[llvm] [Inliner][TTI][AMDGPU] Add size growth TTI hook for unreachable-termed bb (PR #217014)

[Janek van Oirschot via llvm-commits]

================
@@ -0,0 +1,32 @@
+; RUN: opt -mtriple=amdgcn-amd-amdhsa -inline-call-penalty=0 -inline-threshold=1 -passes=inline -S < %s | FileCheck %s
+
+ at g = external global i32
+
+define void @small() {
+  store volatile i32 1, ptr @g
+  ret void
+}
+
+define void @big() {
+  store volatile i32 1, ptr @g
+  store volatile i32 2, ptr @g
+  store volatile i32 3, ptr @g
+  store volatile i32 4, ptr @g
+  store volatile i32 5, ptr @g
+  ret void
+}
+
+define void @caller_small() {
+; CHECK-LABEL: @caller_small
+; CHECK:       store volatile i32 1, ptr @g
+; CHECK-NOT:   call void @small
+  call void @small()
+  unreachable
+}
+
+define void @caller_big() {
+; CHECK-LABEL: @caller_big
+; CHECK:       call void @big
----------------
JanekvO wrote:

It's not inlined for a different reason as it now goes through the cost analysis of the inliner which determines this shouldn't be inlined. The nuance as I have it as mental model: `allowSizeGrowth` can stop a call from being inlined but it cannot force a call to get inlined.

https://github.com/llvm/llvm-project/pull/217014