[Beignet] [PATCH] GBE: workaround register allocation fail caused by custom loop unroll.

[Zhigang Gong]

LGTM, pushed, thanks.

On Mon, Nov 03, 2014 at 03:52:25PM +0800, Ruiling Song wrote:
> As this issue only occurs under strict math, we disable custom
> loop unroll if strict math is enabled.
> 
> Signed-off-by: Ruiling Song <ruiling.song at intel.com>
> ---
>  backend/src/llvm/llvm_to_gen.cpp |   19 ++++++++++++-------
>  1 file changed, 12 insertions(+), 7 deletions(-)
> 
> diff --git a/backend/src/llvm/llvm_to_gen.cpp b/backend/src/llvm/llvm_to_gen.cpp
> index 0c7785b..33aec59 100644
> --- a/backend/src/llvm/llvm_to_gen.cpp
> +++ b/backend/src/llvm/llvm_to_gen.cpp
> @@ -107,7 +107,7 @@ namespace gbe
>      FPM.doFinalization();
>    }
>  
> -  void runModulePass(Module &mod, TargetLibraryInfo *libraryInfo, const DataLayout &DL, int optLevel)
> +  void runModulePass(Module &mod, TargetLibraryInfo *libraryInfo, const DataLayout &DL, int optLevel, bool strictMath)
>    {
>      llvm::PassManager MPM;
>  
> @@ -158,11 +158,16 @@ namespace gbe
>        MPM.add(createGVNPass());                 // Remove redundancies
>      }
>  #if LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR >= 5
> -    MPM.add(createCustomLoopUnrollPass()); //1024, 32, 1024, 512)); //Unroll loops
> -    MPM.add(createLoopUnrollPass()); //1024, 32, 1024, 512)); //Unroll loops
> -    if(optLevel > 0) {
> -      MPM.add(createSROAPass(/*RequiresDomTree*/ false));
> -      MPM.add(createGVNPass());                 // Remove redundancies
> +    // FIXME Workaround: we find that CustomLoopUnroll may increase register pressure greatly,
> +    // and it may even make som cl kernel cannot compile because of limited scratch memory for spill.
> +    // As we observe this under strict math. So we disable CustomLoopUnroll if strict math is enabled.
> +    if (!strictMath) {
> +      MPM.add(createCustomLoopUnrollPass()); //1024, 32, 1024, 512)); //Unroll loops
> +      MPM.add(createLoopUnrollPass()); //1024, 32, 1024, 512)); //Unroll loops
> +      if(optLevel > 0) {
> +        MPM.add(createSROAPass(/*RequiresDomTree*/ false));
> +        MPM.add(createGVNPass());                 // Remove redundancies
> +      }
>      }
>  #endif
>      MPM.add(createMemCpyOptPass());             // Remove memcpy / form memset
> @@ -250,7 +255,7 @@ namespace gbe
>      OUTPUT_BITCODE(AFTER_LINK, mod);
>  
>      runFuntionPass(mod, libraryInfo, DL);
> -    runModulePass(mod, libraryInfo, DL, optLevel);
> +    runModulePass(mod, libraryInfo, DL, optLevel, strictMath);
>      llvm::PassManager passes;
>  #if LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR >= 5
>      passes.add(new DataLayoutPass(DL));
> -- 
> 1.7.10.4
> 
> _______________________________________________
> Beignet mailing list
> Beignet at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/beignet