On Monday, May 20, 2019 6:44:19 PM CEST, Richard Earnshaw (lists) wrote:
On 20/05/2019 15:45, Vicente Bergas wrote:
Hello,
i would like to have a debug build with frame pointers on an ARM
Cortex-M4 MCU.
These are the options passed to GCC 9.1.0:
-fno-omit-frame-pointer
-mthumb
-march=armv7e-m
-mtune=cortex-m4
-mcpu=cortex-m4
-mfloat-abi=hard
-ffreestanding
-std=c11
I also tested several combinations of the following options:
-mabi=aapcs
This is the default these days.
-mabi=atpcs
-mapcs-frame
-mtpcs-frame
-mtpcs-leaf-frame
All of the above are deprecated relics of the old APCS, TPCS and ATPCS ABIs.
-mno-sched-prolog
-O1
-O0
These have no direct effect on use of the frame pointer.
But never got the fp register (r11) set up.
This is one sample function from objdump (fp appears nowhere):
80005c0: b5b0 push {r4, r5, r7, lr}
80005c2: af00 add r7, sp, #0
80005c4: 4605 mov r5, r0
80005c6: 2030 movs r0, #48
80005c8: f7ff ffa6 bl 8000518
80005cc: 2078 movs r0, #120
80005ce: f7ff ffa3 bl 8000518
80005d2: 241c movs r4, #28
80005d4: fa25 f304 lsr.w r3, r5, r4
80005d8: f003 030f and.w r3, r3, #15
80005dc: 2b09 cmp r3, #9
80005de: bfcc ite gt
80005e0: 2037 movgt r0, #55
80005e2: 2030 movle r0, #48
80005e4: 4418 add r0, r3
80005e6: f7ff ff97 bl 8000518
80005ea: 3c04 subs r4, #4
80005ec: f114 0f04 cmn.w r4, #4
80005f0: d1f0 bne.n 80005d4
80005f2: bdb0 pop {r4, r5, r7, pc}
So, all in all, how can it be done?
In answer to your question, it depends on why you want a frame pointer.
If you mean to address local variables via a register other than SP,
then -fno-omit-frame-pointer is sufficient, but GCC will currently use
r7 in this case from thumb code.
If you want to create a frame backtrace structure by which you can walk
the stack without dwarf based debug information being included, then GCC
does not currently have any means to do this from thumb2 code.
Sorry. We're thinking about what can be done here, but this is an ABI
issue and must be resolved there first.
R.
Thanks Richard for such a quick answer.
I wanted to print a backtrace when the HardFault interrupt is triggered.
From what you say, that is not possible and it is not GCCs fault, is a lack
on the ABI. That also explains why Clang neither keeps track of the frame
pointer.
Why is the register r11 reserved as "fp" and the ABI has no use for it?
Anyways, i'll put a stack on the application and push to it at every
function entry point and pop from it at exit. Sigh.
Regards,
Vicenç.
