Bug 77740

There seems to be a 'writes_accumulator' flag that is set to TRUE for instructions that are known to write to the accumulator. The scheduling code currently checks this flag on Gen >= 6 to compute dependencies among instruction that read/write the accumulator.

For Gen < 6, there seems that for some very specific cases where this flag is set to TRUE even if the scheduling code is not going to use it. As far as I can see this is because the 'writes_accumulator' flag can also be used in the optimization passes to remove dead code (and we don't want to eliminate code that writes to the accumulator values that we intend to use).

So, I think for this task the 'writes_accumulator' flag alone is not enough: in the scheduling code we need to know if the instruction can write to the accumulator independently of whether we are expecting to use that value or not and in Gen < 6, as far as I can see, there is a bunch of operations that can write implicitly to the accumulator (mostly opcodes > 63).

I think the way to go about this would be to have:

bool instruction_backend::writes_accumulator_implicitly(int gen);

that we can call from the scheduling code to know if a particular instruction can write implicitly  to the accumulator. In Gen >= 6, it would simply return the value of 'writes_accumulator' but for Gen < 6 it would check also the opcodes.

I think with that operation in place we could unify the scheduling code paths for all Gens.

I'll write a patch for this but I will need help with the testing since I am running on Gen 7.

Created attachment 98117 [details] [review]
Patch

I managed to get an ironlake laptop so I could test the patch myself. I tested with a small program that implemented a simple lightning model demo and seems to work fine.

According to the debug traces in instruction_scheduler::schedule_instructions() I can see that the patch is indeed influencing instruction scheduling (I'll attach the traces).

Might need more thorough testing though, please let me know if there are specific tests that you would like me to run to verify the patch before sending it for review to the mailing list.

Created attachment 98122 [details]
Scheduling plan (before patch)

Scheduling plan before the patch for sample program

Created attachment 98123 [details]
Scheduling plan (after patch)

Scheduling plan after the patch for sample program

A piglit test run raises problems though so this still needs more work.

Fixed patch and sent it to the mailing list for review. Passes piglit tests on  ironlake and ivybridge but I could not test on other gens:

http://lists.freedesktop.org/archives/mesa-dev/2014-May/059001.html

Committed as http://cgit.freedesktop.org/mesa/mesa/commit/?id=520dfa4b5c64a09bb34c7e912b102fb99aef7699