Compiler produces orr instead of bic using (O2 and Os)

[Morty]

The code is more or less form the SMT32 lib, I did some of the preprocesser-replacements to simplify the code.
I noticed this bug using 7-2018-q2 on windows, but research shows it's more wide spread.

Using -O2/-Os and up GCC 6.4 produces the expected bic, while gcc 7.3 and later produces an orr. -O1 seems to be fine:

https://godbolt.org/z/WJW4nc

#include <stdint.h>

typedef struct
{
    volatile uint32_t DIFSEL;
} ADC_TypeDef;

#define WRITE_REG(REG, VAL) ((REG) = (VAL))
#define MODIFY_REG(REG, CLEARMASK, SETMASK) WRITE_REG((REG), (((READ_REG(REG)) & (~(CLEARMASK))) | (SETMASK)))

static inline void LL_ADC_SetChannelSingleDiff(ADC_TypeDef *ADCx, uint32_t Channel, uint32_t SingleDiff)
{
    (((ADCx->DIFSEL)) = ((((((ADCx->DIFSEL))) & (~(Channel & (((0x7FFFFU << (0U))))))) | ((Channel & (((0x7FFFFU << (0U))))) & ((0x7FFFFU << (0U)) << (SingleDiff & ((0x20U << (0U)))))))));
}

int test(ADC_TypeDef * testvar) {

    LL_ADC_SetChannelSingleDiff(testvar,
    (( (0x01U << (26U))) | (((uint32_t)0x00000000U) | (((uint32_t) 3U) << ((uint32_t)20U))) | ((0x00002U << (0U))) ),
          ( (0x7FU << (0U)))
    );
}

gcc-6.4 -Os:
test:
        ldr r3, [r0]
        bic r3, r3, #2
        str r3, [r0]
        bx lr

gcc-7.3 -Os:
test:
        ldr r3, [r0]
        orr r3, r3, #2
        str r3, [r0]
        bx lr

[Dominic Plunkett (dp111)]

I suspect the issue is : 0x7FFFU<< (SingleDiff & ((0x20U << (0U))))

which is a shift by 32 , this is undefined so the compiler can do what it likes

If you change the 0x7FFFFU to 0x7FFFFLL you get what you would expect. I would suggest there are betters to code this function than using 0x7FFFFLL

[Morty (morty-rh)]

I'm afraid you're right: http://c0x.coding-guidelines.com/6.5.7.html#1185

None the less I think this behavior is pretty dangerous without a warning, as I have seen this kind of stuff pretty often.
From the STM32-Hal where I got this code from:

  /* Bits of channels in single or differential mode are set only for */
  /* differential mode (for single mode, mask of bits allowed to be set is */
  /* shifted out of range of bits of channels in single or differential mode. */

As this is a gcc issue and not arm-gcc (the behavior can be seen with x86, too), I discuss that there.

[Dominic Plunkett (dp111)]

Running cppcheck ( static analysis ) would show up this issue as being undefined. I think the STM Hal should be improved.

[Morty (morty-rh)]

I actually run cppcheck on our code once in a while, but there are just too many warnings in the STM HAL. Gcc already throws plenty of warnings.