DFA-scheduler

[Vladimir Makarov]

P. Pavan Kumar wrote:

>Respected sir,
>
> DFA pipeline hazard recognizer is used to avoid structural hazards.
>While intra block(basic block scheduling) scheduling how data hazards are
>avoided . DFA will take care of only structural hazards, but what about
>data hazards.
>
>First cycle mutiple pass scheduling algorithm tries different schedules
>and choose the best one.
>intra block scheduling reorders the instructions within the basic block by
>constructing DAG and based on LIST scheduling algorithm .
>If First cycle mutiple pass algorithm chooses the best schedule then why
>gcc needs intra block scheduling ?
>
The first cycle mutiple pass algorithm tries to increase number of insns 
issued on the same cycle while guaranting that the insn with the highest 
priority (the most important part of which is critical path length) will 
be issued on the same cycle.  It can result in issuing not highest 
priority insn first.  This algorithm is useful for VLIW processors.  For 
example, a VLIW processor (e.g. FRV) requires to put integer insns at 
the start of VLIW and fp ones at the end.  If fp insn has the highest 
priority, the scheduler without the first cycle mutiple pass algorithm 
will issue it first putting nops at the start of VLIW and the integer 
insns in the next VLIW.  It means additional cycle for executing the 
insns.  The scheduler with the algorithm will issue the insns on the 
same cycle (putting them in one VLIW). The algorithm was useful to speed 
up Itanium benchmarks.

You could read more about this in an article about DFA in proceedings of 
the first gcc summit (you can find them on gcc web site).

>Now gcc will do instruction scheduling beyond basic blocks i,e inter block
>scheduling (region scheduling) then why gcc needs intrablock scheduling?
>  
>
BB scheduling is used only after the RA for all targets except for 
Itanium (historically there were no practically interblock movements 
after the RA, and there are even few of them even before the RA). 
 Itanium port uses an extended block scheduling.  There is recent 
Maxim's Kuvyrkov's proposal to make ebb scheduling working for all other 
targets.

>what is command line option for enabling only intra block(with in basic
>block) scheduling. and diabling all remaining  intruction scheduling (
>trace, modulo etc).
>
>  
>
Please, read the documentation.  As I remember modulo scheduling and 
trace scheduling is not on by default.  Forcing BB scheduling before the 
RA can be achieved by prohibiting interblock movements (please see gcc 
documentation).