[llvm] [AArch64] Sanitise pow inputs using a target DAG combine (PR #192958)

[Paul Walker via llvm-commits]

================
@@ -0,0 +1,244 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
+; RUN: llc -mattr=+neon,+sve -mtriple=aarch64-linux-gnu --vector-library=ArmPL < %s -o - | FileCheck --check-prefix=CHECK-ARMPL %s
+
+define <4 x float> @foo_false_is_pow_v4f32(<4 x float> %a, <4 x float> %b, <4 x float> %c, <4 x float> %d) {
+; CHECK-ARMPL-LABEL: foo_false_is_pow_v4f32:
+; CHECK-ARMPL:       // %bb.0:
+; CHECK-ARMPL-NEXT:    str x30, [sp, #-16]! // 8-byte Folded Spill
+; CHECK-ARMPL-NEXT:    .cfi_def_cfa_offset 16
+; CHECK-ARMPL-NEXT:    .cfi_offset w30, -16
----------------
paulwalker-arm wrote:

Perhaps add `nounwind` to the functions to remove the cfi stuff?

https://github.com/llvm/llvm-project/pull/192958