Merge pull request #3306 from afonso360/trampoline-simd-bools

Implement SIMD bools in trampoline

cranelift/codegen/src/ir/types.rs

@@ -127,6 +127,21 @@ impl Type {
         }
     }
 
+    /// Get a type with the same number of lanes as this type, but with the lanes replaced by
+    /// integers of the same size.
+    ///
+    /// Scalar types follow this same rule, but `b1` is converted into `i8`
+    pub fn as_int(self) -> Self {
+        self.replace_lanes(match self.lane_type() {
+            B1 | B8 => I8,
+            B16 => I16,
+            B32 => I32,
+            B64 => I64,
+            B128 => I128,
+            _ => unimplemented!(),
+        })
+    }
+
     /// Get a type with the same number of lanes as this type, but with lanes that are half the
     /// number of bits.
     pub fn half_width(self) -> Option<Self> {
@@ -193,6 +208,11 @@ impl Type {
         }
     }
 
+    /// Is this a vector boolean type?
+    pub fn is_bool_vector(self) -> bool {
+        self.is_vector() && self.lane_type().is_bool()
+    }
+
     /// Is this a scalar integer type?
     pub fn is_int(self) -> bool {
         match self {
@@ -325,6 +345,19 @@ impl Type {
             Err(()) => panic!("unable to determine architecture pointer width"),
         }
     }
+
+    /// Coerces boolean types (scalar and vectors) into their integer counterparts.
+    /// B1 is converted into I8.
+    pub fn coerce_bools_to_ints(self) -> Self {
+        let is_scalar_bool = self.is_bool();
+        let is_vector_bool = self.is_vector() && self.lane_type().is_bool();
+
+        if is_scalar_bool || is_vector_bool {
+            self.as_int()
+        } else {
+            self
+        }
+    }
 }
 
 impl Display for Type {
@@ -530,4 +563,13 @@ mod tests {
         assert_eq!(I32X4.as_bool_pedantic(), B32X4);
         assert_eq!(I32.as_bool_pedantic(), B32);
     }
+
+    #[test]
+    fn as_int() {
+        assert_eq!(B32X4.as_int(), I32X4);
+        assert_eq!(B8X8.as_int(), I8X8);
+        assert_eq!(B1.as_int(), I8);
+        assert_eq!(B8.as_int(), I8);
+        assert_eq!(B128.as_int(), I128);
+    }
 }

cranelift/filetests/filetests/runtests/simd-valltrue.clif

@@ -3,67 +3,41 @@ test run
 target aarch64
 target x86_64 machinst
 
-; TODO: Refactor this once we support simd bools in the trampoline
+function %vall_true_b8x16(b8x16) -> b1 {
-
+block0(v0: b8x16):
-function %vall_true_b8x16() -> b1, b1, b1 {
-block0:
-    v0 = vconst.b8x16 [false false false false false false false false false false false false false false false false]
     v1 = vall_true v0
-
+    return v1
-    v2 = vconst.b8x16 [true false false false false false false false false false false false false false false false]
-    v3 = vall_true v2
-
-    v4 = vconst.b8x16 [true true true true true true true true true true true true true true true true]
-    v5 = vall_true v4
-
-    return v1, v3, v5
 }
-; run: %vall_true_b8x16() == [false, false, true]
+; run: %vall_true_b8x16([false false false false false false false false false false false false false false false false]) == false
+; run: %vall_true_b8x16([true false false false false false false false false false false false false false false false]) == false
+; run: %vall_true_b8x16([true true true true true true true true true true true true true true true true]) == true
 
 
-function %vall_true_b16x8() -> b1, b1, b1 {
+function %vall_true_b16x8(b16x8) -> b1 {
-block0:
+block0(v0: b16x8):
-    v0 = vconst.b16x8 [false false false false false false false false]
     v1 = vall_true v0
-
+    return v1
-    v2 = vconst.b16x8 [true false false false false false false false]
-    v3 = vall_true v2
-
-    v4 = vconst.b16x8 [true true true true true true true true]
-    v5 = vall_true v4
-
-    return v1, v3, v5
 }
-; run: %vall_true_b16x8() == [false, false, true]
+; run: %vall_true_b16x8([false false false false false false false false]) == false
+; run: %vall_true_b16x8([true false false false false false false false]) == false
+; run: %vall_true_b16x8([true true true true true true true true]) == true
 
