Hard fault on boot w/ GCC 14 on Cortex-M

[fabiobaltieri]

Discussed in #74379

^{Originally posted by agesome June 17, 2024}
Hi. When building Zephyr with GCC 14.1.0 (arch linux package), it seems that on startup, z_thread_entry is called with entry = NULL, when normally it would = _main, which leads to a fault.

zephyr/lib/os/thread_entry.c

Line 48 in 4d0a8c1

entry(p1, p2, p3);

If we break in arch_switch_to_main_thread, it looks like this, _main (which is passed to z_thread_entry as entry) is OK:

#0  arch_switch_to_main_thread (main_thread=main_thread@entry=0x20020a50 <z_main_thread>, 
    stack_ptr=stack_ptr@entry=0x20033440 "\356\242^i\205\200\346", <incomplete sequence \333>..., 
    _main=_main@entry=0x1001e47d <bg_thread_main>) at /workspace/zephyr-workspace/zephyr/arch/arm/core/cortex_m/thread.c:575
#1  0x1001e63c in switch_to_main_thread (stack_ptr=0x20033440 "\356\242^i\205\200\346", <incomplete sequence \333>...)
    at /workspace/zephyr-workspace/zephyr/kernel/init.c:585
#2  z_cstart () at /workspace/zephyr-workspace/zephyr/kernel/init.c:684
#3  0x10002390 in z_prep_c () at /workspace/zephyr-workspace/zephyr/arch/arm/core/cortex_m/prep_c.c:196
#4  0x100021f4 in z_arm_reset () at /workspace/zephyr-workspace/zephyr/arch/arm/core/cortex_m/reset.S:169

but if we then break in z_thread_entry, it looks as if arch_switch_to_main_thread had _main = 0

#0  z_thread_entry (entry=0x0, p1=0x0, p2=0x0, p3=0x0) at /workspace/zephyr-workspace/zephyr/lib/os/thread_entry.c:48
#1  0x100023d6 in arch_switch_to_main_thread (main_thread=0x0, stack_ptr=0x0, _main=0x0)
    at /workspace/zephyr-workspace/zephyr/arch/arm/core/cortex_m/thread.c:575
#2  0xc5d30d16 in ?? ()

After some experimenting, I fixed this by commenting

zephyr/arch/arm/core/cortex_m/thread.c

Line 593 in 4d0a8c1

CODE_UNREACHABLE;

With GCC 12 included in Zephyr SDK, this problem does not appear.

So at this point I'm not sure if it's some incompatible change 12->14, or a bug in 14. Any advice?

I also made a comparison of assembly on GCC 14 with/without CODE_UNREACHABLE and GCC 12: https://gist.github.com/agesome/bfc1e15b561df353fd5f340c0413e348

[fabiobaltieri]

cc @stephanosio in case you know something about this already (ARM + GCC 14)

[fabiobaltieri]

@ithinuel could you look into it? We bumped into the issue as well and also verified that removing that line is an effective workaround

[ithinuel]

As far as I can tell from reading the gist, It looks like CODE_UNREACHABLE correctly makes the function do nothing (not even return) after the asm volatile block in gcc14, while it still returned in gcc 12 (with pop {r4, r5, r6, pc}) even if there’s noway to comeback to this point.

The difference that concerns me though is that in GCC 14, it does not set _current to main_thread as expected from line 525.
What shows as

   0x10001eb4 <+36>:    asrs    r4, r6, #9
   0x10001eb6 <+38>:    movs    r0, #0

in the first listing is actually the address of _current: 0x20001274.
Could the compiler fail to see that this is still needed and erroneously optimises this out?

[stephanosio]

Note that this is also reproducible with GCC 14 on mps2/an385 and other Cortex-M platforms.

[ithinuel]

It was pointed to me that _main should be passed to the assembly block as r rather than +r and it may be confusing the compiler. Maybe forcing line 525 to be volatile to prevent compiler optimisation.

I'm afraid this might be a bit too cutting edge for the system I'm working from (NixOS). developer.arm.com only has 13.3.

If you know how to get gcc14 for arm-none-eabi available in NixOS please let me know.

[fabiobaltieri]

@ithinuel I have a setup to test, did this change

index fa500032d3c..00d31c05ee3 100644
--- a/arch/arm/core/cortex_m/thread.c
+++ b/arch/arm/core/cortex_m/thread.c
@@ -559,6 +559,8 @@ void arch_switch_to_main_thread(struct k_thread *main_thread, char *stack_ptr,
 #endif
 #endif /* CONFIG_BUILTIN_STACK_GUARD */
 
+       volatile k_thread_entry_t __main = _main;
+
        /*
         * Set PSP to the highest address of the main stack
         * before enabling interrupts and jumping to main.
@@ -586,7 +588,7 @@ void arch_switch_to_main_thread(struct k_thread *main_thread, char *stack_ptr,
        "mov   r3, #0\n"
        "ldr   r4, =z_thread_entry\n"
        "bx    r4\n"            /* We don’t intend to return, so there is no need to link. */
-       : "+r" (_main)
+       : "+r" (__main)
        : "r" (stack_ptr)
        : "r0", "r1", "r2", "r3", "r4", "ip", "lr");

and it worked, though I won't try to pretend I understand why :-) does it make any sense?

[fabiobaltieri]

index fa500032d3c..695e216a43c 100644
--- a/arch/arm/core/cortex_m/thread.c
+++ b/arch/arm/core/cortex_m/thread.c
@@ -588,7 +588,7 @@ void arch_switch_to_main_thread(struct k_thread *main_thread, char *stack_ptr,
        "bx    r4\n"            /* We don’t intend to return, so there is no need to link. */
        : "+r" (_main)
        : "r" (stack_ptr)
-       : "r0", "r1", "r2", "r3", "r4", "ip", "lr");
+       : "r0", "r1", "r2", "r3", "r4", "ip", "lr", "memory");
 
        CODE_UNREACHABLE;
 }

this alwo works (suggested by @ithinuel over discord)

[fabiobaltieri]

index fa500032d3c..6cd7144e79d 100644
--- a/arch/arm/core/cortex_m/thread.c
+++ b/arch/arm/core/cortex_m/thread.c
@@ -586,9 +586,9 @@ void arch_switch_to_main_thread(struct k_thread *main_thread, char *stack_ptr,
        "mov   r3, #0\n"
        "ldr   r4, =z_thread_entry\n"
        "bx    r4\n"            /* We don’t intend to return, so there is no need to link. */
-       : "+r" (_main)
-       : "r" (stack_ptr)
-       : "r0", "r1", "r2", "r3", "r4", "ip", "lr");
+       :
+       : "r" (_main), "r" (stack_ptr)
+       : "r0", "r1", "r2", "r3", "r4", "ip", "lr", "memory");
 
        CODE_UNREACHABLE;
 }

this also works

[ithinuel]

Seems to be the best solution.