perf(alloc): don't overalign allocations

[mkroening]

This PR makes our overalignment of every allocation to cache line size non-default.
This heavily reduces our memory usage in scenarios with many small allocations (such as deserializing JSON).
This also circumvents SFBdragon/talc#44, which is the cause of #1968.

That issue's JSON benchmark has this performance on my machine:

alloc	time
main	35 s
galloc	200 ms
bump	195 ms
galloc + this PR	120 ms
bump + this PR	#GP
this PR	90 ms
host	10 ms

bump + this PR is broken because it incorrectly aligns allocations.

Depends on #1983.
Closes #1935.
Closes #1940.
Closes #1968.

[github-actions]

Benchmark Results

Benchmark	Current: 8123de4	Previous: 8e08755	Performance Ratio
startup_benchmark Build Time	110.72 s	118.64 s	0.93
startup_benchmark File Size	0.90 MB	0.90 MB	0.99
Startup Time - 1 core	0.89 s (±0.03 s)	0.92 s (±0.03 s)	0.98
Startup Time - 2 cores	0.91 s (±0.03 s)	0.93 s (±0.03 s)	0.98
Startup Time - 4 cores	0.91 s (±0.03 s)	0.90 s (±0.03 s)	1.01
multithreaded_benchmark Build Time	111.78 s	115.52 s	0.97
multithreaded_benchmark File Size	1.00 MB	1.01 MB	0.99
Multithreaded Pi Efficiency - 2 Threads	85.12 % (±9.57 %)	89.14 % (±8.69 %)	0.95
Multithreaded Pi Efficiency - 4 Threads	42.06 % (±3.67 %)	43.33 % (±4.80 %)	0.97
Multithreaded Pi Efficiency - 8 Threads	24.88 % (±1.56 %)	24.89 % (±3.25 %)	1.00
micro_benchmarks Build Time	111.64 s	121.35 s	0.92
micro_benchmarks File Size	1.00 MB	1.01 MB	0.99
Scheduling time - 1 thread	75.70 ticks (±5.38 ticks)	64.78 ticks (±3.48 ticks)	1.17
Scheduling time - 2 threads	43.43 ticks (±5.88 ticks)	39.58 ticks (±5.81 ticks)	1.10
Micro - Time for syscall (getpid)	3.54 ticks (±0.37 ticks)	3.11 ticks (±0.35 ticks)	1.14
Memcpy speed - (built_in) block size 4096	68554.82 MByte/s (±48794.10 MByte/s)	76892.26 MByte/s (±53208.68 MByte/s)	0.89
Memcpy speed - (built_in) block size 1048576	29459.00 MByte/s (±24153.57 MByte/s)	42437.70 MByte/s (±29401.77 MByte/s)	0.69
Memcpy speed - (built_in) block size 16777216	27922.81 MByte/s (±23184.88 MByte/s)	24216.95 MByte/s (±20131.45 MByte/s)	1.15
Memset speed - (built_in) block size 4096	69240.91 MByte/s (±49188.21 MByte/s)	76746.14 MByte/s (±53110.30 MByte/s)	0.90
Memset speed - (built_in) block size 1048576	30241.21 MByte/s (±24590.79 MByte/s)	42677.86 MByte/s (±29565.07 MByte/s)	0.71
Memset speed - (built_in) block size 16777216	28703.15 MByte/s (±23639.72 MByte/s)	24944.26 MByte/s (±20616.34 MByte/s)	1.15
Memcpy speed - (rust) block size 4096	61490.56 MByte/s (±45358.39 MByte/s)	72842.39 MByte/s (±51003.78 MByte/s)	0.84
Memcpy speed - (rust) block size 1048576	29444.81 MByte/s (±24209.84 MByte/s)	42614.04 MByte/s (±29546.77 MByte/s)	0.69
Memcpy speed - (rust) block size 16777216	28129.09 MByte/s (±23368.74 MByte/s)	25125.16 MByte/s (±20845.63 MByte/s)	1.12
Memset speed - (rust) block size 4096	62376.72 MByte/s (±45990.87 MByte/s)	73212.71 MByte/s (±51231.34 MByte/s)	0.85
Memset speed - (rust) block size 1048576	30207.40 MByte/s (±24646.85 MByte/s)	42865.24 MByte/s (±29720.96 MByte/s)	0.70
Memset speed - (rust) block size 16777216	28847.47 MByte/s (±23765.07 MByte/s)	25840.21 MByte/s (±21299.14 MByte/s)	1.12
alloc_benchmarks Build Time	105.91 s	106.38 s	1.00
alloc_benchmarks File Size	0.96 MB	0.97 MB	1.00
Allocations - Allocation success	100.00 %	100.00 %	1
Allocations - Deallocation success	70.05 % (±0.27 %)	70.03 % (±0.28 %)	1.00
Allocations - Pre-fail Allocations	100.00 %	100.00 %	1
Allocations - Average Allocation time	8576.21 Ticks (±445.05 Ticks)	11224.94 Ticks (±674.74 Ticks)	0.76
Allocations - Average Allocation time (no fail)	8576.21 Ticks (±445.05 Ticks)	11224.94 Ticks (±674.74 Ticks)	0.76
Allocations - Average Deallocation time	1106.42 Ticks (±328.69 Ticks)	733.24 Ticks (±64.51 Ticks)	1.51
mutex_benchmark Build Time	105.97 s	111.63 s	0.95
mutex_benchmark File Size	1.01 MB	1.01 MB	0.99
Mutex Stress Test Average Time per Iteration - 1 Threads	12.62 ns (±0.72 ns)	13.10 ns (±0.85 ns)	0.96
Mutex Stress Test Average Time per Iteration - 2 Threads	12.76 ns (±0.97 ns)	14.56 ns (±0.90 ns)	0.88

This comment was automatically generated by workflow using github-action-benchmark.

[jounathaen]

Very nice.

I do think, we should not include the overalign feature, but instead be clever and thread aware on small allocations. This would then also help/solve #1984

[mkroening]

I do think, we should not include the overalign feature, but instead be clever and thread aware on small allocations. This would then also help/solve #1984

I agree. I have completely removed our intermediate layer. We can reintroduce it if it turns out to be useful in the future.