[Mesa-stable] [PATCH] gallivm: Make sure module has the correct data layout when pass manager runs

[Roland Scheidegger]

Makes sense to me.
Reviewed-by: Roland Scheidegger <sroland at vmware.com>

Am 05.05.2017 um 03:25 schrieb Tom Stellard:
> The datalayout for modules was purposely not being set in order to work around
> the fact that the ExecutionEngine requires that the module's datalayout
> matches the datalayout of the TargetMachine that the ExecutionEngine is
> using.
> 
> When the pass manager runs on a module with no datalayout, it uses
> the default datalayout which is little-endian.  This causes problems
> on big-endian targets, because some optimizations that are legal on
> little-endian or illegal on big-endian.
> 
> To resolve this, we set the datalayout prior to running the pass
> manager, and then clear it before creating the ExectionEngine.
> 
> This patch fixes a lot of piglit tests on big-endian ppc64.
> 
> Cc: mesa-stable at lists.freedesktop.org
> ---
>  src/gallium/auxiliary/gallivm/lp_bld_init.c | 34 +++++++++++++++--------------
>  1 file changed, 18 insertions(+), 16 deletions(-)
> 
> diff --git a/src/gallium/auxiliary/gallivm/lp_bld_init.c b/src/gallium/auxiliary/gallivm/lp_bld_init.c
> index ef2580e..9f1ade6 100644
> --- a/src/gallium/auxiliary/gallivm/lp_bld_init.c
> +++ b/src/gallium/auxiliary/gallivm/lp_bld_init.c
> @@ -125,19 +125,6 @@ create_pass_manager(struct gallivm_state *gallivm)
>     LLVMAddTargetData(gallivm->target, gallivm->passmgr);
>  #endif
>  
> -   /* Setting the module's DataLayout to an empty string will cause the
> -    * ExecutionEngine to copy to the DataLayout string from its target
> -    * machine to the module.  As of LLVM 3.8 the module and the execution
> -    * engine are required to have the same DataLayout.
> -    *
> -    * TODO: This is just a temporary work-around.  The correct solution is
> -    * for gallivm_init_state() to create a TargetMachine and pull the
> -    * DataLayout from there.  Currently, the TargetMachine used by llvmpipe
> -    * is being implicitly created by the EngineBuilder in
> -    * lp_build_create_jit_compiler_for_module()
> -    */
> -
> -#if HAVE_LLVM < 0x0308
>     {
>        char *td_str;
>        // New ones from the Module.
> @@ -145,9 +132,6 @@ create_pass_manager(struct gallivm_state *gallivm)
>        LLVMSetDataLayout(gallivm->module, td_str);
>        free(td_str);
>     }
> -#else
> -   LLVMSetDataLayout(gallivm->module, "");
> -#endif
>  
>     if ((gallivm_debug & GALLIVM_DEBUG_NO_OPT) == 0) {
>        /* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
> @@ -628,6 +612,24 @@ gallivm_compile_module(struct gallivm_state *gallivm)
>     }
>  
>     if (use_mcjit) {
> +      /* Setting the module's DataLayout to an empty string will cause the
> +       * ExecutionEngine to copy to the DataLayout string from its target
> +       * machine to the module.  As of LLVM 3.8 the module and the execution
> +       * engine are required to have the same DataLayout.
> +       *
> +       * We must make sure we do this after running the optimization passes,
> +       * because those passes need a correct datalayout string.  For example,
> +       * if those optimization passes see an empty datalayout, they will assume
> +       * this is a little endian target and will do optimizations that break big
> +       * endian machines.
> +       *
> +       * TODO: This is just a temporary work-around.  The correct solution is
> +       * for gallivm_init_state() to create a TargetMachine and pull the
> +       * DataLayout from there.  Currently, the TargetMachine used by llvmpipe
> +       * is being implicitly created by the EngineBuilder in
> +       * lp_build_create_jit_compiler_for_module()
> +       */
> +      LLVMSetDataLayout(gallivm->module, "");
>        assert(!gallivm->engine);
>        if (!init_gallivm_engine(gallivm)) {
>           assert(0);
>