[llvm] [NFC][LV] Improve ee with sideeffects legality test (PR #158275)

[Graham Hunter via llvm-commits]

================
@@ -604,36 +604,31 @@ exit:
 
 ;; ICE was caused by assert for the load used in the uncountable exit condition
 ;; being guaranteed to execute.
- at e = external addrspace(21) global [4 x i8]
-define void @crash_conditional_load_for_uncountable_exit() {
+ at ee.global = external global [4 x i8]
+define void @crash_conditional_load_for_uncountable_exit(ptr dereferenceable(40) noalias %store.area) {
 ; CHECK-LABEL: LV: Checking a loop in 'crash_conditional_load_for_uncountable_exit'
 ; CHECK:       LV: Not vectorizing: Load for uncountable exit not guaranteed to execute.
 entry:
-  br label %cont
-
-handler.out_of_bounds:
-  unreachable
+  br label %for.body
 
-cont:
-  %h.06 = phi i64 [ 0, %entry ], [ %inc, %a.exit ]
-  %arrayidx = getelementptr i8, ptr addrspace(21) @e, i64 %h.06
-  br i1 false, label %cont1, label %handler.type_mismatch
+for.body:
+  %iv = phi i64 [ 0, %entry ], [ %iv.next, %for.inc ]
+  %ee.addr = getelementptr i8, ptr @ee.global, i64 %iv
+  br i1 false, label %ee.block, label %invalid.block
----------------
huntergr-arm wrote:

Providing the condition via argument dropped out earlier with `Cannot vectorize early exit loop with more than one early exit`. Disabling that check bails out at `Cannot vectorize potentially faulting early exit loop`. Disabling that check as well will give us the crash. So I'll add it as part of the second test, but I don't think we would have hit it before.

https://github.com/llvm/llvm-project/pull/158275