runtime (gc_blocks.go): make sweep branchless

Instead of looping over each block, we can use bit hacks to operate on an entire state byte.
This deinterleaves the state bits in order to enable these tricks.

Author: niaow

builder/sizes_test.go

@@ -42,9 +42,9 @@ func TestBinarySize(t *testing.T) {
 	// This is a small number of very diverse targets that we want to test.
 	tests := []sizeTest{
 		// microcontrollers
-		{"hifive1b", "examples/echo", 3568, 280, 0, 2268},
-		{"microbit", "examples/serial", 2630, 342, 8, 2272},
-		{"wioterminal", "examples/pininterrupt", 7175, 1493, 116, 6912},
+		{"hifive1b", "examples/echo", 3524, 296, 0, 2268},
+		{"microbit", "examples/serial", 2598, 358, 8, 2272},
+		{"wioterminal", "examples/pininterrupt", 7095, 1509, 116, 6912},
 
 		// TODO: also check wasm. Right now this is difficult, because
 		// wasm binaries are run through wasm-opt and therefore the

src/runtime/gc_blocks.go

@@ -71,19 +71,20 @@ var zeroSizedAlloc uint8
 type blockState uint8
 
 const (
-	blockStateFree blockState = 0 // 00
-	blockStateHead blockState = 1 // 01
-	blockStateTail blockState = 2 // 10
-	blockStateMark blockState = 3 // 11
-	blockStateMask blockState = 3 // 11
+	blockStateLow  blockState = 1
+	blockStateHigh blockState = 1 << blocksPerStateByte
+
+	blockStateFree blockState = 0
+	blockStateHead blockState = blockStateLow
+	blockStateTail blockState = blockStateHigh
+	blockStateMark blockState = blockStateLow | blockStateHigh
+	blockStateMask blockState = blockStateLow | blockStateHigh
 )
 
+const blockStateEach = 1<<blocksPerStateByte - 1
+
 // The byte value of a block where every block is a 'tail' block.
-const blockStateByteAllTails = 0 |
-	uint8(blockStateTail<<(stateBits*3)) |
-	uint8(blockStateTail<<(stateBits*2)) |
-	uint8(blockStateTail<<(stateBits*1)) |
-	uint8(blockStateTail<<(stateBits*0))
+const blockStateByteAllTails = byte(blockStateTail) * blockStateEach
 
 // String returns a human-readable version of the block state, for debugging.
 func (s blockState) String() string {
@@ -180,7 +181,7 @@ func (b gcBlock) stateByte() byte {
 // Return the block state given a state byte. The state byte must have been
 // obtained using b.stateByte(), otherwise the result is incorrect.
 func (b gcBlock) stateFromByte(stateByte byte) blockState {
-	return blockState(stateByte>>((b%blocksPerStateByte)*stateBits)) & blockStateMask
+	return blockState(stateByte>>(b%blocksPerStateByte)) & blockStateMask
 }
 
 // State returns the current block state.
@@ -193,38 +194,12 @@ func (b gcBlock) state() blockState {
 // from head to mark.
 func (b gcBlock) setState(newState blockState) {
 	stateBytePtr := (*uint8)(unsafe.Add(metadataStart, b/blocksPerStateByte))
-	*stateBytePtr |= uint8(newState << ((b % blocksPerStateByte) * stateBits))
+	*stateBytePtr |= uint8(newState << (b % blocksPerStateByte))
 	if gcAsserts && b.state() != newState {
 		runtimePanic("gc: setState() was not successful")
 	}
 }
 
