[PATCH] D24338: [InlineCost] Remove CallPenalty and change MinSizeThreshold to 5

[Eric Christopher via llvm-commits]

echristo added a subscriber: echristo.
echristo added a comment.

Calls on Power are actually slower than a normal instruction, so inlining is even more important for performance. This probably needs some benchmarking and perhaps a way to conditionalize cost. A few more inline comments as well.

Thanks!

-eric


================
Comment at: lib/Transforms/Scalar/LoopStrengthReduce.cpp:1064
@@ -1063,2 +1063,3 @@
     }
+
     AddRecCost += 1; /// TODO: This should be a function of the stride.
----------------
Whitespace :)

================
Comment at: test/Transforms/Inline/ephemeral.ll:22
@@ -23,2 +21,3 @@
 
 ; @inner() should be inlined for -Oz.
+; CHECK-NOT: call i32 @inner
----------------
Since you're having to change the function here it seems reasonable to document it a bit more with how the structure of the function is affecting the inlining?

================
Comment at: test/Transforms/Inline/inline_minisize.ll:25
@@ -24,3 +24,2 @@
   store i32 0, i32* %i, align 4
-  store i32 0, i32* %i, align 4
   br label %for.cond
----------------
What happened here?

================
Comment at: test/Transforms/Inline/inline_minisize.ll:200
@@ -200,3 +199,3 @@
 
-define i32 @fct3(i32 %c) nounwind uwtable ssp {
+define i32 @fct3(i32 %c) optsize nounwind uwtable ssp {
 entry:
----------------
not minsize?

================
Comment at: test/Transforms/Inline/nonnull.ll:1
@@ -1,2 +1,2 @@
-; RUN: opt -S -inline %s | FileCheck %s
+; RUN: opt -S -inline %s -inline-threshold=20 | FileCheck %s
 
----------------
Not sure what to do about the tests that need a threshold change here, it seems slightly off to add a threshold to them. Thoughts on alternatives?


Repository:
  rL LLVM

https://reviews.llvm.org/D24338