[llvm-commits] [llvm] r151877 - in /llvm/trunk: lib/Target/X86/X86ISelLowering.cpp test/CodeGen/X86/sibcall-5.ll

[Chad Rosier]

Author: mcrosier
Date: Thu Mar  1 20:50:46 2012
New Revision: 151877

URL: http://llvm.org/viewvc/llvm-project?rev=151877&view=rev
Log:
Prevent obscure and incorrect tail-call optimization.

In this instance we are generating the tail-call during legalizeDAG.  The 2nd
floor call can't be a tail call because it clobbers %xmm1, which is defined by
the first floor call.  The first floor call can't be a tail-call because it's
not in the tail position.  The only reasonable way I could think to fix this
in a target-independent manner was to check for glue logic on the copy reg.

rdar://10930395

Modified:
    llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
    llvm/trunk/test/CodeGen/X86/sibcall-5.ll

Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=151877&r1=151876&r2=151877&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Thu Mar  1 20:50:46 2012
@@ -1589,6 +1589,11 @@
       Copy->getOpcode() != ISD::FP_EXTEND)
     return false;
 
+  // If anything is glued to the copy, then we can't safely perform a tail call.
+  if (Copy->getOpcode() == ISD::CopyToReg &&
+      Copy->getNumOperands() == 4)
+    return false;
+
   bool HasRet = false;
   for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
        UI != UE; ++UI) {

Modified: llvm/trunk/test/CodeGen/X86/sibcall-5.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/sibcall-5.ll?rev=151877&r1=151876&r2=151877&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/X86/sibcall-5.ll (original)
+++ llvm/trunk/test/CodeGen/X86/sibcall-5.ll Thu Mar  1 20:50:46 2012
@@ -1,5 +1,6 @@
 ; RUN: llc < %s -mtriple=i386-apple-darwin8 -mattr=+sse2  | FileCheck %s --check-prefix=X32
 ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+sse2 | FileCheck %s --check-prefix=X64
+; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=-sse3 | FileCheck %s --check-prefix=X64_BAD
 
 ; Sibcall optimization of expanded libcalls.
 ; rdar://8707777
@@ -29,3 +30,31 @@
 declare float @sinf(float) nounwind readonly
 
 declare double @sin(double) nounwind readonly
+
+; rdar://10930395
+%0 = type opaque
+
+@"\01L_OBJC_SELECTOR_REFERENCES_2" = external hidden global i8*, section "__DATA, __objc_selrefs, literal_pointers, no_dead_strip"
+
+define hidden { double, double } @foo2(%0* %self, i8* nocapture %_cmd) uwtable optsize ssp {
+; X64_BAD: foo
+; X64_BAD: call
+; X64_BAD: call
+; X64_BAD: call
+  %1 = load i8** @"\01L_OBJC_SELECTOR_REFERENCES_2", align 8, !invariant.load !0
+  %2 = bitcast %0* %self to i8*
+  %3 = tail call { double, double } bitcast (i8* (i8*, i8*, ...)* @objc_msgSend to { double, double } (i8*, i8*)*)(i8* %2, i8* %1) optsize
+  %4 = extractvalue { double, double } %3, 0
+  %5 = extractvalue { double, double } %3, 1
+  %6 = tail call double @floor(double %4) optsize
+  %7 = tail call double @floor(double %5) optsize
+  %insert.i.i = insertvalue { double, double } undef, double %6, 0
+  %insert5.i.i = insertvalue { double, double } %insert.i.i, double %7, 1
+  ret { double, double } %insert5.i.i
+}
+
+declare i8* @objc_msgSend(i8*, i8*, ...)
+
+declare double @floor(double) optsize
+
+!0 = metadata !{}