wazevo(frontend): faster non-imported global access (#1889)

Signed-off-by: Takeshi Yoneda <t.y.mathetake@gmail.com>

internal/engine/compiler/engine.go

@@ -656,6 +656,14 @@ func (e *moduleEngine) ResolveImportedFunction(index, indexInImportedModule wasm
 	e.functions[index] = imported.functions[indexInImportedModule]
 }
 
+// GetGlobalValue implements the same method as documented on wasm.ModuleEngine.
+func (e *moduleEngine) GetGlobalValue(wasm.Index) (lo, hi uint64) {
+	panic("BUG: GetGlobalValue should never be called on compiler mode")
+}
+
+// OwnsGlobals implements the same method as documented on wasm.ModuleEngine.
+func (e *moduleEngine) OwnsGlobals() bool { return false }
+
 // ResolveImportedMemory implements wasm.ModuleEngine.
 func (e *moduleEngine) ResolveImportedMemory(wasm.ModuleEngine) {}
 

internal/engine/interpreter/interpreter.go

@@ -87,6 +87,14 @@ type moduleEngine struct {
 	parentEngine *engine
 }
 
+// GetGlobalValue implements the same method as documented on wasm.ModuleEngine.
+func (e *moduleEngine) GetGlobalValue(wasm.Index) (lo, hi uint64) {
+	panic("BUG: GetGlobalValue should never be called on interpreter mode")
+}
+
+// OwnsGlobals implements the same method as documented on wasm.ModuleEngine.
+func (e *moduleEngine) OwnsGlobals() bool { return false }
+
 // callEngine holds context per moduleEngine.Call, and shared across all the
 // function calls originating from the same moduleEngine.Call execution.
 //

