[llvm] [llvm][NVPTX] Don't reorder MIs that construct a PTX function call (PR #116522)

[Justin Fargnoli via llvm-commits]

================
@@ -0,0 +1,28 @@
+; RUN: llc -O3 -march=nvptx64 -enable-misched %s -o - | FileCheck %s
+
+target datalayout = "e-i64:64-i128:128-v16:16-v32:32-n16:32:64"
+target triple = "nvptx64-nvidia-cuda"
+
+define ptx_kernel void @my_kernel(i32 %arg_0, i32 %arg_3.tr, i32 %"$$i_l40_0_t23.0") {
+Entry_BB:
+  br label %BB1692
+
+BB1692:                                           ; preds = %BB1692, %Entry_BB
+  %"$$i_l40_0_t23.02" = phi i32 [ 0, %Entry_BB ], [ 1, %BB1692 ]
----------------
justinfargnoli wrote:

> Running update_llc_test_checks.py generates too many unnecessary checks as below.
Is there a way to control or remove some unneeded checks?

In some sense, that's the intention of `update_llc_test_checks.py`. It generates checks for the entire function. @Artem-B likes it because you don't have to rely on another person to write the test correctly. I'm a fan of it because it makes it easier to see how my change affects the entire program. 

For better or for worse, it's the existing convention for the vast majority of the tests in `llvm/tests/CodeGen/*`. 

https://github.com/llvm/llvm-project/pull/116522