[LLVMdev] About thread_local in 3.0

[Jianzhou Zhao]

Hi LLVM,

I am using 3.0, and I have a question about the __thread in c and
thread_local in LLVM IR: O1-O4 and the final linked code behave
differently.

////////////////////////////////////
The following C code is from the LLVM testcase
(SingleSource/UnitTests/Threads/2010-12-08-tls.c)

#include <stdio.h>

__thread int a = 4;

int foo (void)
{
  return a;
}

int main (void) {
  printf("a is %d\n", foo());
  return 0;
}

It contains a __thread attribute. Both GCC's output program and
clang's produce "a is 4", and then return. The command line options
are the default.
gcc 2010-12-08-tls.c
clang 2010-12-08-tls.c

////////////////////////////////////
The following code (2010-12-08-tls.O1.ll) is generated by
clang -c -O1 -emit-llvm 2010-12-08-tls.c -o 2010-12-08-tls.O1.bc
llvm-dis 2010-12-08-tls.O1.bc

@a = thread_local global i32 4, align 4
@.str = private unnamed_addr constant [9 x i8] c"a is %d\0A\00", align 1

define i32 @foo() nounwind readonly {
  %1 = load i32* @a, align 4, !tbaa !0
  ret i32 %1
}

define i32 @main() nounwind {
  %1 = tail call i32 @foo()
  %2 = tail call i32 (i8*, ...)* @printf(i8* getelementptr inbounds
([9 x i8]* @.str, i32 0, i32 0), i32 %1) nounwind
  ret i32 0
}

declare i32 @printf(i8* nocapture, ...) nounwind

!0 = metadata !{metadata !"int", metadata !1}
!1 = metadata !{metadata !"omnipotent char", metadata !2}
!2 = metadata !{metadata !"Simple C/C++ TBAA", null}

If I jit the code, say "lli 2010-12-08-tls.O1.bc", I got

0  lli       0x091ac1a8
1  lli       0x091ac7e7
2            0xffffe400  + 0
3  lli       0x08f7a103
llvm::ExecutionEngine::runFunctionAsMain(llvm::Function*,
std::vector<std::string, std::allocator<std::string> > const&, char
const* const*) + 1459
4  lli       0x0861df9e main + 3374
5  libc.so.6 0xb75b7ace __libc_start_main + 254
Stack dump:
0.      Program arguments: lli 2010-12-08-tls.O4.bc
Segmentation fault

If I lli the code, say "lli --force-interpreter=true
2010-12-08-tls.O1.bc", I got
  "a is 24"

////////////////////////////////////
The following code (2010-12-08-tls.O4.ll) is generated by
clang -c -O4 -emit-llvm 2010-12-08-tls.c -o 2010-12-08-tls.O4.bc
llvm-dis 2010-12-08-tls.O4.bc

@a = thread_local global i32 4, align 4
@.str = private unnamed_addr constant [9 x i8] c"a is %d\0A\00", align 1

define i32 @foo() nounwind readonly {
  %1 = load i32* @a, align 4, !tbaa !0
  ret i32 %1
}

define i32 @main() nounwind {
  %1 = load i32* @a, align 4, !tbaa !0
  %2 = tail call i32 (i8*, ...)* @printf(i8* getelementptr inbounds
([9 x i8]* @.str, i32 0, i32 0), i32 %1) nounwind
  ret i32 0
}

declare i32 @printf(i8* nocapture, ...) nounwind

!0 = metadata !{metadata !"int", metadata !1}
!1 = metadata !{metadata !"omnipotent char", metadata !2}
!2 = metadata !{metadata !"Simple C/C++ TBAA", null}

We can see, O4 does inlining, but otherwise its output is the same to
O1's. And I got the same results from jit and lli.

////////////////////////////////////
The following code (2010-12-08-tls.ld.ll) is generated by
opt -std-link-opt 2010-12-08-tls.O1.bc -o 2010-12-08-tls.ld.bc
llvm-dis 2010-12-08-tls.ld.bc

@.str = private unnamed_addr constant [9 x i8] c"a is %d\0A\00", align 1

define i32 @main() nounwind {
  %1 = tail call i32 (i8*, ...)* @printf(i8* getelementptr inbounds
([9 x i8]* @.str, i32 0, i32 0), i32 4) nounwind
  ret i32 0
}

declare i32 @printf(i8* nocapture, ...) nounwind

We can see link-opt value-numbered %1 with 4. Now, jit and lli produce
the results "a is 24", which is same to clang and gcc's native
output's. I am not familiar with thread_local and the __thread in C,
was wondering if this is the expected behavior of thread_local.
Thanks.

-- 
J