clang generates endless loop #9986

llvmbot · 2011-04-03T03:15:12Z


Bugzilla Link	9614
Resolution	FIXED
Resolved on	Mar 07, 2013 16:17
Version	trunk
OS	Linux
Blocks	llvm/llvm-bugzilla-archive#11199
Attachments	glibc wchar.h
Reporter	LLVM Bugzilla Contributor
CC	@ahatanak,@andersk,@shining,@echristo,@efriedma-quic,@foutrelis,@ismail

Extended Description

Clang generates llvm intermediate code with an endless loop, when -O is used.
Here is source .c code:

$ cat test.c
#include <wchar.h>
int main ()
{
return btowc ('\0');
}

And after compilation we have:

(gdb) disas main
Dump of assembler code for function main:
0x00000000004004f0 <+0>: jmp 0x4004f0

End of assembler dump.

Here is LLVM IR:

$ clang -S -emit-llvm test.c; cat test.s
; ModuleID = 'test.c'
target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
target triple = "x86_64-pc-linux-gnu"

define i32 @main() nounwind {
%1 = alloca i32, align 4
store i32 0, i32* %1
%2 = call i32 @btowc(i32 0) nounwind
ret i32 %2
}

declare i32 @btowc(i32) nounwind

$ clang -O -S -emit-llvm test.c; cat test.s
; ModuleID = 'test.c'
target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
target triple = "x86_64-pc-linux-gnu"

define i32 @main() nounwind readnone {
br label %tailrecurse.i

tailrecurse.i: ; preds = %tailrecurse.i, %0
br label %tailrecurse.i
}

I've attached wchar.h just in case.

efriedma-quic · 2011-04-03T22:13:17Z

Testcase boils down to something like the following:

typedef __typeof(+L'a') wint_t;
extern wint_t __btowc_alias (int __c) __asm ("btowc");
__attribute((nothrow)) inline wint_t btowc (int __c) {
return __btowc_alias (__c);
}
int f(int x) { return btowc(x); }

clang resolves the call to __btowc_alias into a call to btowc, so clang essentially sees through the trick to try and make the inline version call the out-of-line version. The function then ends up looking like it's tail-recursive, i.e. an infinite loop.

shining · 2011-05-22T12:55:49Z

I've found another test case with realpath.

extern char *__realpath_alias (__const char *__restrict __name, char
*__restrict __resolved) asm ("" "realpath");

extern char *
realpath (__const char *__restrict __name, char *__restrict __resolved)
{
return __realpath_alias (__name, __resolved);
}

char *f(const char *path, char *buf)
{
return realpath(path, buf);
}

efriedma-quic · 2011-06-23T20:53:38Z

*** Bug llvm/llvm-bugzilla-archive#10179 has been marked as a duplicate of this bug. ***

efriedma-quic · 2011-06-28T01:20:31Z

*** Bug llvm/llvm-bugzilla-archive#10207 has been marked as a duplicate of this bug. ***

llvmbot · 2011-07-17T15:11:11Z

I get an infinite loop for btowc() with -O2 but not -O1 as described in bug #10207.

This bug appears to have quite far reaching consequences. Not only does it make the configure scripts of such software as gnu tar, coreutils, diffutils, findutils, gettext, and grep hang. It might be the reason gnu sed, and awk hang during their tests as well (at least they hang when compiled with -O2 but not with -O1).

d0k · 2011-10-06T22:05:35Z

*** Bug llvm/llvm-bugzilla-archive#10981 has been marked as a duplicate of this bug. ***

llvmbot · 2011-10-11T21:54:52Z

*** Bug llvm/llvm-bugzilla-archive#10160 has been marked as a duplicate of this bug. ***

llvmbot · 2011-10-19T06:06:39Z

Another testcase with getcwd, boils down to (after preprocessing, removing impossible code paths, etc)

typedef typeof(sizeof(int)) size_t;
extern char *__getcwd_alias (char *__buf, size_t __size) asm ("" "getcwd");
extern char *getcwd (char *__buf, size_t __size)
{
return __getcwd_alias (__buf, __size);
}
char *test(char *buf, int len)
{
return getcwd((char *)buf, len);
}

