Skip to content

Commit b2804c0

Browse files
James Marshloganek
James Marsh
authored and
loganek
committed
Fix splat opcodes, add V128 handling in preserve_referenced_local and reserve_block_ret
1 parent 474acd7 commit b2804c0

File tree

2 files changed

+136
-36
lines changed

2 files changed

+136
-36
lines changed

‎core/iwasm/interpreter/wasm_interp_fast.c‎

Lines changed: 32 additions & 13 deletions
Original file line numberDiff line numberDiff line change
@@ -47,7 +47,7 @@ typedef float64 CellType_F64;
4747
&& (app_addr) <= shared_heap_end_off - bytes + 1)
4848

4949
#define shared_heap_addr_app_to_native(app_addr, native_addr) \
50-
native_addr = shared_heap_base_addr + ((app_addr)-shared_heap_start_off)
50+
native_addr = shared_heap_base_addr + ((app_addr) - shared_heap_start_off)
5151

5252
#define CHECK_SHARED_HEAP_OVERFLOW(app_addr, bytes, native_addr) \
5353
if (app_addr_in_shared_heap(app_addr, bytes)) \
@@ -1793,7 +1793,7 @@ wasm_interp_call_func_bytecode(WASMModuleInstance *module,
17931793
else
17941794
cur_func_type = cur_func->u.func->func_type;
17951795

1796-
/* clang-format off */
1796+
/* clang-format off */
17971797
#if WASM_ENABLE_GC == 0
17981798
if (cur_type != cur_func_type) {
17991799
wasm_set_exception(module, "indirect call type mismatch");
@@ -5923,12 +5923,11 @@ wasm_interp_call_func_bytecode(WASMModuleInstance *module,
59235923
uint32 offset, addr;
59245924
offset = read_uint32(frame_ip);
59255925
V128 data = POP_V128();
5926-
addr = POP_I32();
5926+
int32 base = POP_I32();
5927+
offset += base;
5928+
addr = GET_OPERAND(uint32, I32, 0);
59275929

5928-
V128 data;
5929-
data = POP_V128();
5930-
5931-
CHECK_MEMORY_OVERFLOW(16);
5930+
CHECK_MEMORY_OVERFLOW(32);
59325931
STORE_V128(maddr, data);
59335932
break;
59345933
}
@@ -5948,14 +5947,14 @@ wasm_interp_call_func_bytecode(WASMModuleInstance *module,
59485947
case SIMD_v8x16_shuffle:
59495948
{
59505949
V128 indices;
5951-
V128 v2 = POP_V128();
5952-
V128 v1 = POP_V128();
5953-
addr_ret = GET_OFFSET();
5954-
59555950
bh_memcpy_s(&indices, sizeof(V128), frame_ip,
59565951
sizeof(V128));
59575952
frame_ip += sizeof(V128);
59585953

5954+
V128 v2 = POP_V128();
5955+
V128 v1 = POP_V128();
5956+
addr_ret = GET_OFFSET();
5957+
59595958
V128 result;
59605959
for (int i = 0; i < 16; i++) {
59615960
uint8_t index = indices.i8x16[i];
@@ -5983,6 +5982,7 @@ wasm_interp_call_func_bytecode(WASMModuleInstance *module,
59835982
SIMDE_V128_TO_SIMD_V128(simde_result, result);
59845983

59855984
PUT_V128_TO_ADDR(frame_lp + addr_ret, result);
5985+
break;
59865986
}
59875987

59885988
/* Splat */
@@ -6008,7 +6008,15 @@ wasm_interp_call_func_bytecode(WASMModuleInstance *module,
60086008

60096009
case SIMD_i8x16_splat:
60106010
{
6011-
SIMD_SPLAT_OP_I32(simde_wasm_i8x16_splat);
6011+
uint32 val = POP_I32();
6012+
addr_ret = GET_OFFSET();
6013+
6014+
simde_v128_t simde_result = simde_wasm_i8x16_splat(val);
6015+
6016+
V128 result;
6017+
SIMDE_V128_TO_SIMD_V128(simde_result, result);
6018+
6019+
PUT_V128_TO_ADDR(frame_lp + addr_ret, result);
60126020
break;
60136021
}
60146022
case SIMD_i16x8_splat:
@@ -6140,7 +6148,18 @@ wasm_interp_call_func_bytecode(WASMModuleInstance *module,
61406148
/* i8x16 comparison operations */
61416149
case SIMD_i8x16_eq:
61426150
{
6143-
SIMD_DOUBLE_OP(simde_wasm_i8x16_eq);
6151+
V128 v2 = POP_V128();
6152+
V128 v1 = POP_V128();
6153+
addr_ret = GET_OFFSET();
6154+
6155+
simde_v128_t simde_result =
6156+
simde_wasm_i8x16_eq(SIMD_V128_TO_SIMDE_V128(v1),
6157+
SIMD_V128_TO_SIMDE_V128(v2));
6158+
6159+
V128 result;
6160+
SIMDE_V128_TO_SIMD_V128(simde_result, result);
6161+
6162+
PUT_V128_TO_ADDR(frame_lp + addr_ret, result);
61446163
break;
61456164
}
61466165
case SIMD_i8x16_ne:

