[clang] vsetvli/vsetivli does not permit lmul operand to be optional

[ThinkOpenly]

The following instructions should all produce the same encoding:

[vsetvli.s.txt](https://github.com/user-attachments/files/20694598/vsetvli.s.txt)

[vsetvli.s.txt](https://github.com/user-attachments/files/20694603/vsetvli.s.txt)

vsetvli t0, a0, 192
vsetvli t0, a0, e8, m1, ta, ma
vsetvli t0, a0, e8, ta, ma

However, clang (20) does not accept the 3rd syntax that omits the lmul operand "m1", which is the default value:

$ clang -c vsetvli.s
vsetvli.s:21:25: error: operand must be e[8|16|32|64],m[1|2|4|8|f2|f4|f8],[ta|tu],[ma|mu]
        vsetvli t0, a0, e8, ta, ma
                        ^

Originally reported riscv-software-src/riscv-unified-db#838.

[ThinkOpenly]

vsetvli.s.txt

[llvmbot]

@llvm/issue-subscribers-backend-risc-v

Author: Paul Clarke (ThinkOpenly)

The following instructions should all produce the same encoding: ```

vsetvli.s.txt

vsetvli.s.txt

vsetvli t0, a0, 192
vsetvli t0, a0, e8, m1, ta, ma
vsetvli t0, a0, e8, ta, ma

However, clang (20) does not accept the 3rd syntax that omits the lmul operand "m1", which is the default value:

$ clang -c vsetvli.s
vsetvli.s:21:25: error: operand must be e[8|16|32|64],m[1|2|4|8|f2|f4|f8],[ta|tu],[ma|mu]
vsetvli t0, a0, e8, ta, ma
^

Originally reported https://github.com/riscv-software-src/riscv-unified-db/pull/838.
</details>

[topperc]

The RISC-V vector spec used to say that the fields were optional, but it no longer says this. Need to check what it said at the time of ratification.

Here's one of the commits that removed such text riscv/riscv-isa-manual@75e857d

[ThinkOpenly]

Per https://lf-riscv.atlassian.net/wiki/spaces/HOME/pages/16154732/Ratified+Extensions, the ratification date is November 2021.

The 20240411 version of the RISC-V ISA includes the text:

m1 # LMUL=1, assumed if m setting absent

[lukel97]

As much as I wish this was fixed pre-ratification, it looks like binutils accepts LMUL being optional: https://godbolt.org/z/hT53cKs4r

It might then be worthwhile accepting the old syntax just for compatibility?

[wangpc-pp]

There was a PR to support this(#115277, it costed me a long time to find it 😄).

[lukel97]

We discussed this at the LLVM RISC-V sync up and the consensus was that we should try and match binutils behaviour. Also with the caveat that all vtypei operands are optional in binutils, including the policy flags

[lenary]

Given I was an advocate for fixing this in the meeting, and no one else volunteered, I'm going to start looking at fixing this. I'm not sure how long it will be until I have a patch.

[lenary]

The binutils behaviour is... interesting. All operands are optional, but you must have at least one of them: https://godbolt.org/z/f81x1bjME

I think that we can manage the same in LLVM.