Implement i64x2.mul (#2860)

This is the only instruction in the current spec proposal that had not
yet been implemnented in the tools.

9 files changed, 102 insertions, 56 deletions

diff --git a/test/binaryen.js/kitchen-sink.js b/test/binaryen.js/kitchen-sink.js
index 0b693b970..b0b81433d 100644
--- a/test/binaryen.js/kitchen-sink.js
+++ b/test/binaryen.js/kitchen-sink.js
@@ -383,6 +383,7 @@ function test_core() {
     module.i32x4.dot_i16x8_s(module.v128.const(v128_bytes), module.v128.const(v128_bytes)),
     module.i64x2.add(module.v128.const(v128_bytes), module.v128.const(v128_bytes)),
     module.i64x2.sub(module.v128.const(v128_bytes), module.v128.const(v128_bytes)),
+    module.i64x2.mul(module.v128.const(v128_bytes), module.v128.const(v128_bytes)),
     module.f32x4.add(module.v128.const(v128_bytes), module.v128.const(v128_bytes)),
     module.f32x4.sub(module.v128.const(v128_bytes), module.v128.const(v128_bytes)),
     module.f32x4.mul(module.v128.const(v128_bytes), module.v128.const(v128_bytes)),
diff --git a/test/binaryen.js/kitchen-sink.js.txt b/test/binaryen.js/kitchen-sink.js.txt
index 0e93201ff..b27d147b3 100644
--- a/test/binaryen.js/kitchen-sink.js.txt
+++ b/test/binaryen.js/kitchen-sink.js.txt
@@ -1244,6 +1244,12 @@ getExpressionInfo(tuple[3])={"id":14,"type":5,"value":3.7}
        )
       )
       (drop
+       (i64x2.mul
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+       )
+      )
+      (drop
        (f32x4.add
         (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
         (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
@@ -3036,6 +3042,12 @@ getExpressionInfo(tuple[3])={"id":14,"type":5,"value":3.7}
        )
       )
       (drop
+       (i64x2.mul
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+       )
+      )
+      (drop
        (f32x4.add
         (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
         (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
diff --git a/test/example/c-api-kitchen-sink.c b/test/example/c-api-kitchen-sink.c
index 9b3b68cf7..3ea6c4ef0 100644
--- a/test/example/c-api-kitchen-sink.c
+++ b/test/example/c-api-kitchen-sink.c
@@ -550,6 +550,7 @@ void test_core() {
     makeBinary(module, BinaryenMulVecI32x4(), v128),
     makeBinary(module, BinaryenAddVecI64x2(), v128),
     makeBinary(module, BinaryenSubVecI64x2(), v128),
+    makeBinary(module, BinaryenMulVecI64x2(), v128),
     makeBinary(module, BinaryenAddVecF32x4(), v128),
     makeBinary(module, BinaryenSubVecF32x4(), v128),
     makeBinary(module, BinaryenMulVecF32x4(), v128),
diff --git a/test/example/c-api-kitchen-sink.txt b/test/example/c-api-kitchen-sink.txt
index d3e0281ee..ced13ac72 100644
--- a/test/example/c-api-kitchen-sink.txt
+++ b/test/example/c-api-kitchen-sink.txt
@@ -1147,6 +1147,12 @@ BinaryenFeatureAll: 1023
        )
       )
       (drop
+       (i64x2.mul
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+        (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
+       )
+      )
+      (drop
        (f32x4.add
         (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
         (v128.const i32x4 0x04030201 0x08070605 0x0c0b0a09 0x100f0e0d)
diff --git a/test/simd.wast b/test/simd.wast
index fdb6d0309..edd9a2c43 100644
--- a/test/simd.wast
+++ b/test/simd.wast
@@ -790,6 +790,12 @@
    (local.get $1)
   )
  )
+ (func $i64x2.mul (param $0 v128) (param $1 v128) (result v128)
+  (i64x2.mul
+   (local.get $0)
+   (local.get $1)
+  )
+ )
  (func $f32x4.add (param $0 v128) (param $1 v128) (result v128)
   (f32x4.add
    (local.get $0)
diff --git a/test/simd.wast.from-wast b/test/simd.wast.from-wast
index 17b9d7ffd..35643ec04 100644
--- a/test/simd.wast.from-wast
+++ b/test/simd.wast.from-wast
@@ -806,6 +806,12 @@