This bug clang build a hanging vim on Linux unless optimizations are explicitly disabled.

llvmbot · 2011-10-19T06:08:06Z

Another testcase with getcwd, boils down to (after preprocessing, removing
impossible code paths, etc)

typedef typeof(sizeof(int)) size_t;
extern char *__getcwd_alias (char *__buf, size_t __size) asm ("" "getcwd");
extern char *getcwd (char *__buf, size_t __size)
{
return __getcwd_alias (__buf, __size);
}
char *test(char *buf, int len)
{
return getcwd((char *)buf, len);
}

This bug clang build a hanging vim on Linux unless optimizations are explicitly
disabled.

s/clang/causes clang to/

llvmbot · 2011-10-21T13:01:34Z

Dragonegg turns this into

define i32 @btowc(i32 %__c) nounwind uwtable inlinehint {
entry:
%0 = tail call i32 @"\01btowc"(i32 %__c) nounwind
ret i32 %0
}

declare i32 @"\01btowc"(i32)

define i32 @f(i32 %x) nounwind uwtable {
entry:
%0 = tail call i32 @"\01btowc"(i32 %x) nounwind
ret i32 %0
}

which results in f being codegened as

    jmp     btowc

llvmbot · 2011-10-25T09:04:58Z

Is this still a problem for clang? This is what I'm getting now:

$ cat test.c
#include <wchar.h>
int main () {
return btowc ('\0');
}

$ clang test.c -o - -S -emit-llvm -O
; ModuleID = 'test.c'
target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
target triple = "x86_64-apple-macosx10.7.2"

define i32 @main() nounwind uwtable ssp {
entry:
%call = tail call i32 @btowc(i32 0) nounwind
ret i32 %call
}

declare i32 @btowc(i32)

$ cat > test.c
typedef __typeof(+L'a') wint_t;
extern wint_t __btowc_alias (int __c) __asm ("btowc");
__attribute((nothrow)) inline wint_t btowc (int __c) {
return __btowc_alias (__c);
}
int f(int x) { return btowc(x); }

$ clang test.c -o - -S -emit-llvm -O
; ModuleID = 'test.c'
target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
target triple = "x86_64-apple-macosx10.7.2"

define i32 @f(i32 %x) nounwind uwtable ssp {
entry:
%call.i = tail call i32 @"\01btowc"(i32 %x) nounwind
ret i32 %call.i
}

declare i32 @"\01btowc"(i32)

$ cat > test.c
extern char *__realpath_alias (__const char *__restrict __name, char
*__restrict __resolved) asm ("" "realpath");

extern char *
realpath (__const char *__restrict __name, char *__restrict __resolved) {
return __realpath_alias (__name, __resolved);
}

char *f(const char *path, char *buf) {
return realpath(path, buf);
}

$ clang test.c -o - -S -emit-llvm -O
; ModuleID = 'test.c'
target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
target triple = "x86_64-apple-macosx10.7.2"

define i8* @realpath(i8* noalias %__name, i8* noalias %__resolved) nounwind uwtable ssp {
entry:
%call = tail call i8* @"\01realpath"(i8* %__name, i8* %__resolved) nounwind
ret i8* %call
}

declare i8* @"\01realpath"(i8*, i8*)

define i8* @f(i8* %path, i8* %buf) nounwind uwtable ssp {
entry:
%call.i = tail call i8* @"\01realpath"(i8* %path, i8* %buf) nounwind
ret i8* %call.i
}

$ cat > test.c
typedef typeof(sizeof(int)) size_t;
extern char *__getcwd_alias (char *__buf, size_t __size) asm ("" "getcwd");
extern char *getcwd (char *__buf, size_t __size) {
return __getcwd_alias (__buf, __size);
}
char *test(char *buf, int len) {
return getcwd((char *)buf, len);
}

