ssa: memory usage optimization around lastDefinitions (#2233)

This avoids the unnecessary duplicated zero initial value definitions for Wasm function locals. As a result, for certain binaries which have huge number of locals like Zig standard libraries, we can see 1~10% improvements in compilation performance without impacting other cases like the below. ### Zig Stdlib tests ``` │ old_zig.txt │ new_zig.txt │ │ sec/op │ sec/op vs base │ Zig/Compile/test-opt.wasm-10 4.466 ± 2% 4.399 ± 0% -1.50% (p=0.007 n=7) Zig/Run/test-opt.wasm-10 18.86 ± 0% 18.84 ± 0% ~ (p=0.535 n=7) Zig/Compile/test.wasm-10 5.684 ± 1% 5.084 ± 1% -10.55% (p=0.001 n=7) Zig/Run/test.wasm-10 19.28 ± 1% 19.25 ± 1% ~ (p=0.535 n=7) geomean 9.802 9.490 -3.18% │ old_zig.txt │ new_zig.txt │ │ B/op │ B/op vs base │ Zig/Compile/test-opt.wasm-10 395.4Mi ± 0% 396.7Mi ± 0% +0.32% (p=0.001 n=7) Zig/Run/test-opt.wasm-10 741.7Mi ± 0% 741.7Mi ± 0% ~ (p=0.941 n=7) Zig/Compile/test.wasm-10 671.6Mi ± 0% 660.1Mi ± 0% -1.72% (p=0.001 n=7) Zig/Run/test.wasm-10 1.296Gi ± 0% 1.296Gi ± 0% ~ (p=0.363 n=7) geomean 715.1Mi 712.6Mi -0.35% │ old_zig.txt │ new_zig.txt │ │ allocs/op │ allocs/op vs base │ Zig/Compile/test-opt.wasm-10 363.1k ± 0% 363.2k ± 0% ~ (p=0.456 n=7) Zig/Run/test-opt.wasm-10 51.58k ± 0% 51.58k ± 0% ~ (p=0.812 n=7) Zig/Compile/test.wasm-10 516.3k ± 0% 515.5k ± 0% -0.16% (p=0.001 n=7) Zig/Run/test.wasm-10 2.156M ± 0% 2.156M ± 0% ~ (p=0.171 n=7) geomean 380.0k 379.8k -0.03% ``` ### wazero compiled as wasip1 binary ``` goos: darwin goarch: arm64 pkg: github.com/tetratelabs/wazero │ old.txt │ new.txt │ │ sec/op │ sec/op vs base │ Compilation-10 2.418 ± 0% 2.421 ± 0% ~ (p=0.383 n=7) │ old.txt │ new.txt │ │ B/op │ B/op vs base │ Compilation-10 339.9Mi ± 0% 339.9Mi ± 0% -0.01% (p=0.001 n=7) │ old.txt │ new.txt │ │ allocs/op │ allocs/op vs base │ Compilation-10 603.9k ± 0% 604.0k ± 0% ~ (p=0.620 n=7) ``` ### TinyGo ``` goos: darwin goarch: arm64 pkg: github.com/tetratelabs/wazero/internal/integration_test/stdlibs │ old_tinygo.txt │ new_tinygo.txt │ │ sec/op │ sec/op vs base │ TinyGo/Compile/container_heap.test-10 414.1m ± 0% 413.1m ± 1% ~ (p=0.165 n=7) TinyGo/Run/container_heap.test-10 14.48m ± 1% 14.54m ± 1% ~ (p=0.383 n=7) TinyGo/Compile/container_list.test-10 413.1m ± 1% 412.1m ± 1% ~ (p=0.620 n=7) TinyGo/Run/container_list.test-10 14.29m ± 0% 14.33m ± 1% ~ (p=0.259 n=7) TinyGo/Compile/container_ring.test-10 406.8m ± 1% 407.1m ± 1% ~ (p=0.902 n=7) TinyGo/Run/container_ring.test-10 14.28m ± 1% 14.29m ± 1% ~ (p=0.318 n=7) TinyGo/Compile/crypto_des.test-10 421.3m ± 0% 