Re: Question about Gimple FE

[Yinsong Xue]

On 15/4/7 下午4:33, "Richard Biener"  wrote:

>I still like the idea of using C + extensions most.  As well as making the
>-fdump-tree-XXX dumps (more) valid C (+ extensions).  Cut & pasting
>from dump files to generate testcases is currently somewhat awkward,
>mainly due to the issue how we dump labels + goto destinations.
>
>Richard.

What is the issue all about specifically? 
Does that mean we have to translate labels and gotos to conditions and loops?

—
Yinsong

Re: Question about Gimple FE

[Yinsong Xue]

On 15/4/8 下午4:15, "Richard Biener"  wrote:

>No.  Currently we dump
>
>  :
>  # i_57 = PHI 
>  # ivtmp_60 = PHI 
>  _65 = (int) ratio_mult_vf.8_45;
>  tmp.9_64 = i_30 + _65;
>  tmp.10_66 = ivtmp_33 - ratio_mult_vf.8_45;
>  if (niters.6_41 == ratio_mult_vf.8_45)
>goto ;
>  else
>goto ;
>
>  :
>  # i_56 = PHI 
>  # ivtmp_59 = PHI 
>  goto ;
>
>and : isn't a valid label in C.  Dumping sth like _BB11:
>instead would make it a valid label
>and you need to do less editing to get it compile.
>
>Yes, PHIs are another story ;)
>
>Likewise names of temporary variables passes generate are not always
>valid identifiers (containing '.'s).
>
>IMHO trying to get -fdump-tree-cfg (dump after CFG build) emit mostly
>valid C would be nice (that is,
>before you go into SSA form and need to deal with PHIs).
>
>For SSA form we'd need to find a nice syntax for PHI nodes.  I doesn't
>need to be the same as used
>in debugging dumps.  Easy enough to parse would be using a function
>call, thus for bb12 above
>
>_BB12:
>  i_56 = __PHI (tmp$9_64, &_BB11, i_30, &_BB6);
>  ivtmp_59 = __PHI (tmp$10_66, &_BB11, ivtmp_33, &_BB6);
>  goto _BB9;
>
>possibly less conflicting would be to allow '11:' as label as
>extension so we can use literal 11 for
>the __PHI argument.
>
>The point of re-using (parts of) the C frontend is that you get types
>and decls for free.  The GIMPLE
>part is really the easy part of writing a gimple FE ;)
>
>You need to shortcut most of the C FEs specialities such as type
>promotion rules and stuff, of course.
>
>Richard.

That’s pretty much what I thought :)

A potential problem is, the dump format doesn’t contain enough 
Information (like user-defined types) to reconstruct the GIMPLE. 
I guess that’s why we’re considering LLVM’s IR. 

—
Yinsong