[LLVMdev] Re: __main() function and AliasSet
Chris Lattner
sabre at nondot.org
Sun May 14 13:07:35 PDT 2006
On Sun, 14 May 2006, Nai XIA wrote:
> Oh, I appologize that I should not have asked about __main() ---- it appears
> in FAQ.
> But the question remains that why call to __main() can alias stack location?
> I think the memory location pointed by data_X pointers are not visible to
> __main().
> In comparison, calls to printf() do not have similar effect.
First, some background: -steens-aa and -anders-aa work reasonable well,
but aren't production quality. In particular, they both assume that
printf doesn't have side effects, when (in fact) printf can on certain GNU
systems when the right format string is used. This is why they both think
that printf has no side effects: they special case it.
In practice, aliasing is a bunch of heuristics, and you cannot ever be
guaranteed to get an exact answer. As an example of this, both of these
passes are "context insensitive". As such, they don't know anything about
what the effect of a call is, so the call to __main (even though they
could theoretically know) is treated quite conservatively.
There are a couple of different options you have here. The alias passes
can be combined together, so something like this:
opt -globalsmodref-aa -steens-aa ...
should be able to tell that __main has no side effects. globalsmodref-aa
is a production quality pass that does some simple context sensitive
analysis (such as noticing functions with no side effects at all).
Another option is the -ds-aa pass. This pass is very powerful, but is
also the farthest from production quality. That said, it does get almost
all common things right, it just has some bugs in areas like variable
length arrays etc.
-Chris
> On 5/14/06, Nai Xia <nelson.xia at gmail.com> wrote:
>>
>> In a code segment of my pass plugin, I try to gather AliasSets for all
>> StoreInst, LoadInst and CallInst instructions in a function.
>> Some behaviors of the pass puzzled me.
>> Below is the *.ll of the test program which I run the pass on,
>> it was get with "llvm-gcc -Wl,--disable-opt" from a rather simple *.c
>> program.
>>
>> ----------------------------------
>> ; ModuleID = 'ptralias.bc'
>> target endian = little
>> target pointersize = 32
>> target triple = "i686-pc-linux-gnu"
>> deplibs = [ "c", "crtend" ]
>> %.str_1 = internal constant [25 x sbyte] c"ptra=0x ptrb=0x
>> ptrc=0x\0A\00" ; <[25 x sbyte]*> [#uses=1]
>> %ptr = weak global void ()* null ; <void ()**> [#uses=0]
>>
>> implementation ; Functions:
>>
>> declare int %printf(sbyte*, ...)
>>
>> void %foo1() {
>> ret void
>> }
>>
>> void %foo2() {
>> ret void
>> }
>>
>> int %main(int %argc, sbyte** %argv) {
>> entry:
>> %data_b = alloca int ; <int*> [#uses=2]
>> %data_c = alloca int ; <int*> [#uses=1]
>> %data_d = alloca int ; <int*> [#uses=3]
>> %data_e = alloca int ; <int*> [#uses=2]
>> %data_f = alloca int ; <int*> [#uses=2]
>> call void %__main( )
>> store int 2, int* %data_b
>> store int 3, int* %data_c
>> store int 4, int* %data_d
>> store int 5, int* %data_e
>> store int 6, int* %data_f
>> switch int %argc, label %switchexit [
>> int 3, label %label.3
>> int 2, label %then.2
>> int 1, label %label.1
>> int 0, label %endif.2
>> ]
>>
>> label.1: ; preds = %entry
>> br label %switchexit
>>
>> label.3: ; preds = %entry
>> br label %then.2
>>
>> switchexit: ; preds = %label.1, %entry
>> %ptr_b.0 = phi int* [ %data_d, %label.1 ], [ null, %entry ] ;
>> <int*> [#uses=1]
>> br label %endif.2
>>
>> then.2: ; preds = %label.3, %entry
>> %ptr_a.1.0 = phi int* [ %data_f, %label.3 ], [ %data_e, %entry
>> ] ; <int*> [#uses=1]
>> store int 0, int* %ptr_a.1.0
>> br label %then.3
>>
>> endif.2: ; preds = %switchexit, %entry
>> %ptr_b.0.1 = phi int* [ %ptr_b.0, %switchexit ], [ %data_b, %entry
>> ] ; <int*> [#uses=2]
>> %tmp.12 = seteq int* %ptr_b.0.1, null ; <bool> [#uses=1]
>> br bool %tmp.12, label %then.4, label %then.3
>>
>> then.3: ; preds = %endif.2, %then.2
>> %ptr_b.0.2 = phi int* [ %data_d, %then.2 ], [ %ptr_b.0.1, %endif.2
>> ] ; <int*> [#uses=1]
>> store int 0, int* %ptr_b.0.2
>> %tmp.1913 = call int (sbyte*, ...)* %printf( sbyte* getelementptr
>> ([25 x sbyte]* %.str_1, int 0, int 0) ) ; <int> [#uses=0]
>> ret int 0
>>
>> then.4: ; preds = %endif.2
>> %tmp.19 = call int (sbyte*, ...)* %printf( sbyte* getelementptr
>> ([25 x sbyte]* %.str_1, int 0, int 0) ) ; <int> [#uses=0]
>> ret int 0
>> }
>>
>> void %__main() {
>> entry:
>> ret void
>> }
>>
>> ----------------------------------
>> I think the right AliasSet information calculated for this program should
>> be
>>
>> Information for alias set0:
>> pointer0=data_b
>> pointer1=data_d
>> pointer2=ptr_b.0.2
>> Information for alias set1:
>> pointer0=data_c
>> Information for alias set2:
>> Information for alias set3:
>> pointer0=data_e
>> pointer1=data_f
>> pointer2=ptr_a.1.0
>> Information for alias set4:
>> Information for alias set5:
>>
>> ,where the empty AliasSets I think should be "Forwarded".
>>
>> However, the result of the pass was:
>>
>> Information for alias set0:
>> pointer0=data_b
>> pointer1=data_d
>> pointer2=data_e
>> pointer3=data_f
>> pointer4=ptr_a.1.0
>> pointer5=ptr_b.0.2
>> Information for alias set1:
>> pointer0=data_c
>> After I deleted "call void %__main( )" in %main(), the pass get the right
>> answer.
>>
>> So my question is:
>>
>> 1. What is the purpose for call to __main() ? __main() is just a empty
>> function. My .c program only contains main().
>> 2. My explanation for this behavior is that the CallSite of "call void
>> %__main( )" alias some pointers in the program,
>> however, __main() is obviously empty, why should the AliasAnalysis
>> think that it may/must Mod/Ref any stack location of main()?
>>
>> btw: the AliasAnalysis pass I used is -steens-aa and it also the same with
>> -anders-aa.
>>
>> --
>> Regards,
>> Nai
>>
>
-Chris
--
http://nondot.org/sabre/
http://llvm.org/
More information about the llvm-dev
mailing list