Higher optimization levels miscompile consecutive i24 subtractions by retaining carry

[NinjadenMu]

Reproduction:

#include "debug.h"

typedef __INT24_TYPE__ i24;

int main()
{
  volatile int24_t input = 5;
  int24_t foo = input;

  int24_t bar = -foo;
  int24_t quux = -foo + 1;
  dbg_printf("%d", bar);
  dbg_printf("%d", quux);
}

Compiling this under -O0 prints the expected -5, -4, while compiling under -O3 prints -5, -5.

I suspect this is due to the carry bit being retained between computing bar=-foo via sbc 0 foo followed by quux=-foo+1 via sbc 1 foo, which actually computes 1-foo-1 due to the retained carry bit. -O0 avoids this by including or a, a after computing bar.

Should be able to see -O0 clearing the carry flag with or a, a while -O3 doesn't by viewing assembly generated for this:

typedef __INT24_TYPE__ i24;

extern void consume(i24, i24);

void test(i24 foo)
{
  consume(-foo, 1 - foo);
}

On clang version 19.1.0, commit ef28e9c54cd1333a6091ab2ffbd315b465fc5090

[Sightem]

Fixed in 199e8cc