    (local.get $1)
   )
  )
+ (func $i64x2.mul (param $0 v128) (param $1 v128) (result v128)
+  (i64x2.mul
+   (local.get $0)
+   (local.get $1)
+  )
+ )
  (func $f32x4.add (param $0 v128) (param $1 v128) (result v128)
   (f32x4.add
    (local.get $0)
diff --git a/test/simd.wast.fromBinary b/test/simd.wast.fromBinary
index 8ddfb9a6f..40ce99ec2 100644
--- a/test/simd.wast.fromBinary
+++ b/test/simd.wast.fromBinary
@@ -806,6 +806,12 @@
    (local.get $1)
   )
  )
+ (func $i64x2.mul (param $0 v128) (param $1 v128) (result v128)
+  (i64x2.mul
+   (local.get $0)
+   (local.get $1)
+  )
+ )
  (func $f32x4.add (param $0 v128) (param $1 v128) (result v128)
   (f32x4.add
    (local.get $0)
diff --git a/test/simd.wast.fromBinary.noDebugInfo b/test/simd.wast.fromBinary.noDebugInfo
index 20617d09d..85ca0a9f4 100644
--- a/test/simd.wast.fromBinary.noDebugInfo
+++ b/test/simd.wast.fromBinary.noDebugInfo
@@ -807,314 +807,320 @@
   )
  )
  (func $140 (param $0 v128) (param $1 v128) (result v128)
-  (f32x4.add
+  (i64x2.mul
    (local.get $0)
    (local.get $1)
   )
  )
  (func $141 (param $0 v128) (param $1 v128) (result v128)
-  (f32x4.sub
+  (f32x4.add
    (local.get $0)
    (local.get $1)
   )
  )
  (func $142 (param $0 v128) (param $1 v128) (result v128)
-  (f32x4.mul
+  (f32x4.sub
    (local.get $0)
    (local.get $1)
   )
  )
  (func $143 (param $0 v128) (param $1 v128) (result v128)
-  (f32x4.div
+  (f32x4.mul
    (local.get $0)
    (local.get $1)
   )
  )
  (func $144 (param $0 v128) (param $1 v128) (result v128)
-  (f32x4.min
+  (f32x4.div
    (local.get $0)
    (local.get $1)
   )
  )
  (func $145 (param $0 v128) (param $1 v128) (result v128)
-  (f32x4.max
+  (f32x4.min
    (local.get $0)
    (local.get $1)
   )
  )
  (func $146 (param $0 v128) (param $1 v128) (result v128)
-  (f32x4.pmin
+  (f32x4.max
    (local.get $0)
    (local.get $1)
   )
  )
  (func $147 (param $0 v128) (param $1 v128) (result v128)
+  (f32x4.pmin
+   (local.get $0)
+   (local.get $1)
+  )
+ )
+ (func $148 (param $0 v128) (param $1 v128) (result v128)
   (f32x4.pmax
    (local.get $0)
    (local.get $1)
   )
  )
- (func $148 (param $0 v128) (result v128)
+ (func $149 (param $0 v128) (result v128)
   (f32x4.abs
    (local.get $0)
   )
  )
- (func $149 (param $0 v128) (result v128)
+ (func $150 (param $0 v128) (result v128)
   (f32x4.neg
    (local.get $0)
   )
  )
- (func $150 (param $0 v128) (result v128)
+ (func $151 (param $0 v128) (result v128)
   (f32x4.sqrt
    (local.get $0)
   )
  )
- (func $151 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+ (func $152 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
   (f32x4.qfma
    (local.get $0)
    (local.get $1)
    (local.get $2)
   )
  )
- (func $152 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+ (func $153 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
   (f32x4.qfms
    (local.get $0)
    (local.get $1)
    (local.get $2)
   )
  )
- (func $153 (param $0 v128) (param $1 v128) (result v128)
+ (func $154 (param $0 v128) (param $1 v128) (result v128)
   (f64x2.add
    (local.get $0)
    (local.get $1)
   )
  )
- (func $154 (param $0 v128) (param $1 v128) (result v128)
+ (func $155 (param $0 v128) (param $1 v128) (result v128)
   (f64x2.sub
    (local.get $0)
    (local.get $1)
   )
  )
- (func $155 (param $0 v128) (param $1 v128) (result v128)
+ (func $156 (param $0 v128) (param $1 v128) (result v128)
   (f64x2.mul
    (local.get $0)
    (local.get $1)
   )
  )
- (func $156 (param $0 v128) (param $1 v128) (result v128)
+ (func $157 (param $0 v128) (param $1 v128) (result v128)
   (f64x2.div
    (local.get $0)
    (local.get $1)
   )
  )
- (func $157 (param $0 v128) (param $1 v128) (result v128)
+ (func $158 (param $0 v128) (param $1 v128) (result v128)
   (f64x2.min
