pulley: implement vector shuffle (bytecodealliance#9910)

* impl shuffle * raise inst size limit
eduardomourar · Dec 30, 2024 · 8d1c6c3 · 8d1c6c3
1 parent 6a95189
commit 8d1c6c3
Show file tree

Hide file tree

Showing 5 changed files with 26 additions and 5 deletions.
diff --git a/cranelift/codegen/src/isa/pulley_shared/inst/mod.rs b/cranelift/codegen/src/isa/pulley_shared/inst/mod.rs
@@ -507,8 +507,11 @@ where
     }
 
     fn worst_case_size() -> CodeOffset {
-        // `Vconst128 { dst, imm }` is 20 bytes (3 byte opcode + dst + 16-byte imm)
-        20
+        // `VShuffle { dst, src1, src2, imm }` is 22 bytes:
+        // 3-byte opcode
+        // dst, src1, src2
+        // 16-byte immediate
+        22
     }
 
     fn ref_type_regclass(_settings: &settings::Flags) -> RegClass {

diff --git a/cranelift/codegen/src/isa/pulley_shared/lower.isle b/cranelift/codegen/src/isa/pulley_shared/lower.isle
@@ -1396,6 +1396,10 @@
 (rule (lower (scalar_to_vector a @ (value_type $F64)))
   (pulley_vinsertf64 (pulley_vconst128 0) a 0))
 
+;;;; Rules for `shuffle` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule (lower (has_type $I8X16 (shuffle a b (u128_from_immediate mask))))
+  (pulley_vshuffle a b mask))
 
 ;;;; Rules for `swizzle` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 

diff --git a/crates/wast-util/src/lib.rs b/crates/wast-util/src/lib.rs
@@ -404,11 +404,9 @@ impl WastTest {
                 "misc_testsuite/simd/canonicalize-nan.wast",
                 "misc_testsuite/simd/issue_3327_bnot_lowering.wast",
                 "misc_testsuite/simd/v128-select.wast",
-                "spec_testsuite/proposals/annotations/simd_lane.wast",
                 "spec_testsuite/proposals/relaxed-simd/i32x4_relaxed_trunc.wast",
                 "spec_testsuite/proposals/relaxed-simd/relaxed_dot_product.wast",
                 "spec_testsuite/proposals/relaxed-simd/relaxed_madd_nmadd.wast",
-                "spec_testsuite/proposals/memory64/simd_lane.wast",
                 "spec_testsuite/proposals/memory64/relaxed_madd_nmadd.wast",
                 "spec_testsuite/proposals/memory64/relaxed_dot_product.wast",
                 "spec_testsuite/proposals/memory64/i32x4_relaxed_trunc.wast",
@@ -423,7 +421,6 @@ impl WastTest {
                 "spec_testsuite/simd_i32x4_extadd_pairwise_i16x8.wast",
                 "spec_testsuite/simd_i32x4_trunc_sat_f32x4.wast",
                 "spec_testsuite/simd_i32x4_trunc_sat_f64x2.wast",
-                "spec_testsuite/simd_lane.wast",
                 "spec_testsuite/simd_load.wast",
                 "spec_testsuite/simd_splat.wast",
             ];

diff --git a/pulley/src/interp.rs b/pulley/src/interp.rs
@@ -4538,6 +4538,20 @@ impl ExtendedOpVisitor for Interpreter<'_> {
         ControlFlow::Continue(())
     }
 
+    fn vshuffle(&mut self, dst: VReg, src1: VReg, src2: VReg, mask: u128) -> ControlFlow<Done> {
+        let a = self.state[src1].get_u8x16();
+        let b = self.state[src2].get_u8x16();
+        let result = mask.to_le_bytes().map(|m| {
+            if m < 16 {
+                a[m as usize]
+            } else {
+                b[m as usize - 16]
+            }
+        });
+        self.state[dst].set_u8x16(result);
+        ControlFlow::Continue(())
+    }
+
     fn vswizzlei8x16(&mut self, operands: BinaryOperands<VReg>) -> ControlFlow<Done> {
         let src1 = self.state[operands.src1].get_i8x16();
         let src2 = self.state[operands.src2].get_i8x16();

diff --git a/pulley/src/lib.rs b/pulley/src/lib.rs
@@ -1227,6 +1227,9 @@ macro_rules! for_each_extended_op {
             /// `dst = ieee_minimum(src1, src2)`
             vminimumf64x2 = Vminimumf64x2 { operands: BinaryOperands<VReg> };
 
+            /// `dst = shuffle(src1, src2, mask)`
+            vshuffle = VShuffle { dst: VReg, src1: VReg, src2: VReg, mask: u128 };
+
             /// `dst = swizzle(src1, src2)`
             vswizzlei8x16 = Vswizzlei8x16 { operands: BinaryOperands<VReg> };