 
-function %vall_true_b32x4() -> b1, b1, b1 {
+function %vall_true_b32x4(b32x4) -> b1 {
-block0:
+block0(v0: b32x4):
-    v0 = vconst.b32x4 [false false false false]
     v1 = vall_true v0
-
+    return v1
-    v2 = vconst.b32x4 [true false false false]
-    v3 = vall_true v2
-
-    v4 = vconst.b32x4 [true true true true]
-    v5 = vall_true v4
-
-    return v1, v3, v5
 }
-; run: %vall_true_b32x4() == [false, false, true]
+; run: %vall_true_b32x4([false false false false]) == false
+; run: %vall_true_b32x4([true false false false]) == false
+; run: %vall_true_b32x4([true true true true]) == true
 
 
-function %vall_true_b64x2() -> b1, b1, b1 {
+function %vall_true_b64x2(b64x2) -> b1 {
-block0:
+block0(v0: b64x2):
-    v0 = vconst.b64x2 [false false]
     v1 = vall_true v0
-
+    return v1
-    v2 = vconst.b64x2 [true false]
-    v3 = vall_true v2
-
-    v4 = vconst.b64x2 [true true]
-    v5 = vall_true v4
-
-    return v1, v3, v5
 }
-; run: %vall_true_b64x2() == [false, false, true]
+; run: %vall_true_b64x2([false false]) == false
+; run: %vall_true_b64x2([true false]) == false
+; run: %vall_true_b64x2([true true]) == true

cranelift/filetests/filetests/runtests/simd-vanytrue.clif

@@ -3,67 +3,41 @@ test run
 target x86_64 machinst
 ; TODO: The AArch64 backend is producing an illegal instruction for b64x2. See: #3304
 
-; TODO: Refactor this once we support simd bools in the trampoline
+function %vany_true_b8x16(b8x16) -> b1 {
-
+block0(v0: b8x16):
-function %vany_true_b8x16() -> b1, b1, b1 {
-block0:
-    v0 = vconst.b8x16 [false false false false false false false false false false false false false false false false]
     v1 = vany_true v0
-
+    return v1
-    v2 = vconst.b8x16 [true false false false false false false false false false false false false false false false]
-    v3 = vany_true v2
-
-    v4 = vconst.b8x16 [true true true true true true true true true true true true true true true true]
-    v5 = vany_true v4
-
-    return v1, v3, v5
 }
-; run: %vany_true_b8x16() == [false, true, true]
+; run: %vany_true_b8x16([false false false false false false false false false false false false false false false false]) == false
+; run: %vany_true_b8x16([true false false false false false false false false false false false false false false false]) == true
+; run: %vany_true_b8x16([true true true true true true true true true true true true true true true true]) == true
 
 
-function %vany_true_b16x8() -> b1, b1, b1 {
+function %vany_true_b16x8(b16x8) -> b1 {
-block0:
+block0(v0: b16x8):
-    v0 = vconst.b16x8 [false false false false false false false false]
     v1 = vany_true v0
-
+    return v1
-    v2 = vconst.b16x8 [true false false false false false false false]
-    v3 = vany_true v2
-
-    v4 = vconst.b16x8 [true true true true true true true true]
-    v5 = vany_true v4
-
-    return v1, v3, v5
 }
-; run: %vany_true_b16x8() == [false, true, true]
+; run: %vany_true_b16x8([false false false false false false false false]) == false
+; run: %vany_true_b16x8([true false false false  false false false false]) == true
+; run: %vany_true_b16x8([true true true true true true true true]) == true
 
