https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128684

            Bug ID: 128684
           Summary: powerpc: gcc 16.2.0 fails to generate exit condition
                    from loop
           Product: gcc
           Version: 16.2.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: mark.cave-ayland at ilande dot co.uk
  Target Milestone: ---

Created attachment 66145
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=66145&action=edit
zip file containing the example files above

Whilst testing with Fedora 44 (gcc 16.2.0), a bug has been discovered whereby
the powerpc binaries for OpenBIOS are being generated incorrectly such that the
generated code enters an infinite loop during startup.

The code itself looks like this:

static void patch_rfi_bad(void)
{
    uint32_t* ptr;
    uint32_t* vec_start = (uint32_t*) 0x100ULL;
    uint32_t* vec_end = (uint32_t*) (__vectors_end - __vectors);

    for (ptr = vec_start; ptr != vec_end; ptr++)  {
        if (*ptr == RFI)
            *ptr = RFID;
    }
}

What happens at startup is that OpenBIOS takes the vector table at 0xffff0000,
relocates it down to 0x0 and then if it detects a 64-bit CPU, patches all
instances of rfi to rfid using the function above.


Problem
=======

With gcc 16.2.0 the generated code for the for() loop above contains no exit
condition, causing an infinite loop. In my local test case here the generated
assembly for patch_rfi_bad() looks like this:

fff0104c:       3d 40 ff f0     lis     r10,-16
fff01050:       39 4a 03 00     addi    r10,r10,768
fff01054:       3d 20 ff f0     lis     r9,-16
fff01058:       39 29 00 00     addi    r9,r9,0
fff0105c:       7d 29 50 50     subf    r9,r9,r10
fff01060:       2c 09 01 00     cmpwi   r9,256
fff01064:       4d 82 00 20     beqlr
fff01068:       39 20 01 00     li      r9,256
fff0106c:       3d 00 4c 00     lis     r8,19456
fff01070:       61 08 00 64     ori     r8,r8,100
fff01074:       3c e0 4c 00     lis     r7,19456
fff01078:       60 e7 00 24     ori     r7,r7,36
fff0107c:       48 00 00 0c     b       fff01088 <setup_mmu+0x88>
fff01080:       90 e9 00 00     stw     r7,0(r9)
fff01084:       39 29 00 04     addi    r9,r9,4
fff01088:       81 49 00 00     lwz     r10,0(r9)
fff0108c:       7c 0a 40 00     cmpw    r10,r8
fff01090:       40 82 ff f4     bne     fff01084 <setup_mmu+0x84>
fff01094:       4b ff ff ec     b       fff01080 <setup_mmu+0x80>

Here we see that the loop gets stuck between fff01084 and fff01094. Compare
this with a working version patch_rfi_good() which looks like this:

static void patch_rfi_good(void)
{
    uint32_t* ptr;
    uint32_t* vec_start = (uint32_t*) 0x100ULL;
    uint32_t* vec_end = (uint32_t*) 0x300ULL;

    for (ptr = vec_start; ptr != vec_end; ptr++)  {
        if (*ptr == RFI)
            *ptr = RFID;
    }
}

and its corresponding assembly:

fff01000:       39 20 01 00     li      r9,256
fff01004:       3d 00 4c 00     lis     r8,19456
fff01008:       61 08 00 64     ori     r8,r8,100
fff0100c:       3c e0 4c 00     lis     r7,19456
fff01010:       60 e7 00 24     ori     r7,r7,36
fff01014:       39 40 00 80     li      r10,128
fff01018:       7d 49 03 a6     mtctr   r10
fff0101c:       48 00 00 10     b       fff0102c <setup_mmu+0x2c>
fff01020:       90 e9 00 00     stw     r7,0(r9)
fff01024:       39 29 00 04     addi    r9,r9,4
fff01028:       42 40 00 14     bdz     fff0103c <setup_mmu+0x3c>
fff0102c:       81 49 00 00     lwz     r10,0(r9)
fff01030:       7c 0a 40 00     cmpw    r10,r8
fff01034:       40 82 ff f0     bne     fff01024 <setup_mmu+0x24>
fff01038:       4b ff ff e8     b       fff01020 <setup_mmu+0x20>

The difference here is that ctr is loaded at fff01018 and the bdz instruction
at  fff01028 decrements ctr and takes the branch to exit the loop when ctr
reaches zero.


Analysis
========

The issue seems to be related to the use of arithmetic with the __vectors_end
and __vectors globals to calculate vec_end. patch_rfi_good() hardcodes vec_end
with its numeric constant value which appears to solve the issue.

As a reference point my local working OpenBIOS builds which still use
patch_rfi_bad() are generated with gcc 14.2.0 (Debian).


Reproducer
==========

The following files are included in a zip file attached to this bug report:

  - test.c: C program containing patch_rfi_bad() and patch_rfi_good()
  - testasm.s: PPC asm code containing the vector table and its globals
  - ldscript: a linker script that allows for testing in QEMU
  - build.sh: a simple script to build the reproducer

Reply via email to