[MC][AsmParser] Check the list of supported condition codes

[atrosinenko]

At now, a number of conditional code names are supported across the instructions. These are short forms like .eq, .le, .ge as well as the longer forms mentioned in the VM Primer (such as .if_eq, .if_le, .if_ge, .if_gt_or_eq). The mapping of strings to EraVMCC constants is listed in parseExplicitCondition function in EraVMAsmParser.cpp file (permalink at the time of writing).

After redesigning the assembly syntax, only the shorter form is likely to remain, so parseExplicitCondition function should be updated accordingly.

[asl]

@sayon Will you please comment on what should be supported?

[sayon]

The longer form (like .if_eq) is unnecessary, the short forms are sufficient to us. So,

• eq
• lt
• gt
• ne
• ge
• le

@akiramenai @hedgar2017 do you think we should support something like of as an alias to lt?

[hedgar2017]

@sayon I dislike unnecessary aliases personally.

[akiramenai]

The issue here is that our lt is not quite lt, it's lt for sub. But it's more ISA issue than asm, we reuse flags for different purposes in different context. Whatever naming we choose, it's misleading for some instructions. Maybe we could rename eq -> z as it works like a conventional zero flag almost everywhere. Maybe, we want to rename lt to some single letter same as gt and make xz, yz suffixes to check 2 flags combinations. An abstract 2nd flag and 3rd flag with no associations with the name might be the best solution in our case.

[asl]

@akiramenai @hedgar2017 @sayon Ok, let me assign this issue to you, so you will decide on the intended functionality :)

[akiramenai]

I think we can go as @sayon said. of was reverted from the backend, and I'm ok to keep the things as they are.

[akiramenai]

As no one objects, I think we can close the issue.

[atrosinenko]

As this needs to be actually updated in the code (just a few lines), reopening it and assigning to myself.

[atrosinenko]

@akiramenai Is "gt or lt" predicate still used (see encoding) and which name should it have?

[atrosinenko]

The last remaining question to clarify is: the spec on encoding mentions 8 predicates:

Definition encode_predicate (p:predicate) : BITS 3 :=
  # match p with
    | IfAlways ⇒ 0
    | IfGT ⇒ 1
    | IfLT ⇒ 2
    | IfEQ ⇒ 3
    | IfGE ⇒ 4
    | IfLE ⇒ 5
    | IfNotEQ ⇒ 6
    | IfGTOrLT ⇒ 7
    end
.

The asm parser understands 7 of them: 6 explicit .<code> suffixes and an implicit COND_NONE if neither suffix is specified. Other patches for code emission and disassembler assume 7 possible predicates, too. Is it correct that the last condition code, IfGTOrLT ⇒ 7, is not used anymore?

Reassigning the task - please feel free to close it if IfGTOrLT is actually unused or assign it back to me if this condition code has to be accounted for.

[akiramenai]

@atrosinenko it looks like we mislead you wrt IfGTOrLT. It's only unused by current code generator, but we plan to keep on VM level. For now we can use gl mnemonic for it (or whatever name the current assembler accepts if it's more convenient), but we plan to rename our flags as it confuses even us internally. We are still discussing the final naming scheme.

[atrosinenko]

Added support of .gtlt predicate in #559.