https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96717
Bug ID: 96717
Summary: -flifetime-dse=2 breaks webkit-gtk-2.28.4
Product: gcc
Version: 11.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: slyfox at gcc dot gnu.org
Target Milestone: ---
Created attachment 49084
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=49084&action=edit
part-of-webkit-gtk-2.28.4.tar.gz
Initially I noticed the bug on liferea which uses webkit-gtk-2.28.4 as a
rendering engine.
I was not able to extract minimal example because I got lost in SFINATE
instance selections. I am not sure if it's a webkit bug or gcc bug.
I'm leaning to gcc bug. gcc-10 used to work. gcc-11 generates code that hits
'ud2'.
The webkit client code looks simple: we create a hash set of raw pointers and
add pointers there to see how hashset resizes:
"""
#define USE_SYSTEM_MALLOC 1 /* avoid bmalloc */
#include "wtf/HashSet.h"
#include "wtf/Threading.h"
using namespace WTF;
int main() {
HashSet<Thread*> hst;
for (int i = 1; i < 1000; ++i) {
Thread * v = (Thread*)(long)(i * 128);
hst.add(v);
}
}
"""
But headers take almost 2 megabytes.
Building the example with -fno-lifetime-dse produces working code. Building
without breaks on first hash table resize:
$ ./mk.bash
./a
./mk.bash: line 50: 1054935 Illegal instruction (core dumped) ./a
132
./a-nodse
0
The 'ud2' is encountered in HashTable::rehash() method:
https://github.com/WebKit/webkit/blob/master/Source/WTF/wtf/HashTable.h#L1304
Attaching self-contained directory with sources and headers.
Can you help me rule out gcc's misbehaviour?
