# Crystal spec is VERY slow and very high CPU usage when the first time run \`crystal spec -O1\` **URL:** \ **Category:** Help & Support\ **Created:** [January 11, 2026, 9:22am UTC](https://forum.crystal-lang.org/t/crystal-spec-is-very-slow-and-very-high-cpu-usage-when-the-first-time-run-crystal-spec-o1/8646 "2026-01-11T09:22:41Z")\ **Posts on this page:** 10\ **Page:** 1
**Author:** ![zw963](https://yyz2.discourse-cdn.com/flex036/user_avatar/forum.crystal-lang.org/zw963/32/1623_2.png) [@zw963](https://forum.crystal-lang.org/u/zw963)\ **Post date:** [January 11, 2026, 9:22am UTC](https://forum.crystal-lang.org/t/crystal-spec-is-very-slow-and-very-high-cpu-usage-when-the-first-time-run-crystal-spec-o1/8646/1 "2026-01-11T09:22:41Z")
Please check following screenshot. ![image](https://canada1.discourse-cdn.com/flex036/uploads/crystal_lang/original/2X/c/cf9f8d490055e2422c0acc9d320900ba57c36bef.png) Why the first time run spec with -O1 is so slow?(15 min on a very little shard) I almost think about crystal spec is in the dead loop. You can reproduce it use [GitHub - crystal-china/captcha: Crystal library that generates image CAPTCHAs.](https://github.com/crystal-china/captcha) and run spec with: `crystal spec -O1` ---
**Author:** ![jgaskins](https://yyz2.discourse-cdn.com/flex036/user_avatar/forum.crystal-lang.org/jgaskins/32/2449_2.png) [@jgaskins](https://forum.crystal-lang.org/u/jgaskins)\ **Post date:** [January 11, 2026, 8:08pm UTC](https://forum.crystal-lang.org/t/crystal-spec-is-very-slow-and-very-high-cpu-usage-when-the-first-time-run-crystal-spec-o1/8646/2 "2026-01-11T20:08:58Z")
I don’t usually run specs with `-O1`, but that does take an exceptionally long time to compile. I let it run for 27 minutes on my machine before killing it. ```plaintext ➜ captcha git:(master) time csp -O1 ^C crystal spec -O1 24.24s user 0.60s system 1% cpu 27:48.08 total ``` Without `-O1`, it takes significantly less time even on a cold compilation: ```plaintext ➜ captcha git:(master) crystal clear_cache ➜ captcha git:(master) time csp ... Finished in 557.48 milliseconds 3 examples, 0 failures, 0 errors, 0 pending crystal spec 10.23s user 0.76s system 126% cpu 8.715 total ``` Even `--release` wasn’t bad: ```plaintext ➜ captcha git:(master) crystal clear_cache ➜ captcha git:(master) time csp --release ... Finished in 101.54 milliseconds 3 examples, 0 failures, 0 errors, 0 pending crystal spec --release 14.58s user 1.38s system 99% cpu 16.127 total ``` ---
**Author:** ![kojix2](https://yyz2.discourse-cdn.com/flex036/user_avatar/forum.crystal-lang.org/kojix2/32/1583_2.png) [@kojix2](https://forum.crystal-lang.org/u/kojix2)\ **Post date:** [January 12, 2026, 8:57am UTC](https://forum.crystal-lang.org/t/crystal-spec-is-very-slow-and-very-high-cpu-usage-when-the-first-time-run-crystal-spec-o1/8646/3 "2026-01-12T08:57:15Z")
Since the second run took 2 seconds, I suspect the cache mechanism is involved. Adding the `-O1` option would likely require recompiling many standard library classes into objects again. However, I’m not sure if that alone explains the slow compilation. ---
**Author:** ![zw963](https://yyz2.discourse-cdn.com/flex036/user_avatar/forum.crystal-lang.org/zw963/32/1623_2.png) [@zw963](https://forum.crystal-lang.org/u/zw963)\ **Post date:** [January 12, 2026, 9:06am UTC](https://forum.crystal-lang.org/t/crystal-spec-is-very-slow-and-very-high-cpu-usage-when-the-first-time-run-crystal-spec-o1/8646/4 "2026-01-12T09:06:07Z")
Created [issue](https://github.com/crystal-lang/crystal/issues/16564) ---
**Author:** ![kojix2](https://yyz2.discourse-cdn.com/flex036/user_avatar/forum.crystal-lang.org/kojix2/32/1583_2.png) [@kojix2](https://forum.crystal-lang.org/u/kojix2)\ **Post date:** [January 12, 2026, 10:28am UTC](https://forum.crystal-lang.org/t/crystal-spec-is-very-slow-and-very-high-cpu-usage-when-the-first-time-run-crystal-spec-o1/8646/5 "2026-01-12T10:28:05Z")
ysbaddaden has already commented on the GitHub issue, but I was also able to reproduce the problem on my side. ```sh rm -r tmp/* CRYSTAL_CACHE_DIR=/home/kojix2/Crystal/captcha/tmp crystal spec -O1 --stats ``` ```console Parse: 00:00:00.000041077 ( 0.98MB) Semantic (top level): 00:00:00.379616584 ( 151.17MB) Semantic (new): 00:00:00.001428824 ( 151.17MB) Semantic (type declarations): 00:00:00.030816626 ( 159.17MB) Semantic (abstract def check): 00:00:00.009114052 ( 167.17MB) Semantic (restrictions augmenter): 00:00:00.010037323 ( 167.17MB) Semantic (ivars initializers): 00:00:00.014306303 ( 175.17MB) Semantic (cvars initializers): 00:00:00.188233005 ( 207.23MB) Semantic (main): 00:00:00.446909827 ( 287.36MB) Semantic (cleanup): 00:00:00.000357046 ( 287.36MB) Semantic (recursive struct check): 00:00:00.000592775 ( 287.36MB) Codegen (crystal): 00:00:00.544523850 ( 335.36MB) Codegen (bc+obj): 00:48:03.114563685 ( 343.36MB) Codegen (linking): 00:00:00.431805899 ( 343.36MB) Codegen (bc+obj): - no previous .o files were reused ... Finished in 206.41 milliseconds 3 examples, 0 failures, 0 errors, 0 pending ``` As shown above, Codegen (bc+obj) took 48 minutes and 3 seconds. Next, I inspected the timestamps in the cache directory: ```auto ls -ltr --time-style="+%H:%M:%S" tmp/home-kojix2-Crystal-captcha-spec/ | tail -n 5 ``` ```console -rw-r--r--. 1 kojix2 kojix2 4316 18:17:53 S-lice40U-I-nt3241.o1.bc -rw-r--r--. 1 kojix2 kojix2 4248 18:17:53 S-lice40U-I-nt3241.o1.o -rw-r--r--. 1 kojix2 kojix2 690024 19:05:28 S-taticA-rray40U-I-nt3244321638441.o1.o -rw-r--r--. 1 kojix2 kojix2 2596 19:05:28 C-rI-mage5858T-I-F-F-.o1.bc -rw-r--r--. 1 kojix2 kojix2 2912 19:05:28 C-rI-mage5858T-I-F-F-.o1.o ``` It appears that generating `S-taticA-rray40U-I-nt3244321638441.o1.o` alone took 47 minutes and 35 seconds. ---
**Author:** ![ysbaddaden](https://yyz2.discourse-cdn.com/flex036/user_avatar/forum.crystal-lang.org/ysbaddaden/32/2252_2.png) [@ysbaddaden](https://forum.crystal-lang.org/u/ysbaddaden)\ **Post date:** [January 12, 2026, 11:13am UTC](https://forum.crystal-lang.org/t/crystal-spec-is-very-slow-and-very-high-cpu-usage-when-the-first-time-run-crystal-spec-o1/8646/6 "2026-01-12T11:13:24Z")
Oh, nice idea to use file timestamps! ---
**Author:** ![kojix2](https://yyz2.discourse-cdn.com/flex036/user_avatar/forum.crystal-lang.org/kojix2/32/1583_2.png) [@kojix2](https://forum.crystal-lang.org/u/kojix2)\ **Post date:** [January 12, 2026, 2:39pm UTC](https://forum.crystal-lang.org/t/crystal-spec-is-very-slow-and-very-high-cpu-usage-when-the-first-time-run-crystal-spec-o1/8646/7 "2026-01-12T14:39:43Z")
To be honest, what I still don’t quite understand is why the `llc` command doesn’t take much time to compile the .bc file. ```auto llc-18 -O1 -time-passes -filetype=obj S-taticA-rray40U-I-nt3244321638441.o1.bc -o out.o ``` This takes only 29 seconds. What is the difference between the Crystal compiler and the llc command? ---
**Author:** ![naqvis](https://yyz2.discourse-cdn.com/flex036/user_avatar/forum.crystal-lang.org/naqvis/32/906_2.png) [@naqvis](https://forum.crystal-lang.org/u/naqvis)\ **Post date:** [January 13, 2026, 3:02am UTC](https://forum.crystal-lang.org/t/crystal-spec-is-very-slow-and-very-high-cpu-usage-when-the-first-time-run-crystal-spec-o1/8646/8 "2026-01-13T03:02:57Z")
> [@kojix2](#): > > What is the difference between the Crystal compiler and the llc command? `llc -O1` only runs backend optimizations like instruction selection and register allocation. It does NOT run middle-end optimization passes, which is why it’s fast. Crystal’s compiler runs the full LLVM optimization pipeline via `LLVM.run_passes` with “default”. This is equivalent to running `opt -O1` first, then doing what `llc` does. ---
**Author:** ![kojix2](https://yyz2.discourse-cdn.com/flex036/user_avatar/forum.crystal-lang.org/kojix2/32/1583_2.png) [@kojix2](https://forum.crystal-lang.org/u/kojix2)\ **Post date:** [January 13, 2026, 1:44pm UTC](https://forum.crystal-lang.org/t/crystal-spec-is-very-slow-and-very-high-cpu-usage-when-the-first-time-run-crystal-spec-o1/8646/9 "2026-01-13T13:44:16Z")
That’s right. But, when I run the `opt` command, it takes less than a second. ```sh opt-18 -O1 -time-passes S-taticA-rray40U-I-nt3244321638441.o1.bc -o /dev/null ``` So it’s hard to tell which part of the pipeline is actually slow. In fact, `llc` seems to be taking longer. My understanding of compiler internal is fairly limited. And investigation is based entirely on AI support. Can anyone reproduce the long object-file generation time from `.bc` files, and explain why? ---
**Author:** ![zw963](https://yyz2.discourse-cdn.com/flex036/user_avatar/forum.crystal-lang.org/zw963/32/1623_2.png) [@zw963](https://forum.crystal-lang.org/u/zw963)\ **Post date:** [January 13, 2026, 3:04pm UTC](https://forum.crystal-lang.org/t/crystal-spec-is-very-slow-and-very-high-cpu-usage-when-the-first-time-run-crystal-spec-o1/8646/10 "2026-01-13T15:04:46Z")
Just for clarify, because crimage use Slice instead of StaticArray in [this](https://github.com/naqvis/crimage/commit/b65f67df27f4cb8807a34d94a90dfd4162acf45a) recent commit, so, if someone want to reproduce, need use code before this commit instead.