pulley: Implement simd vector negation

Filling out some more miscellaneous `*.wast` tests
alexcrichton · Dec 19, 2024 · a9304c9 · a9304c9
1 parent 4178766
commit a9304c9
Show file tree

Hide file tree

Showing 5 changed files with 44 additions and 4 deletions.
diff --git a/cranelift/codegen/src/isa/pulley_shared/lower.isle b/cranelift/codegen/src/isa/pulley_shared/lower.isle
@@ -1162,6 +1162,13 @@
 (rule 0 (lower (has_type (fits_in_32 _) (ineg a))) (pulley_xneg32 (sext32 a)))
 (rule 1 (lower (has_type $I64 (ineg a))) (pulley_xneg64 a))
 
+;; vector negation
+
+(rule 1 (lower (has_type $I8X16 (ineg a))) (pulley_vneg8x16 a))
+(rule 1 (lower (has_type $I16X8 (ineg a))) (pulley_vneg16x8 a))
+(rule 1 (lower (has_type $I32X4 (ineg a))) (pulley_vneg32x4 a))
+(rule 1 (lower (has_type $I64X2 (ineg a))) (pulley_vneg64x2 a))
+
 ;;;; Rules for `fabs` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
 (rule (lower (has_type $F32 (fabs a))) (pulley_fabs32 a))

diff --git a/cranelift/filetests/filetests/runtests/simd-ineg.clif b/cranelift/filetests/filetests/runtests/simd-ineg.clif
@@ -7,6 +7,10 @@ target x86_64 skylake
 set enable_multi_ret_implicit_sret
 target riscv64 has_v
 target riscv64 has_v has_c has_zcb
+target pulley32
+target pulley32be
+target pulley64
+target pulley64be
 
 function %ineg_i8x16(i8x16) -> i8x16 {
 block0(v0: i8x16):

diff --git a/crates/wast-util/src/lib.rs b/crates/wast-util/src/lib.rs
@@ -428,20 +428,16 @@ impl WastTest {
                 "spec_testsuite/simd_f64x2_cmp.wast",
                 "spec_testsuite/simd_f64x2_pmin_pmax.wast",
                 "spec_testsuite/simd_f64x2_rounding.wast",
-                "spec_testsuite/simd_i16x8_arith.wast",
                 "spec_testsuite/simd_i16x8_arith2.wast",
                 "spec_testsuite/simd_i16x8_extadd_pairwise_i8x16.wast",
                 "spec_testsuite/simd_i16x8_q15mulr_sat_s.wast",
                 "spec_testsuite/simd_i16x8_sat_arith.wast",
-                "spec_testsuite/simd_i32x4_arith.wast",
                 "spec_testsuite/simd_i32x4_arith2.wast",
                 "spec_testsuite/simd_i32x4_dot_i16x8.wast",
                 "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_i64x2_arith.wast",
                 "spec_testsuite/simd_i64x2_arith2.wast",
-                "spec_testsuite/simd_i8x16_arith.wast",
                 "spec_testsuite/simd_i8x16_arith2.wast",
                 "spec_testsuite/simd_i8x16_sat_arith.wast",
                 "spec_testsuite/simd_lane.wast",

diff --git a/pulley/src/interp.rs b/pulley/src/interp.rs
@@ -4011,4 +4011,28 @@ impl ExtendedOpVisitor for Interpreter<'_> {
         self.state[operands.dst].set_u64x2(c);
         ControlFlow::Continue(())
     }
+
+    fn vneg8x16(&mut self, dst: VReg, src: VReg) -> ControlFlow<Done> {
+        let a = self.state[src].get_i8x16();
+        self.state[dst].set_i8x16(a.map(|i| i.wrapping_neg()));
+        ControlFlow::Continue(())
+    }
+
+    fn vneg16x8(&mut self, dst: VReg, src: VReg) -> ControlFlow<Done> {
+        let a = self.state[src].get_i16x8();
+        self.state[dst].set_i16x8(a.map(|i| i.wrapping_neg()));
+        ControlFlow::Continue(())
+    }
+
+    fn vneg32x4(&mut self, dst: VReg, src: VReg) -> ControlFlow<Done> {
+        let a = self.state[src].get_i32x4();
+        self.state[dst].set_i32x4(a.map(|i| i.wrapping_neg()));
+        ControlFlow::Continue(())
+    }
+
+    fn vneg64x2(&mut self, dst: VReg, src: VReg) -> ControlFlow<Done> {
+        let a = self.state[src].get_i64x2();
+        self.state[dst].set_i64x2(a.map(|i| i.wrapping_neg()));
+        ControlFlow::Continue(())
+    }
 }
diff --git a/pulley/src/lib.rs b/pulley/src/lib.rs
@@ -1097,6 +1097,15 @@ macro_rules! for_each_extended_op {
             vult64x2 = Vult64x2 { operands: BinaryOperands<VReg> };
             /// `dst = src <= dst` (unsigned)
             vulteq64x2 = Vulteq64x2 { operands: BinaryOperands<VReg> };
+
+            /// `dst = -src`
+            vneg8x16 = Vneg8x16 { dst: VReg, src: VReg };
+            /// `dst = -src`
+            vneg16x8 = Vneg16x8 { dst: VReg, src: VReg };
+            /// `dst = -src`
+            vneg32x4 = Vneg32x4 { dst: VReg, src: VReg };
+            /// `dst = -src`
+            vneg64x2 = Vneg64x2 { dst: VReg, src: VReg };
         }
     };
 }