[Boost regression] Miscompile with private constant array

[DougGregor]

Bugzilla Link	9416
Resolution	FIXED
Resolved on	Jan 15, 2013 01:00
Version	trunk
OS	All
Attachments	Example C++ source, Preprocessed output, LLVM IR generated by Clang - fails, LLVM IR generated with clang, replacing "private" with "internal" linkage for the array in question - passes
CC	@lattner,@echristo

Extended Description

With the given example file (t.cpp), the std::copy instantiation appears to be miscompiled when built with Clang for x86-64. The LLVM IR looks reasonable, in the resulting assembly the std::copy caller pushes the [first, last) arguments into [%rdi, %rsi) but the callee seems to pick them up as [%rsi, %di). At least, the subtraction looks backwards (?). The net effect is that the loop inside std::copy is never executed.

(More attachments and comments coming)

[DougGregor]

The Clang-generated LLVM IR is in t-with-private.s. Note that the array weekday_long_names has private linkage:

@​_ZZ4mainE18weekday_long_names = private constant [7 x i8*] [i8* getelementptr inbounds ([6 x i8]* @.str, i32 0, i32 0), i8* getelementptr inbounds ([6 x i8]* @.str1, i32 0, i32 0), i8* getelementptr inbounds ([6 x i8]* @.str2, i32 0, i32 0), i8* getelementptr inbounds ([6 x i8]* @.str3, i32 0, i32 0), i8* getelementptr inbounds ([6 x i8]* @.str4, i32 0, i32 0), i8* getelementptr inbounds ([6 x i8]* @.str5, i32 0, i32 0), i8* getelementptr inbounds ([6 x i8]* @.str6, i32 0, i32 0)], align 16

if we change this to "internal" linkage, we get the right results (as in t-with-internal.s).

[DougGregor]

This is the source of the remaining regression in Boost.DateTime.

[echristo]

Bisect seems to believe:

[yendi:~] echristo% ~comp/bin/smooshlab bisect --sandbox=bar -b clang-x86_64-darwin10-RA /bin/sh -c '"%(path)s/bin/clang++" t.cpp && ./a.out'
FAIL: clang-r127133-b664
FAIL: clang-r127132-b663
FAIL: clang-r127130-b661
FAIL: clang-r127126-b657
FAIL: clang-r127118-b649
FAIL: clang-r127090-b632
FAIL: clang-r127038-b597
FAIL: clang-r126930-b535
FAIL: clang-r126598-b394
PASS: clang-r126050-b135
PASS: clang-r126313-b264
FAIL: clang-r126471-b330
PASS: clang-r126381-b298
FAIL: clang-r126436-b314
PASS: clang-r126391-b306
PASS: clang-r126421-b310
FAIL: clang-r126426-b312
FAIL: clang-r126422-b311
clang-r126421-b310: first working build
clang-r126422-b311: next failing build

[echristo]

This, of course, doesn't tell us anything we don't already know:

Index: test/CodeGen/decl.c

--- test/CodeGen/decl.c (revision 126421)
+++ test/CodeGen/decl.c (revision 126422)
@@ -1,6 +1,6 @@
// RUN: %clang_cc1 -w -emit-llvm < %s | FileCheck %s

-// CHECK: @​test1.x = internal constant [12 x i32] [i32 1
+// CHECK: @​test1.x = private constant [12 x i32] [i32 1
// CHECK: @​test2.x = internal constant [13 x i32] [i32 1,
// CHECK: @​test5w = global %0 { i32 2, [4 x i8] undef }
// CHECK: @​test5y = global %union.test5u { double 7.300000e+0{{[0]*}}1 }
Index: lib/CodeGen/CGDecl.cpp

--- lib/CodeGen/CGDecl.cpp (revision 126421)
+++ lib/CodeGen/CGDecl.cpp (revision 126422)
@@ -642,7 +642,7 @@ CodeGenFunction::EmitAutoVarAlloca(const
// candidate nor a __block variable, emit it as a global instead.
if (CGM.getCodeGenOpts().MergeAllConstants && Ty.isConstQualified() &&
!NRVO && !isByRef) {

•      EmitStaticVarDecl(D, llvm::GlobalValue::InternalLinkage);
  

•      EmitStaticVarDecl(D, llvm::GlobalValue::PrivateLinkage);
  
       emission.Address = 0; // signal this condition to later callbacks
       assert(emission.wasEmittedAsGlobal());
  

[echristo]

This is caused by the patch from:

http://llvm.org/bugs/show_bug.cgi?id=9301

[lattner]

Rafael, what's going on here? We can't generate random temporaries with private linkage on darwin. Why not just make it be internal?

[lattner]

I'm pretty sure we compile Boost correctly. Is this still broken?

[DougGregor]

This has been fixed "forever".

[lattner]

fwew, I was worried there for a microsecond. Thanks! :)