[PATCH] D43859: [GlobalOpt] don't change CC of musttail calle(e|r) PR36546

[Fedor Indutny via Phabricator via llvm-commits]

indutny updated this revision to Diff 136358.
indutny marked an inline comment as done.
indutny added a comment.

Address inline comments.

---

As far as I can see it, the documentation doesn't say anything about `musttail` in particular. It says that `tail` calls require `fastcc` to be optimized, but even this seems to be rather stale.

Changing CC for the whole chain is totally a valid alternative. Would it make sense to you if I'll add a `FIXME` comment there?


https://reviews.llvm.org/D43859

Files:
  lib/Transforms/IPO/GlobalOpt.cpp
  test/Transforms/GlobalOpt/musttail_cc.ll


Index: test/Transforms/GlobalOpt/musttail_cc.ll
===================================================================
--- /dev/null
+++ test/Transforms/GlobalOpt/musttail_cc.ll
@@ -0,0 +1,34 @@
+; RUN: opt < %s -globalopt -S | FileCheck %s
+; PR36546
+
+; Check that musttail callee preserves its calling convention
+
+define i32 @test(i32 %a) {
+  ; CHECK: %ca = musttail call i32 @foo(i32 %a)
+  %ca = musttail call i32 @foo(i32 %a)
+  ret i32 %ca
+}
+
+; CHECK-LABEL: define internal i32 @foo(i32 %a)
+define internal i32 @foo(i32 %a) {
+  ret i32 %a
+}
+
+; Check that musttail caller preserves its calling convention
+
+define i32 @test2(i32 %a) {
+  %ca = call i32 @foo1(i32 %a)
+  ret i32 %ca
+}
+
+; CHECK-LABEL: define internal i32 @foo1(i32 %a)
+define internal i32 @foo1(i32 %a) {
+  ; CHECK: %ca = musttail call i32 @foo2(i32 %a)
+  %ca = musttail call i32 @foo2(i32 %a)
+  ret i32 %ca
+}
+
+; CHECK-LABEL: define internal i32 @foo2(i32 %a)
+define i32 @foo2(i32 %a) {
+  ret i32 %a
+}
Index: lib/Transforms/IPO/GlobalOpt.cpp
===================================================================
--- lib/Transforms/IPO/GlobalOpt.cpp
+++ lib/Transforms/IPO/GlobalOpt.cpp
@@ -2115,8 +2115,29 @@
 /// GHC, or anyregcc.
 static bool hasChangeableCC(Function *F) {
   CallingConv::ID CC = F->getCallingConv();
+
   // FIXME: Is it worth transforming x86_stdcallcc and x86_fastcallcc?
-  return CC == CallingConv::C || CC == CallingConv::X86_ThisCall;
+  if (CC != CallingConv::C && CC != CallingConv::X86_ThisCall)
+    return false;
+
+  // Can't change CC of the function that either has musttail calls, or is a
+  // musttail callee itself
+  for (User *U : F->users()) {
+    if (isa<BlockAddress>(U))
+      continue;
+    CallInst* CI = dyn_cast<CallInst>(U);
+    if (!CI)
+      continue;
+
+    if (CI->isMustTailCall())
+      return false;
+  }
+
+  for (BasicBlock &BB : *F)
+    if (BB.getTerminatingMustTailCall())
+      return false;
+
+  return true;
 }
 
 /// Return true if the block containing the call site has a BlockFrequency of


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D43859.136358.patch
Type: text/x-patch
Size: 2055 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20180228/430eab59/attachment.bin>