Invalid work with floating point if stack is not aligned to 8 bytes

[Vitaliy Bortsov]

I check it for RM46L852 chip, with software floating point support.

#include <stdio.h>

int main()
{
    float f = 0.000187F;
    char buf[100];
    snprintf(buf, 100, "%f %f", f, (int)(f * 1e6f));
    return 0;
}

If top of stack (at least before calling __libc_init() from my startup code) is not aligned to 8 bytes then buf will be invalid for both numbers.
If top of stack is aligned to 8 bytes then buf will be correct.

Bug is detected for v5.2 2015q4, v4.9 2015q1.
Compilation flags:
arm-none-eabi-gcc -Og -mcpu=cortex-r4f -marm -g -mfloat-abi=soft -fmessage-length=0 -fno-builtin -c -std=c99 main.cpp -o main.o

Linking flags:
arm-none-eabi-g++ -Og -mcpu=cortex-r4f -marm -g -mfloat-abi=soft -fmessage-length=0 -fno-builtin -Wl,-T,rom.lds -Wl,--gc-sections --specs=nosys.specs main.o sys_startup.o -o fw.elf

[Vitaliy Bortsov (bortsov)]

Additional information.
According to ARM documentation, SP must be aligned to 8 bytes for AAPCS and ABI. See http://infocenter.arm.com/help/topic/com.arm.doc.ihi0046b/IHI0046B_ABI_Advisory_1.pdf
As I suppose code in newlib rely for 8-bytes alignment.

Is all calls in newlib done according to ARM ABI, AAPCS and 8-bytes alignment on stack? It seems that no.

Should I close this "bug"?

[Tejas Belagod (belagod-tejas)]

The 32-bit AAPCS mandates the SP to be aligned to 8-byte boundary for all AAPCS-conforming public interfaces. So, if the stack isn't aligned to 8-bytes, the result of the program is unpredictable. This is not a bug. Thanks.