Re: [Combine] Unusual behaviour in combine

[Segher Boessenkool <segher at kernel dot crashing dot org>]

On Mon, Jun 17, 2019 at 04:47:29PM +0000, Joel Hutton wrote:
> Hi Segher,
> 
>  > If you want the const_vector for r97, you should look at the combination
>  > that tries *that* insn together with 10 and 12.  Did it try that?  
> Why not?
> 
> It did not attempt to combine all three, I'm not sure why not, I would 
> have expected
> it to, they are all in the same basic block.

I'll need some compilable source code to investigate this.  That means
compilable for *me*, so with some target that is in trunk.

> I don't understand why in the other example
> two insns were combined using a constant set by a third insn, but in 
> this case three
> insns must be combined, perhaps because in this example they are in the
> same basic block and in the other example the constant is set outside 
> the basic block?

In this case you are *not* combining three insns: you would be "combining"
*one* instruction.  It would be changing it to try to set its destination
reg to what is said in the REG_EQUIV some other way than using the insns
that set it currently.  And combine can already do that, when it tries to
combine at least two of the insns.

> insn_cost 4 for    33: r100:DI=x0:DI
>       REG_DEAD x0:DI
> insn_cost 4 for     2: r94:DI=r100:DI
>       REG_DEAD r100:DI
> insn_cost 4 for    34: r101:DI=x1:DI
>       REG_DEAD x1:DI
> insn_cost 4 for     3: r95:DI=r101:DI
>       REG_DEAD r101:DI
> insn_cost 4 for     5: r91:DI=0
> insn_cost 4 for    10: r98:V4SF=const_vector
> insn_cost 8 for     9: r96:V4SF=[r94:DI+r91:DI]
> insn_cost 8 for    11: r97:V4SF=r96:V4SF*r98:V4SF
>       REG_DEAD r96:V4SF
>       REG_EQUAL r96:V4SF*const_vector

So this insn 5 makes it hard to combine with insn 9, because you'll quickly
need a four insn combination, and those are restricted somewhat.  But 10+11
should do something.  But it seems it isn't tried at all?  Huh.


Segher