‎core/iwasm/interpreter/wasm_loader.c‎

Lines changed: 104 additions & 23 deletions
Original file line numberDiff line numberDiff line change
@@ -9125,6 +9125,9 @@ preserve_referenced_local(WASMLoaderContext *loader_ctx, uint8 opcode,
91259125

91269126
if (is_32bit_type(cur_type))
91279127
i++;
9128+
else if (cur_type == VALUE_TYPE_V128) {
9129+
i += 4;
9130+
}
91289131
else
91299132
i += 2;
91309133
}
@@ -9155,7 +9158,10 @@ preserve_local_for_block(WASMLoaderContext *loader_ctx, uint8 opcode,
91559158
return false;
91569159
}
91579160

9158-
if (is_32bit_type(cur_type)) {
9161+
if (cur_type == VALUE_TYPE_V128) {
9162+
i += 4;
9163+
}
9164+
else if (is_32bit_type(cur_type)) {
91599165
i++;
91609166
}
91619167
else {
@@ -9498,6 +9504,8 @@ wasm_loader_get_const_offset(WASMLoaderContext *ctx, uint8 type, void *value,
94989504
|| (type == VALUE_TYPE_EXTERNREF
94999505
&& *(int32 *)value == c->value.i32)
95009506
#endif
9507+
|| (type == VALUE_TYPE_V128
9508+
&& (0 == memcmp(value, &(c->value.v128), sizeof(V128))))
95019509
|| (type == VALUE_TYPE_F64
95029510
&& (0 == memcmp(value, &(c->value.f64), sizeof(float64))))
95039511
|| (type == VALUE_TYPE_F32
@@ -9508,6 +9516,9 @@ wasm_loader_get_const_offset(WASMLoaderContext *ctx, uint8 type, void *value,
95089516
}
95099517
if (is_32bit_type(c->value_type))
95109518
operand_offset += 1;
9519+
else if (c->value_type == VALUE_TYPE_V128) {
9520+
operand_offset += 4;
9521+
}
95119522
else
95129523
operand_offset += 2;
95139524
}
@@ -9559,6 +9570,10 @@ wasm_loader_get_const_offset(WASMLoaderContext *ctx, uint8 type, void *value,
95599570
c->value.i32 = *(int32 *)value;
95609571
ctx->const_cell_num++;
95619572
break;
9573+
case VALUE_TYPE_V128:
9574+
bh_memcpy_s(&(c->value.v128), sizeof(WASMValue), value,
9575+
sizeof(V128));
9576+
ctx->const_cell_num++;
95629577
#if WASM_ENABLE_REF_TYPES != 0 && WASM_ENABLE_GC == 0
95639578
case VALUE_TYPE_EXTERNREF:
95649579
case VALUE_TYPE_FUNCREF:
@@ -9760,17 +9775,22 @@ reserve_block_ret(WASMLoaderContext *loader_ctx, uint8 opcode,
97609775
block_type, &return_types, &reftype_maps, &reftype_map_count);
97619776
#endif
97629777

9763-
/* If there is only one return value, use EXT_OP_COPY_STACK_TOP/_I64 instead
9764-
* of EXT_OP_COPY_STACK_VALUES for interpreter performance. */
9778+
/* If there is only one return value, use EXT_OP_COPY_STACK_TOP/_I64/V128
9779+
* instead of EXT_OP_COPY_STACK_VALUES for interpreter performance. */
97659780
if (return_count == 1) {
97669781
uint8 cell = (uint8)wasm_value_type_cell_num(return_types[0]);
9767-
if (cell <= 2 /* V128 isn't supported whose cell num is 4 */
9768-
&& block->dynamic_offset != *(loader_ctx->frame_offset - cell)) {
9782+
if (block->dynamic_offset != *(loader_ctx->frame_offset - cell)) {
97699783
/* insert op_copy before else opcode */
97709784
if (opcode == WASM_OP_ELSE)
97719785
skip_label();
9772-
emit_label(cell == 1 ? EXT_OP_COPY_STACK_TOP
9773-
: EXT_OP_COPY_STACK_TOP_I64);
9786+
9787+
if (cell == 4) {
9788+
emit_label(EXT_OP_COPY_STACK_TOP_V128);
9789+
}
9790+
else {
9791+
emit_label(cell == 1 ? EXT_OP_COPY_STACK_TOP
9792+
: EXT_OP_COPY_STACK_TOP_I64);
9793+
}
97749794
emit_operand(loader_ctx, *(loader_ctx->frame_offset - cell));
97759795
emit_operand(loader_ctx, block->dynamic_offset);
97769796

