     let a = &Operand::new("a", Mem).with_doc("Value loaded");
     let Offset =

cranelift/codegen/src/isa/aarch64/lower_inst.rs

@@ -1354,11 +1354,17 @@ pub(crate) fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(ctx: &mut C, insn: IRIns
         | Opcode::ScalarToVector
         | Opcode::Swizzle
         | Opcode::Uload8x8
+        | Opcode::Uload8x8Complex
         | Opcode::Sload8x8
+        | Opcode::Sload8x8Complex
         | Opcode::Uload16x4
+        | Opcode::Uload16x4Complex
         | Opcode::Sload16x4
+        | Opcode::Sload16x4Complex
         | Opcode::Uload32x2
-        | Opcode::Sload32x2 => {
+        | Opcode::Uload32x2Complex
+        | Opcode::Sload32x2
+        | Opcode::Sload32x2Complex => {
             // TODO
             panic!("Vector ops not implemented.");
         }