internal/engine/wazevo/backend/backend_test.go

@@ -1937,34 +1937,26 @@ L2:
 L1 (SSA Block: blk0):
 	mov x128?, x0
 	mov x129?, x1
-	ldr x130?, [x129?, #0x8]
-	ldr w131?, [x130?, #0x8]
-	ldr x132?, [x129?, #0x10]
-	ldr x133?, [x132?, #0x8]
-	ldr x134?, [x129?, #0x18]
-	ldr s135?, [x134?, #0x8]
-	ldr x136?, [x129?, #0x20]
-	ldr d137?, [x136?, #0x8]
+	ldr w130?, [x129?, #0x8]
+	ldr x131?, [x129?, #0x18]
+	ldr s132?, [x129?, #0x28]
+	ldr d133?, [x129?, #0x38]
 	str x129?, [x128?, #0x8]
 	mov x0, x128?
 	mov x1, x129?
 	bl f1
-	ldr x138?, [x129?, #0x8]
-	ldr w139?, [x138?, #0x8]
-	ldr x140?, [x129?, #0x10]
-	ldr x141?, [x140?, #0x8]
-	ldr x142?, [x129?, #0x18]
-	ldr s143?, [x142?, #0x8]
-	ldr x144?, [x129?, #0x20]
-	ldr d145?, [x144?, #0x8]
-	mov v3.8b, v145?.8b
-	mov v2.8b, v143?.8b
-	mov x3, x141?
-	mov x2, x139?
-	mov v1.8b, v137?.8b
-	mov v0.8b, v135?.8b
-	mov x1, x133?
-	mov x0, x131?
+	ldr w134?, [x129?, #0x8]
+	ldr x135?, [x129?, #0x18]
+	ldr s136?, [x129?, #0x28]
+	ldr d137?, [x129?, #0x38]
+	mov v3.8b, v137?.8b
+	mov v2.8b, v136?.8b
+	mov x3, x135?
+	mov x2, x134?
+	mov v1.8b, v133?.8b
+	mov v0.8b, v132?.8b
+	mov x1, x131?
+	mov x0, x130?
 	ret
 `,
 			afterFinalizeARM64: `
@@ -1974,29 +1966,21 @@ L1 (SSA Block: blk0):
 	orr x27, xzr, #0x20
 	str x27, [sp, #-0x10]!
 	str x1, [sp, #0x10]
-	ldr x8, [x1, #0x8]
-	ldr w8, [x8, #0x8]
+	ldr w8, [x1, #0x8]
 	str w8, [sp, #0x2c]
-	ldr x9, [x1, #0x10]
-	ldr x9, [x9, #0x8]
+	ldr x9, [x1, #0x18]
 	str x9, [sp, #0x24]
-	ldr x10, [x1, #0x18]
-	ldr s8, [x10, #0x8]
+	ldr s8, [x1, #0x28]
 	str s8, [sp, #0x20]
-	ldr x10, [x1, #0x20]
-	ldr d9, [x10, #0x8]
+	ldr d9, [x1, #0x38]
 	str d9, [sp, #0x18]
 	str x1, [x0, #0x8]
 	bl f1
 	ldr x8, [sp, #0x10]
-	ldr x9, [x8, #0x8]
-	ldr w9, [x9, #0x8]
-	ldr x10, [x8, #0x10]
-	ldr x10, [x10, #0x8]
-	ldr x11, [x8, #0x18]
-	ldr s8, [x11, #0x8]
-	ldr x8, [x8, #0x20]
-	ldr d9, [x8, #0x8]
+	ldr w9, [x8, #0x8]
+	ldr x10, [x8, #0x18]
+	ldr s8, [x8, #0x28]
+	ldr d9, [x8, #0x38]
 	mov v3.8b, v9.8b
 	mov v2.8b, v8.8b
 	mov x3, x10
@@ -2022,36 +2006,28 @@ L1 (SSA Block: blk0):
 			afterLoweringARM64: `
 L1 (SSA Block: blk0):
 	mov x129?, x1
-	ldr x131?, [x129?, #0x8]
-	orr w141?, wzr, #0x1
-	str w141?, [x131?, #0x8]
-	ldr x133?, [x129?, #0x10]
-	orr x140?, xzr, #0x2
-	str x140?, [x133?, #0x8]
-	ldr x135?, [x129?, #0x18]
-	ldr s139?, #8; b 8; data.f32 3.000000
-	str s139?, [x135?, #0x8]
-	ldr x137?, [x129?, #0x20]
-	ldr d138?, #8; b 16; data.f64 4.000000
-	str d138?, [x137?, #0x8]
+	orr w137?, wzr, #0x1
+	str w137?, [x129?, #0x8]
+	orr x136?, xzr, #0x2
+	str x136?, [x129?, #0x18]
+	ldr s135?, #8; b 8; data.f32 3.000000
+	str s135?, [x129?, #0x28]
+	ldr d134?, #8; b 16; data.f64 4.000000
+	str d134?, [x129?, #0x38]
 	ret
 `,
 			afterFinalizeARM64: `
 L1 (SSA Block: blk0):
 	stp x30, xzr, [sp, #-0x10]!
 	str xzr, [sp, #-0x10]!
-	ldr x8, [x1, #0x8]
-	orr w9, wzr, #0x1
-	str w9, [x8, #0x8]
-	ldr x8, [x1, #0x10]
-	orr x9, xzr, #0x2
-	str x9, [x8, #0x8]
-	ldr x8, [x1, #0x18]
+	orr w8, wzr, #0x1
+	str w8, [x1, #0x8]
+	orr x8, xzr, #0x2
+	str x8, [x1, #0x18]
 	ldr s8, #8; b 8; data.f32 3.000000
-	str s8, [x8, #0x8]
-	ldr x8, [x1, #0x20]
+	str s8, [x1, #0x28]
 	ldr d8, #8; b 16; data.f64 4.000000
-	str d8, [x8, #0x8]
+	str d8, [x1, #0x38]
 	add sp, sp, #0x10
 	ldr x30, [sp], #0x10
 	ret

internal/engine/wazevo/frontend/frontend_test.go

@@ -1181,15 +1181,11 @@ blk0: (exec_ctx:i64, module_ctx:i64, v2:i32)
 			m:    testcases.GlobalsGet.Module,
 			exp: `
 blk0: (exec_ctx:i64, module_ctx:i64)
-	v2:i64 = Load module_ctx, 0x8
-	v3:i32 = Load v2, 0x8
-	v4:i64 = Load module_ctx, 0x10
-	v5:i64 = Load v4, 0x8
-	v6:i64 = Load module_ctx, 0x18
-	v7:f32 = Load v6, 0x8
-	v8:i64 = Load module_ctx, 0x20
-	v9:f64 = Load v8, 0x8
-	Jump blk_ret, v3, v5, v7, v9
+	v2:i32 = Load module_ctx, 0x8
+	v3:i64 = Load module_ctx, 0x18
+	v4:f32 = Load module_ctx, 0x28
+	v5:f64 = Load module_ctx, 0x38
+	Jump blk_ret, v2, v3, v4, v5
 `,
 		},
 		{
@@ -1198,18 +1194,14 @@ blk0: (exec_ctx:i64, module_ctx:i64)
 			exp: `
 blk0: (exec_ctx:i64, module_ctx:i64)
 	v2:i32 = Iconst_32 0x1
-	v3:i64 = Load module_ctx, 0x8
-	Store v2, v3, 0x8
-	v4:i64 = Iconst_64 0x2
-	v5:i64 = Load module_ctx, 0x10
-	Store v4, v5, 0x8
-	v6:f32 = F32const 3.000000
-	v7:i64 = Load module_ctx, 0x18
-	Store v6, v7, 0x8
-	v8:f64 = F64const 4.000000
-	v9:i64 = Load module_ctx, 0x20
-	Store v8, v9, 0x8
-	Jump blk_ret, v2, v4, v6, v8
+	Store v2, module_ctx, 0x8
+	v3:i64 = Iconst_64 0x2
+	Store v3, module_ctx, 0x18
+	v4:f32 = F32const 3.000000
+	Store v4, module_ctx, 0x28
+	v5:f64 = F64const 4.000000
+	Store v5, module_ctx, 0x38
+	Jump blk_ret, v2, v3, v4, v5
 `,
 		},
 		{
@@ -1220,50 +1212,34 @@ signatures:
 	sig1: i64i64_v
 
