[marc-mabe]

Whether reallocating a small block keeps it in place was decided by old_bin_num > 0 && size <= bin_data_size[old_bin_num - 1], which needs a test of the bin number, an index computation and a compare on every in-place reallocation.

Store the smallest size a block of each bin keeps in a table, bin_keep_size[], which is one more than the size of the bin below, and 0 for the first bin, so the check becomes a single load and compare. The behavior is unchanged.

As each entry comes from the row of the bin below, the table has one entry more than there are bins, and the last one is unused. This does not grow the binary: the table is aligned to at least 16 bytes, so 30 entries (120 bytes) would be padded to 128 bytes anyway, and the extra 4 bytes fit into that padding.

Mark the truncation as unexpected: a block reallocated to a smaller size usually stays in its bin. Without the hint, GCC merges the epilogue of the in-place return with another return path, which adds a branch to every in-place reallocation of a large run.

Benchmark	arm64	amd64 (emulated)	i386 (emulated)	arm32 (emulated)
Zend/bench.php	+0.000% (−0.0000%)	−0.002% (−0.0000%)	−0.000% (+0.0000%)	−0.000% (+0.0000%)
Zend/bench.php JIT	+0.000% (+0.0000%)	−0.008% (−0.0000%)	not run	not run
Symfony Demo	−0.023% (+0.0000%)	−0.019% (+0.0000%)	−0.003% (+0.0000%)	not run
Symfony Demo JIT	−0.025% (+0.0001%)	−0.022% (−0.0000%)	not run	not run
WordPress 6.2	−0.009% (−0.0000%)	−0.007% (−0.0000%)	+0.001% (−0.0047%)	not run
WordPress 6.2 JIT	−0.007% (+0.0085%)	−0.007% (+0.0010%)	not run	not run

Instructions per request (callgrind) vs. master (first run).
in parentheses master vs. master as a run-to-run control.

[iluuu1994]

bin_data_size[old_bin_num] is accessed above and so bin_data_size[old_bin_num - 1] is very likely to be in cache. I think this should be benchmarked on real hardware to make sure there's no regression.

[ndossche]

Right, now this patch adds another table that must be loaded from memory.
The instruction count from valgrind doesn't take into account how modern processors actually work.
This is also where micro benchmarks are deceptive: since micro benchmarks execute everything in a hot loop, it's likely that table will be cached too.
My gut feeling says this isn't a good change.

[marc-mabe]

I'll run a real world bench with perf on real amd64/arm64. Just not sure when I can find the time for it.

[marc-mabe]

@iluuu1994 @ndossche Seems you are right

Benchmark diff

• Base: c65321c70f7d (HEAD) on mbennewitz, native, amd64, 11th Gen Intel(R) Core(TM) i7-1165G7 @ 2.80GHz -- 20261014-212054-mbennewitz-amd64-c65321c70f7d
• Head: c9ae38d585d4 (mm-realloc-keep-size) on mbennewitz, native, amd64, 11th Gen Intel(R) Core(TM) i7-1165G7 @ 2.80GHz -- 20261014-212717-mbennewitz-amd64-c9ae38d585d4

Change from base to head; lower is better everywhere. The ± figure is the threshold a change must exceed to count as significant (bold): 2× the runs' combined noise for perf, a measured noise floor for callgrind and memory. ? means a single perf run gave no noise estimate.

perf: per request

Workload	duration_time	task-clock	cycles	instructions	branches	branch-misses	cache-references	cache-misses
Zend/bench.php	+13.62% ±2.90 ❌	+13.72% ±2.93 ❌	+0.22% ±0.98	-0.01% ±0.02	-0.00% ±0.02	-1.98% ±6.95	-1.38% ±1.12 ✅	+15.67% ±5.67 ❌
Symfony Demo 2.2.3	+10.08% ±6.90 ❌	+10.06% ±6.93 ❌	-1.87% ±6.67	-3.50% ±7.12	-3.76% ±7.72	-3.50% ±4.89	-4.19% ±4.39	+19.10% ±28.46
Wordpress 6.2	+14.46% ±2.60 ❌	+15.90% ±2.56 ❌	-0.04% ±0.58	-0.01% ±0.00 ✅	-0.01% ±0.00 ✅	+0.77% ±0.87	+0.83% ±2.15	+10.86% ±3.89 ❌
Zend/bench.php JIT	+7.99% ±4.29 ❌	+8.27% ±4.30 ❌	-0.69% ±1.02	-0.05% ±0.02 ✅	-0.04% ±0.03 ✅	-7.75% ±14.32	-1.82% ±1.15 ✅	+0.40% ±3.48
Symfony Demo 2.2.3 JIT	+7.24% ±6.23 ❌	+7.30% ±6.38 ❌	-2.89% ±5.47	-6.26% ±6.36	-6.92% ±7.00	-4.53% ±4.64	-3.54% ±3.69	+26.31% ±6.67 ❌
Wordpress 6.2 JIT	+13.61% ±1.93 ❌	+17.39% ±2.57 ❌	+0.81% ±0.38 ❌	+0.00% ±0.00	-0.01% ±0.00 ✅	-0.25% ±0.81	+1.41% ±0.78 ❌	+35.93% ±1.86 ❌

