https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83654
--- Comment #2 from Florian Weimer <fw at gcc dot gnu.org> --- I forgot to add a compiler barrier to f2 for the executable test case, so it is not strictly equivalent. With it, valgrind reports: ==375147== Invalid read of size 4 ==375147== at 0x8048403: f2 (in /root/build-32/a.out) ==375147== by 0x80482CA: main (in /root/build-32/a.out) ==375147== Address 0xfebb3d0c is on thread 1's stack ==375147== 4 bytes below stack pointer valgrind warns about the final probe. Here's the disassembly with addresses: 080483e0 <f2>: 80483e0: 55 push %ebp 80483e1: 89 e5 mov %esp,%ebp 80483e3: 83 ec 08 sub $0x8,%esp 80483e6: 8d 84 24 00 f0 ff ff lea -0x1000(%esp),%eax 80483ed: 39 c4 cmp %eax,%esp 80483ef: 74 12 je 8048403 <f2+0x23> 80483f1: 81 ec 00 10 00 00 sub $0x1000,%esp 80483f7: 83 8c 24 fc 0f 00 00 orl $0x0,0xffc(%esp) 80483fe: 00 80483ff: 39 c4 cmp %eax,%esp 8048401: 75 ee jne 80483f1 <f2+0x11> 8048403: 83 4c 24 fc 00 orl $0x0,-0x4(%esp) 8048408: 83 ec 0c sub $0xc,%esp 804840b: 50 push %eax 804840c: e8 bf ff ff ff call 80483d0 <f1> 8048411: c9 leave 8048412: c3 ret
