[LLVMdev] Alias analysis issue with structs on PPC

[Daniel Berlin]

On Fri, Mar 13, 2015 at 2:39 PM Olivier H Sallenave <ohsallen at us.ibm.com>
wrote:

> Hi,
>
> I have the following C loop to vectorize:
>
> struct box {
>     double* source;
> };
>
> void test(double* restrict result, struct box my_struct, int len)
> {
>     for (int i=0 ; i<len; i++) {
>         result[i] = my_struct.source[i] * my_struct.source[i];
>     }
> }
>
> There are two references in the loop, result[i] (restrict) and
> my_struct.source[i] (readonly). The compiler should easily figure out that
> they do not alias.
>
> Compiling for x86, the loop alias analysis works just fine:
>   AST: Alias Set Tracker: 2 alias sets for 2 pointer values.
>   AliasSet[0x7fd8e2f32290, 1] must alias, No access Pointers: (double*
> %arrayidx5, 18446744073709551615)
>   AliasSet[0x7fd8e2f322e0, 1] must alias, No access Pointers: (double*
> %arrayidx, 18446744073709551615)
>
> Compiling for PPC with -target powerpc64le-ibm-linux-gnu, the two
> addresses now alias:
>   AST: Alias Set Tracker: 1 alias sets for 2 pointer values.
>   AliasSet[0x7f931bd5bdc0, 2] may alias, No access Pointers: (double*
> %arrayidx5, 18446744073709551615), (double* %arrayidx, 18446744073709551615)
>
> BasicAA is used for both targets by default. The difference is that in
> PPC, the IR obtained from Clang takes an i64 as parameter instead of a
> double* for my_struct.
>

I don't even want to know why this would be the case :)


> This parameter is then coerced into double* using an inttoptr instruction.
> The code in BasicAliasAnalysis.cpp which is triggered for x86 is the
> following:
>
>     // Function arguments can't alias with things that are known to be
>     // unambigously identified at the function level.
>     if ((isa<Argument>(O1) && isIdentifiedFunctionLocal(O2)) ||
>         (isa<Argument>(O2) && isIdentifiedFunctionLocal(O1)))
>       return NoAlias;
>
> isIdentifiedFunctionLocal(V) returns true for a noalias argument (such as
> result), but the other address (my_struct) must be a function argument in
> order to return NoAlias, which is not the case anymore for PPC (since
> my_struct is now the result from an inttoptr instruction). If I understand,
> the problem is that we cannot trust the fact that locals do not alias with
> restrict parameters (because the compiler could generate some locals which
> alias)?
>
Yes, because pointers *based on* the noalias'd argument are legal aliases.

So if you don't know it's an argument or an identified local, it could be
based on the restricted pointer, and thus, alias it.


If someone has suggestions about this, that would help a lot.
>

The only way you could prove something in this case would be to walk the
chain and prove the value comes directly from an argument with no
modification.

That is expensive to do in the general case, and inttoptr is generally not
going to be wonderful for performance (for example, outside of BasicAA,
more advanced AA's like CFL-AA will give up on anything that comes from or
goes through ptrtoint/inttoptr). So what's the issue that makes you pass
this as i64 in the first place?
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150313/46717bad/attachment.html>