[cfe-dev] CFG temporary objects destructors

[Ted Kremenek]

Thanks John.

On Oct 22, 2010, at 4:17 PM, John McCall wrote:

> 
> On Oct 22, 2010, at 3:49 PM, Ted Kremenek wrote:
> 
>> 
>> On Oct 22, 2010, at 10:59 AM, Marcin Świderski wrote:
>> 
>>> W dniu 22 października 2010 11:22 użytkownik Zhongxing Xu <xuzhongxing at gmail.com> napisał:
>>> When the binary operator is logical operator, why we need to replicate the control flow when adding the dtors in LHS and RHS? Could we just add the dtor right after where the LHS and RHS are evaluated?
>>> 
>>> That is, for code
>>> 
>>> A && B
>>> 
>>> we generate CFG like this:
>>> 
>>> A && B
>>> ~A()
>>> |      \
>>> |       B
>>> |       ~B()
>>> |        |
>>> A && B
>>> 
>>> Since in logical binary operator, we only need the boolean value. The temporary object can be destroyed right after it is evaluated.
>>> 
>>> This depends on how much we want to simulate real control flow of expression. C++ standard states that destructors of temporaries should be called at the end of full expression and in reverse order of their construction. Your example does not satisfy this. It only guaranties that destructor for temporary will be called.
>> 
>> 
>> I think Zhongxing is right.  Here is what the compiler does:
> 
> Marcin is right;  the temporary is destroyed conditionally at the end of the full expression.  It just happens to be the case that that immediately follows the call to 'B::operator bool()' in your example.  You can see that easily in the following modification to your example:
> 
> void baz(bool);
> int test() {
>   foo(foo() || bar());
> }
> 
> test():                              ## @_Z4testv
> Leh_func_begin0:
> ## BB#0:                                ## %entry
> 	pushq	%rbp
> Ltmp0:
> 	movq	%rsp, %rbp
> Ltmp1:
> 	pushq	%r14
> 	pushq	%rbx
> 	subq	$16, %rsp
> Ltmp2:
> 	leaq	-24(%rbp), %rbx
> 	movq	%rbx, %rdi
> 	callq	foo()
> 	movq	%rbx, %rdi
> 	callq	A::operator bool()
> 	cmpb	$1, %al
> 	jne	LBB0_2
> ## BB#1:                                ## %lor.end.thread7
> 	movl	$1, %edi
> 	callq	baz(bool)
> 	jmp	LBB0_3
> LBB0_2:                                 ## %temp.cond-dtor.call
> 	leaq	-32(%rbp), %rbx
> 	movq	%rbx, %rdi
> 	callq	bar()
> 	movq	%rbx, %rdi
> 	callq	B::operator bool()
> 	movzbl	%al, %edi
> 	callq	baz(bool)
> 	movl	%eax, %r14d
> 	movq	%rbx, %rdi
> 	callq	B::~B()
> 	movl	%r14d, %eax
> LBB0_3:                                 ## %temp.cond-dtor.cont
> 	movl	%eax, %ebx
> 	leaq	-24(%rbp), %rdi
> 	callq	A::~A()
> 	movl	%ebx, %eax
> 	addq	$16, %rsp
> 	popq	%rbx
> 	popq	%r14
> 	popq	%rbp
> 	ret
> 
> Er, actually, the optimizer has helpfully cloned the call to 'baz' so as to obscure my point, but I think you get it.
> 
> John.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20101022/05c32f0e/attachment.html>