Runtime.stackAlloc should grow down for wasm (#1073)

* Runtime.stackAlloc should grow down for wasm

* stackAlloc should align properly; update tests

Author: jgravelle-google

src/wasm-emscripten.cpp

@@ -119,16 +119,12 @@ void generateStackAllocFunction(LinkerObject& linker) {
   SetLocal* setStackLocal = builder.makeSetLocal(1, loadStack);
   GetLocal* getStackLocal = builder.makeGetLocal(1, i32);
   GetLocal* getSizeArg = builder.makeGetLocal(0, i32);
-  Binary* add = builder.makeBinary(AddInt32, getStackLocal, getSizeArg);
+  Binary* sub = builder.makeBinary(SubInt32, getStackLocal, getSizeArg);
   const static uint32_t bitAlignment = 16;
   const static uint32_t bitMask = bitAlignment - 1;
-  Const* addConst = builder.makeConst(Literal(bitMask));
-  Binary* maskedAdd = builder.makeBinary(
-    AndInt32,
-    builder.makeBinary(AddInt32, add, addConst),
-    builder.makeConst(Literal(~bitMask))
-  );
-  Store* storeStack = generateStoreStackPointer(builder, linker, maskedAdd);
+  Const* subConst = builder.makeConst(Literal(~bitMask));
+  Binary* maskedSub = builder.makeBinary(AndInt32, sub, subConst);
+  Store* storeStack = generateStoreStackPointer(builder, linker, maskedSub);
 
   Block* block = builder.makeBlock();
   block->list.push_back(setStackLocal);

test/dot_s/alias.wast

@@ -43,12 +43,9 @@
   (i32.store offset=4
    (i32.const 0)
    (i32.and
-    (i32.add
-     (i32.add
-      (get_local $1)
-      (get_local $0)
-     )
-     (i32.const 15)
+    (i32.sub
+     (get_local $1)
+     (get_local $0)
     )
     (i32.const -16)
    )

test/dot_s/alternate-lcomm.wast

@@ -19,12 +19,9 @@
   (i32.store offset=4
    (i32.const 0)
    (i32.and
-    (i32.add
-     (i32.add
-      (get_local $1)
-      (get_local $0)
-     )
-     (i32.const 15)
+    (i32.sub
+     (get_local $1)
+     (get_local $0)
     )
     (i32.const -16)
    )

test/dot_s/asm_const.wast

@@ -31,12 +31,9 @@
   (i32.store offset=4
    (i32.const 0)
    (i32.and
-    (i32.add
-     (i32.add
-      (get_local $1)
-      (get_local $0)
-     )
-     (i32.const 15)
+    (i32.sub
+     (get_local $1)
+     (get_local $0)
     )
     (i32.const -16)
    )

test/dot_s/basics.wast

@@ -112,12 +112,9 @@
   (i32.store offset=4
    (i32.const 0)
    (i32.and
-    (i32.add
-     (i32.add
-      (get_local $1)
-      (get_local $0)
-     )
-     (i32.const 15)
+    (i32.sub
+     (get_local $1)
+     (get_local $0)
     )
     (i32.const -16)
    )

test/dot_s/bcp-1.wast

@@ -324,12 +324,9 @@
   (i32.store offset=4
    (i32.const 0)
    (i32.and
-    (i32.add
-     (i32.add
-      (get_local $1)
-      (get_local $0)
-     )
-     (i32.const 15)
+    (i32.sub
+     (get_local $1)
+     (get_local $0)
     )
     (i32.const -16)
    )

test/dot_s/data-offset-folding.wast

@@ -21,12 +21,9 @@
   (i32.store offset=4
    (i32.const 0)
    (i32.and
-    (i32.add
-     (i32.add
-      (get_local $1)
-      (get_local $0)
-     )
-     (i32.const 15)
+    (i32.sub
+     (get_local $1)
+     (get_local $0)
     )
     (i32.const -16)
    )

test/dot_s/debug.wast

@@ -71,12 +71,9 @@
   (i32.store offset=4
    (i32.const 0)
    (i32.and
-    (i32.add
-     (i32.add
-      (get_local $1)
-      (get_local $0)
-     )
-     (i32.const 15)
+    (i32.sub
+     (get_local $1)
+     (get_local $0)
     )
     (i32.const -16)
    )

test/dot_s/dso_handle.wast

@@ -25,12 +25,9 @@
   (i32.store offset=4
    (i32.const 0)
    (i32.and
-    (i32.add
-     (i32.add
-      (get_local $1)
-      (get_local $0)
-     )
-     (i32.const 15)
+    (i32.sub
+     (get_local $1)
+     (get_local $0)
     )
     (i32.const -16)
    )

test/dot_s/dyncall.wast

@@ -65,12 +65,9 @@
   (i32.store offset=4
    (i32.const 0)
    (i32.and
-    (i32.add
-     (i32.add
-      (get_local $1)
-      (get_local $0)
-     )
-     (i32.const 15)
+    (i32.sub
+     (get_local $1)
+     (get_local $0)
     )
     (i32.const -16)
    )