00:05 <italove> austriancoder: tomeu: I'm not sure if it's helps, but to give a bit of context, I do remember a test where `dst.use` was needed for a `store` and it was when both the `addr` and `val` were both uniform regs, which is why that line for "inst.dst.use = [...]" is the way it is currently.

00:07 <italove> that being said, I don't remember which test was it :(, but perhaps removing it will cause piglit regressions? maybe it's worth a try

05:51 <tomeu> italove: yeah, but that assumption broke in the tests I was looking at

05:51 <tomeu> eg. 0x00811033, 0x00000c04, 0x50000040, 0x76ffdfff, /* store.s32mem.x___, t0.xxxx, 0, -1 */; use or amode or reg bitfields are nonzero for store dest (amode=0,use=1,reg=1) tex not used but fields non-zero (id=0,amode=4,swiz=128)

05:52 <tomeu> I think I will write a script to run all of the piglit CL tests and dump the cmdstream and shaders to files

05:56 <austriancoder> italove: tomeu: if you can provide that opencl kernel if found that would be great.

05:57 <tomeu> oh, one such kernel is the one I pasted above, but I guess on your HW the blob doesn't set the flag

06:00 <tomeu> austriancoder: this is the full dump: https://paste.debian.net/1267649/

06:00 <tomeu> so I guess they reused that bit in newer HW revisions?

06:20 <austriancoder> tomeu: we need to figure it out.. will compare feature flags .

06:37 <tomeu> \o/