llvm dev - Aug 2015 - LibCallAliasAnalysis class instantiation parameter

Hello,

[sorry for the duplicate post, attempting to fix the formatting of the previous
post]

Can anybody tell me the intent of the LibCallAliasAnalysis class? I see that to
instantiate it, it requires a class derived from the pure virtual class
LibCallInfo, but I cannot find any classes in the source tree that derive from
LibCallInfo. Is this derived class intended to be supplied from a compiler
front-end, library writer, or the llvm backend?

I was looking at a case similar to the following:

              #include <math.h>

     double x[1000], y[1000];
     double angle;

     void rotate(int point_count) {
        int i;

            for (i=0; i < point_count; i++) {
                    double tmp_x = x[i];
             x[i] = tmp_x * cos(angle) - y[i] * sin(angle);
                    y[i] = y[i] * cos(angle) + tmp_x * sin(angle);
         }
     }

In this case, the queries done by getModRefInfo are unable to avoid loading the
‘angle’ variable from memory for each call to sin/cos. Based on the comments in
LibCallSemantics.h, it appears that LibCallAliasAnalysis should be able to
handle this case by recognizing that sin/cos only affect errno. However, I don’t
see any way of enabling this analysis without an appropriate derivation of
LibCallInfo. Does anybody have info related to this?

Thanks,
Chris

Hi Chris,

LibCallAliasAnalysis was recently deleted from trunk because it was never really
used or supported. We do need a way to reason about errno, but we don't have
it yet.

To your particular problem, if you can compile your code with -fno-math-errno
(or -ffast-math which implies it), then the sin/cos functions will be marked as
readnone and AA will understand that they don't have side effects.

 -Hal

----- Original Message -----
> From: "Chris Chrulski via llvm-dev" <llvm-dev at
lists.llvm.org>
> To: llvm-dev at lists.llvm.org
> Sent: Thursday, August 6, 2015 11:04:26 AM
> Subject: Re: [llvm-dev] LibCallAliasAnalysis class instantiation parameter
> 
> Hello,
> 
> [sorry for the duplicate post, attempting to fix the formatting of
> the previous post]
> 
> Can anybody tell me the intent of the LibCallAliasAnalysis class? I
> see that to instantiate it, it requires a class derived from the
> pure virtual class LibCallInfo, but I cannot find any classes in the
> source tree that derive from LibCallInfo. Is this derived class
> intended to be supplied from a compiler front-end, library writer,
> or the llvm backend?
> 
> I was looking at a case similar to the following:
> 
>               #include <math.h>
> 
>      double x[1000], y[1000];
>      double angle;
> 
>      void rotate(int point_count) {
>         int i;
> 
>             for (i=0; i < point_count; i++) {
>                     double tmp_x = x[i];
>              x[i] = tmp_x * cos(angle) - y[i] * sin(angle);
>                     y[i] = y[i] * cos(angle) + tmp_x * sin(angle);
>          }
>      }
> 
> In this case, the queries done by getModRefInfo are unable to avoid
> loading the ‘angle’ variable from memory for each call to sin/cos.
> Based on the comments in LibCallSemantics.h, it appears that
> LibCallAliasAnalysis should be able to handle this case by
> recognizing that sin/cos only affect errno. However, I don’t see any
> way of enabling this analysis without an appropriate derivation of
> LibCallInfo. Does anybody have info related to this?
> 
> Thanks,
> Chris
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org         http://llvm.cs.uiuc.edu
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
> 
-- 
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory

Thank you, Hal

Using the -fno-math-errno option was able to trigger the desired optimizations.

Chris




On Saturday, August 15, 2015 7:31 PM, Hal Finkel <hfinkel at anl.gov>
wrote:
Hi Chris,

LibCallAliasAnalysis was recently deleted from trunk because it was never really
used or supported. We do need a way to reason about errno, but we don't have
it yet.

To your particular problem, if you can compile your code with -fno-math-errno
(or -ffast-math which implies it), then the sin/cos functions will be marked as
readnone and AA will understand that they don't have side effects.

-Hal


----- Original Message -----
> From: "Chris Chrulski via llvm-dev" <llvm-dev at
lists.llvm.org>
> To: llvm-dev at lists.llvm.org
> Sent: Thursday, August 6, 2015 11:04:26 AM
> Subject: Re: [llvm-dev] LibCallAliasAnalysis class instantiation parameter
> 
> Hello,
> 
> [sorry for the duplicate post, attempting to fix the formatting of
> the previous post]
> 
> Can anybody tell me the intent of the LibCallAliasAnalysis class? I
> see that to instantiate it, it requires a class derived from the
> pure virtual class LibCallInfo, but I cannot find any classes in the
> source tree that derive from LibCallInfo. Is this derived class
> intended to be supplied from a compiler front-end, library writer,
> or the llvm backend?
> 
> I was looking at a case similar to the following:
> 
>               #include <math.h>
> 
>      double x[1000], y[1000];
>      double angle;
> 
>      void rotate(int point_count) {
>         int i;
> 
>             for (i=0; i < point_count; i++) {
>                     double tmp_x = x[i];
>              x[i] = tmp_x * cos(angle) - y[i] * sin(angle);
>                     y[i] = y[i] * cos(angle) + tmp_x * sin(angle);
>          }
>      }
> 
> In this case, the queries done by getModRefInfo are unable to avoid
> loading the ‘angle’ variable from memory for each call to sin/cos.
> Based on the comments in LibCallSemantics.h, it appears that
> LibCallAliasAnalysis should be able to handle this case by
> recognizing that sin/cos only affect errno. However, I don’t see any
> way of enabling this analysis without an appropriate derivation of
> LibCallInfo. Does anybody have info related to this?
> 
> Thanks,
> Chris
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org        http://llvm.cs.uiuc.edu
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
> 
-- 
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory