M0 : bad C11 atomics

I can still reproduce with GCC 9.2.1 (9-2019-q4-major):

main:
 @ args = 0, pretend = 0, frame = 0
 @ frame_needed = 1, uses_anonymous_args = 0
 push {r7, lr} @
 add r7, sp, #0 @,,
.L3:
@ at.c:7: while (atomic_flag_test_and_set( &aflag));
 nop
.L2:
@ at.c:7: while (atomic_flag_test_and_set( &aflag));
 ldr r3, .L4 @ tmp112,
 ldrb r2, [r3] @ tmp113,
 movs r1, #1 @ tmp114,
 strb r1, [r3] @ tmp115,
 uxtb r3, r2 @ _1, tmp113
@ at.c:7: while (atomic_flag_test_and_set( &aflag));
 cmp r3, #0 @ _1,
 bne .L2 @,
@ at.c:8: atomic_flag_clear( &aflag);
 ldr r3, .L4 @ tmp116,
 dmb ish
 movs r2, #0 @ tmp117,
 strb r2, [r3] @ tmp118,
 dmb ish
@ at.c:7: while (atomic_flag_test_and_set( &aflag));
 b .L3 @
.L5:
 .align 2
.L4:
 .word aflag
 .size main, .-main
 .ident "GCC: (GNU Tools for Arm Embedded Processors 9-2019-q4-major) 9.2.1 20191025 (release) [ARM/arm-9-branch revision 277599]"

Just checking in for the yearly test. Still present in 10.2.0 ;

main:
 @ args = 0, pretend = 0, frame = 0
 @ frame_needed = 1, uses_anonymous_args = 0
 push {r7, lr} @
 add r7, sp, #0 @,,
.L3:
@ at.c:7: while (atomic_flag_test_and_set( &aflag));
 nop
.L2:
@ at.c:7: while (atomic_flag_test_and_set( &aflag));
 ldr r3, .L4 @ tmp115,
 ldrb r2, [r3] @ tmp116,
 movs r1, #1 @ tmp117,
 strb r1, [r3] @ tmp118,
 uxtb r3, r2 @ _1, tmp116
@ at.c:7: while (atomic_flag_test_and_set( &aflag));
 cmp r3, #0 @ _1,
 bne .L2 @,
@ at.c:8: atomic_flag_clear( &aflag);
 ldr r3, .L4 @ tmp119,
 dmb ish
 movs r2, #0 @ tmp120,
 strb r2, [r3] @ tmp121,
 dmb ish
@ at.c:7: while (atomic_flag_test_and_set( &aflag));
 b .L3 @
.L5:
 .align 2
.L4:
 .word aflag
 .size main, .-main
 .ident "GCC: (Arch Repository) 10.2.0"

The C11/C++11 atomic support on gcc for embedded ARM is almost completely broken - it is worse than useless because it silently gives incorrect code. For simple read or write access to 32-bit or smaller atomics, the code is correct but inefficient (you don't need DSB and similar barriers on single-core Cortex-M devices). For anything bigger or more complex, you get non-atomic sequences or calls to library functions that have spin locks that cannot work on a single core device.