    (local.get $0)
    (local.get $1)
   )
  )
- (func $158 (param $0 v128) (param $1 v128) (result v128)
+ (func $159 (param $0 v128) (param $1 v128) (result v128)
   (f64x2.max
    (local.get $0)
    (local.get $1)
   )
  )
- (func $159 (param $0 v128) (param $1 v128) (result v128)
+ (func $160 (param $0 v128) (param $1 v128) (result v128)
   (f64x2.pmin
    (local.get $0)
    (local.get $1)
   )
  )
- (func $160 (param $0 v128) (param $1 v128) (result v128)
+ (func $161 (param $0 v128) (param $1 v128) (result v128)
   (f64x2.pmax
    (local.get $0)
    (local.get $1)
   )
  )
- (func $161 (param $0 v128) (result v128)
+ (func $162 (param $0 v128) (result v128)
   (f64x2.abs
    (local.get $0)
   )
  )
- (func $162 (param $0 v128) (result v128)
+ (func $163 (param $0 v128) (result v128)
   (f64x2.neg
    (local.get $0)
   )
  )
- (func $163 (param $0 v128) (result v128)
+ (func $164 (param $0 v128) (result v128)
   (f64x2.sqrt
    (local.get $0)
   )
  )
- (func $164 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+ (func $165 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
   (f64x2.qfma
    (local.get $0)
    (local.get $1)
    (local.get $2)
   )
  )
- (func $165 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
+ (func $166 (param $0 v128) (param $1 v128) (param $2 v128) (result v128)
   (f64x2.qfms
    (local.get $0)
    (local.get $1)
    (local.get $2)
   )
  )
- (func $166 (param $0 v128) (result v128)
+ (func $167 (param $0 v128) (result v128)
   (i32x4.trunc_sat_f32x4_s
    (local.get $0)
   )
  )
- (func $167 (param $0 v128) (result v128)
+ (func $168 (param $0 v128) (result v128)
   (i32x4.trunc_sat_f32x4_u
    (local.get $0)
   )
  )
- (func $168 (param $0 v128) (result v128)
+ (func $169 (param $0 v128) (result v128)
   (i64x2.trunc_sat_f64x2_s
    (local.get $0)
   )
  )
- (func $169 (param $0 v128) (result v128)
+ (func $170 (param $0 v128) (result v128)
   (i64x2.trunc_sat_f64x2_u
    (local.get $0)
   )
  )
- (func $170 (param $0 v128) (result v128)
+ (func $171 (param $0 v128) (result v128)
   (f32x4.convert_i32x4_s
    (local.get $0)
   )
  )
- (func $171 (param $0 v128) (result v128)
+ (func $172 (param $0 v128) (result v128)
   (f32x4.convert_i32x4_u
    (local.get $0)
   )
  )
- (func $172 (param $0 v128) (result v128)
+ (func $173 (param $0 v128) (result v128)
   (f64x2.convert_i64x2_s
    (local.get $0)
   )
  )
- (func $173 (param $0 v128) (result v128)
+ (func $174 (param $0 v128) (result v128)
   (f64x2.convert_i64x2_u
    (local.get $0)
   )
  )
- (func $174 (param $0 i32) (result v128)
+ (func $175 (param $0 i32) (result v128)
   (v8x16.load_splat
    (local.get $0)
   )
  )
- (func $175 (param $0 i32) (result v128)
+ (func $176 (param $0 i32) (result v128)
   (v16x8.load_splat
    (local.get $0)
   )
  )
- (func $176 (param $0 i32) (result v128)
+ (func $177 (param $0 i32) (result v128)
   (v32x4.load_splat
    (local.get $0)
   )
  )
- (func $177 (param $0 i32) (result v128)
+ (func $178 (param $0 i32) (result v128)
   (v64x2.load_splat
    (local.get $0)
   )
  )
- (func $178 (param $0 v128) (param $1 v128) (result v128)
+ (func $179 (param $0 v128) (param $1 v128) (result v128)
   (i8x16.narrow_i16x8_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $179 (param $0 v128) (param $1 v128) (result v128)
+ (func $180 (param $0 v128) (param $1 v128) (result v128)
   (i8x16.narrow_i16x8_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $180 (param $0 v128) (param $1 v128) (result v128)
+ (func $181 (param $0 v128) (param $1 v128) (result v128)
   (i16x8.narrow_i32x4_s
    (local.get $0)
    (local.get $1)
   )
  )
- (func $181 (param $0 v128) (param $1 v128) (result v128)
+ (func $182 (param $0 v128) (param $1 v128) (result v128)
   (i16x8.narrow_i32x4_u
