`conj` seems to compile incorrectly on 1.10

[kshyatt]

QuantumKitHub/Strided.jl#64 is fixing this for me -- using conj in a AMD device kernel seems to not work, but writing my own function which does the same thing does work. This problem doesn't occur on 1.11 or 1.12.

AMDGPU v2.5.1
GPUCompiler v1.21.0
KernelAbstractions v0.9.41
LLVM v9.9.0

What's weird is that something like map(conj, A) on a ROCArray seems to work fine...

[vchuravy]

Do you have a MWE that one can run?

Or @device_code dir=conj dumps?

[kshyatt]

The MWE isn't very minimal, sadly :(. I'll try to get a dump!

[kshyatt]

Here are the dumps for the failing version https://gist.github.com/kshyatt/8d31435e7569d17bbb6837c7557a0a44

[kshyatt]

And for the working one https://gist.github.com/kshyatt/b2b00bfc80b69b3601bc49d4def1755c

[luraess]

Thanks for reporting. I recall this relates to a discussion thread on Slack. Could you add some context as to allow investigation (besides the dumps)? As suggested, some kind of MWE would be very helpful as well.

[kshyatt]

I'll try to find a MWE. This was kind of hard to provoke but I'll see what I can do tomorrow!

[luraess]

In any case, it would be helpful to get the pointer to what makes it failing in order to reproduce and debug (or put a robot on it).

[vchuravy]

Can you post how to execute the code? What GPUs did you use?
What are the numerical inputs and results?

[kshyatt]

I’m going to as soon as I have some time to make a MWe? Sorry, did I run over your dog without realizing?

…

On Thu, Jun 18, 2026 at 11:25 AM Valentin Churavy ***@***.***> wrote: *vchuravy* left a comment (JuliaGPU/AMDGPU.jl#931) <#931 (comment)> Can you post how to execute the code? What GPUs did you use? What are the numerical inputs and results? — Reply to this email directly, view it on GitHub <#931?email_source=notifications&email_token=AAGKJY25DRHGFJJDU7OHUED5AOYRLA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTINZUGAZDSOJZG44KM4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KYZTPN52GK4S7MNWGSY3L#issuecomment-4740299978>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAGKJY6KRRQSRMJO7LXLM5T5AOYRLAVCNFSNUABFKJSXA33TNF2G64TZHMZDONRWHEZTANRYHNEXG43VMU5TINRXGMYTMNJTGI32C5QC> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/AAGKJY55G25PNTDYNVD3A3D5AOYRLA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTINZUGAZDSOJZG44KM4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KUZTPN52GK4S7NFXXG> and Android <https://github.com/notifications/mobile/android/AAGKJYZVUUN2GQWJLBGACCL5AOYRLA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTINZUGAZDSOJZG44KM4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2K4ZTPN52GK4S7MFXGI4TPNFSA>. Download it today! You are receiving this because you authored the thread.Message ID: ***@***.***>

[vchuravy]

Sorry, I meant instead of preparing an MWE. Compiler bugs are like toothache.

I was looking through the compiler dumps and I can't see any obvious differences.

[kshyatt]

I have an idea for how to replicate this that doesn’t involve a lot of our custom types which I’ll try to run later today if I can

…

On Thu, Jun 18, 2026 at 11:40 AM Valentin Churavy ***@***.***> wrote: *vchuravy* left a comment (JuliaGPU/AMDGPU.jl#931) <#931 (comment)> Sorry, I meant instead of preparing an MWE. Compiler bugs are like tooth ache. — Reply to this email directly, view it on GitHub <#931?email_source=notifications&email_token=AAGKJY7GJUACTGX4USK7EZ35AO2KHA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTINZUGA2TCMRSGYYKM4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KYZTPN52GK4S7MNWGSY3L#issuecomment-4740512260>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAGKJY7JUXFNVRQ6CXGVXCT5AO2KHAVCNFSNUABFKJSXA33TNF2G64TZHMZDONRWHEZTANRYHNEXG43VMU5TINRXGMYTMNJTGI32C5QC> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/AAGKJY26QKV57KROBV3AC5D5AO2KHA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTINZUGA2TCMRSGYYKM4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KUZTPN52GK4S7NFXXG> and Android <https://github.com/notifications/mobile/android/AAGKJY7URLQZZ6NIHTTJI7L5AO2KHA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTINZUGA2TCMRSGYYKM4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2K4ZTPN52GK4S7MFXGI4TPNFSA>. Download it today! You are receiving this because you authored the thread.Message ID: ***@***.***>

[kshyatt]

Alternatively if you want to do some archaeology there’s the discussion and logs for the latest PR at Strided.jl

…

On Thu, Jun 18, 2026 at 11:41 AM Katie H ***@***.***> wrote: I have an idea for how to replicate this that doesn’t involve a lot of our custom types which I’ll try to run later today if I can On Thu, Jun 18, 2026 at 11:40 AM Valentin Churavy < ***@***.***> wrote: > *vchuravy* left a comment (JuliaGPU/AMDGPU.jl#931) > <#931 (comment)> > > Sorry, I meant instead of preparing an MWE. Compiler bugs are like tooth > ache. > > — > Reply to this email directly, view it on GitHub > <#931?email_source=notifications&email_token=AAGKJY7GJUACTGX4USK7EZ35AO2KHA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTINZUGA2TCMRSGYYKM4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KYZTPN52GK4S7MNWGSY3L#issuecomment-4740512260>, > or unsubscribe > <https://github.com/notifications/unsubscribe-auth/AAGKJY7JUXFNVRQ6CXGVXCT5AO2KHAVCNFSNUABFKJSXA33TNF2G64TZHMZDONRWHEZTANRYHNEXG43VMU5TINRXGMYTMNJTGI32C5QC> > . > Triage notifications, keep track of coding agent tasks and review pull > requests on the go with GitHub Mobile for iOS > <https://github.com/notifications/mobile/ios/AAGKJY26QKV57KROBV3AC5D5AO2KHA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTINZUGA2TCMRSGYYKM4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KUZTPN52GK4S7NFXXG> > and Android > <https://github.com/notifications/mobile/android/AAGKJY7URLQZZ6NIHTTJI7L5AO2KHA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTINZUGA2TCMRSGYYKM4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2K4ZTPN52GK4S7MFXGI4TPNFSA>. > Download it today! > You are receiving this because you authored the thread.Message ID: > ***@***.***> >

[vchuravy]

With @kshyatt help and looking through Strided.jl I managed a reproducer:

https://gist.github.com/vchuravy/84a7773b6b00553447f887183a741dee#file-minimal_repro-jl

vchuravy@loki ~/s/T/mwe_gist (main)> julia +1.10 --project=. minimal_repro.jl
Julia 1.10.11 | HIPDevice(id=1, name=AMD Radeon 780M Graphics, gcn_arch=gfx1103)
[identity] ok = true   (0 / 360 differ)
[conj    ] ok = false   (360 / 360 differ)
    @ 1: cpu=1.1597548f0 + 0.9957996f0im  gpu=0.042156354f0 + 1.3683695f0im
[_conj   ] ok = true   (0 / 360 differ)
vchuravy@loki ~/s/T/mwe_gist (main)> julia +1.11 --project=. minimal_repro.jl
Julia 1.11.9 | HIPDevice(id=1, name=AMD Radeon 780M Graphics, gcn_arch=gfx1103)
^[[A[identity] ok = true   (0 / 360 differ)
[conj    ] ok = false   (360 / 360 differ)
    @ 1: cpu=-1.1009178f0 - 1.230011f0im  gpu=-0.8775279f0 - 0.889241f0im
[_conj   ] ok = true   (0 / 360 differ)
vchuravy@loki ~/s/T/mwe_gist (main)> julia +1.12 --project=. minimal_repro.jl
Julia 1.12.6 | HIPDevice(id=1, name=AMD Radeon 780M Graphics, gcn_arch=gfx1103)
[identity] ok = true   (0 / 360 differ)
[conj    ] ok = true   (0 / 360 differ)
[_conj   ] ok = true   (0 / 360 differ)

I am trying to narrow down if it is in the backend, middle-end or frontend.

[vchuravy]

Likely a backend issue fixed by LLVM18, trying a bisect.