runtime: Mildly over-allocate when growing slices

This refactors runtime.sliceAppend() and runtime.sliceGrow() so, not
only sliceAppend() uses sliceGrow() if needed, sliceGrow() only mildly
over-allocates a new buffer using a sequence similar to what CPython
uses for list objects.

Signed-off-by: L. Pereira <[email protected]>

Author: lpereira

src/runtime/slice.go

@@ -8,37 +8,19 @@ import (
 
 // Builtin append(src, elements...) function: append elements to src and return
 // the modified (possibly expanded) slice.
-func sliceAppend(srcBuf, elemsBuf unsafe.Pointer, srcLen, srcCap, elemsLen uintptr, elemSize uintptr) (unsafe.Pointer, uintptr, uintptr) {
-	if elemsLen == 0 {
-		// Nothing to append, return the input slice.
-		return srcBuf, srcLen, srcCap
-	}
+func sliceAppend(srcBuf, elemsBuf unsafe.Pointer, srcLen, srcCap, elemsLen, elemSize uintptr) (unsafe.Pointer, uintptr, uintptr) {
+	newLen := srcLen + elemsLen
 
-	if srcLen+elemsLen > srcCap {
-		// Slice does not fit, allocate a new buffer that's large enough.
-		srcCap = srcCap * 2
-		if srcCap == 0 { // e.g. zero slice
-			srcCap = 1
-		}
-		for srcLen+elemsLen > srcCap {
-			// This algorithm may be made more memory-efficient: don't multiply
-			// by two but by 1.5 or something. As far as I can see, that's
-			// allowed by the Go language specification (but may be observed by
-			// programs).
-			srcCap *= 2
-		}
-		buf := alloc(srcCap*elemSize, nil)
+	if elemsLen != 0 {
+		// Allocate a new slice with capacity for elemsLen more elements, if necessary;
+		// otherwise, reuse the passed slice.
+		srcBuf, srcLen, srcCap = sliceGrow(srcBuf, srcLen, srcCap, newLen, elemSize)
 
-		// Copy the old slice to the new slice.
-		if srcLen != 0 {
-			memmove(buf, srcBuf, srcLen*elemSize)
-		}
-		srcBuf = buf
+		// Append the new elements in-place.
+		memmove(unsafe.Add(srcBuf, srcLen*elemSize), elemsBuf, elemsLen*elemSize)
 	}
 
-	// The slice fits (after possibly allocating a new one), append it in-place.
-	memmove(unsafe.Add(srcBuf, srcLen*elemSize), elemsBuf, elemsLen*elemSize)
-	return srcBuf, srcLen + elemsLen, srcCap
+	return srcBuf, newLen, srcCap
 }
 
 // Builtin copy(dst, src) function: copy bytes from dst to src.
@@ -54,29 +36,28 @@ func sliceCopy(dst, src unsafe.Pointer, dstLen, srcLen uintptr, elemSize uintptr
 
 // sliceGrow returns a new slice with space for at least newCap elements
 func sliceGrow(oldBuf unsafe.Pointer, oldLen, oldCap, newCap, elemSize uintptr) (unsafe.Pointer, uintptr, uintptr) {
-
-	// TODO(dgryski): sliceGrow() and sliceAppend() should be refactored to share the base growth code.
-
 	if oldCap >= newCap {
 		// No need to grow, return the input slice.
 		return oldBuf, oldLen, oldCap
 	}
 
-	// allow nil slice
-	if oldCap == 0 {
-		oldCap++
-	}
+	// Mildly over-allocate using a similar integer sequence to what CPython uses when
+	// allocating lists.  This still helps amortizing repeated appends to this slice,
+	// but is much more conservative than doubling the capacity every time a reallocation
+	// is necessary.
+	// Sequence goes as follows: 0, 8, 16, 24, 32, 40, 48, 56, 64, 74, 80, ...
+	newCap = (newCap + (newCap >> 3) + 6) & ^uintptr(7)
 
-	// grow capacity
-	for oldCap < newCap {
-		oldCap *= 2
+	// Account for nil slices.
+	if newCap == 0 {
+		newCap++
 	}
 
-	buf := alloc(oldCap*elemSize, nil)
+	buf := alloc(newCap*elemSize, nil)
 	if oldLen > 0 {
 		// copy any data to new slice
 		memmove(buf, oldBuf, oldLen*elemSize)
 	}
 
-	return buf, oldLen, oldCap
+	return buf, oldLen, newCap
 }

testdata/slice.txt

@@ -8,11 +8,11 @@ bar: len=3 cap=5 data: 1 2 4
 slice is nil? true true
 grow: len=0 cap=0 data:
 grow: len=1 cap=1 data: 42
-grow: len=3 cap=4 data: 42 -1 -2
+grow: len=3 cap=8 data: 42 -1 -2
 grow: len=7 cap=8 data: 42 -1 -2 1 2 4 5
 grow: len=7 cap=8 data: 42 -1 -2 1 2 4 5
 grow: len=14 cap=16 data: 42 -1 -2 1 2 4 5 42 -1 -2 1 2 4 5
-bytes: len=6 cap=6 data: 1 2 3 102 111 111
+bytes: len=6 cap=8 data: 1 2 3 102 111 111
 slice to array pointer: 1 -2 20 4
 unsafe.Add array: 1 5 8 4
 unsafe.Slice array: 3 3 9 15 4