[LLVMdev] [Polly] Summary of some expensive compiler passes, especially PollyDependence

[Tobias Grosser]

On 08/08/2013 01:29 AM, Star Tan wrote:
> Hi all,
>
>
> I have summarized the top 10 compiler passes for Polly when compiling LLVM test-ssuite. Results can be viewed on:
>      https://gist.github.com/tanstar/581bcea1e4e03498f935/raw/f6a4ec4e8565f7a7bbdb924cd59fcf145caac039/Polly-top10
>
>
> Based on the comparison between "clang -O3" and "polly -O3" listed on:
>      http://188.40.87.11:8000/db_default/v4/nts/18?compare_to=14&baseline=14

Please compare against clang -O3 -load LLVMPolly.so, otherwise 
especially the compile time of the small binaries includes the overhead 
of loading the Polly shared object file.

> we can see Polly's compile-time overhead is mainly resulted by some expensive Polly passes such as PollyDependence, PollyOptimization and PollyCodegen. Especially, I notice that the PollyDependence can lead to significant extra compile-time overhead. Its compile-time percentage for some expensive benchmarks can be summarized as:
>      nestedloop:  41.4% (Polly - Calculate dependence)
>      salsa20:       98.5% (Polly - Calculate dependence)
>      seidel-2d:     72.1% (Polly - Calculate dependence)
>      multiplies:     54.3% (Poly -  Calculate dependence)
>      Puzzle:         22.8% (Poly -  Calculate dependence)
>
>
> As a result, it is critical to improve the PollyDependence pass.  I have previously committed a patch file to split start value from SCEVAddRecExpr (r187728), which can improve the PollyDependence pass for some benchmarks as follows:
>       http://188.40.87.11:8000/db_default/v4/nts/25?baseline=18&compare_to=18

Those are very nice results. Removing 80% of the compile time for the lu 
benchmark is a very nice outcome.

> Obviously, it is not enough.

Sure, there is still plenty left to be optimised. However, when Polly 
detects a code region that it can optimise it seems fair that we spend 
some time in analysing and optimising it. Especially if our 
transformation improves the run-time performance.

