Use i1 as result type for llvm.vp.icmp

[dkurt]

icmp return type should be i1 but not i32: https://llvm.org/docs/LangRef.html#llvm-vp-icmp-intrinsics

This patch fixed an error which I got on RISC-V:

Intrinsic has incorrect return type!
ptr @llvm.vp.icmp.nxv2i32

[steven-johnson]

The buildbots seem to disagree with you, this is failing literally everywhere

[dkurt]

@steven-johnson, sorry, forgot that type is still used in the if statement. Fixed that.

[steven-johnson]

Looks like it has broken Hexagon/HVX codegen (see https://buildbot.halide-lang.org/master/#/builders/42/builds/549/steps/29/logs/correctness_simd_op_check_hvx):

Failed: vdelta(v*,v*) did not generate for target=hexagon-32-noos-hvx-hvx_128-no_asserts-no_bounds_query-no_runtime vector_width=512. Instead we got:

attn @pranavb-ca for suggestions

[dkurt]

Thanks! I've tried debug Hexagon locally but I'm not able to download the SDK.

[steven-johnson]

Thanks! I've tried debug Hexagon locally but I'm not able to download the SDK.

Update: this may be unrelated to your PR, I'm getting an apparently-similar failure in main with top-of-tree LLVM. Investigating.

[pranavb-ca]

@aankit-ca - fyi, since I am temporarily working in a timezone ~12 hrs away.

[steven-johnson]

The "failure" can be ignored, see https://reviews.llvm.org/rG073d5e5945c428e20db0884943e6dcb7ff2158df -- output is different but better, I will update the testcase today. In the meantime, this PR is good to go.

[steven-johnson]

LGTM but adding @abadams as he knows this better than I

[abadams]

Actually I think we want @zvookin

[dkurt]

Is RISC-V instance on buildbot a good option for tests? QEMU can be used to run JIT and AOT tests. I've tried collect some brief proposals here: halide/build_bot#214.

[zvookin]

I'll try to get to this shortly, but I don't think this is the right thing. (The right thing is for the internals of that function to always use i1 and to cast the result to the requested result type. Also, needs a test, which is on me.)

[dkurt]

see #7205