[NVPTX] Improve lowering of v4i8

[Artem-B]

Make it a legal type and plumb through lowering of relevant instructions.

[ThomasRaoux]

LGTM

[ThomasRaoux]

Actually trying out the patch in triton it causes some invalid ptx to get emitted (see comment)

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[Artem-B]

I still need to test it on live code, though we do not have much code that would end up using v4i8.

The generated PTX checked in llvm/test/CodeGen/NVPTX/i8x4-instructions.ll could use extra scrutiny.

[ThomasRaoux]

I ran the patch on our triton kernels and I don't see any functional problems left.

[ThomasRaoux]

Looks like it required quite a lot of cases to be handled :(
Thanks for doing this, it solves some of the problems triton had with latest LLVM. Changes look good to me.

[Artem-B]

I see one suspicious failure in tensorflow tests. I suspect I've messed something up in v4i8 comparison.

[Artem-B]

I see one suspicious failure in tensorflow tests. I suspect I've messed something up in v4i8 comparison.

Yup, there is a problem:

Successfully custom legalized node
 ... replacing: t10: v4i8 = BUILD_VECTOR Constant:i16<-128>, Constant:i16<-128>, Constant:i16<-128>, Constant:i16<-128>
     with:      t17: v4i8 = bitcast Constant:i32<-128>

[Artem-B]

I see one suspicious failure in tensorflow tests. I suspect I've messed something up in v4i8 comparison.

Yup, there is a problem:

Successfully custom legalized node
 ... replacing: t10: v4i8 = BUILD_VECTOR Constant:i16<-128>, Constant:i16<-128>, Constant:i16<-128>, Constant:i16<-128>
     with:      t17: v4i8 = bitcast Constant:i32<-128>

Resolved by 9821e90

[Artem-B]

Found another issue. We merge four independent byte loads with align 1 into a 32-bit load, which fails at runtime on misaligned pointers.

%t0 = type { [17 x i8] }

@shared_storage = linkonce_odr local_unnamed_addr addrspace(3) global %t0 undef, align 1

define <4 x i8> @in_v4i8(<4 x i8> %x, <4 x i8> %y) nounwind {
  %v = load <4 x i8>, ptr getelementptr inbounds (i8, ptr addrspacecast (ptr addrspace(3) @shared_storage to ptr), i64 9), align 1
  ret <4 x i8> %v
}

        mov.u64         %rd1, shared_storage;
        cvta.shared.u64         %rd2, %rd1;
        ld.u32  %r1, [%rd2+9];
        st.param.b32    [func_retval0+0], %r1;
        ret;

[Artem-B]

clang-format failure on GitHub is weird -- it just silently exits with an error.
I ran the same command locally and fixed one place it was not happy about.

The buildkite failure somewhere in RISC-V appears to be unrelated.

[abadams]

I believe this may be causing failures for u/srem. See #69124

Edit: Also causing failures when you sign-extend the result of a <4 x i8> comparison.