Absolute values

Workload	Metric	Base	Head	Change
Zend/bench.php	duration_time	176.942 ms	201.032 ms	+13.62% ±2.90 ❌
Zend/bench.php	task-clock	176.040 ms	200.200 ms	+13.72% ±2.93 ❌
Zend/bench.php	cycles	708,719,746	710,291,282	+0.22% ±0.98
Zend/bench.php	instructions	2,259,805,444	2,259,695,718	-0.01% ±0.02
Zend/bench.php	branches	442,355,130	442,348,828	-0.00% ±0.02
Zend/bench.php	branch-misses	1,044,447	1,023,809	-1.98% ±6.95
Zend/bench.php	cache-references	1,361,591	1,342,848	-1.38% ±1.12 ✅
Zend/bench.php	cache-misses	585,156	676,854	+15.67% ±5.67 ❌
Symfony Demo 2.2.3	duration_time	21.406 ms	23.565 ms	+10.08% ±6.90 ❌
Symfony Demo 2.2.3	task-clock	21.381 ms	23.533 ms	+10.06% ±6.93 ❌
Symfony Demo 2.2.3	cycles	87,099,227	85,468,201	-1.87% ±6.67
Symfony Demo 2.2.3	instructions	134,592,704	129,881,682	-3.50% ±7.12
Symfony Demo 2.2.3	branches	27,437,485	26,405,396	-3.76% ±7.72
Symfony Demo 2.2.3	branch-misses	809,366	781,015	-3.50% ±4.89
Symfony Demo 2.2.3	cache-references	1,413,155	1,353,980	-4.19% ±4.39
Symfony Demo 2.2.3	cache-misses	250,331	298,144	+19.10% ±28.46
Wordpress 6.2	duration_time	46.917 ms	53.701 ms	+14.46% ±2.60 ❌
Wordpress 6.2	task-clock	40.351 ms	46.767 ms	+15.90% ±2.56 ❌
Wordpress 6.2	cycles	169,922,069	169,853,412	-0.04% ±0.58
Wordpress 6.2	instructions	282,140,720	282,120,585	-0.01% ±0.00 ✅
Wordpress 6.2	branches	60,394,029	60,388,204	-0.01% ±0.00 ✅
Wordpress 6.2	branch-misses	1,596,233	1,608,569	+0.77% ±0.87
Wordpress 6.2	cache-references	2,423,071	2,443,101	+0.83% ±2.15
Wordpress 6.2	cache-misses	303,405	336,344	+10.86% ±3.89 ❌
Zend/bench.php JIT	duration_time	65.041 ms	70.240 ms	+7.99% ±4.29 ❌
Zend/bench.php JIT	task-clock	64.080 ms	69.380 ms	+8.27% ±4.30 ❌
Zend/bench.php JIT	cycles	263,643,027	261,836,008	-0.69% ±1.02
Zend/bench.php JIT	instructions	680,929,619	680,573,942	-0.05% ±0.02 ✅
Zend/bench.php JIT	branches	139,498,295	139,437,647	-0.04% ±0.03 ✅
Zend/bench.php JIT	branch-misses	811,338	748,420	-7.75% ±14.32
Zend/bench.php JIT	cache-references	1,621,589	1,592,030	-1.82% ±1.15 ✅
Zend/bench.php JIT	cache-misses	652,160	654,736	+0.40% ±3.48
Symfony Demo 2.2.3 JIT	duration_time	23.543 ms	25.246 ms	+7.24% ±6.23 ❌
Symfony Demo 2.2.3 JIT	task-clock	23.495 ms	25.211 ms	+7.30% ±6.38 ❌
Symfony Demo 2.2.3 JIT	cycles	94,691,559	91,952,824	-2.89% ±5.47
Symfony Demo 2.2.3 JIT	instructions	136,701,531	128,149,821	-6.26% ±6.36
Symfony Demo 2.2.3 JIT	branches	27,343,893	25,451,587	-6.92% ±7.00
Symfony Demo 2.2.3 JIT	branch-misses	922,847	881,032	-4.53% ±4.64
Symfony Demo 2.2.3 JIT	cache-references	1,695,078	1,635,113	-3.54% ±3.69
Symfony Demo 2.2.3 JIT	cache-misses	350,020	442,095	+26.31% ±6.67 ❌
Wordpress 6.2 JIT	duration_time	45.605 ms	51.812 ms	+13.61% ±1.93 ❌
Wordpress 6.2 JIT	task-clock	38.856 ms	45.611 ms	+17.39% ±2.57 ❌
Wordpress 6.2 JIT	cycles	163,493,060	164,816,878	+0.81% ±0.38 ❌
Wordpress 6.2 JIT	instructions	244,891,350	244,892,192	+0.00% ±0.00
Wordpress 6.2 JIT	branches	50,577,685	50,573,582	-0.01% ±0.00 ✅
Wordpress 6.2 JIT	branch-misses	1,569,591	1,565,693	-0.25% ±0.81
Wordpress 6.2 JIT	cache-references	3,079,068	3,122,441	+1.41% ±0.78 ❌
Wordpress 6.2 JIT	cache-misses	414,422	563,341	+35.93% ±1.86 ❌

[iluuu1994]

Thanks very much for confirming!