    (local.get $0)
    (local.get $1)
   )
  )
- (func $182 (param $0 v128) (result v128)
+ (func $183 (param $0 v128) (result v128)
   (i16x8.widen_low_i8x16_s
    (local.get $0)
   )
  )
- (func $183 (param $0 v128) (result v128)
+ (func $184 (param $0 v128) (result v128)
   (i16x8.widen_high_i8x16_s
    (local.get $0)
   )
  )
- (func $184 (param $0 v128) (result v128)
+ (func $185 (param $0 v128) (result v128)
   (i16x8.widen_low_i8x16_u
    (local.get $0)
   )
  )
- (func $185 (param $0 v128) (result v128)
+ (func $186 (param $0 v128) (result v128)
   (i16x8.widen_high_i8x16_u
    (local.get $0)
   )
  )
- (func $186 (param $0 v128) (result v128)
+ (func $187 (param $0 v128) (result v128)
   (i32x4.widen_low_i16x8_s
    (local.get $0)
   )
  )
- (func $187 (param $0 v128) (result v128)
+ (func $188 (param $0 v128) (result v128)
   (i32x4.widen_high_i16x8_s
    (local.get $0)
   )
  )
- (func $188 (param $0 v128) (result v128)
+ (func $189 (param $0 v128) (result v128)
   (i32x4.widen_low_i16x8_u
    (local.get $0)
   )
  )
- (func $189 (param $0 v128) (result v128)
+ (func $190 (param $0 v128) (result v128)
   (i32x4.widen_high_i16x8_u
    (local.get $0)
   )
  )
- (func $190 (param $0 i32) (result v128)
+ (func $191 (param $0 i32) (result v128)
   (i16x8.load8x8_u
    (local.get $0)
   )
  )
- (func $191 (param $0 i32) (result v128)
+ (func $192 (param $0 i32) (result v128)
   (i16x8.load8x8_s
    (local.get $0)
   )
  )
- (func $192 (param $0 i32) (result v128)
+ (func $193 (param $0 i32) (result v128)
   (i32x4.load16x4_s
    (local.get $0)
   )
  )
- (func $193 (param $0 i32) (result v128)
+ (func $194 (param $0 i32) (result v128)
   (i32x4.load16x4_u
    (local.get $0)
   )
  )
- (func $194 (param $0 i32) (result v128)
+ (func $195 (param $0 i32) (result v128)
   (i64x2.load32x2_s
    (local.get $0)
   )
  )
- (func $195 (param $0 i32) (result v128)
+ (func $196 (param $0 i32) (result v128)
   (i64x2.load32x2_u
    (local.get $0)
   )
  )
- (func $196 (param $0 v128) (param $1 v128) (result v128)
+ (func $197 (param $0 v128) (param $1 v128) (result v128)
   (v8x16.swizzle
    (local.get $0)
    (local.get $1)
diff --git a/test/spec/simd.wast b/test/spec/simd.wast
index d6549f486..26da927a9 100644
--- a/test/spec/simd.wast
+++ b/test/spec/simd.wast
@@ -167,6 +167,7 @@
  (func (export "i64x2.shr_u") (param $0 v128) (param $1 i32) (result v128) (i64x2.shr_u (local.get $0) (local.get $1)))
  (func (export "i64x2.add") (param $0 v128) (param $1 v128) (result v128) (i64x2.add (local.get $0) (local.get $1)))
  (func (export "i64x2.sub") (param $0 v128) (param $1 v128) (result v128) (i64x2.sub (local.get $0) (local.get $1)))
+ (func (export "i64x2.mul") (param $0 v128) (param $1 v128) (result v128) (i64x2.mul (local.get $0) (local.get $1)))
  (func (export "f32x4.abs") (param $0 v128) (result v128) (f32x4.abs (local.get $0)))
  (func (export "f32x4.neg") (param $0 v128) (result v128) (f32x4.neg (local.get $0)))
  (func (export "f32x4.sqrt") (param $0 v128) (result v128) (f32x4.sqrt (local.get $0)))
@@ -783,6 +784,7 @@
 (assert_return (invoke "i64x2.shr_u" (v128.const i64x2 1 0x8000000000000000) (i32.const 64)) (v128.const i64x2 1 0x8000000000000000))
 (assert_return (invoke "i64x2.add" (v128.const i64x2 0x8000000000000001 42) (v128.const i64x2 0x8000000000000001 0)) (v128.const i64x2 2 42))
 (assert_return (invoke "i64x2.sub" (v128.const i64x2 2 42) (v128.const i64x2 0x8000000000000001 0)) (v128.const i64x2 0x8000000000000001 42))
+(assert_return (invoke "i64x2.mul" (v128.const i64x2 2 42) (v128.const i64x2 0x8000000000000001 0)) (v128.const i64x2 2 0))
 
 ;; f32x4 arithmetic
 (assert_return (invoke "f32x4.abs" (v128.const f32x4 -0 nan -infinity 5)) (v128.const f32x4 0 nan infinity 5))