-// markFree sets the block state to free, no matter what state it was in before.
-func (b gcBlock) markFree() {
-	stateBytePtr := (*uint8)(unsafe.Add(metadataStart, b/blocksPerStateByte))
-	*stateBytePtr &^= uint8(blockStateMask << ((b % blocksPerStateByte) * stateBits))
-	if gcAsserts && b.state() != blockStateFree {
-		runtimePanic("gc: markFree() was not successful")
-	}
-	if gcAsserts {
-		*(*[wordsPerBlock]uintptr)(unsafe.Pointer(b.address())) = [wordsPerBlock]uintptr{}
-	}
-}
-
-// unmark changes the state of the block from mark to head. It must be marked
-// before calling this function.
-func (b gcBlock) unmark() {
-	if gcAsserts && b.state() != blockStateMark {
-		runtimePanic("gc: unmark() on a block that is not marked")
-	}
-	clearMask := blockStateMask ^ blockStateHead // the bits to clear from the state
-	stateBytePtr := (*uint8)(unsafe.Add(metadataStart, b/blocksPerStateByte))
-	*stateBytePtr &^= uint8(clearMask << ((b % blocksPerStateByte) * stateBits))
-	if gcAsserts && b.state() != blockStateHead {
-		runtimePanic("gc: unmark() was not successful")
-	}
-}
-
 // objHeader is a structure prepended to every heap object to hold metadata.
 type objHeader struct {
 	// next is the next object to scan after this.
@@ -633,36 +608,69 @@ func markRoot(addr, root uintptr) {
 // Sweep goes through all memory and frees unmarked memory.
 // It returns how many bytes are free in the heap after the sweep.
 func sweep() (freeBytes uintptr) {
-	freeCurrentObject := false
-	var freed uint64
-	for block := gcBlock(0); block < endBlock; block++ {
-		switch block.state() {
-		case blockStateHead:
-			// Unmarked head. Free it, including all tail blocks following it.
-			block.markFree()
-			freeCurrentObject = true
-			gcFrees++
-			freed++
-		case blockStateTail:
-			if freeCurrentObject {
-				// This is a tail object following an unmarked head.
-				// Free it now.
-				block.markFree()
-				freed++
-			}
-		case blockStateMark:
-			// This is a marked object. The next tail blocks must not be freed,
-			// but the mark bit must be removed so the next GC cycle will
-			// collect this object if it is unreferenced then.
-			block.unmark()
-			freeCurrentObject = false
-		case blockStateFree:
-			freeBytes += bytesPerBlock
-		}
-	}
-	gcFreedBlocks += freed
-	freeBytes += uintptr(freed) * bytesPerBlock
-	return
+	endBlock := endBlock
+	metadataEnd := unsafe.Add(metadataStart, (endBlock+(blocksPerStateByte-1))/blocksPerStateByte)
+	var oldFreeBlocks, freedHeads, freedTails uintptr
+	var carry byte
+	// Pre-subtract the blocks that do not actually exist from oldFreeBlocks.
+	oldFreeBlocks -= (blocksPerStateByte - 1) - uintptr(endBlock+(blocksPerStateByte-1))%blocksPerStateByte
+	for meta := metadataStart; meta != metadataEnd; meta = unsafe.Add(meta, 1) {
+		// Fetch the state byte.
+		stateBytePtr := (*byte)(unsafe.Pointer(meta))
+		stateByte := *stateBytePtr
+
+		// Count existing free blocks in the state byte.
+		lowState := stateByte & blockStateEach
+		highState := stateByte >> blocksPerStateByte
+		freeBlocks := lowState | highState
+		oldFreeBlocks += uintptr(count4LUT[freeBlocks])
+
+		// Count unmarked heads in the state byte.
+		unmarkedHeads := lowState &^ highState
+		freedHeads += uintptr(count4LUT[unmarkedHeads])
+
+		// Identify and seperate live and free tails.
+		// Adding 1 to a run of bits will clear the run.
+		// We can use this to clear tails after a freed head.
+		tails := highState &^ lowState
+		tailClear := tails + (unmarkedHeads << 1) + carry
+		carry = tailClear >> blocksPerStateByte
+		freedTails += uintptr(count4LUT[tails&^tailClear])
+		tails &= tailClear
+
+		// Construct the new state byte.
+		markedHeads := highState & lowState
+		*stateBytePtr = markedHeads | (tails << blocksPerStateByte)
+	}
+
+	// Update the GC metrics.
+	gcFrees += uint64(freedHeads)
+	freedBlocks := freedHeads + freedTails
+	gcFreedBlocks += uint64(freedBlocks)
+	freeBlocks := oldFreeBlocks + freedBlocks
+
+	return freeBlocks * bytesPerBlock
+}
+
+// count4LUT is a lookup table used to count set bits in a 4-bit mask.
+// TODO: replace with popcnt when available
+var count4LUT = [16]uint8{
+	0b0000: 0,
+	0b0001: 1,
+	0b0010: 1,
+	0b0011: 2,
+	0b0100: 1,
+	0b0101: 2,
+	0b0110: 2,
+	0b0111: 3,
+	0b1000: 1,
+	0b1001: 2,
+	0b1010: 2,
+	0b1011: 3,
+	0b1100: 2,
+	0b1101: 3,
+	0b1110: 3,
+	0b1111: 4,
 }
 
 // dumpHeap can be used for debugging purposes. It dumps the state of each heap