[llvm-commits] [llvm] r62338 - in /llvm/trunk: lib/Target/X86/X86ISelLowering.cpp test/CodeGen/X86/ret-addr.ll

[Bill Wendling]

They are there because of the "Frame Address" lowering. They should go  
away, but I haven't investigated that yet.

$ cat a.c

void *x() {
   return __builtin_frame_address(0);
}

void *y() {
   return __builtin_frame_address(1);
}

void *z() {
   return __builtin_frame_address(2);
}

$ llvm-gcc -S -o - a.c -Os

	.text
	.align	1,0x90
	.globl	_z
_z:
	pushl	%ebp
	movl	%esp, %ebp
	pushl	%ebp
	movl	(%ebp), %eax
	movl	(%eax), %eax
	popl	%ebp
	popl	%ebp
	ret


	.align	1,0x90
	.globl	_y
_y:
	pushl	%ebp
	movl	%esp, %ebp
	pushl	%ebp
	movl	(%ebp), %eax
	popl	%ebp
	popl	%ebp
	ret


	.align	1,0x90
	.globl	_x
_x:
	pushl	%ebp
	movl	%esp, %ebp
	pushl	%ebp
	movl	%ebp, %eax
	popl	%ebp
	popl	%ebp
	ret

On Jan 16, 2009, at 11:49 AM, Evan Cheng wrote:

> Where are the extra push and pop coming from?
>
> Evan
> On Jan 16, 2009, at 11:43 AM, Bill Wendling wrote:
>
>> For Depth > 1 (2 in this case), I get this from LLVM:
>>
>> 	.align	1,0x90
>> 	.globl	_h
>> _h:
>> 	pushl	%ebp
>> 	movl	%esp, %ebp
>> 	pushl	%ebp
>> 	movl	(%ebp), %eax
>> 	movl	(%eax), %eax
>> 	movl	4(%eax), %eax
>> 	popl	%ebp
>> 	popl	%ebp
>> 	ret
>>
>> And this from GCC:
>>
>> .globl _h
>> _h:
>> 	pushl	%ebp
>> 	movl	%esp, %ebp
>> 	movl	0(%ebp), %eax
>> 	movl	(%eax), %eax
>> 	movl	4(%eax), %eax
>> 	leave
>> 	ret
>>
>> Except for the extraneous push/pop, these are the same, right? Or  
>> am I
>> missing something?
>>
>> -bw
>>
>> On Fri, Jan 16, 2009 at 11:37 AM, Dale Johannesen <dalej at apple.com>  
>> wrote:
>>> This doesn't do the right thing for Depth > 1.  I don't think we  
>>> need
>>> to short term, but the test should be ==1 and a comment should
>>> indicate this.
>>>
>>> On Jan 16, 2009, at 11:25 AMPST, Bill Wendling wrote:
>>>>
>>>> +  if (Depth > 0) {
>>>> +    SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
>>>> +    SDValue Offset =
>>>> +      DAG.getConstant(TD->getPointerSize(),
>>>> +                      Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
>>>> +    return DAG.getLoad(getPointerTy(), DAG.getEntryNode(),
>>>> +                       DAG.getNode(ISD::ADD, getPointerTy(),
>>>> FrameAddr, Offset),
>>>> +                       NULL, 0);
>>>> +  }
>>>
>>> _______________________________________________
>>> llvm-commits mailing list
>>> llvm-commits at cs.uiuc.edu
>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>>>
>> _______________________________________________
>> llvm-commits mailing list
>> llvm-commits at cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>