[llvm] [NVPTX] Fix crash caused by ComputePTXValueVTs (PR #104524)

[Artem Belevich via llvm-commits]

================
@@ -0,0 +1,61 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc < %s -march=nvptx64 -mcpu=sm_20 | FileCheck %s
+
+target triple = "nvptx-nvidia-cuda"
+
+define <6 x half> @half6() {
----------------
Artem-B wrote:

We do want to have a more exhaustive coverage here for non-power-of-2 sized vectors.

For the sake of completeness I'd enumerate all variants for i1/i8/i16/i32/i64 vectors with the total size <= 24 bytes (16 bytes to cover the max possible vector size we may ever lower natively + one element, so we cover cases of power-of-2 vector+remainder). Vectorized lowering of loads/stores is pretty hairy, so it would be nice to know that we're not missing any other uncommon corner cases.


https://github.com/llvm/llvm-project/pull/104524