calculate_global_regs_live

[Eric Botcazou <ebotcazou at libertysurf dot fr>]

Hi,

Is it a feature or an oversight that calculate_global_regs_live doesn't clear 
the regsets containing the liveness information for each BB before calculing 
them again?

The algorithm is explained by the following comment:

  /* We work through the queue until there are no more blocks.  What
     is live at the end of this block is precisely the union of what
     is live at the beginning of all its successors.  So, we set its
     GLOBAL_LIVE_AT_END field based on the GLOBAL_LIVE_AT_START field
     for its successors.  Then, we compute GLOBAL_LIVE_AT_START for
     this block by walking through the instructions in this block in
     reverse order and updating as we go.  If that changed
     GLOBAL_LIVE_AT_START, we add the predecessors of the block to the
     queue; they will now need to recalculate GLOBAL_LIVE_AT_END.

     We are guaranteed to terminate, because GLOBAL_LIVE_AT_START
     never shrinks.  If a register appears in GLOBAL_LIVE_AT_START, it
     must either be live at the end of the block, or used within the
     block.  In the latter case, it will certainly never disappear
     from GLOBAL_LIVE_AT_START.  In the former case, the register
     could go away only if it disappeared from GLOBAL_LIVE_AT_START
     for one of the successor blocks.  By induction, that cannot
     occur.  */

If GLOBAL_LIVE_AT_START starts as non-zero, it can shrink and the reasoning 
above doesn't apply anymore.  I have a testcase which causes an infinite 
oscillation in the function because of this problem.

-- 
Eric Botcazou