Draft: implement malloc/free within the bump allocator, passing size with realloc and free by rainers · Pull Request #23924 · dlang/dmd

rainers · 2026-09-25T19:47:23Z

For my usual testcase (all phobos unittests compiled with -o-), this reduces compilation time by about 9% (26.8 s -> 24.6 s) and reduces process memory by about 2.5 % (12212 MB -> 11943 MB).

Not sure if that's worth having to pass allocation sizes to mem.xfree and mem.xrealloc.

Let's see what the perf runner reports...

github-actions · 2026-09-25T20:04:51Z

DMD perf check

Metric	Base	PR	Δ
compile hello.d (instr)	203.6 M	204.6 M	+0.498%
compile hello.d -O -release (instr)	221.2 M	222.2 M	+0.469%
dmd binary size (stripped)	8.17 MB	8.18 MB	+0.14%
peak RSS (compile Phobos)	610.2 MB	596.0 MB	-2.32%
peak RSS (compile vibe.d)	1880 MB	1833 MB	-2.48%
page faults (compile Phobos)	151,717	148,639	-2.03%
page faults (compile vibe.d)	470,239	458,873	-2.42%

Breakdown — compile hello.d

Phase (wall, self time)	Base	PR	Δ
parse	17.8 ms	19.3 ms	+8.78%
sema_other	11.7 ms	12.0 ms	+2.36%
sema1	7.4 ms	7.1 ms	-3.57%
sema3	3.6 ms	3.7 ms	+4.45%
codegen	1.5 ms	1.5 ms	+3.19%

All measurements

Metric	Base	PR	Δ
compile hello.d (instr)	203.6 M	204.6 M	+0.498%
compile hello.d -O -release (instr)	221.2 M	222.2 M	+0.469%
compile Phobos (instr)	4,692.6 M	4,694.9 M	+0.047%
compile Phobos codegen (instr)	1,417.1 M	1,418.5 M	+0.099%
compile vibe.d (instr)	13,560.5 M	13,560.3 M	-0.002%
dmd binary size (stripped)	8.17 MB	8.18 MB	+0.14%
hello binary size (stripped)	0.72 MB	0.72 MB	0.00%
peak RSS (compile hello.d)	44.37 MB	43.92 MB	-1.02%
peak RSS (compile Phobos)	610.2 MB	596.0 MB	-2.32%
peak RSS (compile vibe.d)	1880 MB	1833 MB	-2.48%
page faults (compile hello.d)	8,730	8,607	-1.41%
page faults (compile Phobos)	151,717	148,639	-2.03%
page faults (compile vibe.d)	470,239	458,873	-2.42%
compile dmd itself (wall)	7.7 s	7.7 s	+0.44%
compile hello.d (wall)	41.9 ms	43.7 ms	+4.24%
compile Phobos (wall)	916 ms	908 ms	-0.90%

_{c01a0b2 vs merge-base f302b81 · about these metrics}

ibuclaw · 2026-09-27T08:17:21Z

+    enum pureBumpMalloc = cast(void* function(size_t) pure nothrow)&bumpMalloc;
+    enum pureBumpFree  = cast(void function(void*p, size_t) pure nothrow)&bumpFree;


Could also fake it with pragma(mangle), which has support on older compilers.

dmd/druntime/src/core/memory.d

Line 1234 in 807e8a2

extern (C) private pure @system @nogc nothrow

Indeed, that seems to work across all versions. Updated.

LDC is not happy with this change: Function type does not match previously declared function with the same mangled name. Restored.

nordlow · 2026-09-28T10:18:14Z

Cherry-picking this on master gives merge conflicts.

Update: Ahh, merge works. Nevermind this.

dkorpel · 2026-09-29T10:36:07Z

Does this supersede #23914 ?

Not sure if that's worth having to pass allocation sizes to mem.xfree and mem.xrealloc.

I'd say definitely yes!

dkorpel · 2026-09-29T11:47:19Z

+    enum smallAlignmentShift = 4;
+    enum smallAlignment = 1 << smallAlignmentShift;
+    enum maxSmallSize = 1024;
+    __gshared void*[maxSmallSize >> smallAlignmentShift] firstSmallFree;
+
+    enum largeAlignmentShift = 8;
+    enum largeAlignment = 1 << largeAlignmentShift;
+    enum maxLargeSize = 64 * 1024;
+    __gshared void*[maxLargeSize >> largeAlignmentShift] firstLargeFree;
+
+    enum hugeAlignment = 4096;
+    __gshared void* firstHugeFree;


Is this bucket strategy tuned for dmd?

Is this bucket strategy tuned for dmd?

Not really, just my first guess. I suspect allocation sizes are mostly rather small, but I can add some statistics on bucket usage.

