Full LTO seems to cause a regression so this uses ThinLTO only on macOS (Clang) and only on ARM64. Here are the results of a few runs (the Intel runners just seem terrible):

---

Before

ARM64

encode: 145.9584 / decode: 351.2980

https://github.com/jcrist/msgspec/actions/runs/19412999871/job/55537240877

Name (time in us)	Min	Max	Mean	StdDev	Median	IQR	Outliers	OPS (Kops/s)	Rounds	Iterations
encode	130.4833 (1.0)	1,755.7709 (1.0)	153.0818 (1.0)	42.0774 (1.0)	145.9584 (1.0)	14.0292 (1.0)	446;834	6.5325 (1.0)	10000	10
decode	324.3583 (2.49)	10,951.2000 (6.24)	374.9843 (2.45)	154.3808 (3.67)	351.2980 (2.41)	20.2022 (1.44)	246;1309	2.6668 (0.41)	10000	10

encode: 143.2750 / decode: 354.1458

https://github.com/jcrist/msgspec/actions/runs/19412999871/job/55539651179

Name (time in us)	Min	Max	Mean	StdDev	Median	IQR	Outliers	OPS (Kops/s)	Rounds	Iterations
encode	130.3250 (1.0)	2,596.3542 (1.0)	154.4214 (1.0)	50.5128 (1.0)	143.2750 (1.0)	13.8417 (1.0)	471;1029	6.4758 (1.0)	10000	10
decode	323.7542 (2.48)	4,192.2167 (1.61)	375.4091 (2.43)	119.6400 (2.37)	354.1458 (2.47)	34.9271 (2.52)	225;631	2.6638 (0.41)	10000	10

x86_64

encode: 228.6268 / decode: 729.5701

https://github.com/jcrist/msgspec/actions/runs/19412999871/job/55537240884

Name (time in us)	Min	Max	Mean	StdDev	Median	IQR	Outliers	OPS (Kops/s)	Rounds	Iterations
encode	206.4159 (1.0)	7,320.7824 (1.61)	250.7420 (1.0)	94.9931 (1.0)	228.6268 (1.0)	36.5537 (1.0)	722;1198	3.9882 (1.0)	10000	10
decode	694.9477 (3.37)	4,550.2093 (1.0)	749.2888 (2.99)	110.4365 (1.16)	729.5701 (3.19)	49.4375 (1.35)	498;662	1.3346 (0.33)	10000	10

encode: 230.9933 / decode: 757.6821

https://github.com/jcrist/msgspec/actions/runs/19412999871/job/55539352389

Name (time in us)	Min	Max	Mean	StdDev	Median	IQR	Outliers	OPS (Kops/s)	Rounds	Iterations
encode	210.2697 (1.0)	5,402.5395 (1.0)	251.5679 (1.0)	100.4256 (1.0)	230.9933 (1.0)	43.8295 (1.0)	340;569	3.9751 (1.0)	10000	10
decode	697.6369 (3.32)	14,816.0169 (2.74)	807.6615 (3.21)	318.6815 (3.17)	757.6821 (3.28)	103.2592 (2.36)	235;698	1.2381 (0.31)	10000	10

---

After

ARM64

encode: 137.4250 / decode: 341.0833

https://github.com/jcrist/msgspec/actions/runs/19413142685/job/55537726515

Name (time in us)	Min	Max	Mean	StdDev	Median	IQR	Outliers	OPS (Kops/s)	Rounds	Iterations
encode	125.5625 (1.0)	1,878.1959 (1.26)	146.3765 (1.0)	53.2579 (1.0)	137.4250 (1.0)	3.5416 (1.0)	378;1460	6.8317 (1.0)	10000	10
decode	309.8791 (2.47)	1,487.6791 (1.0)	351.2185 (2.40)	71.5229 (1.34)	341.0833 (2.48)	21.1479 (5.97)	509;788	2.8472 (0.42)	10000	10

encode: 143.0917 / decode: 341.9875

https://github.com/jcrist/msgspec/actions/runs/19413142685/job/55539646141

Name (time in us)	Min	Max	Mean	StdDev	Median	IQR	Outliers	OPS (Kops/s)	Rounds	Iterations
encode	125.4333 (1.0)	1,788.0458 (1.0)	148.8024 (1.0)	41.2471 (1.0)	143.0917 (1.0)	14.7687 (1.0)	396;575	6.7203 (1.0)	10000	10
decode	303.9750 (2.42)	7,741.4625 (4.33)	375.3218 (2.52)	174.8312 (4.24)	341.9875 (2.39)	24.7896 (1.68)	493;1304	2.6644 (0.40)	10000	10

x86_64

encode: 224.2774 / decode: 690.9108

https://github.com/jcrist/msgspec/actions/runs/19413142685/job/55537726518

Name (time in us)	Min	Max	Mean	StdDev	Median	IQR	Outliers	OPS (Kops/s)	Rounds	Iterations
encode	207.8972 (1.0)	8,588.2149 (1.0)	253.7614 (1.0)	145.7307 (1.0)	224.2774 (1.0)	52.2173 (1.83)	290;703	3.9407 (1.0)	10000	10
decode	684.0644 (3.29)	14,789.9789 (1.72)	749.2749 (2.95)	312.8542 (2.15)	690.9108 (3.08)	28.4929 (1.0)	386;1463	1.3346 (0.34)	10000	10

encode: 339.4682 / decode: 924.6652

https://github.com/jcrist/msgspec/actions/runs/19413142685/job/55539486694

Name (time in us)	Min	Max	Mean	StdDev	Median	IQR	Outliers	OPS	Rounds	Iterations
encode	256.4145 (1.0)	29,541.7157 (3.10)	415.5266 (1.0)	584.9506 (1.74)	339.4682 (1.0)	129.0764 (1.0)	243;649	2,406.5850 (1.0)	10000	10
decode	780.7286 (3.04)	9,538.8675 (1.0)	1,033.4723 (2.49)	336.9571 (1.0)	924.6652 (2.72)	247.0121 (1.91)	866;452	967.6118 (0.40)	10000	10

ofek marked this pull request as ready for review

November 17, 2025 06:27