Significant compilation time regression starting in v1.95.0-nightly (bevy)

[gpoblon]

View all comments

Summary

Compilation time regression from 1.95 nightly onward.
1.94 nightly (nightly-2026-01-15) is not affected.

EDIT: this issue affects both LLVm and cranelift backends.

Command used

cargo build in a fresh (cargo init + cargo add bevy@0.18.1) bevy repository. Nothing fancy.

Expected vs Actual behaviour

• 1.94 stable cold compilation time: 48.5 seconds.
• 1.94 nightly cold compilation time: 46.9 seconds.
• 1.95 nightly cold compilation time: 85.8 seconds.
• 1.96 nightly cold compilation time: 86 seconds.

Configuration

[unstable] [build], [target] are empty. No env / local crate override. Built with time cargo build --profile x.

My .cargo/config.toml:

[profile.x]
inherits = "dev"
debug-assertions = true
incremental = false
# default, will change
opt-level = 0
debug = true
split-debuginfo = 'off'
strip = "none"
lto = false
panic = 'unwind'
codegen-units = 256

Afterwards, I toggled these options one by one (+ cranelift + Ctarget-cpu + Zthreads + Zshare-generics + linker + mold + wild) and results are consistent: compilation time roughly doubles in all scenarios, it does not seem to be related to an option.

Operating system

EndeavourOS (up to date)

HEAD

(version 1.94 nightly)

rustup toolchain install nightly-2026-01-15
rustup override set nightly-2026-01-15

(version 1.95 nightly)

rustup toolchain install nightly-2026-02-25
rustup override set nightly-2026-02-25

Additional context

Build comparison table

Stable 1.94

stable	time (secs)	bin size (mo)	Disk deps (go)	Disk deps (go) w/ loop	bin size (mo) w/ loop	exec speed(ms)w/loop	Notes
REFERENCE	48.5	3.8	5.6	6.9	1100	158	The reference for all tests
codegen-units=1	53.3	3.8	4	4.7	784	158
codegen-units=16	46.7	3.8	4.4	5.5	887	158	Sweet spot.
codegen-units=32	50	3.8	4	6.7	939	158
Ctarget-cpu=native	48.7	3.8	6.9	6.9	1100	160
debug-assertions=false	47.7	-	-	6.8	1100	141.6
debug-assertions=false + debug=false	42.5	-	-	2.3	87	130
debug=false	42.5	0.43	1.8	2.3	91	152
debug="line-tables-only	43.6	0.43	1.9	3	271	152.4	exec speed anomaly
split-debuginfo="unpacked"	48.25	-	-	5.8	221	158
opt-level=1	83.7	3.8	5.1	7.7	1400	11.9
opt-level=2	95.8	3.8	5.7	8.7	1500	11.2
opt-level=3	96.4	3.8	5.8	8.9	1600	10.8
opt-level=3 + codegen-units=1	148	-	-	5.1	945	10.8	Affects compilation time
opt-level=3 + codegen-units=16	92	-	-	7.3	1300	10.8
opt-level="s"	78.9	3.8	4.4	6.5	1100	14.2
lto="fat"	(46 / ) 897	4.4	4.4	9.9	791	158	Takes forever to build local crate
lto="thin"	(50 / ) 62.4	2.3	5.3	5	312	158	Big local crates make compilation very long.
lto="thin" + codegen-units=16	60	-	-	4.5	255	158	Big local crates make compilation very long.
panic="abort"	37 / 51.95	3.8	5.5	7.6	1100	153	Surprisingly, improves performance
strip="debuginfo"	47.7	0.43	5.5	5	90	158
strip="symbols"	47.8	0.345	4.8	4.9	55	158.4
REF1 = opt-level = 3 +lto = "thin" +debug = false +strip = "symbols" +panic = "abort" +codegen-units = 1 +split-debuginfo = "off" +Ctarget-cpu=native	103			1.5	23	11.2	BiS for release builds
REF1 but codegen-units = 16	72.8			1.7	37	11.1
REF1 + built from RAM (build.target-dir = "/tmp/cargo-target")	103			1.5	23	11.2	df /tmp outputs tmpfs which indicates /tmp is in my RAM.No difference here means my SSD is not a bottleneck.
REF1 + mold linker	103			1.5	23	11.2	unaffected
REF1 + wild linker	103			1.5	23	11.2	unaffected

