[llvm-exegesis] Add support for instruction with memory operands (uop mode)

[llvmbot]

Bugzilla Link	36906
Resolution	FIXED
Resolved on	Nov 09, 2018 02:53
Version	trunk
OS	Linux
Blocks	#35023 #35432
Reporter	LLVM Bugzilla Contributor
CC	@adibiagio,@legrosbuffle,@gchatelet,@gregbedwell,@LebedevRI,@RKSimon,@rotateright

[llvmbot]

assigned to @gchatelet

[RKSimon]

Just seeing if there is much movement on this - latency mode might be trickier to accomplish but non-dependent code snippets for uops/throughput are less problematic.

It would need to keep to a small memory footprint to assume all L1D hits (assume maybe just 16k?) and preferably avoid the same cachelines for at least the number of buffer entries in the LS unit - ideally all this info would be in MCExtraProcessorInfo (some of it is currently dumped in X86TargetTransformInfo.cpp).

[llvmbot]

I plan to work on this as soon as I'm done with updating the SKX model for non-memory instructions.
I already did this in our pre-llvm experiment (https://github.com/google/EXEgesis/blob/master/exegesis/itineraries/compute_itineraries.cc), I just have to port it to llvm-exegesis, which should not be too much work.

[LebedevRI]

I thought this worked before, but now it no longer works:
$ ./bin/llvm-exegesis -num-repetitions=10000 -mode=uops -opcode-name=VFRCZPSrm
VFRCZPSrm: unsupported opcode: access segment memory

[legrosbuffle]

It should have :) Guillaume added an extra check on friday. Assigning to him for investigation.

[gchatelet]

Yes it does not work (on purpose) because we don't know how to handle memory segment related instruction yet.
Before it would do garbage, now it returns an error.

[LebedevRI]

Before it would do garbage, now it returns an error.

Oh, ok, that makes so much more sense now :)
Thanks.

[legrosbuffle]

Actually the check that was added is too restrictive: we are explicitly not using the segment register for instructions that have 5 operands.