[llvm] [IR] Reject invalid partial.reduce intrinsics (PR #161831)

[Sander de Smalen via llvm-commits]

================
@@ -0,0 +1,19 @@
+; RUN: not llvm-as %s -o /dev/null 2>&1 | FileCheck %s
+
+define void @element_count_mismatch() {
+  ; CHECK: Invalid vector widths for partial reduction. The width of the input vector must be a known integer multiple of the width of the accumulator vector.
+  call <3 x i32> @llvm.vector.partial.reduce.add(<3 x i32> poison, <8 x i32> poison)
+
+  ; CHECK: Invalid vector widths for partial reduction. The width of the input vector must be a known integer multiple of the width of the accumulator vector.
+  call <vscale x 4 x i32> @llvm.vector.partial.reduce.add(<vscale x 4 x i32> poison, <8 x i32> poison)
+
+  ; CHECK: Invalid vector widths for partial reduction. The width of the input vector must be a known integer multiple of the width of the accumulator vector.
+  call <4 x i32> @llvm.vector.partial.reduce.add(<4 x i32> poison, <vscale x 8 x i32> poison)
----------------
sdesmalen-arm wrote:

I agree it could be lowered and is desirable to do so, but that's out of the scope of this PR because the practical interpretation of "known" by LLVM is that this is not supported. To support it, I'd also argue removing "known" from the LangRef, which currently states:

> The second argument is a vector with a length that is a known integer multiple of the result’s type

where "known" can be interpreted as "known at compile-time". I'm just locking this down for now so that it isn't accidentally used, but we can relax this restriction in a follow-up.

https://github.com/llvm/llvm-project/pull/161831