 blk0: (exec_ctx:i64, module_ctx:i64)
-	v2:i64 = Load module_ctx, 0x8
-	v3:i32 = Load v2, 0x8
-	v4:i64 = Load module_ctx, 0x10
-	v5:i64 = Load v4, 0x8
-	v6:i64 = Load module_ctx, 0x18
-	v7:f32 = Load v6, 0x8
-	v8:i64 = Load module_ctx, 0x20
-	v9:f64 = Load v8, 0x8
+	v2:i32 = Load module_ctx, 0x8
+	v3:i64 = Load module_ctx, 0x18
+	v4:f32 = Load module_ctx, 0x28
+	v5:f64 = Load module_ctx, 0x38
 	Store module_ctx, exec_ctx, 0x8
 	Call f1:sig1, exec_ctx, module_ctx
-	v10:i64 = Load module_ctx, 0x8
-	v11:i32 = Load v10, 0x8
-	v12:i64 = Load module_ctx, 0x10
-	v13:i64 = Load v12, 0x8
-	v14:i64 = Load module_ctx, 0x18
-	v15:f32 = Load v14, 0x8
-	v16:i64 = Load module_ctx, 0x20
-	v17:f64 = Load v16, 0x8
-	Jump blk_ret, v3, v5, v7, v9, v11, v13, v15, v17
+	v6:i32 = Load module_ctx, 0x8
+	v7:i64 = Load module_ctx, 0x18
+	v8:f32 = Load module_ctx, 0x28
+	v9:f64 = Load module_ctx, 0x38
+	Jump blk_ret, v2, v3, v4, v5, v6, v7, v8, v9
 `,
 			expAfterOpt: `
 signatures:
 	sig1: i64i64_v
 
 blk0: (exec_ctx:i64, module_ctx:i64)
-	v2:i64 = Load module_ctx, 0x8
-	v3:i32 = Load v2, 0x8
-	v4:i64 = Load module_ctx, 0x10
-	v5:i64 = Load v4, 0x8
-	v6:i64 = Load module_ctx, 0x18
-	v7:f32 = Load v6, 0x8
-	v8:i64 = Load module_ctx, 0x20
-	v9:f64 = Load v8, 0x8
+	v2:i32 = Load module_ctx, 0x8
+	v3:i64 = Load module_ctx, 0x18
+	v4:f32 = Load module_ctx, 0x28
+	v5:f64 = Load module_ctx, 0x38
 	Store module_ctx, exec_ctx, 0x8
 	Call f1:sig1, exec_ctx, module_ctx
-	v10:i64 = Load module_ctx, 0x8
-	v11:i32 = Load v10, 0x8
-	v12:i64 = Load module_ctx, 0x10
-	v13:i64 = Load v12, 0x8
-	v14:i64 = Load module_ctx, 0x18
-	v15:f32 = Load v14, 0x8
-	v16:i64 = Load module_ctx, 0x20
-	v17:f64 = Load v16, 0x8
-	Jump blk_ret, v3, v5, v7, v9, v11, v13, v15, v17
+	v6:i32 = Load module_ctx, 0x8
+	v7:i64 = Load module_ctx, 0x18
+	v8:f32 = Load module_ctx, 0x28
+	v9:f64 = Load module_ctx, 0x38
+	Jump blk_ret, v2, v3, v4, v5, v6, v7, v8, v9
 `,
 		},
 		{

internal/engine/wazevo/frontend/lower.go

@@ -3459,21 +3459,25 @@ func (c *Compiler) reloadMemoryBaseLen() {
 	c.clearSafeBounds()
 }
 