$ clang test.c -o - -S -O -emit-llvm
; ModuleID = 'test.c'
target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
target triple = "x86_64-apple-macosx10.7.2"

define i8* @getcwd(i8* %__buf, i64 %__size) nounwind uwtable ssp {
entry:
%call = tail call i8* @"\01getcwd"(i8* %__buf, i64 %__size) nounwind
ret i8* %call
}

declare i8* @"\01getcwd"(i8*, i64)

define i8* @test(i8* %buf, i32 %len) nounwind uwtable ssp {
entry:
%conv = sext i32 %len to i64
%call.i = tail call i8* @"\01getcwd"(i8* %buf, i64 %conv) nounwind
ret i8* %call.i
}

ismail · 2011-10-25T09:08:19Z

Is this still a problem for clang? This is what I'm getting now:

Still reproducable on Linux with

[/kutu/stuff]> clang -v
SUSE Linux clang version 3.0 (branches/release_30 142341) (based on LLVM 3.0)
Target: x86_64-unknown-linux-gnu
Thread model: posix

And AFAIK it only affects Linux.

llvmbot · 2011-10-25T09:18:57Z

Is this still a problem for clang? This is what I'm getting now:

Still reproducable on Linux with

[/kutu/stuff]> clang -v
SUSE Linux clang version 3.0 (branches/release_30 142341) (based on LLVM 3.0)
Target: x86_64-unknown-linux-gnu
Thread model: posix

And AFAIK it only affects Linux.

Interesting. When I took your IR example from the fist comment and ran it through 'opt' it generated IR which looked just fine. And the result was a 'jmp btowc' command, like Duncan's.

Can you investigate further? I don't have easy access to a Linux box..

ismail · 2011-10-25T09:24:22Z

I can only reproduce the problem with my original testcase from bug 10160 , http://llvm.org/bugs/attachment.cgi?id=6754

llvmbot · 2011-10-25T09:33:54Z

Interesting. When I took your IR example from the fist comment and ran it
through 'opt' it generated IR which looked just fine. And the result was a 'jmp
btowc' command, like Duncan's.

Can you investigate further? I don't have easy access to a Linux box..

I just remembered that I do have access to a Linux box. All of the tests I ran give the result as "ret i8* undef" (or some variation on that) with `-O'. Almost certainly not what you're looking for.

ismail · 2011-10-25T09:36:16Z

Interesting. When I took your IR example from the fist comment and ran it
through 'opt' it generated IR which looked just fine. And the result was a 'jmp
btowc' command, like Duncan's.

Can you investigate further? I don't have easy access to a Linux box..

I just remembered that I do have access to a Linux box. All of the tests I
ran give the result as "ret i8* undef" (or some variation on that) with `-O'.
Almost certainly not what you're looking for.

Yes, please try with http://llvm.org/bugs/attachment.cgi?id=6754 and you need to do

clang -O2 -D_FORTIFY_SOURCE=2 mmap.c

since the definition of _FORTIFY_SOURCE=2 along with -O2 triggers the problem. bug 10160 had a nicer discussion of this bug but its marked as a duplicate :(

llvmbot · 2011-10-26T20:00:10Z

Another reduced testcase for which clang, when passed -D_FORTIFY_SOURCE=2, turns into an infinite loop, on current 64-bit Debian testing, in both 64-bit and 32-bit mode:

$ cat bug.c
#include <stdint.h>
#include <stdio.h>

int fread_n_bytes(FILE * f, int n, uint8_t *s)
{
int i;

if (s == NULL)
for (i = 0; i < n; i++)
fgetc(f);
else
if(fread(s, 1, n, f) < (size_t)n) // infinite loop here.
return -1;

return 0;
}

int main(int argc, char * argv[]) {
uint8_t buf[8];
return fread_n_bytes((FILE *)0x12345678, 8, buf);
}

