[Mesa-stable] [Mesa-dev] [PATCH] radeon/compute: Unconditionally inline all functions

[Tom Stellard]

On Thu, Nov 14, 2013 at 11:38:06AM -0600, Aaron Watry wrote:
> On Thu, Nov 14, 2013 at 10:29 AM, Tom Stellard <tom at stellard.net> wrote:
> > From: Tom Stellard <thomas.stellard at amd.com>
> >
> > We need to do this until function calls are supported.
> >
> > https://bugs.freedesktop.org/show_bug.cgi?id=64225
> >
> > CC: "10.0" <mesa-stable at lists.freedesktop.org>
> > ---
> >  src/gallium/drivers/radeon/radeon_llvm_util.c | 24 ++++++++++++++++++++++--
> >  1 file changed, 22 insertions(+), 2 deletions(-)
> >
> > diff --git a/src/gallium/drivers/radeon/radeon_llvm_util.c b/src/gallium/drivers/radeon/radeon_llvm_util.c
> > index 7192dee..02a11b5 100644
> > --- a/src/gallium/drivers/radeon/radeon_llvm_util.c
> > +++ b/src/gallium/drivers/radeon/radeon_llvm_util.c
> > @@ -30,6 +30,7 @@
> >  #include <llvm-c/BitReader.h>
> >  #include <llvm-c/Core.h>
> >  #include <llvm-c/Target.h>
> > +#include <llvm-c/Transforms/IPO.h>
> >  #include <llvm-c/Transforms/PassManagerBuilder.h>
> >
> >  LLVMModuleRef radeon_llvm_parse_bitcode(const unsigned char * bitcode,
> > @@ -58,9 +59,28 @@ static void radeon_llvm_optimize(LLVMModuleRef mod)
> >         LLVMTargetDataRef TD = LLVMCreateTargetData(data_layout);
> >         LLVMPassManagerBuilderRef builder = LLVMPassManagerBuilderCreate();
> >         LLVMPassManagerRef pass_manager = LLVMCreatePassManager();
> > -       LLVMAddTargetData(TD, pass_manager);
> >
> > -       LLVMPassManagerBuilderUseInlinerWithThreshold(builder, 1000000000);
> > +       /* Functions calls are not supported yet, so we need to inline
> > +        * everything.  The most efficient way to do this is to add
> > +        * the always_inline attribute to all non-kernel functions
> > +        * and then run the Always Inline pass.  The Always Inline
> > +        * pass will automaically inline functions with this attribute
> > +        * and does not perform the expensive cost analysis that the normal
> > +        * inliner does.
> > +        */
> > +
> > +       LLVMValueRef fn;
> > +       LLVMValueRef last_fn = LLVMGetLastFunction(mod);
> > +       for (fn = LLVMGetFirstFunction(mod); fn != last_fn;
> > +                                               fn = LLVMGetNextFunction(fn)) {
> 
> Will this actually run for the last function? Is there reason to
> believe that the last function isn't affected?
> 
> Consider:
> 
> int testFn(int4 arg);
> 
> kernel void testKernel(global int* data){
>  int4 input;
>  .... stuff ....
>  int val = testFn(input);
> }
> 
> int testFn(int4 arg){
>   return 0;
> }
> 
> I'm assuming that functions will exist in the order that they were
> defined (not declared) when looping over the functions in the module.
> Given that LLVMGetNextFunction returns 0 when it hits the end of the
> list, couldn't we just check that?
> 

Nice catch, this patch is wrong.  I will send a new one.

-Tom

> 
> 
> > +               /* All the non-kernel functions have internal linkage */
> > +               if (LLVMGetLinkage(fn) == LLVMInternalLinkage) {
> > +                       LLVMAddFunctionAttr(fn, LLVMAlwaysInlineAttribute);
> > +               }
> > +       }
> > +
> > +       LLVMAddTargetData(TD, pass_manager);
> > +       LLVMAddAlwaysInlinerPass(pass_manager);
> >         LLVMPassManagerBuilderPopulateModulePassManager(builder, pass_manager);
> >
> >         LLVMRunPassManager(pass_manager, mod);
> > --
> > 1.8.1.4
> >
> > _______________________________________________
> > mesa-dev mailing list
> > mesa-dev at lists.freedesktop.org
> > http://lists.freedesktop.org/mailman/listinfo/mesa-dev