422.4m ± 1% ~ (p=0.209 n=7) TinyGo/Run/crypto_des.test-10 18.32m ± 1% 18.38m ± 2% ~ (p=0.209 n=7) TinyGo/Compile/crypto_md5.test-10 418.7m ± 1% 419.8m ± 0% ~ (p=0.318 n=7) TinyGo/Run/crypto_md5.test-10 21.78m ± 5% 20.65m ± 2% ~ (p=0.165 n=7) TinyGo/Compile/crypto_rc4.test-10 421.8m ± 4% 404.3m ± 1% -4.16% (p=0.001 n=7) TinyGo/Run/crypto_rc4.test-10 162.6m ± 1% 162.3m ± 1% ~ (p=0.318 n=7) TinyGo/Compile/crypto_sha1.test-10 420.5m ± 1% 419.4m ± 1% ~ (p=0.535 n=7) TinyGo/Run/crypto_sha1.test-10 16.04m ± 1% 16.12m ± 1% +0.50% (p=0.038 n=7) TinyGo/Compile/crypto_sha256.test-10 426.8m ± 0% 427.4m ± 1% ~ (p=1.000 n=7) TinyGo/Run/crypto_sha256.test-10 16.24m ± ∞ ¹ 16.38m ± ∞ ¹ ~ (p=1.000 n=1) ² geomean 95.31m 94.88m -0.45% ¹ need >= 6 samples for confidence interval at level 0.95 ² need >= 4 samples to detect a difference at alpha level 0.05 │ old_tinygo.txt │ new_tinygo.txt │ │ B/op │ B/op vs base │ TinyGo/Compile/container_heap.test-10 48.58Mi ± 0% 48.55Mi ± 0% -0.06% (p=0.001 n=7) TinyGo/Run/container_heap.test-10 16.63Mi ± 0% 16.63Mi ± 0% ~ (p=0.620 n=7) TinyGo/Compile/container_list.test-10 48.56Mi ± 0% 48.54Mi ± 0% -0.05% (p=0.001 n=7) TinyGo/Run/container_list.test-10 16.40Mi ± 0% 16.40Mi ± 0% ~ (p=0.535 n=7) TinyGo/Compile/container_ring.test-10 47.81Mi ± 0% 47.78Mi ± 0% -0.05% (p=0.001 n=7) TinyGo/Run/container_ring.test-10 16.30Mi ± 0% 16.30Mi ± 0% ~ (p=0.871 n=7) TinyGo/Compile/crypto_des.test-10 48.70Mi ± 0% 48.67Mi ± 0% -0.05% (p=0.001 n=7) TinyGo/Run/crypto_des.test-10 16.76Mi ± 0% 16.76Mi ± 0% ~ (p=0.119 n=7) TinyGo/Compile/crypto_md5.test-10 48.75Mi ± 0% 48.73Mi ± 0% -0.03% (p=0.001 n=7) TinyGo/Run/crypto_md5.test-10 44.97Mi ± 0% 44.97Mi ± 0% ~ (p=0.333 n=7) TinyGo/Compile/crypto_rc4.test-10 47.79Mi ± 0% 47.76Mi ± 0% -0.06% (p=0.001 n=7) TinyGo/Run/crypto_rc4.test-10 29.28Mi ± 0% 29.28Mi ± 0% ~ (p=0.274 n=7) TinyGo/Compile/crypto_sha1.test-10 48.99Mi ± 0% 48.97Mi ± 0% -0.03% (p=0.001 n=7) TinyGo/Run/crypto_sha1.test-10 17.44Mi ± 0% 17.44Mi ± 0% ~ (p=0.456 n=7) TinyGo/Compile/crypto_sha256.test-10 48.82Mi ± 0% 48.81Mi ± 0% -0.03% (p=0.004 n=7) TinyGo/Run/crypto_sha256.test-10 17.53Mi ± ∞ ¹ 17.53Mi ± ∞ ¹ ~ (p=1.000 n=1) ² geomean 31.45Mi 31.45Mi -0.02% ¹ need >= 6 samples for confidence interval at level 0.95 ² need >= 4 samples to detect a difference at alpha level 0.05 │ old_tinygo.txt │ new_tinygo.txt │ │ allocs/op │ allocs/op vs base │ TinyGo/Compile/container_heap.test-10 83.64k ± 0% 83.63k ± 0% ~ (p=0.365 n=7) TinyGo/Run/container_heap.test-10 374.9k ± 0% 374.9k ± 0% ~ (p=1.000 n=7) TinyGo/Compile/container_list.test-10 