This is an archive of the discontinued LLVM Phabricator instance.

[llvm-exegesis][AArch64] Handle OPERAND_UNKNOWN
Needs ReviewPublic

Authored by kpdev42 on Apr 7 2023, 2:38 AM.

Details

Summary

Patch allows running many instructions on arm64, including ALU, loads, stores, etc.

Diff Detail

Event Timeline

kpdev42 created this revision.Apr 7 2023, 2:38 AM
kpdev42 requested review of this revision.Apr 7 2023, 2:38 AM

ping
@courbet could you please take a look?

I wanted gchatelet@ to have a look at this, because he wrote the Instruction/Operand code. In particular, I'm wondering why we're not using the info on the actual Operand value rather than relying on the MCOperandInfo for the operand. Specifically, we do know that this is a register since we return true for isReg, so we could just bake this knowledge into getExplicitOperandInfo in a target-independant way.

That being said, I don't know how the AArch64 backend is design, but it feels weird that the target info for that opearnd is OPERAND_UNKNOWN if we know that it's a register.

I wanted gchatelet@ to have a look at this, because he wrote the Instruction/Operand code. In particular, I'm wondering why we're not using the info on the actual Operand value rather than relying on the MCOperandInfo for the operand. Specifically, we do know that this is a register since we return true for isReg, so we could just bake this knowledge into getExplicitOperandInfo in a target-independant way.

That being said, I don't know how the AArch64 backend is design, but it feels weird that the target info for that opearnd is OPERAND_UNKNOWN if we know that it's a register.

On aarch64 OPERAND_UNKNOWN can be register, shift, extend or noop. If it is register we simply pass it further for randomizeMCOperand to handle it accordingly. If it is immediate then we set it to noop for most ofvindtructions, except those excepting non zero shift value (MOVIv and friends). I am not familiar with other platforms so I'm not handling OPERAND_UNKNOWN for them at all (even when it is register). We have better randomization for aarch64 immediate, which can handle shift/extend which we plan to submit step by step if this one goes in.