[llvm] [ARM][SLP] Fix cost function for SLP Vectorization of ZExt/SExt (PR #122713)

[David Green via llvm-commits]

================
@@ -0,0 +1,354 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt < %s -passes=slp-vectorizer -S -mtriple=arm-none-eabi --mattr=+mve | FileCheck %s
+
+
+define dso_local i64 @vadd(ptr noundef %0) #0 {
+; CHECK-LABEL: define dso_local i64 @vadd(
+; CHECK-SAME: ptr noundef [[TMP0:%.*]]) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:    [[TMP2:%.*]] = alloca ptr, align 4
+; CHECK-NEXT:    store ptr [[TMP0]], ptr [[TMP2]], align 4
+; CHECK-NEXT:    [[TMP3:%.*]] = load ptr, ptr [[TMP2]], align 4
+; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i32, ptr [[TMP3]], i32 0
+; CHECK-NEXT:    [[TMP5:%.*]] = load i32, ptr [[TMP4]], align 4
+; CHECK-NEXT:    [[TMP6:%.*]] = sext i32 [[TMP5]] to i64
+; CHECK-NEXT:    [[TMP7:%.*]] = load ptr, ptr [[TMP2]], align 4
+; CHECK-NEXT:    [[TMP8:%.*]] = getelementptr inbounds i32, ptr [[TMP7]], i32 1
+; CHECK-NEXT:    [[TMP9:%.*]] = load i32, ptr [[TMP8]], align 4
+; CHECK-NEXT:    [[TMP10:%.*]] = sext i32 [[TMP9]] to i64
+; CHECK-NEXT:    [[TMP11:%.*]] = add nsw i64 [[TMP6]], [[TMP10]]
+; CHECK-NEXT:    [[TMP12:%.*]] = load ptr, ptr [[TMP2]], align 4
+; CHECK-NEXT:    [[TMP13:%.*]] = getelementptr inbounds i32, ptr [[TMP12]], i32 2
+; CHECK-NEXT:    [[TMP14:%.*]] = load i32, ptr [[TMP13]], align 4
+; CHECK-NEXT:    [[TMP15:%.*]] = sext i32 [[TMP14]] to i64
+; CHECK-NEXT:    [[TMP16:%.*]] = add nsw i64 [[TMP11]], [[TMP15]]
+; CHECK-NEXT:    [[TMP17:%.*]] = load ptr, ptr [[TMP2]], align 4
+; CHECK-NEXT:    [[TMP18:%.*]] = getelementptr inbounds i32, ptr [[TMP17]], i32 3
+; CHECK-NEXT:    [[TMP19:%.*]] = load i32, ptr [[TMP18]], align 4
+; CHECK-NEXT:    [[TMP20:%.*]] = sext i32 [[TMP19]] to i64
+; CHECK-NEXT:    [[TMP21:%.*]] = add nsw i64 [[TMP16]], [[TMP20]]
+; CHECK-NEXT:    ret i64 [[TMP21]]
+;
+  %2 = alloca ptr, align 4
----------------
davemgreen wrote:

Can you run these tests through opt -O1 (or sroa/mem2reg) to clean up the allocas. They might be confusing the SLP vectorizer on their own.

https://github.com/llvm/llvm-project/pull/122713