Here's a summary of the bin usage of my phobos unittest benchmark:

Size range Mallocs Frees Approx. live memory Interpretation

0x01–0x10 8,587,199 3,549,229 ~50 MB Very high count

0x11–0x20 11,973,412 188,068 ~377 MB Huge count + memory

0x21–0x30 535,439 84,989 ~21 MB Moderate

0x31–0x40 2,076,679 17,630 ~131 MB High count + memory

0x41–0x50 7,573,155 36,854 ~452 MB Huge count + memory

0x51–0x60 108,464 5,100 ~6 MB Low impact

0x61–0x100 ~1k–26k/bucket ~0.5k–11k/bucket <10 MB each Low impact

0x101–0x300 ~700–6,000/bucket ~400–1,100/bucket <1 MB each Negligible individually

0x301–0x3000 hundreds–thousands tens–thousands <1 MB/bucket Low

0x3001–0xffff tens–thousands generally similar Generally small Low

>= 0x10000 89 57 ~47 MB net reported Few allocations, but expensive

rainers · 2026-09-29T20:57:14Z

Does this supersede #23914 ?

No, it replaces the C malloc calls by using the memory of the bump allocator, that can benefit additionally from #23914 when using larger pages. The measurements in #23914 (comment) were made on top of this PR.

…alloc and free

…strlen

rainers marked this pull request as draft September 25, 2026 19:47

rainers force-pushed the bumpMallocFree branch 2 times, most recently from a152334 to e038f21 Compare September 26, 2026 07:54

rainers mentioned this pull request Sep 26, 2026

Tune bump allocator for 2MB huge page size #23914

Open

ibuclaw reviewed Sep 27, 2026

View reviewed changes

rainers force-pushed the bumpMallocFree branch from 12359e1 to 9502df6 Compare September 27, 2026 09:27

dkorpel reviewed Sep 29, 2026

View reviewed changes

This was referenced Oct 3, 2026

refactor: use endloc instead of endlinnum #23962

Merged

implement mem.xmalloc/xfree within the bump allocator #23999

Merged

implement malloc/free within the bump allocator, passing size with re…

a335419

…alloc and free

rainers force-pushed the bumpMallocFree branch from 9199c40 to a335419 Compare October 9, 2026 15:15

rainers added 2 commits October 9, 2026 17:37

save size before block in debug builds and verify in free

16135e1

Filename.buildPath: do not over-allocate, free size is determined by …

c01a0b2

…strlen

Navigation Menu

Repository navigation

Draft: implement malloc/free within the bump allocator, passing size with realloc and free - #23924

Draft: implement malloc/free within the bump allocator, passing size with realloc and free#23924
rainers wants to merge 3 commits into
dlang:masterfrom
rainers:bumpMallocFree

rainers commented Sep 25, 2026

Uh oh!

github-actions Bot commented Sep 25, 2026 •

edited

Loading

Uh oh!

ibuclaw Sep 27, 2026 •

edited

Loading

Uh oh!

rainers Sep 27, 2026

Uh oh!

rainers Sep 27, 2026 •

edited

Loading

Uh oh!

nordlow commented Sep 28, 2026 •

edited

Loading

Uh oh!

dkorpel commented Sep 29, 2026

Uh oh!

dkorpel Sep 29, 2026

Uh oh!

rainers Sep 29, 2026

Uh oh!

rainers Oct 2, 2026

Uh oh!

rainers commented Sep 29, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

		enum pureBumpMalloc = cast(void* function(size_t) pure nothrow)&bumpMalloc;
		enum pureBumpFree = cast(void function(void*p, size_t) pure nothrow)&bumpFree;

Size range	Mallocs	Frees	Approx. live memory	Interpretation
`0x01–0x10`	8,587,199	3,549,229	~50 MB	Very high count
`0x11–0x20`	11,973,412	188,068	~377 MB	Huge count + memory
`0x21–0x30`	535,439	84,989	~21 MB	Moderate
`0x31–0x40`	2,076,679	17,630	~131 MB	High count + memory
`0x41–0x50`	7,573,155	36,854	~452 MB	Huge count + memory
`0x51–0x60`	108,464	5,100	~6 MB	Low impact
`0x61–0x100`	~1k–26k/bucket	~0.5k–11k/bucket	<10 MB each	Low impact
`0x101–0x300`	~700–6,000/bucket	~400–1,100/bucket	<1 MB each	Negligible individually
`0x301–0x3000`	hundreds–thousands	tens–thousands	<1 MB/bucket	Low
`0x3001–0xffff`	tens–thousands	generally similar	Generally small	Low
`>= 0x10000`	89	57	~47 MB net reported	Few allocations, but expensive

Conversation

rainers commented Sep 25, 2026

Uh oh!

github-actions Bot commented Sep 25, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

DMD perf check

Uh oh!

ibuclaw Sep 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

rainers Sep 27, 2026

Choose a reason for hiding this comment

Uh oh!

rainers Sep 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

nordlow commented Sep 28, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

dkorpel commented Sep 29, 2026

Uh oh!

dkorpel Sep 29, 2026

Choose a reason for hiding this comment

Uh oh!

rainers Sep 29, 2026

Choose a reason for hiding this comment

Uh oh!

rainers Oct 2, 2026

Choose a reason for hiding this comment

Uh oh!

rainers commented Sep 29, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

github-actions Bot commented Sep 25, 2026 •

edited

Loading

ibuclaw Sep 27, 2026 •

edited

Loading

rainers Sep 27, 2026 •

edited

Loading

nordlow commented Sep 28, 2026 •

edited

Loading