[Bug c/66547] New: arm-none-eabi-gcc - stack misaligned when calling va_arg function

[Bug c/66547] New: arm-none-eabi-gcc - stack misaligned when calling va_arg function

[matt at hpamotorsport dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66547

            Bug ID: 66547
           Summary: arm-none-eabi-gcc - stack misaligned when calling
                    va_arg function
           Product: gcc
           Version: 4.9.3
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: matt at hpamotorsport dot com
  Target Milestone: ---

Created attachment 35785
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=35785&action=edit
File from project

I'm working on a project that uses ChibiOS
(https://github.com/ChibiOS/ChibiOS/tree/stable_2.6.x) where I have a function
that calls a va_arg function but the stack is not aligned after the calling
function allocates it's local resources. This causes 64-bit variables (long
long as well as double) to be incorrectly placed.

The assembly generated shows that the stack pointer is moved by 116 where it
should be moved by 120 to keep the alignment to 8-bytes.




arm-none-eabi-gcc -c -mcpu=cortex-m3 -Os -ggdb -fsingle-precision-constant
-mno-unaligned-access -mabi=aapcs -mtune=cortex-m3 -save-temps
-ffunction-sections -fdata-sections -fno-common  -Wall -Wextra
-Wstrict-prototypes -Wa,-alms=build/lst/performance.lst  -DCORTEX_USE_FPU=FALSE
-DBOARD_HALDEX_STM32_P105_A1=1 -DOSC_HAS_64BIT -DOSC_HAS_FLOAT
-DOSC_HAS_HEARTBEAT -DSLP_HAS_USB -D__BUILD_NUMBER=`cat build-number.txt`
-DTHUMB_PRESENT -mno-thumb-interwork -DTHUMB_NO_INTERWORKING -MD -MP -MF
.dep/performance.o.d -mthumb -DTHUMB -I.
-I../ChibiOS_2.6.7/os/ports/common/ARMCMx/CMSIS/include
-I../ChibiOS_2.6.7/os/ports/common/ARMCMx
-I../ChibiOS_2.6.7/os/ports/GCC/ARMCMx
-I../ChibiOS_2.6.7/os/ports/GCC/ARMCMx/STM32F1xx
-I../ChibiOS_2.6.7/os/kernel/include -I../ChibiOS_2.6.7/os/hal/include
-I../ChibiOS_2.6.7/os/hal/platforms/STM32F1xx
-I../ChibiOS_2.6.7/os/hal/platforms/STM32
-I../ChibiOS_2.6.7/os/hal/platforms/STM32/GPIOv1
-I../ChibiOS_2.6.7/os/hal/platforms/STM32/I2Cv1
-I../ChibiOS_2.6.7/os/hal/platforms/STM32/RTCv1
-I../ChibiOS_2.6.7/os/hal/platforms/STM32/SPIv1
-I../ChibiOS_2.6.7/os/hal/platforms/STM32/TIMv1
-I../ChibiOS_2.6.7/os/hal/platforms/STM32/USARTv1
-I../ChibiOS_2.6.7/os/hal/platforms/STM32/OTGv1 -I../HPA_Libraries/board
-I../ChibiOS_2.6.7/os/various -I../HPA_Libraries performance.c -o
build/obj/performance.o

[Bug target/66547] arm-none-eabi-gcc - stack misaligned when calling va_arg function

[matt at hpamotorsport dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66547

--- Comment #1 from Matthew Peters <matt at hpamotorsport dot com> ---
Adding some notes.

The stack is a local stack generated with "static WORK_AREA(...)" from ChibiOS.
I've checked and the stack is aligned at the beginning of
performance_suite_thread.

I've been unable to make a test that produces the error.

[Bug target/66547] arm-none-eabi-gcc - stack misaligned when calling va_arg function

[rearnsha at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66547

Richard Earnshaw <rearnsha at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|---                         |INVALID

--- Comment #2 from Richard Earnshaw <rearnsha at gcc dot gnu.org> ---
I presume you are talking about the function performance_suite_thread (the only
mention of a stack adjustment by 116 bytes).

By my calculations this function does not misalign the stack.

The sub sp, sp, #116 has to be read in context with the initial push
instruction.  116 is 29 words, the push instruction pushes 9 registers (36
bytes), making for a total of 38 words (152 bytes) of data: that's an even
multiple of 8 bytes, so if SP was aligned on entry to that function it will
still be aligned once the prologue code has completed.

Note that the code reads:

performance_suite_thread:
        // ABI Assertion: SP % 8 = 0

.LFB66:
        .loc 1 86 0
        .cfi_startproc
        @ args = 0, pretend = 0, frame = 72
        @ frame_needed = 0, uses_anonymous_args = 0
.LVL4:
        push    {r4, r5, r6, r7, r8, r9, r10, fp, lr}  // 9 words
        .cfi_def_cfa_offset 36
        .cfi_offset 4, -36
        .cfi_offset 5, -32
        .cfi_offset 6, -28
        .cfi_offset 7, -24
        .cfi_offset 8, -20
        .cfi_offset 9, -16
        .cfi_offset 10, -12
        .cfi_offset 11, -8
        .cfi_offset 14, -4
        .loc 1 115 0
        ldr     r3, .L37
        .loc 1 86 0
        sub     sp, sp, #116                         // + 29 words = 38 total
        .cfi_def_cfa_offset 152
        // ABI Assertion: SP % 8 = 0 if calls are made from this function.

Note that the .cfi_def_cfa_offset of 152 would appear to confirm this analysis:
the total change in the frame offsets is 152 bytes.

[Bug target/66547] arm-none-eabi-gcc - stack misaligned when calling va_arg function

[matt at hpamotorsport dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66547

--- Comment #3 from Matthew Peters <matt at hpamotorsport dot com> ---
Thanks for looking into that; you are correct.

And I feel rather stupid as, when testing to double-check your assessment, I
found that the stack was not aligned before the function starts.