[Beignet] [PATCH OCL20 V2 1/4] Backend: Chang scan limit for GVN pass

[Yang, Rong R]

This patch would be reverted when LLVM have another solution instead of break GVN silently.
The patchset LGTM, pushed, thanks.

> -----Original Message-----
> From: Beignet [mailto:beignet-bounces at lists.freedesktop.org] On Behalf Of
> Xiuli Pan
> Sent: Friday, April 29, 2016 11:25
> To: beignet at lists.freedesktop.org
> Cc: Pan, Xiuli <xiuli.pan at intel.com>
> Subject: [Beignet] [PATCH OCL20 V2 1/4] Backend: Chang scan limit for GVN
> pass
> 
> From: Pan Xiuli <xiuli.pan at intel.com>
> 
> Set memdep-block-scan-limit into llvm context to avoid unfinished GVN pass.
> V2: Revert remove praser llvm first
> 
> Signed-off-by: Pan Xiuli <xiuli.pan at intel.com>
> ---
>  backend/src/backend/program.cpp | 22 +++++++++++++++++++++-
>  1 file changed, 21 insertions(+), 1 deletion(-)
> 
> diff --git a/backend/src/backend/program.cpp
> b/backend/src/backend/program.cpp index bbb8bd0..8eab1f6 100644
> --- a/backend/src/backend/program.cpp
> +++ b/backend/src/backend/program.cpp
> @@ -531,12 +531,21 @@ namespace gbe {
>      // Arguments to pass to the clang frontend
>      vector<const char *> args;
>      bool bFastMath = false;
> +    static bool ifsetllvm = false;
> 
>      for (auto &s : options) {
>        args.push_back(s.c_str());
>      }
> 
>      args.push_back("-cl-kernel-arg-info");
> +    // The ParseCommandLineOptions used for mllvm args can not be used
> with multithread
> +    // and GVN now have a 100 inst limit on block scan. Now only pass a
> bigger limit
> +    // for each context only once, this can also fix multithread bug.
> +    if(!ifsetllvm) {
> +      args.push_back("-mllvm");
> +      args.push_back("-memdep-block-scan-limit=200");
> +      ifsetllvm = true;
> +    }
>  #ifdef GEN7_SAMPLER_CLAMP_BORDER_WORKAROUND
>      args.push_back("-DGEN7_SAMPLER_CLAMP_BORDER_WORKAROUND");
>  #endif
> @@ -605,7 +614,18 @@ if(sizeof(int*) == 8) {
>      clang::LangOptions & lang_opts = Clang.getLangOpts();
>      lang_opts.OpenCL = 1;
> 
> -    GBE_ASSERT(Clang.getFrontendOpts().LLVMArgs.empty() && "We do not
> have llvm args now");
> +    //llvm flags need command line parsing to take effect
> +    if (!Clang.getFrontendOpts().LLVMArgs.empty()) {
> +      unsigned NumArgs = Clang.getFrontendOpts().LLVMArgs.size();
> +      const char **Args = new const char*[NumArgs + 2];
> +      Args[0] = "clang (LLVM option parsing)";
> +      for (unsigned i = 0; i != NumArgs; ++i){
> +        Args[i + 1] = Clang.getFrontendOpts().LLVMArgs[i].c_str();
> +      }
> +      Args[NumArgs + 1] = 0;
> +      llvm::cl::ParseCommandLineOptions(NumArgs + 1, Args);
> +      delete [] Args;
> +    }
> 
>      // Create an action and make the compiler instance carry it out
>      std::unique_ptr<clang::CodeGenAction> Act(new
> clang::EmitLLVMOnlyAction(llvm_ctx));
> --
> 2.5.0
> 
> _______________________________________________
> Beignet mailing list
> Beignet at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/beignet