1// Copyright 2016 The Go Authors. All rights reserved.
2// Use of this source code is governed by a BSD-style
3// license that can be found in the LICENSE file.
4
5// +build ignore
6
7// Generate tables for small malloc size classes.
8//
9// See malloc.go for overview.
10//
11// The size classes are chosen so that rounding an allocation
12// request up to the next size class wastes at most 12.5% (1.125x).
13//
14// Each size class has its own page count that gets allocated
15// and chopped up when new objects of the size class are needed.
16// That page count is chosen so that chopping up the run of
17// pages into objects of the given size wastes at most 12.5% (1.125x)
18// of the memory. It is not necessary that the cutoff here be
19// the same as above.
20//
21// The two sources of waste multiply, so the worst possible case
22// for the above constraints would be that allocations of some
23// size might have a 26.6% (1.266x) overhead.
24// In practice, only one of the wastes comes into play for a
25// given size (sizes < 512 waste mainly on the round-up,
26// sizes > 512 waste mainly on the page chopping).
27// For really small sizes, alignment constraints force the
28// overhead higher.
29
30package main
31
32import (
33	"bytes"
34	"flag"
35	"fmt"
36	"go/format"
37	"io"
38	"io/ioutil"
39	"log"
40	"os"
41)
42
43// Generate msize.go
44
45var stdout = flag.Bool("stdout", false, "write to stdout instead of sizeclasses.go")
46
47func main() {
48	flag.Parse()
49
50	var b bytes.Buffer
51	fmt.Fprintln(&b, "// Code generated by mksizeclasses.go; DO NOT EDIT.")
52	fmt.Fprintln(&b, "//go:generate go run mksizeclasses.go")
53	fmt.Fprintln(&b)
54	fmt.Fprintln(&b, "package runtime")
55	classes := makeClasses()
56
57	printComment(&b, classes)
58
59	printClasses(&b, classes)
60
61	out, err := format.Source(b.Bytes())
62	if err != nil {
63		log.Fatal(err)
64	}
65	if *stdout {
66		_, err = os.Stdout.Write(out)
67	} else {
68		err = ioutil.WriteFile("sizeclasses.go", out, 0666)
69	}
70	if err != nil {
71		log.Fatal(err)
72	}
73}
74
75const (
76	// Constants that we use and will transfer to the runtime.
77	maxSmallSize = 32 << 10
78	smallSizeDiv = 8
79	smallSizeMax = 1024
80	largeSizeDiv = 128
81	pageShift    = 13
82
83	// Derived constants.
84	pageSize = 1 << pageShift
85)
86
87type class struct {
88	size   int // max size
89	npages int // number of pages
90
91	mul    int
92	shift  uint
93	shift2 uint
94	mask   int
95}
96
97func powerOfTwo(x int) bool {
98	return x != 0 && x&(x-1) == 0
99}
100
101func makeClasses() []class {
102	var classes []class
103
104	classes = append(classes, class{}) // class #0 is a dummy entry
105
106	align := 8
107	for size := align; size <= maxSmallSize; size += align {
108		if powerOfTwo(size) { // bump alignment once in a while
109			if size >= 2048 {
110				align = 256
111			} else if size >= 128 {
112				align = size / 8
113			} else if size >= 16 {
114				align = 16 // required for x86 SSE instructions, if we want to use them
115			}
116		}
117		if !powerOfTwo(align) {
118			panic("incorrect alignment")
119		}
120
121		// Make the allocnpages big enough that
122		// the leftover is less than 1/8 of the total,
123		// so wasted space is at most 12.5%.
124		allocsize := pageSize
125		for allocsize%size > allocsize/8 {
126			allocsize += pageSize
127		}
128		npages := allocsize / pageSize
129
130		// If the previous sizeclass chose the same
131		// allocation size and fit the same number of
132		// objects into the page, we might as well
133		// use just this size instead of having two
134		// different sizes.