Nightly 1.94

nightly	time (secs)	bin size (mo)	Disk deps (go)	Disk deps (go)main	main bin size (mo)	exec speed(ms)	Notes
REFERENCE	46.9			6.8	1.1	158
codegen-units=1	52.9			5.2	912	157
Zthreads=1	46.9	4.2	5.2	7.6	1200	158	Seems to be the default value
Zthreads=8	33.8	4.2	5.2	7.6	1200	158	!= 1.95: here sweet spot in 16
Zthreads=8 + codegen-units=16	31.4	4.2	5.2	7.6	1200	158
Zthreads=16	33.4	4.2	5.2	7.6	1200	158	9950X has 16 cores
Zthreads=32	40	4.2	5.2	7.6	1200	158	9950X has 32 threads
Zshare-generics=y	48.6	4.2	5.2	6.1	1000	-	No effect on REFERENCE configuration.
Ctarget-cpu=native	49.2	4.2	7.6	7.6	1.2	166
debug=false	43.15	0.44	2.1	2.3	107	155
opt-level=1	84	4.2	5.3	8.1	1500	12
opt-level=2	97.5	4.2	6	9.2	1600	10.8
opt-level=3	97.7	4.2	6	9.3	1700	11.1
REF1 = opt-level = 3 +lto = "thin" +debug = false +debug-assertions=falsestrip = "symbols" +panic = "abort" +codegen-units = 1 +split-debuginfo = "off"+ Ctarget-cpu=native+ Zthreads=8				1.6	23	10.9	(release profile)
REF1 + Zshare-generics=y				1.6	21	11.1
opt-level = 0 +lto = "off" +debug = line-tables-only +strip = false +panic = "unwind" +codegen-units = 16 +split-debuginfo = "unpacked"+ Ctarget-cpu=native+ Zthreads=8+ Zshare-generics=y				3.2	186	163	(dev profile)
REF2 = opt-level = 0debug = falsedebug-assertions=falsesplit-debuginfo = 'off'strip = falselto = "off"panic = 'abort'codegen-units = 16+ Ctarget-cpu=native+ Zthreads=8+ Zshare-generics=y	29.34			2	84	117.6
REF2 + debug-assertions=false	29.7			2.1	88	145.2
REF2 + mold	25.74			2.1	116	118
REF2 + wild	24.93			2	85	116.8
REF2 + cranelift	29.5			2.7	376	195
REF2 + cranelift + wild	24.7			2.7	366	196

Nightly 1.95 (and 1.96)

nightly	time (secs)	bin size (mo)	Disk deps (go)	Disk deps (go)main	main bin size (mo)	exec speed(ms)	Notes
REFERENCE	86	4.2	5.2	-	-	-	A lot longer than stable!
codegen-units=1
Zthreads=1	86	4.2	5.2	-	-	-	Seems to be the default value
Zthreads=8	66	4.2	5.2	-	-	-	Sweet spot.From there : the higher the value, the longer it took.
Zthreads=16	86	4.2	5.2	-	-	-	9950X has 16 cores
Zthreads=32	95	4.2	5.2	-	-	-	9950X has 32 threads
Zshare-generics=y	86	4.2	5.2	-	-	-	No effect on REFERENCE configuration.
Ctarget-cpu=native	86	4.2	7.6	7.6	1.2	166
debug=true	94			7.5	1200	157.3
debug=false	79.8	0.44	2.1	2.3	103	155
opt-level=1	116.7	4.2	5.3	8.1	1500	12
opt-level=2	128	4.2	6	9.2	1600	10.8
opt-level=3	129.7	4.2	6	9.3	1700	11.1
REF1 = opt-level = 3 +lto = "thin" +debug = false +strip = "symbols" +panic = "abort" +codegen-units = 1 +split-debuginfo = "off"+ Ctarget-cpu=native+ Zthreads=8	123			1.6	23	10.9	(release profile)
REF1 + Zshare-generics=y	106.5			1.6	21	11.1
opt-level = 0 +lto = "off" +debug = line-tables-only +strip = false +panic = "unwind" +codegen-units = 16 +split-debuginfo = "unpacked"+ Ctarget-cpu=native+ Zthreads=8+ Zshare-generics=y	65			3.2	186	163	(dev profile)
REF2 = opt-level = 0debug = falsesplit-debuginfo = 'off'strip = falselto = "off"panic = 'abort'codegen-units = 16+ Ctarget-cpu=native+ Zthreads=8+ Zshare-generics=y	60.1			2.1	87	145.4	(dev profile best of both worlds)
REF2 + mold	56.3			2.2	121	146.8
REF2 + wild	56.2			2.1	88	145.8
REF2 + cranelift	61.7			2.8	402	195
REF2 + cranelift + wild	55.18			2.8	387	196

[mati865]

Likely related to LLVM 22 update #150722