[Bug middle-end/83654] New: -fstack-clash-protection probes below the stack pointer for VLA with constant size

[fw at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83654

            Bug ID: 83654
           Summary: -fstack-clash-protection probes below the stack
                    pointer for VLA with constant size
           Product: gcc
           Version: 8.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: fw at gcc dot gnu.org
  Target Milestone: ---

Created attachment 43010
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=43010&action=edit
const-vla.c

On i386, the reproducer compiles to:

f2:
.LFB0:
        .cfi_startproc
        pushl   %ebp
        .cfi_def_cfa_offset 8
        .cfi_offset 5, -8
        movl    %esp, %ebp
        .cfi_def_cfa_register 5
        subl    $8, %esp
        leal    -4096(%esp), %eax
        cmpl    %eax, %esp
        je      .L5
.L7:
        subl    $4096, %esp
        orl     $0, 4092(%esp)
        cmpl    %eax, %esp
        jne     .L7
.L5:
        orl     $0, -4(%esp)

The final probe accesses memory below the stack pointer.  i386 does not have a
red zone, so this results in a valgrind warning:

==375071== Invalid read of size 4
==375071==    at 0x80482EE: main (in /root/build-32/a.out)
==375071==  Address 0xfeb83d0c is on thread 1's stack
==375071==  4 bytes below stack pointer

I don't think this issue is observable without valgrind, so the severity is
low.  However, it affects glibc's getaddrinfo implementation, so it seems
worthwhile addressing this.

The issue might not be target-specific.  x86-64 results in similar code, but
valgrind won't warn because of the red zone.