[compiler-rt] Undefined Behavior in compiler-rt builtin __divdf3

[nvetrini]

In the implementation of __divdf3, which defines macro DOUBLE_PRECISION, the macro invocation REP_C(-1) is expanded (see https://github.com/llvm/llvm-project/blame/ca69a8d2f403de3617970dcfa2f84756f7f336dd/compiler-rt/lib/builtins/fp_div_impl.inc#L29). In this context REP_C is defined as UINT64_C, for which the behavior with the constant -1 as argument is undefined for all versions of the C Standard. Given the intent of the problematic use of REP_C, UINT64_MAX could be used instead.

[pinskia]

It is NOT UB but rather implementation defined behavior ...

[nvetrini]

From my interpretation of the C23 standard (I expect earlier ones to be analogous in this respect), this falls under UB 150

(150) The argument in an instance of one of the integer-constant macros is not a decimal, octal, or
hexadecimal constant, or it has a value that exceeds the limits for the corresponding type
(7.22.4).

It could be argued that -1UL can be represented as an unsigned integer, and certainly the construct does behave as expected in all reasonable compilers, though it is also easy to avoid the IDB/UB.