[WebAssembly] Add fence in __wasm_init_memory

[tlively]

Summary

memory.init is not atomic, so without the fence there is no
guarantee that the initialized memory is visible on other threads once
they have been woken up by the initializing thread.

[aheejin]

lld/wasm/Writer.cpp

733

In sequentially consistent models, wouldn't this store serve as the fence anyway?

[tlively]

lld/wasm/Writer.cpp

733

Yes, you're correct. I confused myself.

lld/test/wasm/data-segments.ll

 ; PASSIVE-NEXT:    StartFunction:   1
 ; PASSIVE-LABEL: - Type:            CODE
 ; PASSIVE-NEXT:    Functions:
 ; PASSIVE-NEXT:      - Index:           0
 ; PASSIVE-NEXT:        Locals:          []
 ; PASSIVE-NEXT:        Body:            0B
 ; PASSIVE-NEXT:      - Index:           1
 ; PASSIVE-NEXT:        Locals:          []
-; PASSIVE-NEXT:        Body:            41B4D60041004101FE480200044041B4D6004101427FFE0102001A0541800841004114FC08000041940841004190CE00FC08010041A4D6004100410DFC08020041B4D6004102FE17020041B4D600417FFE0002001A0BFC0900FC0901FC09020B
+; PASSIVE-NEXT:        Body:            41B4D60041004101FE480200044041B4D6004101427FFE0102001A0541800841004114FC08000041940841004190CE00FC08010041A4D6004100410DFC080200FE030041B4D6004102FE17020041B4D600417FFE0002001A0BFC0900FC0901FC09020B
 ; PASSIVE-NEXT:  - Index:           2
 ; PASSIVE-NEXT:    Locals:          []
 ; PASSIVE-NEXT:    Body:            0B
 ; PASSIVE-NEXT:  - Type:            DATA
 ; PASSIVE-NEXT:    Segments:
 ; PASSIVE-NEXT:      - SectionOffset:   3
 ; PASSIVE-NEXT:        InitFlags:       1
 ; PASSIVE-NEXT:        Content:         68656C6C6F00676F6F646279650000002A000000

lld/wasm/Writer.cpp

if (segments.size()) {
       //      (i32.const $__init_memory_flag)
       //      (i32.const 1)
       //      (i32.const -1)
       //     )
       //    )
       //   )
       //   (else
       //    ( ... initialize data segments ... )
+      //    (atomic.fence)
       //    (i32.atomic.store align=2 offset=0
       //     (i32.const $__init_memory_flag)
       //     (i32.const 2)
       //    )
       //    (drop
       //     (i32.atomic.notify align=2 offset=0
       //      (i32.const $__init_memory_flag)
       //      (i32.const -1u)
       //     )
       //    )
       //   )
       //  )
if (segments.size()) {
           // memory.init instruction
           writeU8(os, WASM_OPCODE_MISC_PREFIX, "bulk-memory prefix");
           writeUleb128(os, WASM_OPCODE_MEMORY_INIT, "memory.init");
           writeUleb128(os, s->index, "segment index immediate");
           writeU8(os, 0, "memory index immediate");
         }
       }
 
+      // Fence to commit stores from memory.init instructions
+      writeU8(os, WASM_OPCODE_ATOMICS_PREFIX, "atomics prefix");
+      writeUleb128(os, WASM_OPCODE_ATOMIC_FENCE, "atomic.fence");
+      writeU8(os, 0, "memory ordering");
+
       // Set flag to 2 to mark end of initialization
       writeI32Const(os, flagAddress, "flag address");
       writeI32Const(os, 2, "flag value");
       writeU8(os, WASM_OPCODE_ATOMICS_PREFIX, "atomics prefix");
       writeUleb128(os, WASM_OPCODE_I32_ATOMIC_STORE, "i32.atomic.store");
       writeMemArg(os, 2, 0);
 
       // Notify any waiters that memory initialization is complete

llvm/include/llvm/BinaryFormat/Wasm.h

enum : unsigned {
   WASM_OPCODE_ELSE = 0x05,
   WASM_OPCODE_DROP = 0x1a,
   WASM_OPCODE_MISC_PREFIX = 0xfc,
   WASM_OPCODE_MEMORY_INIT = 0x08,
   WASM_OPCODE_DATA_DROP = 0x09,
   WASM_OPCODE_ATOMICS_PREFIX = 0xfe,
   WASM_OPCODE_ATOMIC_NOTIFY = 0x00,
   WASM_OPCODE_I32_ATOMIC_WAIT = 0x01,
+  WASM_OPCODE_ATOMIC_FENCE = 0x03,
   WASM_OPCODE_I32_ATOMIC_STORE = 0x17,
   WASM_OPCODE_I32_RMW_CMPXCHG = 0x48,
 };
 
 enum : unsigned {
   WASM_LIMITS_FLAG_HAS_MAX = 0x1,
   WASM_LIMITS_FLAG_IS_SHARED = 0x2,
 };