135		if len(classes) > 1 && npages == classes[len(classes)-1].npages && allocsize/size == allocsize/classes[len(classes)-1].size {
136			classes[len(classes)-1].size = size
137			continue
138		}
139		classes = append(classes, class{size: size, npages: npages})
140	}
141
142	// Increase object sizes if we can fit the same number of larger objects
143	// into the same number of pages. For example, we choose size 8448 above
144	// with 6 objects in 7 pages. But we can well use object size 9472,
145	// which is also 6 objects in 7 pages but +1024 bytes (+12.12%).
146	// We need to preserve at least largeSizeDiv alignment otherwise
147	// sizeToClass won't work.
148	for i := range classes {
149		if i == 0 {
150			continue
151		}
152		c := &classes[i]
153		psize := c.npages * pageSize
154		new_size := (psize / (psize / c.size)) &^ (largeSizeDiv - 1)
155		if new_size > c.size {
156			c.size = new_size
157		}
158	}
159
160	if len(classes) != 67 {
161		panic("number of size classes has changed")
162	}
163
164	for i := range classes {
165		computeDivMagic(&classes[i])
166	}
167
168	return classes
169}
170
171// computeDivMagic computes some magic constants to implement
172// the division required to compute object number from span offset.
173// n / c.size is implemented as n >> c.shift * c.mul >> c.shift2
174// for all 0 <= n < c.npages * pageSize
175func computeDivMagic(c *class) {
176	// divisor
177	d := c.size
178	if d == 0 {
179		return
180	}
181
182	// maximum input value for which the formula needs to work.
183	max := c.npages*pageSize - 1
184
185	if powerOfTwo(d) {
186		// If the size is a power of two, heapBitsForObject can divide even faster by masking.
187		// Compute this mask.
188		if max >= 1<<16 {
189			panic("max too big for power of two size")
190		}
191		c.mask = 1<<16 - d
192	}
193
194	// Compute pre-shift by factoring power of 2 out of d.
195	for d%2 == 0 {
196		c.shift++
197		d >>= 1
198		max >>= 1
199	}
200
201	// Find the smallest k that works.
202	// A small k allows us to fit the math required into 32 bits
203	// so we can use 32-bit multiplies and shifts on 32-bit platforms.
204nextk:
205	for k := uint(0); ; k++ {
206		mul := (int(1)<<k + d - 1) / d //  ⌈2^k / d⌉
207
208		// Test to see if mul works.
209		for n := 0; n <= max; n++ {
210			if n*mul>>k != n/d {
211				continue nextk
212			}
213		}
214		if mul >= 1<<16 {
215			panic("mul too big")
216		}
217		if uint64(mul)*uint64(max) >= 1<<32 {
218			panic("mul*max too big")
219		}
220		c.mul = mul
221		c.shift2 = k
222		break
223	}
224
225	// double-check.
226	for n := 0; n <= max; n++ {
227		if n*c.mul>>c.shift2 != n/d {
228			fmt.Printf("d=%d max=%d mul=%d shift2=%d n=%d\n", d, max, c.mul, c.shift2, n)
229			panic("bad multiply magic")
230		}
231		// Also check the exact computations that will be done by the runtime,
232		// for both 32 and 64 bit operations.
