suspecting a bug in Clang

All,

I am facing a problem with LLVM 3.5 Compiler and suspect it is a bug in the clang. I earlier posted this query to llvm-dev list.

Here is my observation.

I am trying to support a target specific function attribute for a particular target machine.

Having followed the steps specified in the compiler documentation, I see the attribute is passed to the backend if we include the attribute in the function definition. But this attribute is not propagate to the backend if I include it only in the function’s prototype/declaration (in the C program).

This virtual function (given below) invocation in clang/lib/CodeGen/CodeGenModule.cpp in clang is responsible for attaching the target specific attributes to the backend CodeGen objects (functions, variables etc.)

getTargetCodeGenInfo().SetTargetAttributes(D, GO, *this)

I found that this virtual function is getting invoked only for function definitions and not for function declarations.

Actual requirement:

Along with regular C functions we have special functions supported in the target and its definition and invocation will always come in different compilation units. There must be a way to recognize these special functions at the call-sites during codegen in the backend. By attaching some info in the function prototype, we can recognize these special functions at the call-site during backend codegen. That’s why I thought of a new target specific attribute to achieve this purpose.

Is it really a bug?

When I compile a program having function declarations/prototypes with generic attributes like attribute((const)), clang will add it to that function’s attributelist in the IR and it will be available in the backend.

Similarly I expected the target specific attribute should also be added to the attributelist available in the backend and I should be able to identify it using F.hasFnAttribute(“XYZ”), where XYZ is the new attribute name.

I understand that a look-up in the prototype for function-attrs is very rare at codegen. But I was keen to ask this when I found that the standard attributes (‘const’, for example) are actually propagated to the backend through function prototype.

In the sample program below I have declared function foo with attribute 'const’ and you can see 'readnone’ is added to the attribute list.

Sample Program:

extern void attribute((const)) foo (int);

int main ()
{
int my_var = 45;

foo(my_var);
return 0;
}

All,

I am facing a problem with LLVM 3.5 Compiler and suspect it is a bug in the clang. I earlier posted this query to llvm-dev list.

Here is my observation.

I am trying to support a target specific function attribute for a particular target machine.

Having followed the steps specified in the compiler documentation, I see the attribute is passed to the backend if we include the attribute in the function definition. But this attribute is not propagate to the backend if I include it only in the function’s prototype/declaration (in the C program).

This virtual function (given below) invocation in clang/lib/CodeGen/CodeGenModule.cpp in clang is responsible for attaching the target specific attributes to the backend CodeGen objects (functions, variables etc.)

getTargetCodeGenInfo().SetTargetAttributes(D, GO, *this)

I found that this virtual function is getting invoked only for function definitions and not for function declarations.

Actual requirement:

Along with regular C functions we have special functions supported in the target and its definition and invocation will always come in different compilation units. There must be a way to recognize these special functions at the call-sites during codegen in the backend. By attaching some info in the function prototype, we can recognize these special functions at the call-site during backend codegen. That’s why I thought of a new target specific attribute to achieve this purpose.

Is it really a bug?

When I compile a program having function declarations/prototypes with generic attributes like attribute((const)), clang will add it to that function’s attributelist in the IR and it will be available in the backend.

Similarly I expected the target specific attribute should also be added to the attributelist available in the backend and I should be able to identify it using F.hasFnAttribute(“XYZ”), where XYZ is the new attribute name.

I understand that a look-up in the prototype for function-attrs is very rare at codegen. But I was keen to ask this when I found that the standard attributes (‘const’, for example) are actually propagated to the backend through function prototype.

In the sample program below I have declared function foo with attribute 'const’ and you can see 'readnone’ is added to the attribute list.

Sample Program:

extern void attribute((const)) foo (int);

int main ()
{
int my_var = 45;

foo(my_var);
return 0;
}

Generated IR:

; Function Attrs: nounwind
define i16 @main() #0 {
%1 = alloca i16, align 2
%my_var = alloca i16, align 2
store i16 0, i16* %1
call void @llvm.dbg.declare(metadata !{i16* %my_var}, metadata !13), !dbg !14
store i16 45, i16* %my_var, align 2, !dbg !15
%2 = load i16* %my_var, align 2, !dbg !16
call void @foo(i16 %2) #1, !dbg !16
ret i16 0, !dbg !17
}

; Function Attrs: nounwind readnone
declare void @llvm.dbg.declare(metadata, metadata) #1

; Function Attrs: nounwind readnone
declare void @foo(i16) #2

I think you are right: SetTargetAttributes doesn’t get called for function declarations. GetOrCreateLLVMFunction is called to generate the LLVM IR for a function declaration, which eventually calls ConstructAttributeList, in which readnone is added. If you want to add target attributes to function declarations, I believe you’ll have to somehow call SetTargetAttributes in ConstructAttributeList.

Hi Akira,

I have applied this patch in my workspace to add the attributes to function prototypes.