@@ -9805,11 +9825,37 @@ reserve_block_ret(WASMLoaderContext *loader_ctx, uint8 opcode,
98059825
for (i = (int32)return_count - 1; i >= 0; i--) {
98069826
uint8 cells = (uint8)wasm_value_type_cell_num(return_types[i]);
98079827

9808-
frame_offset -= cells;
9809-
dynamic_offset -= cells;
9810-
if (dynamic_offset != *frame_offset) {
9811-
value_count++;
9812-
total_cel_num += cells;
9828+
if (frame_offset - cells < loader_ctx->frame_offset_bottom) {
9829+
set_error_buf(error_buf, error_buf_size, "frame offset underflow");
9830+
goto fail;
9831+
}
9832+
9833+
if (cells == 4) {
9834+
bool needs_copy = false;
9835+
int16 v128_dynamic = dynamic_offset - cells;
9836+
9837+
for (int j = 0; j < 4; j++) {
9838+
if (*(frame_offset - j - 1) != (v128_dynamic + j)) {
9839+
needs_copy = true;
9840+
break;
9841+
}
9842+
}
9843+
9844+
if (needs_copy) {
9845+
value_count++;
9846+
total_cel_num += cells;
9847+
}
9848+
9849+
frame_offset -= cells;
9850+
dynamic_offset = v128_dynamic;
9851+
}
9852+
else {
9853+
frame_offset -= cells;
9854+
dynamic_offset -= cells;
9855+
if (dynamic_offset != *frame_offset) {
9856+
value_count++;
9857+
total_cel_num += cells;
9858+
}
98139859
}
98149860
}
98159861

@@ -9845,19 +9891,50 @@ reserve_block_ret(WASMLoaderContext *loader_ctx, uint8 opcode,
98459891
dynamic_offset = dynamic_offset_org;
98469892
for (i = (int32)return_count - 1, j = 0; i >= 0; i--) {
98479893
uint8 cell = (uint8)wasm_value_type_cell_num(return_types[i]);
9848-
frame_offset -= cell;
9849-
dynamic_offset -= cell;
9850-
if (dynamic_offset != *frame_offset) {
9851-
/* cell num */
9852-
cells[j] = cell;
9853-
/* src offset */
9854-
src_offsets[j] = *frame_offset;
9855-
/* dst offset */
9856-
dst_offsets[j] = dynamic_offset;
9857-
j++;
9894+
9895+
if (cell == 4) {
9896+
bool needs_copy = false;
9897+
int16 v128_dynamic = dynamic_offset - cell;
9898+
9899+
for (int k = 0; k < 4; k++) {
9900+
if (*(frame_offset - k - 1) != (v128_dynamic + k)) {
9901+
needs_copy = true;
9902+
break;
9903+
}
9904+
}
9905+
9906+
if (needs_copy) {
9907+
cells[j] = cell;
9908+
src_offsets[j] = *(frame_offset - cell);
9909+
dst_offsets[j] = v128_dynamic;
9910+
j++;
9911+
}
9912+
9913+
frame_offset -= cell;
9914+
dynamic_offset = v128_dynamic;
9915+
}
9916+
else {
9917+
frame_offset -= cell;
9918+
dynamic_offset -= cell;
9919+
if (dynamic_offset != *frame_offset) {
9920+
cells[j] = cell;
9921+
/* src offset */
9922+
src_offsets[j] = *frame_offset;
9923+
/* dst offset */
9924+
dst_offsets[j] = dynamic_offset;
9925+
j++;
9926+
}
98589927
}
9928+
98599929
if (opcode == WASM_OP_ELSE) {
9860-
*frame_offset = dynamic_offset;
9930+
if (cell == 4) {
9931+
for (int k = 0; k < cell; k++) {
9932+
*(frame_offset + k) = dynamic_offset + k;
9933+
}
9934+
}
9935+
else {
9936+
*frame_offset = dynamic_offset;
9937+
}
98619938
}
98629939
else {
98639940
loader_ctx->frame_offset = frame_offset;
@@ -13031,6 +13108,10 @@ wasm_loader_prepare_bytecode(WASMModule *module, WASMFunction *func,
1303113108
emit_label(EXT_OP_TEE_LOCAL_FAST);
1303213109
emit_byte(loader_ctx, (uint8)local_offset);
1303313110
}
13111+
else if (local_type == VALUE_TYPE_V128) {
13112+
emit_label(EXT_OP_TEE_LOCAL_FAST_V128);
13113+
emit_byte(loader_ctx, (uint8)local_offset);
13114+
}
1303413115
else {
1303513116
emit_label(EXT_OP_TEE_LOCAL_FAST_I64);
1303613117
emit_byte(loader_ctx, (uint8)local_offset);

0 commit comments

Comments
 (0)