Possible mis-matched return types in AArch64 back-end

[rod-chapman]

In dev/aarch64_clean/src/arith_native_aarch64.h we see:

#define mld_rej_uniform_asm MLD_NAMESPACE(rej_uniform_asm)
uint64_t mld_rej_uniform_asm(int32_t *r, const uint8_t *buf, unsigned buflen,
                             const uint8_t *table);

#define mld_rej_uniform_eta2_asm MLD_NAMESPACE(rej_uniform_eta2_asm)
unsigned mld_rej_uniform_eta2_asm(int32_t *r, const uint8_t *buf,
                                  unsigned buflen, const uint8_t *table);

#define mld_rej_uniform_eta4_asm MLD_NAMESPACE(rej_uniform_eta4_asm)
unsigned mld_rej_uniform_eta4_asm(int32_t *r, const uint8_t *buf,
                                  unsigned buflen, const uint8_t *table);

Note that the return types are not the same. In the corresponding .S files for these functions, they all return a 64-bit value via register x0, suggesting that the latter 2 C function prototypes are not correct. Should they also return uint64_t rather than unsigned ?

[mkannwischer]

I don't think there is an actual problem here as all values returned are very small (0-256), so unsigned, uint64_t, int all have the same representation anyway.
It's a bit inconsistent, though, and I think we should switch to int as a part of the runtime dispatch that is being worked on in #607 - we still have some problems with that one though.

[rod-chapman]

I believe that the range and representation of int is implementation-defined, so not guaranteed to be 64-bit on all platforms?

[hanno-becker]

Thanks @rod-chapman for spotting this.

I agree with @mkannwischer: This is an inconsistency that we may get rid of, but not a functional issue.

I believe that the range and representation of int is implementation-defined, so not guaranteed to be 64-bit on all platforms?

Note that int/unsigned are typically 32-bit, even on 64-bit platforms. But yes, guaranteed is only that it's >= 16 bit, nothing else.

[rod-chapman]

I think we're just getting a silent truncation from 64 to 32-bit in a CPU register. It works, but possibly a latent foot-gun?

[rod-chapman]

Confirmed - sizeof(int) and sizeof(unsigned) are both 4, on macOS and AArch64/Linux, using both gcc and clang from within the NIX shell.