This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH][rtl-optimization] Fix PR23726

David Edelsohn wrote on Dienstag, 6. September 2005 17:25 :
> 	This looks like a good catch, but do you really think it is a good
> idea to have every port modify its patterns to add REG_EQUAL notes when
> multiple results are produced.  Shouldn't CSE look inside the PARALLELs or
> the generated code for the PARALLEL automatically add REG_EQUAL notes?
> David
IMO we should distinguish two issues. The double-set issue and the 
"sequence_with_two_results" issue. I completely agree that CSE should look 
into the inside of parallels.
IMO, this will help much for individual insn that produce two results. It 
could not, however, make the REG_EQUAL notes obsolete for expanded 
multi-value sequences. The key difficulty I recognize is that it is 
frequently not possible to express the side effect sufficiently detailed for 
CSE, only by using the input operands of the last insn of the sequence that 
generates multiple results. E.g. If you think about expanding a minus:DI into 
one minus:SI and one minus_with_borrow:SI, you you will need the input 
operands of *both* insn in order to describe what value resides in the CCreg 
after the sequence. 
Also when one of the results of an expanded sequence does not end up in a 
pseudo but in a hard regs, CSE most probably will not find the common 
expression. I expect this to be true also for the case of condition code 
values that end up in hard regs although I have not verified this so far.



Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]