Add x86 SIMD vall_true

In order to implement SIMD's all_true (https://github.com/WebAssembly/simd/blob/master/proposals/simd/SIMD.md#all-lanes-true), we must legalize some instruction (I chose `vall_true`) to a comparison against 0 and a similar reduction as vany_true using `PTEST` and `SETNZ`. Since `icmp` only allows integers but `vall_true` could allow more vector types, `raw_bitcast` is used to convert the lane types into integers, e.g. b32x4 to i32x4. To do so without runtime type-checking, the `raw_bitcast` instruction (which emits no instruction) can now bitcast from any vector type to the same type, e.g. i32x4 to i32x4.

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

@@ -39,6 +39,7 @@ pub(crate) fn define(shared: &mut SharedDefinitions, x86_instructions: &Instruct
     let fmax = insts.by_name("fmax");
     let fmin = insts.by_name("fmin");
     let iadd = insts.by_name("iadd");
+    let icmp = insts.by_name("icmp");
     let iconst = insts.by_name("iconst");
     let imul = insts.by_name("imul");
     let ineg = insts.by_name("ineg");
@@ -62,6 +63,7 @@ pub(crate) fn define(shared: &mut SharedDefinitions, x86_instructions: &Instruct
     let urem = insts.by_name("urem");
     let ushr = insts.by_name("ushr");
     let vconst = insts.by_name("vconst");
+    let vall_true = insts.by_name("vall_true");
     let vany_true = insts.by_name("vany_true");
 
     let x86_bsf = x86_instructions.by_name("x86_bsf");
@@ -459,6 +461,40 @@ pub(crate) fn define(shared: &mut SharedDefinitions, x86_instructions: &Instruct
         );
     }
 
+    // SIMD vall_true
+    let zeroes = constant(vec![0x00; 16]);
+    let eq = Literal::enumerator_for(&imm.intcc, "eq");
+    for ty in ValueType::all_lane_types().filter(allowed_simd_type) {
+        let vall_true = vall_true.bind(vector(ty, sse_vector_size));
+        if ty.is_int() {
+            // In the common case (Wasm's integer-only all_true), we do not require a bitcast.
+            narrow.legalize(
+                def!(y = vall_true(x)),
+                vec![
+                    def!(a = vconst(zeroes)),
+                    def!(c = icmp(eq, x, a)),
+                    def!(d = x86_ptest(c, c)),
+                    def!(y = trueif(eq, d)),
+                ],
+            );
+        } else {
+            // However, to support other types we must bitcast them to an integer vector to use
+            // icmp.
+            let lane_type_as_int = LaneType::int_from_bits(ty.lane_bits() as u16);
+            let raw_bitcast_to_int = raw_bitcast.bind(vector(lane_type_as_int, sse_vector_size));
+            narrow.legalize(
+                def!(y = vall_true(x)),
+                vec![
+                    def!(a = vconst(zeroes)),
+                    def!(b = raw_bitcast_to_int(x)),
+                    def!(c = icmp(eq, b, a)),
+                    def!(d = x86_ptest(c, c)),
+                    def!(y = trueif(eq, d)),
+                ],
+            );
+        }
+    }
+
     narrow.custom_legalize(shuffle, "convert_shuffle");
     narrow.custom_legalize(extractlane, "convert_extractlane");
     narrow.custom_legalize(insertlane, "convert_insertlane");

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

@@ -1639,6 +1639,20 @@ pub(crate) fn define(
         .operands_out(vec![s]),
     );
 
+    ig.push(
+        Inst::new(
+            "vall_true",
+            r#"
+        Reduce a vector to a scalar boolean.
+
+        Return a scalar boolean true if all lanes in ``i`` are non-zero, false otherwise.
+        "#,
+            &formats.unary,
+        )
+        .operands_in(vec![a])
+        .operands_out(vec![s]),
+    );
+
     let x = &operand("x", &TxN.lane_of());
 
     ig.push(

cranelift/filetests/filetests/isa/x86/simd-logical-legalize.clif

@@ -17,3 +17,13 @@ ebb0(v0: b32x4):
     ; nextln: v1 = trueif ne v2
     return v1
 }
+
+function %vall_true_i64x2(i64x2) -> b1 {
+ebb0(v0: i64x2):
+    v1 = vall_true v0
+    ; check: v2 = vconst.i64x2 0x00
+    ; nextln: v3 = icmp eq v0, v2
+    ; nextln: v4 = x86_ptest v3, v3
+    ; nextln: v1 = trueif eq v4
+    return v1
+}

cranelift/filetests/filetests/isa/x86/simd-logical-run.clif

@@ -39,3 +39,21 @@ ebb0:
     return v3
 }
 ; run
+
+function %vall_true_i16x8() -> b1 {
+ebb0:
+    v0 = vconst.i16x8 [1 0 0 0 0 0 0 0]
+    v1 = vall_true v0
+    v2 = bint.i32 v1
+    v3 = icmp_imm eq v2, 0
+    return v3
+}
+; run
+
+function %vall_true_b32x4() -> b1 {
+ebb0:
+    v0 = vconst.b32x4 [true true true true]
+    v1 = vall_true v0
+    return v1
+}
+; run