[PATCH] D100249: [Inliner] Propagate SROA analysis through invariant group intrinsics

Arthur Eubanks via Phabricator via llvm-commits llvm-commits at lists.llvm.org
Sat Apr 10 12:02:23 PDT 2021


aeubanks created this revision.
aeubanks added reviewers: rnk, mtrofin.
Herald added subscribers: haicheng, hiraditya, eraman, Prazek.
aeubanks requested review of this revision.
Herald added a project: LLVM.
Herald added a subscriber: llvm-commits.

SROA can handle invariant group intrinsics, let the inliner know that
for better heuristics when the intrinsics are present.

This fixes size issues in a couple files when turning on
-fstrict-vtable-pointers in Chrome.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D100249

Files:
  llvm/lib/Analysis/InlineCost.cpp
  llvm/test/Transforms/Inline/invariant-group-sroa.ll


Index: llvm/test/Transforms/Inline/invariant-group-sroa.ll
===================================================================
--- /dev/null
+++ llvm/test/Transforms/Inline/invariant-group-sroa.ll
@@ -0,0 +1,26 @@
+; RUN: opt -passes='print<inline-cost>' -disable-output %s | FileCheck %s
+
+; SROA analysis should yield non-zero savings for allocas passed through invariant group intrinsics
+; CHECK: SROACostSavings: 10
+
+declare i8* @llvm.launder.invariant.group.p0i8(i8*)
+declare i8* @llvm.strip.invariant.group.p0i8(i8*)
+
+declare void @b()
+
+define i32 @f() {
+  %a = alloca i32
+  %r = call i32 @g(i32* %a)
+  ret i32 %r
+}
+
+define i32 @g(i32* %a) {
+  %a_i8 = bitcast i32* %a to i8*
+  %a_inv_i8 = call i8* @llvm.launder.invariant.group.p0i8(i8* %a_i8)
+  %a_inv = bitcast i8* %a_inv_i8 to i32*
+  %i1 = load i32, i32* %a_inv
+  %i2 = load i32, i32* %a_inv
+  %i3 = add i32 %i1, %i2
+  %t = call i8* @llvm.strip.invariant.group.p0i8(i8* %a_inv_i8)
+  ret i32 %i3
+}
Index: llvm/lib/Analysis/InlineCost.cpp
===================================================================
--- llvm/lib/Analysis/InlineCost.cpp
+++ llvm/lib/Analysis/InlineCost.cpp
@@ -1884,6 +1884,11 @@
     case Intrinsic::vastart:
       InitsVargArgs = true;
       return false;
+    case Intrinsic::launder_invariant_group:
+    case Intrinsic::strip_invariant_group:
+      if (auto *SROAArg = getSROAArgForValueOrNull(II->getOperand(0)))
+        SROAArgValues[II] = SROAArg;
+      return true;
     }
   }
 


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D100249.336627.patch
Type: text/x-patch
Size: 1499 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20210410/88461014/attachment.bin>


More information about the llvm-commits mailing list