Re: [sdcc-devel] killDeadCode() after computeLiveRanges () fails

[Erik P. <epe...@iv...>]

On Sat, 29 Jun 2013, Philipp Klaus Krause wrote:

> I want to use information on live ranges in lospre. To do that, I
> changed the old call to computeLiveRanges() to recomputeLiveRanges() and
> inserted the following code in line 2206 of SDCCopt.c:
>
> computeControlFlow (ebbi);
> loops = createLoopRegions (ebbi);
> computeDataFlow (ebbi);
> computeLiveRanges (ebbi->bbOrder, ebbi->count, TRUE);
>
> But then I get a fatal error on the call to killDeadCode() in in line
> 2218. What's the problem here?

computeLiveRanges() calls sequenceiCode() which assigns the sequence 
values to the individual iCodes as well as initializing the first and last 
sequence value for the eBBlock. After the call to lospre() is:

   ebbi = iCodeBreakDown (ic);

This generates a completely new set of eBBlock structs with the first and 
last sequence value for the block defaulting to 0. Now the data structures 
are in a state that remiCodeFromeBBlock() cannot handle; if the iCodes 
have been assigned sequence values, then it is expected that the eBBlocks 
will have them too. I think the easiest fix would be to clear the iCode 
sequence values when grouping them into new blocks in iCode2eBBlock() by 
inserting:

   ic->seq = 0; // around SDCCBBlock.c line 315

   loop->seq = 0; // around SDCCBBlock.c line 364

The alternative would be to have something update the first/last sequence 
values in the eBBlock from the first/last iCode sequence values, but this 
could run into problems if a new iCode had been inserted into the block 
without a proper sequence value assigned. This is why I think zeroing the 
iCode sequence values is preferable.

> I do not get that error when commenting out the line
> computeLiveRanges (ebbi->bbOrder, ebbi->count, TRUE);

Yes, then both the iCodes and the eBBlocks still have the default sequence 
value of zero and remiCodeFromeBBlock() is happy with this case.

   Erik