2013/11/7 Jeff Law <l...@redhat.com>: > On 10/31/13 03:15, Ilya Enkovich wrote: >> >> Hi, >> >> Here is a patch which adds support for bound constant to be used as >> DECL_INITIAL for constant static bounds generated by compiler. >> >> Thanks, >> Ilya >> -- >> >> gcc/ >> >> 2013-10-23 Ilya Enkovich <ilya.enkov...@intel.com> >> >> * emit-rtl.c (immed_double_const): Support MODE_POINTER_BOUNDS. >> * explow.c (trunc_int_for_mode): Likewise. >> * varpool.c (ctor_for_folding): Do not fold constant >> bounds vars. > > I'm having a bit of trouble reconciling "add support for bound constant to > be used as DECL_INITIAL" rationale text and the actual patch. > > From reading the patch it appears that you want to allow generation of > immediate constants for objects with MODE_POINTER_BOUNDS. OK, I can see how > that is useful. > > I can kindof see how you want to error out if someone asks for a constant to > be truncated to MODE_POINTER_BOUNDS. Did this trip in practice or is it > preemptive?
As far as I remember change in trunc_int_mode was required to expand bound constants on 32bit target. Size of the constant is equal to size of the HOST_WIDE_INT and thus constant generation goes through gen_int_mode and trunc_int_for_mode. > > > >> diff --git a/gcc/varpool.c b/gcc/varpool.c >> index 2eb1fc1..d9c08c1 100644 >> --- a/gcc/varpool.c >> +++ b/gcc/varpool.c >> @@ -254,6 +254,12 @@ ctor_for_folding (tree decl) >> && TREE_CODE (decl) != CONST_DECL) >> return error_mark_node; >> >> + /* Static constant bounds are created to be >> + used instead of constants and therefore >> + do not let folding it. */ >> + if (POINTER_BOUNDS_P (decl)) >> + return error_mark_node; > > Here's the part I'm struggling a bit with. Why did you need this? > > Isn't this going to prevent that DECL from being used in folding? The > bounds shouldn't really affect that AFAICT. Bounds constants were introduced only for initialization of constant bound vars. Such vars are used to hold commonly used zero bounds (for cases when bounds are unknown) values and null bounds (for null pointers). Usage of such vars is optional and is controlled via compiler flag. It is used to try to decrease overhead on bounds creation. E.g. for MPX we need two instructions to create zero bounds and also it require one GPR. One of these instructions does not become nop when MPX is off which additionally increases overhead. Having constant var we can just load bounds using one MPX instruction. And if I do not prevent folding for these vars then all constant bounds vars usages are replaced with immediate bounds constant usages and I do not get desired effect. Since there are no instructions working with bounds immediates, I do not see reasons for folding. Thanks, Ilya > > jeff
