On 01/02/2018 11:03 PM, Prathamesh Kulkarni wrote:
> Hi,
> malloc_candidate_p() in ipa-pure-const misses detecting that a
> function is malloc-like if the return value is result of PHI and one
> of the arguments of PHI is 0.
> For example:
>
> void g(unsigned n)
> {
> return (n) ? __builtin_malloc (n) : 0;
> }
>
> The reason is that the following check:
> if (TREE_CODE (arg) != SSA_NAME)
> DUMP_AND_RETURN ("phi arg is not SSA_NAME.")
>
> fails for arg with constant value 0 and malloc_candidate_p returns false.
> The patch simply skips the arg if it equals null_pointer_node.
> Does it look OK ?
>
> One concern I have is that with the patch, malloc_candidate_p will
> return true if all the args to PHI are NULL:
> retval = PHI<0, 0>
> return retval
>
> However I expect that PHI with all 0 args would be constant folded to
> 0 earlier, so this case shouldn't occur in practice ?
>
> Bootstrapped+tested on x86_64-unknown-linux-gnu.
> Cross-testing on arm*-*-* and aarch64*-*-* in progress.
A degenerate PHI should be folded/propagated, but you can't absolutely
depend on it -- ie, you must handle it gracefully.
I think the way to do that here is verify that at least one argument is
an SSA_NAME.
jeff