83.38k ± 0% 83.39k ± 0% ~ (p=0.805 n=7) TinyGo/Run/container_list.test-10 370.0k ± 0% 370.0k ± 0% ~ (p=0.633 n=7) TinyGo/Compile/container_ring.test-10 83.30k ± 0% 83.35k ± 0% ~ (p=1.000 n=7) TinyGo/Run/container_ring.test-10 367.6k ± 0% 367.6k ± 0% ~ (p=0.617 n=7) TinyGo/Compile/crypto_des.test-10 83.67k ± 0% 83.70k ± 0% ~ (p=0.805 n=7) TinyGo/Run/crypto_des.test-10 378.1k ± 0% 378.1k ± 0% ~ (p=0.078 n=7) TinyGo/Compile/crypto_md5.test-10 83.81k ± 0% 83.84k ± 0% ~ (p=0.805 n=7) TinyGo/Run/crypto_md5.test-10 393.3k ± 0% 393.3k ± 0% ~ (p=0.690 n=7) TinyGo/Compile/crypto_rc4.test-10 83.39k ± 0% 83.33k ± 0% ~ (p=0.097 n=7) TinyGo/Run/crypto_rc4.test-10 367.1k ± 0% 367.1k ± 0% ~ (p=0.232 n=7) TinyGo/Compile/crypto_sha1.test-10 84.00k ± 0% 84.06k ± 0% ~ (p=0.154 n=7) TinyGo/Run/crypto_sha1.test-10 392.7k ± 0% 392.7k ± 0% ~ (p=1.000 n=7) TinyGo/Compile/crypto_sha256.test-10 83.85k ± 0% 83.86k ± 0% ~ (p=0.620 n=7) TinyGo/Run/crypto_sha256.test-10 394.5k ± ∞ ¹ 394.5k ± ∞ ¹ ~ (p=1.000 n=1) ² ``` Signed-off-by: Takeshi Yoneda <t.y.mathetake@gmail.com>
2024-06-07 07:57:15 -07:00
parent 0c5d97141e
commit f47fd2e46e
5 changed files with 44 additions and 29 deletions
--- a/internal/engine/wazevo/backend/backend_test.go
+++ b/internal/engine/wazevo/backend/backend_test.go
@@ -458,11 +458,11 @@ L2 (SSA Block: blk2):
 			name: "single_predecessor_local_refs", m: testcases.SinglePredecessorLocalRefs.Module,
 			afterLoweringARM64: `
 L1 (SSA Block: blk0):
-	mov x132?, xzr
-	cbz w132?, (L2)
-L3 (SSA Block: blk1):
 	mov x131?, xzr
-	mov x0, x131?
+	cbz w131?, (L2)
+L3 (SSA Block: blk1):
+	mov x130?, xzr
+	mov x0, x130?
 	ret
 L2 (SSA Block: blk2):
 L4 (SSA Block: blk3):
--- a/internal/engine/wazevo/frontend/frontend.go
+++ b/internal/engine/wazevo/frontend/frontend.go
@@ -301,26 +301,7 @@ func (c *Compiler) declareWasmLocals(entry ssa.BasicBlock) {
 		st := WasmTypeToSSAType(typ)
 		variable := c.ssaBuilder.DeclareVariable(st)
 		c.setWasmLocalVariable(wasm.Index(i)+localCount, variable)
-
-		zeroInst := c.ssaBuilder.AllocateInstruction()
-		switch st {
-		case ssa.TypeI32:
-			zeroInst.AsIconst32(0)
-		case ssa.TypeI64:
-			zeroInst.AsIconst64(0)
-		case ssa.TypeF32:
-			zeroInst.AsF32const(0)
-		case ssa.TypeF64:
-			zeroInst.AsF64const(0)
-		case ssa.TypeV128:
-			zeroInst.AsVconst(0, 0)
-		default:
-			panic("TODO: " + wasm.ValueTypeName(typ))
-		}
-
-		c.ssaBuilder.InsertInstruction(zeroInst)
-		value := zeroInst.Return()
-		c.ssaBuilder.DefineVariable(variable, value, entry)
+		c.ssaBuilder.InsertZeroValue(st)
 	}
 }

