[LLVMdev] DAG optimization and lowering algorithm

Demikhovsky, Elena elena.demikhovsky at intel.com
Tue Feb 7 00:38:46 PST 2012


At the beginning, I have the following chain: LOAD -> TRUNCATE -> ZERO_EXTEND.
After Combine(BeforeLegalizeTypes) the optimization of ZERO_EXTEND  gives me the new chain LOAD -> ANY_EXTEND -> AND.

I want to optimize ANY_EXTEND but is not analyzed in the same Combine().

Combine(AfterLegalizeTypes) is no called at all.


- Elena

-----Original Message-----
From: Eli Friedman [mailto:eli.friedman at gmail.com] 
Sent: Tuesday, February 07, 2012 10:15
To: Demikhovsky, Elena
Cc: llvmdev at cs.uiuc.edu
Subject: Re: [LLVMdev] DAG optimization and lowering algorithm

On Mon, Feb 6, 2012 at 11:54 PM, Demikhovsky, Elena
<elena.demikhovsky at intel.com> wrote:
> Hi,
>
> I'm trying to build code for very short function and I encounter with a problem (or bug) in DAG selection algotithm.
> I have a node that was created in Combine(BeforeLegalizeTypes) and should be optimized in Combine(AfterLegalizeTypes). But LegalizeTypes() did not change anything and Combine(AfterLegalizeTypes) was not called.
> Vector legalization that comes afterwards just scalarized the operation.
>
> SelectionDAGISel::CodeGenAndEmitDAG()
> ..
>   CurDAG->Combine(BeforeLegalizeTypes, *AA, OptLevel);
>
>   bool Changed;
>   Changed = CurDAG->LegalizeTypes();
>
>   if (Changed) {
>      CurDAG->Combine(AfterLegalizeTypes, *AA, OptLevel);
>    }
>
> The problem does not exist if the function is bigger and something was changed on LegalizeTypes() stage.
>
> I think that Combine(AfterLegalizeTypes) should be called anyway, even if LegalizeTypes() did not change anything.

We skip the extra run as a performance optimization... we really don't
want extra runs of DAGCombine if we can avoid it.  If there is some
optimization that currently only runs after type legalization, perhaps
we can change it?

-Eli
---------------------------------------------------------------------
Intel Israel (74) Limited

This e-mail and any attachments may contain confidential material for
the sole use of the intended recipient(s). Any review or distribution
by others is strictly prohibited. If you are not the intended
recipient, please contact the sender and delete all copies.





More information about the llvm-dev mailing list