Fix validating wasm stores of boolean vector results (#3202)

Previously cranelift's wasm code generator would emit a raw `store`
instruction for all wasm types, regardless of what the cranelift operand
type was. Cranelift's `store` instruction, however, isn't valid for
boolean vector types. This commit fixes this issue by inserting a
bitcast specifically for the store instruction if a boolean vector type
is being stored, continuing to avoid the bitcast for all other vector types.

Closes #3099

Author: alexcrichton

cranelift/wasm/src/code_translator.rs

@@ -2407,8 +2407,16 @@ fn translate_store<FE: FuncEnvironment + ?Sized>(
     state: &mut FuncTranslationState,
     environ: &mut FE,
 ) -> WasmResult<()> {
-    let val = state.pop1();
-    let val_ty = builder.func.dfg.value_type(val);
+    let mut val = state.pop1();
+    let mut val_ty = builder.func.dfg.value_type(val);
+
+    // Boolean-vector types don't validate with a `store` instruction, so
+    // bitcast them to a vector type which is compatible with the store
+    // instruction.
+    if val_ty.is_vector() && val_ty.lane_type().is_bool() {
+        val = builder.ins().raw_bitcast(I8X16, val);
+        val_ty = I8X16;
+    }
 
     let (flags, base, offset) =
         prepare_addr(memarg, mem_op_size(opcode, val_ty), builder, state, environ)?;

cranelift/wasm/wasmtests/simd-store.wat

@@ -0,0 +1,83 @@
+(module
+  (func (param v128)
+    (v128.store (i32.const 0) (i8x16.eq (local.get 0) (local.get 0))))
+  (func (param v128)
+    (v128.store (i32.const 0) (i16x8.eq (local.get 0) (local.get 0))))
+  (func (param v128)
+    (v128.store (i32.const 0) (i32x4.eq (local.get 0) (local.get 0))))
+  (func (param v128)
+    (v128.store (i32.const 0) (i64x2.eq (local.get 0) (local.get 0))))
+
+  (func (param v128)
+    (v128.store (i32.const 0) (i8x16.ne (local.get 0) (local.get 0))))
+  (func (param v128)
+    (v128.store (i32.const 0) (i16x8.ne (local.get 0) (local.get 0))))
+  (func (param v128)
+    (v128.store (i32.const 0) (i32x4.ne (local.get 0) (local.get 0))))
+  (func (param v128)
+    (v128.store (i32.const 0) (i64x2.ne (local.get 0) (local.get 0))))
+
+  (func (param v128)
+    (v128.store (i32.const 0) (i8x16.lt_s (local.get 0) (local.get 0))))
+  (func (param v128)
+    (v128.store (i32.const 0) (i16x8.lt_s (local.get 0) (local.get 0))))
+  (func (param v128)
+    (v128.store (i32.const 0) (i32x4.lt_s (local.get 0) (local.get 0))))
+  (func (param v128)
+    (v128.store (i32.const 0) (i64x2.lt_s (local.get 0) (local.get 0))))
+
+  (func (param v128)
+    (v128.store (i32.const 0) (i8x16.lt_u (local.get 0) (local.get 0))))
+  (func (param v128)
+    (v128.store (i32.const 0) (i16x8.lt_u (local.get 0) (local.get 0))))
+  (func (param v128)
+    (v128.store (i32.const 0) (i32x4.lt_u (local.get 0) (local.get 0))))
+
+  (func (param v128)
+    (v128.store (i32.const 0) (i8x16.gt_s (local.get 0) (local.get 0))))
+  (func (param v128)
+    (v128.store (i32.const 0) (i16x8.gt_s (local.get 0) (local.get 0))))
+  (func (param v128)
+    (v128.store (i32.const 0) (i32x4.gt_s (local.get 0) (local.get 0))))
+  (func (param v128)
+    (v128.store (i32.const 0) (i64x2.gt_s (local.get 0) (local.get 0))))
+
+  (func (param v128)
+    (v128.store (i32.const 0) (i8x16.gt_u (local.get 0) (local.get 0))))
+  (func (param v128)
+    (v128.store (i32.const 0) (i16x8.gt_u (local.get 0) (local.get 0))))
+  (func (param v128)
+    (v128.store (i32.const 0) (i32x4.gt_u (local.get 0) (local.get 0))))
+
+  (func (param v128)
+    (v128.store (i32.const 0) (f32x4.eq (local.get 0) (local.get 0))))
+  (func (param v128)
+    (v128.store (i32.const 0) (f64x2.eq (local.get 0) (local.get 0))))
+
+  (func (param v128)
+    (v128.store (i32.const 0) (f32x4.ne (local.get 0) (local.get 0))))
+  (func (param v128)
+    (v128.store (i32.const 0) (f64x2.ne (local.get 0) (local.get 0))))
+
+  (func (param v128)
+    (v128.store (i32.const 0) (f32x4.lt (local.get 0) (local.get 0))))
+  (func (param v128)
+    (v128.store (i32.const 0) (f64x2.lt (local.get 0) (local.get 0))))
+
+  (func (param v128)
+    (v128.store (i32.const 0) (f32x4.le (local.get 0) (local.get 0))))
+  (func (param v128)
+    (v128.store (i32.const 0) (f64x2.le (local.get 0) (local.get 0))))
+
+  (func (param v128)
+    (v128.store (i32.const 0) (f32x4.gt (local.get 0) (local.get 0))))
+  (func (param v128)
+    (v128.store (i32.const 0) (f64x2.gt (local.get 0) (local.get 0))))
+
+  (func (param v128)
+    (v128.store (i32.const 0) (f32x4.ge (local.get 0) (local.get 0))))
+  (func (param v128)
+    (v128.store (i32.const 0) (f64x2.ge (local.get 0) (local.get 0))))
+
+  (memory 0)
+)