[RFC] [PATCH V2]: RE: [RFC] [Patch] Relax tree-if-conv.c trap assumptions.

[Bernhard Reutner-Fischer]

On November 10, 2015 1:02:57 PM GMT+01:00, Richard Biener <richard.guenther@gmail.com> wrote:
>On Sat, Nov 7, 2015 at 12:41 PM, Kumar, Venkataramanan
><Venkataramanan.Kumar@amd.com> wrote:
>> Hi Richard,
>>
>> I have now implemented storing of DR and references using hash maps.
>> Please find attached patch.
>>
>> As discussed, I am now storing the ref, DR  and baseref, DR pairs
>along with unconditional read/write information  in  hash tables while
>iterating over DR during its initialization.
>> Then during checking for possible traps for if-converting,  just
>check if the memory reference for a gimple statement is read/written
>unconditionally by querying the hash table instead of quadratic walk.
>>
>> Boot strapped and regression tested on x86_64.
>
>@@ -592,137 +598,153 @@ struct ifc_dr {
>
>   /* -1 when not initialized, 0 when false, 1 when true.  */
>   int rw_unconditionally;
>+
>+  tree ored_result;
>+
>
>excess vertical space at the end.  A better name would be simply
>"predicate".
>
>+  if (!exsist1)

s/exsist/exists/g

Also watch out for wrong spaces around assignments and comparisons (one "=3" in the testcase and one "=0" in the code). Not sure offhand if check_GCC_style.sh catches those.

Thanks,