clang invocations that produce bad programs:
$ clang -D_FORTIFY_SOURCE=2 -Os -g3 -Wall -W -Wno-unused-parameter -Wshadow -Wwrite-strings -Wredundant-decls -Wdeclaration-after-statement -fstack-protector-all -Wstack-protector -o bad1 bug.c
$ clang -m32 -D_FORTIFY_SOURCE=2 -Os -g3 -Wall -W -Wno-unused-parameter -Wshadow -Wwrite-strings -Wredundant-decls -Wdeclaration-after-statement -fstack-protector-all -Wstack-protector -o bad2 bug.c
$ clang -D_FORTIFY_SOURCE=2 -Os -g3 -Wall -W -Wno-unused-parameter -Wshadow -Wwrite-strings -Wredundant-decls -Wdeclaration-after-statement -o bad3 bug.c

Invocations that produce good programs:
$ clang -Os -g3 -Wall -W -Wno-unused-parameter -Wshadow -Wwrite-strings -Wredundant-decls -Wdeclaration-after-statement -o good1 bug.c
$ clang -Os -g3 -Wall -W -Wno-unused-parameter -Wshadow -Wwrite-strings -Wredundant-decls -Wdeclaration-after-statement -fstack-protector-all -Wstack-protector -o good2 bug.c

$ gdb bad1
GNU gdb (GDB) 7.3-debian
Copyright (C) 2011 Free Software Foundation, Inc.
License GPLv3+: GNU GPL version 3 or later http://gnu.org/licenses/gpl.html
This is free software: you are free to change and redistribute it.
There is NO WARRANTY, to the extent permitted by law. Type "show copying"
and "show warranty" for details.
This GDB was configured as "x86_64-linux-gnu".
For bug reporting instructions, please see:
http://www.gnu.org/software/gdb/bugs/...
Reading symbols from /home/a515273/essai_clang/bad1...done.
(gdb) run
Starting program: /home/a515273/essai_clang/bad1
^C
Program received signal SIGINT, Interrupt.
fread_n_bytes (f=, n=, s=) at bug.c:12
12 if(fread(s, 1, n, f) < (size_t)n)
(gdb) bt
#0 fread_n_bytes (f=, n=, s=) at bug.c:12
#1 0x00000000004005cf in main (argc=, argv=) at bug.c:20
(gdb) disassemble
Dump of assembler code for function fread_n_bytes:
0x0000000000400554 <+0>: push %rbp
0x0000000000400555 <+1>: mov %rsp,%rbp
0x0000000000400558 <+4>: push %r14
0x000000000040055a <+6>: push %rbx
0x000000000040055b <+7>: sub $0x10,%rsp
0x000000000040055f <+11>: mov %esi,%ebx
0x0000000000400561 <+13>: mov %rdi,%r14
0x0000000000400564 <+16>: mov %fs:0x28,%rax
0x000000000040056d <+25>: mov %rax,-0x18(%rbp)
0x0000000000400571 <+29>: test %rdx,%rdx
0x0000000000400574 <+32>: jne 0x40059a <fread_n_bytes+70>
0x0000000000400576 <+34>: test %ebx,%ebx
0x0000000000400578 <+36>: jle 0x400586 <fread_n_bytes+50>
0x000000000040057a <+38>: mov %r14,%rdi
0x000000000040057d <+41>: callq 0x400450 fgetc@plt
0x0000000000400582 <+46>: dec %ebx
0x0000000000400584 <+48>: jne 0x40057a <fread_n_bytes+38>
0x0000000000400586 <+50>: mov %fs:0x28,%rax
0x000000000040058f <+59>: cmp -0x18(%rbp),%rax
0x0000000000400593 <+63>: je 0x40059c <fread_n_bytes+72>
0x0000000000400595 <+65>: callq 0x400440 __stack_chk_fail@plt
=> 0x000000000040059a <+70>: jmp 0x40059a <fread_n_bytes+70>
0x000000000040059c <+72>: xor %eax,%eax
0x000000000040059e <+74>: add $0x10,%rsp
0x00000000004005a2 <+78>: pop %rbx
0x00000000004005a3 <+79>: pop %r14
0x00000000004005a5 <+81>: pop %rbp
0x00000000004005a6 <+82>: retq
End of assembler dump.