> I have investigated  the salsa20 benchmark, in which the PollyDependence pass consumes 98.5% of total compile-time. The key code of salsa20 benchmark consists of a serial of array operations:
>    #define R(a,b) (((a) << (b)) | ((a) >> (32 - (b))))
>    for (i = 20;i > 0;i -= 2) {
>      x[ 4] ^= R(x[ 0]+x[12], 7);  x[ 8] ^= R(x[ 4]+x[ 0], 9);
>      x[12] ^= R(x[ 8]+x[ 4],13);  x[ 0] ^= R(x[12]+x[ 8],18);
>      ....
>      x[14] ^= R(x[13]+x[12],13);  x[15] ^= R(x[14]+x[13],18);
>    }
> which would be translated into a serial of affine functions:
>        ReadAccess :=    { Stmt_for_body5[i0] -> MemRef_x[12] };
>        ReadAccess :=    { Stmt_for_body5[i0] -> MemRef_x[4] };
>        MustWriteAccess :=    { Stmt_for_body5[i0] -> MemRef_x[4] };
>        ReadAccess :=    { Stmt_for_body5[i0] -> MemRef_x[0] };
>        ...
> Consequently, the PollyDependence pass would produce very complex Read/Write/MayWrite as follows:
>
>
>        Read: { Stmt_for_body[i0] -> MemRef_in[i0] : i0 >= 0 and i0 <=
> 15; Stmt_for_body5[i0] -> MemRef_x[15] : (i0 >= 0 and i0 <= 9) or (i0
>> = 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9) or
> (i0 >= 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9); Stmt_for_body5[i0] ->
> MemRef_x[9] : (i0 >= 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9) or (i0 >=
> 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9); Stmt_for_body5[i0] ->
> MemRef_x[4] : (i0 >= 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9) or (i0 >=
> 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9); Stmt_for_body5[i0] ->
> MemRef_x[11] : (i0 >= 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9) or (i0 >=
> 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9); Stmt_for_body5[i0] ->
> MemRef_x[2] : (i0 >= 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9) or (i0 >=
> 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9); Stmt_for_body5[i0] ->
> MemRef_x[6] : (i0 >= 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9) or (i0 >=
> 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9); Stmt_for_body5[i0] ->
> MemRef_x[14] : (i0 >= 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9) or (i0 >=
> 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9); Stmt_for_body5[i0] ->
> MemRef_x[8] : (i0 >= 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9) or (i0 >=
> 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9); Stmt_for_body5[i0] ->
> MemRef_x[10] : (i0 >= 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9) or (i0 >=
> 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9) or (i0
>> = 0 and i0 <= 9); Stmt_for_body5[i0] -> MemRef_x[0] : (i0 >= 0 and i0
> <= 9) or (i0 >= 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9) or (i0 >= 0 and
> i0 <= 9) or (i0 >= 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9);
> Stmt_for_body5[i0] -> MemRef_x[13] : (i0 >= 0 and i0 <= 9) or (i0 >= 0
> and i0 <= 9) or (i0 >= 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9);
> Stmt_for_body5[i0] -> MemRef_x[1] : (i0 >= 0 and i0 <= 9) or (i0 >= 0
> and i0 <= 9) or (i0 >= 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9);
> Stmt_for_body5[i0] -> MemRef_x[3] : (i0 >= 0 and i0 <= 9) or (i0 >= 0
> and i0 <= 9) or (i0 >= 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9);
> Stmt_for_body5[i0] -> MemRef_x[7] : (i0 >= 0 and i0 <= 9) or (i0 >= 0
> and i0 <= 9) or (i0 >= 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9);
> Stmt_for_body5[i0] -> MemRef_x[12] : (i0 >= 0 and i0 <= 9) or (i0 >= 0
> and i0 <= 9) or (i0 >= 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9);
> Stmt_for_body5[i0] -> MemRef_x[5] : (i0 >= 0 and i0 <= 9) or (i0 >= 0
> and i0 <= 9) or (i0 >= 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9) or (i0
>> = 0 and i0 <= 9) or (i0 >= 0 and i0 <= 9) }
>        Write: ...
>
>
> As we can see, the reason why PollyDependence leads to expensive compile-time lies in the very complex Read/Write/MayWrite structure.  In fact, we can run isl_union_map_coalesce() on the Read/MayWrite/MustWrite before feeding them into ISL analysis:

Very nice.

> diff --git a/lib/Analysis/Dependences.cpp b/lib/Analysis/Dependences.cpp index 9f918f3..39c3fb6 100644 --- a/lib/Analysis/Dependences.cpp +++ b/lib/Analysis/Dependences.cpp @@ -95,6 +95,10 @@ void Dependences::calculateDependences(Scop &S) { collectInfo(S, &Read, &Write, &MayWrite, &Schedule); + Read = isl_union_map_coalesce(Read); + Write = isl_union_map_coalesce(Write); + MayWrite = isl_union_map_coalesce(MayWrite); + DEBUG(dbgs() << "Read: " << Read << "\n"

This patch is unreadable in the mail. However, the one you submitted 
looked good and was committed.

> With this patch file, we can reduce the compile-time percentage of PollyDependence from 98.5% to 15.3%. Unfortunately, the compile-time percentage of PollyDependence for benchmarks, such as "nestedloop", is still very high.

It would be good to get an up-to-date comparison with the latest patch 
having gone into Polly.

I did not yet look at the nestedloop benchmark, but it sounds basically 
like a benchmark only consisting of loop nests that we can optimise. 
This is definitely interesting to look into. Both in respect of how fast
we can analyse it faster, but also if we are able to improve the 
performance of the generated code. Especially if we improve the 
execution performance some additional compile time is justified.

> My plan is to continue investigating why PollyDependence pass leads to such high compile-time overhead. Could anyone give me some comments or suggestions?

You already made nice progress. I propose to just continue by measuring 
the remaining overhead, taking one of the top cases and analysing it. We 
then see on a case by case basis what can be done.

Cheers,
Tobias