Problem with constant argument to v_cndmask_b32 on gfx1010

[Lolliedieb]

Hello :)

I have a kernel that was compiled with the drivers internal compiler and that I wanted to tweak using clrxasm. My problem is that when I disassemble it, one of the lines is for example

/000000000300: 020212ff 70eac201/ v_cndmask_b32 v1, 0x70eac201, v9, vcc_lo

so the VOP2 variant of the operation attached with an additional constant. Sadly the assembler rejects now this command with error message

Error: Literal in VOP3 encoding is illegal

but well this is no VOP3 (obviously) - at least not for Navi and the kernel was processed without complain with the disasm before (with parameters -aC).

Edit: If am allowed to ask a question here, too:
Is it save to remove the entries for .kernel_code_entry_offset to let them recompute when changing lines in the kernel? Or do I have to compute the new addresses by hand?

[matszpk]

1. In this case (this line) CLRX assembler when waves32 is disabled, an assembler should display different error: "Error: Required 2 scalar registers". I don't know what are using version of the CLRX. Maybe this is an older version with bug. If in this case is latest repository version then it is new bug. Maybe, you are using CLRX in different platform. I tried to assemble that line and no error happened when waves32 enabled. Moreover, the disassembler should add .wave32 directive if you include "-aC" options.
2. You can omit .kernel_code_entry_offset - CLRX automatically compute and set value of that entry.

[Lolliedieb]

The disassembly has a line " .use_wave32" included - so the disassembler realizes this.
The exact line for the assembler then is
./clrxasm -g gfx1010 --wave32 -o ./test.o ./kernel.asm

The clrx version reports itself as 0.1.9, I did build it from sources (master branch), build date (and so code version) was June 30th this year.

Note: when I remove the .use_wave32 & the --wave32 then I am getting exactly the "Error: Required 2 scalar registers" error message you predicted.

[matszpk]

I was using some sample code disassembled and I added exactly that instruction to source. Ofcourse code has been compiled for WAVE32. When I tried to assemble code, no error happened (even using .use_wave32 in sources and command line option --wave32). A disassembled code has been assembled correctly for Navi. I am guessing that in this unique case an assembler incorrectly detect vcc_lo argument as non VCC argument when WAVE32 enabled (that causes VOP3 encoding in some circumstances). However, that shouldn't be happened. I can't respond whether it is some compiler bug or unexpected memory bug in my program. If you can give some info about platform for which you build CLRX, maybe that can help to find that bug.