--- a/internal/engine/wazevo/frontend/frontend_test.go
+++ b/internal/engine/wazevo/frontend/frontend_test.go
@@ -375,13 +375,11 @@ blk3: () <-- (blk1)
 			exp: `
 blk0: (exec_ctx:i64, module_ctx:i64)
 	v2:i32 = Iconst_32 0x0
-	v3:i32 = Iconst_32 0x0
-	v4:i32 = Iconst_32 0x0
 	Brz v2, blk2
 	Jump blk1

 blk1: () <-- (blk0)
-	Return v4
+	Return v2

 blk2: () <-- (blk0)
 	Jump blk3
@@ -392,12 +390,11 @@ blk3: () <-- (blk2)
 			expAfterPasses: `
 blk0: (exec_ctx:i64, module_ctx:i64)
 	v2:i32 = Iconst_32 0x0
-	v4:i32 = Iconst_32 0x0
 	Brz v2, blk2
 	Jump fallthrough

 blk1: () <-- (blk0)
-	Return v4
+	Return v2

 blk2: () <-- (blk0)
 	Jump fallthrough
--- a/internal/engine/wazevo/ssa/builder.go
+++ b/internal/engine/wazevo/ssa/builder.go
@@ -127,7 +127,11 @@ type Builder interface {
 	// Idom returns the immediate dominator of the given BasicBlock.
 	Idom(blk BasicBlock) BasicBlock

+	// VarLengthPool returns the VarLengthPool of Value.
 	VarLengthPool() *wazevoapi.VarLengthPool[Value]
+
+	// InsertZeroValue inserts a zero value constant instruction of the given type.
+	InsertZeroValue(t Type)
 }

 // NewBuilder returns a new Builder implementation.
@@ -203,6 +207,32 @@ type builder struct {
 	donePostBlockLayoutPasses bool

 	currentSourceOffset SourceOffset
+
+	// zeros are the zero value constants for each type.
+	zeros [typeEnd]Value
+}
+
+// InsertZeroValue implements Builder.InsertZeroValue.
+func (b *builder) InsertZeroValue(t Type) {
+	if b.zeros[t].Valid() {
+		return
+	}
+	zeroInst := b.AllocateInstruction()
+	switch t {
+	case TypeI32:
+		zeroInst.AsIconst32(0)
+	case TypeI64:
+		zeroInst.AsIconst64(0)
+	case TypeF32:
+		zeroInst.AsF32const(0)
+	case TypeF64:
+		zeroInst.AsF64const(0)
+	case TypeV128:
+		zeroInst.AsVconst(0, 0)
+	default:
+		panic("TODO: " + t.String())
+	}
+	b.zeros[t] = zeroInst.Insert(b).Return()
 }

 func (b *builder) VarLengthPool() *wazevoapi.VarLengthPool[Value] {
@@ -218,6 +248,7 @@ func (b *builder) ReturnBlock() BasicBlock {
 func (b *builder) Init(s *Signature) {
 	b.nextVariable = 0
 	b.currentSignature = s
+	b.zeros = [typeEnd]Value{ValueInvalid, ValueInvalid, ValueInvalid, ValueInvalid, ValueInvalid, ValueInvalid}
 	resetBasicBlock(b.returnBlk)
 	b.instructionsPool.Reset()
 	b.basicBlocksPool.Reset()
@@ -486,6 +517,9 @@ func (b *builder) findValue(typ Type, variable Variable, blk *basicBlock) Value
 			value:    value,
 		})
 		return value
+	} else if blk.EntryBlock() {
+		// If this is the entry block, we reach the uninitialized variable which has zero value.
+		return b.zeros[b.definedVariableType(variable)]
 	}

 	if pred := blk.singlePred; pred != nil {
--- a/internal/engine/wazevo/ssa/type.go
+++ b/internal/engine/wazevo/ssa/type.go
@@ -21,6 +21,9 @@ const (

 	// TypeV128 represents 128-bit SIMD vectors.
 	TypeV128
+
+	// -- Do not add new types after this line. ----
+	typeEnd
 )

 // String implements fmt.Stringer.