-// globalInstanceValueOffset is the offsetOf .Value field of wasm.GlobalInstance.
-const globalInstanceValueOffset = 8
-
 func (c *Compiler) setWasmGlobalValue(index wasm.Index, v ssa.Value) {
 	variable := c.globalVariables[index]
-	instanceOffset := c.offset.GlobalInstanceOffset(index)
+	opaqueOffset := c.offset.GlobalInstanceOffset(index)
 
 	builder := c.ssaBuilder
-	loadGlobalInstPtr := builder.AllocateInstruction()
-	loadGlobalInstPtr.AsLoad(c.moduleCtxPtrValue, uint32(instanceOffset), ssa.TypeI64)
-	builder.InsertInstruction(loadGlobalInstPtr)
+	if index < c.m.ImportGlobalCount {
+		loadGlobalInstPtr := builder.AllocateInstruction()
+		loadGlobalInstPtr.AsLoad(c.moduleCtxPtrValue, uint32(opaqueOffset), ssa.TypeI64)
+		builder.InsertInstruction(loadGlobalInstPtr)
 
-	store := builder.AllocateInstruction()
-	store.AsStore(ssa.OpcodeStore, v, loadGlobalInstPtr.Return(), uint32(globalInstanceValueOffset))
-	builder.InsertInstruction(store)
+		store := builder.AllocateInstruction()
+		store.AsStore(ssa.OpcodeStore, v, loadGlobalInstPtr.Return(), uint32(0))
+		builder.InsertInstruction(store)
+
+	} else {
+		store := builder.AllocateInstruction()
+		store.AsStore(ssa.OpcodeStore, v, c.moduleCtxPtrValue, uint32(opaqueOffset))
+		builder.InsertInstruction(store)
+	}
 
 	// The value has changed to `v`, so we record it.
 	builder.DefineVariableInCurrentBB(variable, v)
@@ -3482,7 +3486,7 @@ func (c *Compiler) setWasmGlobalValue(index wasm.Index, v ssa.Value) {
 func (c *Compiler) getWasmGlobalValue(index wasm.Index, forceLoad bool) ssa.Value {
 	variable := c.globalVariables[index]
 	typ := c.globalVariablesTypes[index]
-	instanceOffset := c.offset.GlobalInstanceOffset(index)
+	opaqueOffset := c.offset.GlobalInstanceOffset(index)
 
 	builder := c.ssaBuilder
 	if !forceLoad {
@@ -3491,16 +3495,21 @@ func (c *Compiler) getWasmGlobalValue(index wasm.Index, forceLoad bool) ssa.Valu
 		}
 	}
 
-	loadGlobalInstPtr := builder.AllocateInstruction()
-	loadGlobalInstPtr.AsLoad(c.moduleCtxPtrValue, uint32(instanceOffset), ssa.TypeI64)
-	builder.InsertInstruction(loadGlobalInstPtr)
+	var load *ssa.Instruction
+	if index < c.m.ImportGlobalCount {
+		loadGlobalInstPtr := builder.AllocateInstruction()
+		loadGlobalInstPtr.AsLoad(c.moduleCtxPtrValue, uint32(opaqueOffset), ssa.TypeI64)
+		builder.InsertInstruction(loadGlobalInstPtr)
+		load = builder.AllocateInstruction().
+			AsLoad(loadGlobalInstPtr.Return(), uint32(0), typ)
+	} else {
+		load = builder.AllocateInstruction().
+			AsLoad(c.moduleCtxPtrValue, uint32(opaqueOffset), typ)
+	}
 
-	load := builder.AllocateInstruction()
-	load.AsLoad(loadGlobalInstPtr.Return(), uint32(globalInstanceValueOffset), typ)
-	builder.InsertInstruction(load)
-	ret := load.Return()
-	builder.DefineVariableInCurrentBB(variable, ret)
-	return ret
+	v := load.Insert(builder).Return()
+	builder.DefineVariableInCurrentBB(variable, v)
+	return v
 }
 
 const (

internal/engine/wazevo/frontend/lower_test.go

@@ -9,8 +9,6 @@ import (
 )
 
 func Test_Offsets(t *testing.T) {
-	var globalInstance wasm.GlobalInstance
-	require.Equal(t, int(unsafe.Offsetof(globalInstance.Val)), globalInstanceValueOffset)
 	var memInstance wasm.MemoryInstance
 	require.Equal(t, int(unsafe.Offsetof(memInstance.Buffer)), memoryInstanceBufOffset)
 	var tableInstance wasm.TableInstance