echristo · 2011-10-26T20:23:51Z

Looking like the source fortification is necessary? Maybe it's interfering with an optimization that the code expects to happen?

ismail · 2011-10-26T20:34:33Z

Potential fix here: http://comments.gmane.org/gmane.comp.compilers.clang.scm/40901

llvmbot · 2011-10-26T22:56:56Z

Fixed in r143049.

llvmbot · 2011-10-27T07:56:12Z

Well, the testcase I posted in comment #18 still fails (infinite loops) for me after r143049, when -D_FORTIFY_SOURCE=2 is used:

$ clang --version
clang version 3.1 (trunk 143096)
Target: x86_64-unknown-linux-gnu
Thread model: posix
a515273@aofr42198:/essai_clang$ rm bad* good*
a515273@aofr42198:/essai_clang$ sh build.sh
a515273@aofr42198:~/essai_clang$ ls -1
bad1
bad2
bad3
bug.c
build.sh
good1
good2
$ ./good1
Segmentation fault
$ ./good2
Segmentation fault
$ ./bad1
^C
$ ./bad2
^C
$ ./bad3
^C

$ cat build.sh
#!/bin/sh

The base invocation, reduced from my original testcase.

clang -D_FORTIFY_SOURCE=2 -Os -g3 -Wall -W -Wno-unused-parameter -Wshadow -Wwrite-strings -Wredundant-decls -Wdeclaration-after-statement -fstack-protector-all -Wstack-protector -o bad1 bug.c

It ain't better in 32-bit mode.

clang -m32 -D_FORTIFY_SOURCE=2 -Os -g3 -Wall -W -Wno-unused-parameter -Wshadow -Wwrite-strings -Wredundant-decls -Wdeclaration-after-statement -fstack-protector-all -Wstack-protector -o bad2 bug.c

Disabling stack protector switches doesn't help.

clang -D_FORTIFY_SOURCE=2 -Os -g3 -Wall -W -Wno-unused-parameter -Wshadow -Wwrite-strings -Wredundant-decls -Wdeclaration-after-statement -o bad3 bug.c

Disabling FORTIFY_SOURCE helps.

clang -Os -g3 -Wall -W -Wno-unused-parameter -Wshadow -Wwrite-strings -Wredundant-decls -Wdeclaration-after-statement -o good1 bug.c

No problem if stack protector switches are activated but FORTIFY_SOURCE is not set.

clang -Os -g3 -Wall -W -Wno-unused-parameter -Wshadow -Wwrite-strings -Wredundant-decls -Wdeclaration-after-statement -fstack-protector-all -Wstack-protector -o good2 bug.c

$ cat bug.c
#include <stdint.h>
#include <stdio.h>

int fread_n_bytes(FILE * f, int n, uint8_t *s)
{
int i;

if (s == NULL)
for (i = 0; i < n; i++)
fgetc(f);
else
if(fread(s, 1, n, f) < (size_t)n)
return -1;

return 0;
}

int main(int argc, char * argv[]) {
uint8_t buf[8];
return fread_n_bytes((FILE *)0x12345678, 8, buf);
}

ismail · 2011-10-27T11:06:42Z

Fixed in r143049.

Problem is still producable with the testcase from bug 10160 ,
http://llvm.org/bugs/attachment.cgi?id=6754

llvmbot · 2011-10-28T22:20:06Z

Reduced testcase that still fails:

extern int foo_alias (void) asm ("foo");
inline attribute ((always_inline)) int foo (void) {
return foo_alias ();
}
int bar() {
foo();
}

llvmbot · 2011-10-28T22:29:04Z

Dragonegg produces this for the testcase in comment 24:

define i32 @foo() nounwind uwtable inlinehint alwaysinline {
entry:
%0 = tail call i32 @"\01foo"() nounwind
ret i32 %0
}

declare i32 @"\01foo"()

define i32 @bar(...) nounwind uwtable {
entry:
%0 = call i32 @"\01foo"() nounwind
ret i32 undef
}

llvmbot · 2011-10-28T22:45:42Z

fixed in r143222.

llvmbot · 2011-10-28T23:23:07Z

fixed in r143222.

