[LLVMdev] Forcing function inline not working

[Javier Martinez]

I got the code to work in a LLVM 2.5. I think the reason your code wasn't
working is that the function definition and declaration for foo were
missing 'void'. This is perfectly legal C but maybe there's a bug in the
inliner pass that sees the function in the module and the symbol table as
different. Here are the files I used:

[test.c]
int a;

void foo(void);

int main(int argc, int *argv) {
  foo();
  
  return 0;
}

__attribute__((always_inline)) void foo(void) {
  a++;
}


[test.ll]
; ModuleID = 'test.c'
target datalayout =
"e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:32:32"
target triple = "i386-pc-win32"
	%struct.__block_descriptor = type { i32, i32 }
	%struct.__block_literal_generic = type { i8*, i32, i32, i8*,
%struct.__block_descriptor* }
@a = common global i32 0, align 4		; <i32*> [#uses=2]

define i32 @main(i32 %argc, i32* %argv) nounwind {
entry:
	call void @foo()
	ret i32 0
}

define void @foo() nounwind alwaysinline {
entry:
	%tmp = load i32* @a		; <i32> [#uses=1]
	%inc = add i32 %tmp, 1		; <i32> [#uses=1]
	store i32 %inc, i32* @a
	ret void
}

[test_out.ll]
; ModuleID = 'test_out.bc'
target datalayout =
"e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:32:32"
target triple = "i386-pc-win32"
	%struct.__block_descriptor = type { i32, i32 }
	%struct.__block_literal_generic = type { i8*, i32, i32, i8*,
%struct.__block_descriptor* }
@a = common global i32 0, align 4		; <i32*> [#uses=4]

define i32 @main(i32 %argc, i32* %argv) nounwind {
entry:
	%tmp.i = load i32* @a		; <i32> [#uses=1]
	%inc.i = add i32 %tmp.i, 1		; <i32> [#uses=1]
	store i32 %inc.i, i32* @a
	ret i32 0
}

define void @foo() nounwind alwaysinline {
entry:
	%tmp = load i32* @a		; <i32> [#uses=1]
	%inc = add i32 %tmp, 1		; <i32> [#uses=1]
	store i32 %inc, i32* @a
	ret void
}

For bonus points, if you want the definition of foo to be removed after it
has been inlined you have to add linkonce_odr to foo's llvm definition (see
below). I don't know how to get the same result from the .c file.
[test2.ll]
; ModuleID = 'test.c'
target datalayout =
"e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:32:32"
target triple = "i386-pc-win32"
	%struct.__block_descriptor = type { i32, i32 }
	%struct.__block_literal_generic = type { i8*, i32, i32, i8*,
%struct.__block_descriptor* }
@a = common global i32 0, align 4		; <i32*> [#uses=2]

define i32 @main(i32 %argc, i32* %argv) nounwind {
entry:
	call void @foo()
	ret i32 0
}

define linkonce_odr void @foo() nounwind alwaysinline {
entry:
	%tmp = load i32* @a		; <i32> [#uses=1]
	%inc = add i32 %tmp, 1		; <i32> [#uses=1]
	store i32 %inc, i32* @a
	ret void
}

Javier

On Thu, 17 Sep 2009 09:45:14 -0700 (PDT), madiyaan
<ahmadsharif at hotmail.com>
wrote:
> Thanks for your message. However, even with using version 2.5 of llvm the
> function is not inlined. Can you try this with your llvm with the same
opt
> flags, which are as follows?
> 
> -O3 -debug-only=inline -mem2reg
> 
> If it does work for you, can you tell me the exact flags you used for
clang
> or llvm-gcc to produce the .bc file and the exact flags you specified
when
> you invoked opt?
> 
> Does the inlining happen at clang level or after opt?
> 
> 
> Dale Johannesen wrote:
>> 
>> 
>> On Sep 16, 2009, at 6:55 PM, madiyaan wrote:
>> 
>>>
>>> I have the following code:
>>>
>>> static inline void foo() __attribute((always_inline));
>>>
>>> int a;
>>> static inline void foo()
>>> {
>>>  a++;
>>> }
>>>
>>> int main(int argc, char *argv[])
>>> {
>>>  foo();
>>>  return 0;
>>> }
>> 
>> This works fine in current sources.  You should upgrade; 2.5 has been  
>> out for a while and 2.6 will be soon.
>> 
>>> However, the code generated by llvm 2.4 toolchain does not inline this
>>> function. opt retains the function call. Here is the output when I  
>>> try to
>>> compile with -debug-only=inline:
>>>
>>> ..\..\..\win32\bin\win32\debug\opt.exe -O3 -debug-only=inline - 
>>> mem2reg -f -o
>>> test_opt.bc test.bc
>>> Inliner visiting SCC: foo: 0 call sites.
>>> Inliner visiting SCC: main: 1 call sites.
>>>    Inlining: cost=always, Call:        call void (...)* @foo()
>>> Inliner visiting SCC: INDIRECTNODE: 0 call sites.
>>>
>>> Here is the .ll file:
>>>
>>> ; ModuleID = 'test_opt.bc'
>>> target datalayout =
>>> "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32- 
>>> f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:32:32"
>>> target triple = "i386-pc-win32"
>>> @a = common global i32 0, align 4		; <i32*> [#uses=2]
>>>
>>> define i32 @main(i32 %argc, i8** %argv) nounwind {
>>> entry:
>>> 	tail call void (...)* @foo()
>>> 	ret i32 0
>>> }
>>>
>>> define internal void @foo(...) nounwind alwaysinline {
>>> entry:
>>> 	%tmp = load i32* @a		; <i32> [#uses=1]
>>> 	%inc = add i32 %tmp, 1		; <i32> [#uses=1]
>>> 	store i32 %inc, i32* @a
>>> 	ret void
>>> }
>>>
>>>
>>> What am I doing wrong here? Is there a way to force a function to be  
>>> inlined
>>> by opt?
>>>
>>> Best Regards,
>>> -- 
>>> View this message in context:
>>>
http://www.nabble.com/Forcing-function-inline-not-working-tp25483934p25483934.html
>>> Sent from the LLVM - Dev mailing list archive at Nabble.com.
>>>
>>> _______________________________________________
>>> LLVM Developers mailing list
>>> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>> 
>> _______________________________________________
>> LLVM Developers mailing list
>> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>> 
>>