 
-function %vany_true_b32x4() -> b1, b1, b1 {
+function %vany_true_b32x4(b32x4) -> b1 {
-block0:
+block0(v0: b32x4):
-    v0 = vconst.b32x4 [false false false false]
     v1 = vany_true v0
-
+    return v1
-    v2 = vconst.b32x4 [true false false false]
-    v3 = vany_true v2
-
-    v4 = vconst.b32x4 [true true true true]
-    v5 = vany_true v4
-
-    return v1, v3, v5
 }
-; run: %vany_true_b32x4() == [false, true, true]
+; run: %vany_true_b32x4([false false false false]) == false
+; run: %vany_true_b32x4([true false false false]) == true
+; run: %vany_true_b32x4([true true true true]) == true
 
 
-function %vany_true_b64x2() -> b1, b1, b1 {
+function %vany_true_b64x2(b64x2) -> b1 {
-block0:
+block0(v0: b64x2):
-    v0 = vconst.b64x2 [false false]
     v1 = vany_true v0
-
+    return v1
-    v2 = vconst.b64x2 [true false]
-    v3 = vany_true v2
-
-    v4 = vconst.b64x2 [true true]
-    v5 = vany_true v4
-
-    return v1, v3, v5
 }
-; run: %vany_true_b64x2() == [false, true, true]
+; run: %vany_true_b64x2([false false]) == false
+; run: %vany_true_b64x2([true false]) == true
+; run: %vany_true_b64x2([true true]) == true

cranelift/filetests/src/function_runner.rs

@@ -2,7 +2,7 @@
 use core::mem;
 use cranelift_codegen::binemit::{NullRelocSink, NullStackMapSink, NullTrapSink};
 use cranelift_codegen::data_value::DataValue;
-use cranelift_codegen::ir::{condcodes::IntCC, Function, InstBuilder, Signature, Type};
+use cranelift_codegen::ir::{condcodes::IntCC, Function, InstBuilder, Signature};
 use cranelift_codegen::isa::{BackendVariant, TargetIsa};
 use cranelift_codegen::{ir, settings, CodegenError, Context};
 use cranelift_frontend::{FunctionBuilder, FunctionBuilderContext};
@@ -297,13 +297,8 @@ fn make_trampoline(signature: &ir::Signature, isa: &dyn TargetIsa) -> Function {
         .enumerate()
         .map(|(i, param)| {
             // Calculate the type to load from memory, using integers for booleans (no encodings).
-            let ty = if param.value_type.is_bool() {
+            let ty = param.value_type.coerce_bools_to_ints();
-                Type::int(max(param.value_type.bits(), 8)).expect(
+
-                    "to be able to convert any boolean type to its equal-width integer type",
-                )
-            } else {
-                param.value_type
-            };
             // Load the value.
             let loaded = builder.ins().load(
                 ty,
@@ -311,11 +306,16 @@ fn make_trampoline(signature: &ir::Signature, isa: &dyn TargetIsa) -> Function {
                 values_vec_ptr_val,
                 (i * UnboxedValues::SLOT_SIZE) as i32,
             );
+
             // For booleans, we want to type-convert the loaded integer into a boolean and ensure
             // that we are using the architecture's canonical boolean representation (presumably
             // comparison will emit this).
             if param.value_type.is_bool() {
                 builder.ins().icmp_imm(IntCC::NotEqual, loaded, 0)
+            } else if param.value_type.is_bool_vector() {
+                let zero_constant = builder.func.dfg.constants.insert(vec![0; 16].into());
+                let zero_vec = builder.ins().vconst(ty, zero_constant);
+                builder.ins().icmp(IntCC::NotEqual, loaded, zero_vec)
             } else {
                 loaded
             }
@@ -332,9 +332,8 @@ fn make_trampoline(signature: &ir::Signature, isa: &dyn TargetIsa) -> Function {
     let results = builder.func.dfg.inst_results(call).to_vec();
     for ((i, value), param) in results.iter().enumerate().zip(&signature.returns) {
         // Before storing return values, we convert booleans to their integer representation.
-        let value = if param.value_type.is_bool() {
+        let value = if param.value_type.lane_type().is_bool() {
-            let ty = Type::int(max(param.value_type.bits(), 8))
+            let ty = param.value_type.lane_type().as_int();
-                .expect("to be able to convert any boolean type to its equal-width integer type");
             builder.ins().bint(ty, *value)
         } else {
             *value