233		if uint32(n)*uint32(c.mul)>>uint8(c.shift2) != uint32(n/d) {
234			fmt.Printf("d=%d max=%d mul=%d shift2=%d n=%d\n", d, max, c.mul, c.shift2, n)
235			panic("bad 32-bit multiply magic")
236		}
237		if uint64(n)*uint64(c.mul)>>uint8(c.shift2) != uint64(n/d) {
238			fmt.Printf("d=%d max=%d mul=%d shift2=%d n=%d\n", d, max, c.mul, c.shift2, n)
239			panic("bad 64-bit multiply magic")
240		}
241	}
242}
243
244func printComment(w io.Writer, classes []class) {
245	fmt.Fprintf(w, "// %-5s  %-9s  %-10s  %-7s  %-10s  %-9s\n", "class", "bytes/obj", "bytes/span", "objects", "tail waste", "max waste")
246	prevSize := 0
247	for i, c := range classes {
248		if i == 0 {
249			continue
250		}
251		spanSize := c.npages * pageSize
252		objects := spanSize / c.size
253		tailWaste := spanSize - c.size*(spanSize/c.size)
254		maxWaste := float64((c.size-prevSize-1)*objects+tailWaste) / float64(spanSize)
255		prevSize = c.size
256		fmt.Fprintf(w, "// %5d  %9d  %10d  %7d  %10d  %8.2f%%\n", i, c.size, spanSize, objects, tailWaste, 100*maxWaste)
257	}
258	fmt.Fprintf(w, "\n")
259}
260
261func printClasses(w io.Writer, classes []class) {
262	fmt.Fprintln(w, "const (")
263	fmt.Fprintf(w, "_MaxSmallSize = %d\n", maxSmallSize)
264	fmt.Fprintf(w, "smallSizeDiv = %d\n", smallSizeDiv)
265	fmt.Fprintf(w, "smallSizeMax = %d\n", smallSizeMax)
266	fmt.Fprintf(w, "largeSizeDiv = %d\n", largeSizeDiv)
267	fmt.Fprintf(w, "_NumSizeClasses = %d\n", len(classes))
268	fmt.Fprintf(w, "_PageShift = %d\n", pageShift)
269	fmt.Fprintln(w, ")")
270
271	fmt.Fprint(w, "var class_to_size = [_NumSizeClasses]uint16 {")
272	for _, c := range classes {
273		fmt.Fprintf(w, "%d,", c.size)
274	}
275	fmt.Fprintln(w, "}")
276
277	fmt.Fprint(w, "var class_to_allocnpages = [_NumSizeClasses]uint8 {")
278	for _, c := range classes {
279		fmt.Fprintf(w, "%d,", c.npages)
280	}
281	fmt.Fprintln(w, "}")
282
283	fmt.Fprintln(w, "type divMagic struct {")
284	fmt.Fprintln(w, "  shift uint8")
285	fmt.Fprintln(w, "  shift2 uint8")
286	fmt.Fprintln(w, "  mul uint16")
287	fmt.Fprintln(w, "  baseMask uint16")
288	fmt.Fprintln(w, "}")
289	fmt.Fprint(w, "var class_to_divmagic = [_NumSizeClasses]divMagic {")
290	for _, c := range classes {
291		fmt.Fprintf(w, "{%d,%d,%d,%d},", c.shift, c.shift2, c.mul, c.mask)
292	}
293	fmt.Fprintln(w, "}")
294
295	// map from size to size class, for small sizes.
296	sc := make([]int, smallSizeMax/smallSizeDiv+1)
297	for i := range sc {
298		size := i * smallSizeDiv
299		for j, c := range classes {
300			if c.size >= size {
301				sc[i] = j
302				break
303			}
304		}
305	}
306	fmt.Fprint(w, "var size_to_class8 = [smallSizeMax/smallSizeDiv+1]uint8 {")
307	for _, v := range sc {
308		fmt.Fprintf(w, "%d,", v)
309	}
310	fmt.Fprintln(w, "}")
311
312	// map from size to size class, for large sizes.
313	sc = make([]int, (maxSmallSize-smallSizeMax)/largeSizeDiv+1)
314	for i := range sc {
315		size := smallSizeMax + i*largeSizeDiv
316		for j, c := range classes {
317			if c.size >= size {
318				sc[i] = j
319				break
320			}
321		}
322	}
323	fmt.Fprint(w, "var size_to_class128 = [(_MaxSmallSize-smallSizeMax)/largeSizeDiv+1]uint8 {")
324	for _, v := range sc {
325		fmt.Fprintf(w, "%d,", v)
326	}
327	fmt.Fprintln(w, "}")
328}
329