This doesn't apply cleanly to the 3.0 release branch:

http://llvm.org/svn/llvm-project/cfe/branches/release_30

llvmbot · 2011-10-30T12:05:21Z

With clang r143304, the reduced testcase I posted in comments #18 and #22 now segfaults (which is the expected behaviour) even when compiled with -D_FORTIFY_SOURCE=2, and the original application from which I extracted it doesn't contain an infinite loop anymore, either.
Thanks :)

andersk · 2013-03-08T00:17:36Z

This fix doesn’t work in C++ mode, where the fortify wrappers use attribute ((always_inline)) but not attribute ((gnu_inline)):

http://llvm.org/bugs/show_bug.cgi?id=10276#c9

llvmbot transferred this issue from llvm/llvm-bugzilla-archive Dec 3, 2021

This issue was closed.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

clang generates endless loop #9986

clang generates endless loop #9986

llvmbot commented Apr 3, 2011

efriedma-quic commented Apr 3, 2011

shining mannequin commented May 22, 2011

efriedma-quic commented Jun 23, 2011

efriedma-quic commented Jun 28, 2011

llvmbot commented Jul 17, 2011

d0k commented Oct 6, 2011

llvmbot commented Oct 11, 2011

llvmbot commented Oct 19, 2011

llvmbot commented Oct 19, 2011

llvmbot commented Oct 21, 2011

llvmbot commented Oct 25, 2011

ismail commented Oct 25, 2011

llvmbot commented Oct 25, 2011

ismail commented Oct 25, 2011

llvmbot commented Oct 25, 2011

ismail commented Oct 25, 2011

llvmbot commented Oct 26, 2011

echristo commented Oct 26, 2011

ismail commented Oct 26, 2011

llvmbot commented Oct 26, 2011

llvmbot commented Oct 27, 2011

ismail commented Oct 27, 2011

llvmbot commented Oct 28, 2011

llvmbot commented Oct 28, 2011

llvmbot commented Oct 28, 2011

llvmbot commented Oct 28, 2011

llvmbot commented Oct 30, 2011

andersk mannequin commented Mar 8, 2013

clang generates endless loop #9986

clang generates endless loop #9986

Comments

llvmbot commented Apr 3, 2011

Extended Description

efriedma-quic commented Apr 3, 2011

shining mannequin commented May 22, 2011

efriedma-quic commented Jun 23, 2011

efriedma-quic commented Jun 28, 2011

llvmbot commented Jul 17, 2011

d0k commented Oct 6, 2011

llvmbot commented Oct 11, 2011

llvmbot commented Oct 19, 2011

llvmbot commented Oct 19, 2011

llvmbot commented Oct 21, 2011

llvmbot commented Oct 25, 2011

ismail commented Oct 25, 2011

llvmbot commented Oct 25, 2011

ismail commented Oct 25, 2011

llvmbot commented Oct 25, 2011

ismail commented Oct 25, 2011

llvmbot commented Oct 26, 2011

echristo commented Oct 26, 2011

ismail commented Oct 26, 2011

llvmbot commented Oct 26, 2011

llvmbot commented Oct 27, 2011

The base invocation, reduced from my original testcase.

It ain't better in 32-bit mode.

Disabling stack protector switches doesn't help.

Disabling FORTIFY_SOURCE helps.

No problem if stack protector switches are activated but FORTIFY_SOURCE is not set.

ismail commented Oct 27, 2011

llvmbot commented Oct 28, 2011

llvmbot commented Oct 28, 2011

llvmbot commented Oct 28, 2011

llvmbot commented Oct 28, 2011

llvmbot commented Oct 30, 2011

andersk mannequin commented Mar 8, 2013