Improve converting between 128-bit modes that use the same format

  mprove converting between 128-bit modes that use the same format.

This patch improves the insns used for converting between two modes using
the 128-bit floating point format (i.e. converting between KFmode and TFmode if
-mabi=ieeelongdouble is used, and converting between IFmode and TFmode if
-mabi=ibmlongdouble is used).  The new insns have the correct insn type and
instruction length for the move involved.

Previously, the two different moves were lumped together (i.e. converting
between IEEE 128-bit and IEEE 128-bit was matched by the same insns as
converting between IBM 128-bit and IBM 128-bit moves).

I have tested this patch on the following systems:

    1)	LE Power10 using --with-cpu=power10 --with-long-double-format=ieee
    2)	LE Power10 using --with-cpu=power9  --with-long-double-format=ibm
    3)	LE Power10 using --with-cpu=power8  --with-long-double-format=ibm
    4)	LE Power10 using --with-cpu=power10 --with-long-double-format=ibm
    5)	LE Power9  using --with-cpu=power9  --with-long-double-format=ibm
    6)	BE Power8  using --with-cpu=power8  --with-long-double-format=ibm
    7)	BE Power8  using --with-cpu=power5  --with-long-double-format=ibm

There were no regressions in the build or in the tests.  On the power10 with
long double using the IEEE 128-bit format, pr105334.c now runs where it
previously failed.

Can I check this patch into the trunk?

Did we want to backport this to earlier GCC releases?

2022-08-17   Michael Meissner  <meissner@linux.ibm.com>

gcc/

	* config/rs6000/rs6000.md (IFKF): Delete.
	(IFKF_reg): Delete.
	(extendkfif2): New define_expand.
	(trunckfif2): New define_expand.
	(extend<mode>tf2_internal): Split into extendiftf2_internal and
	extendkftf2_internal.  Update the insns to use the correct insn type and
	length attributes based on whether the move uses IEEE 128-bit floating
	point or IBM 128-bit floating point type.
	(extendiftf2_internal): Likewise.
	(extendkftf2_internal): Likewise.
	(extendtf<mode>2_internal): Split into extendtfif2_internal and
	extendtfkf2_internal.  Update the insns to use the correct insn type and
	length attributes based on whether the move uses IEEE 128-bit floating e
	point or IBM 128-bit floating point type.
	(extendtfif2_internal): Likewise.
	(extendtfkf2_internal): Likewise.
---
 gcc/config/rs6000/rs6000.md | 94 +++++++++++++++++++++++++++++--------
 1 file changed, 74 insertions(+), 20 deletions(-)
  

Please do not send new patches as replies to other patches.

On Thu, Aug 18, 2022 at 05:48:29PM -0400, Michael Meissner wrote:
> mprove converting between 128-bit modes that use the same format.

You are missing some characters?  But this is an edited version of the
subject anyway.  Just don't do that (neither the copying or the
editing), it just confuses things.

Please factor this patch into more pieces, pieces that can be reviewed
more easily, pieces that change one thing only.

As is you are just rewriting the lot, and it is not an improvement at
all this way.  No doubt there are many good pieces in it, but mixed with
a non-trivial amount of bad pieces I cannot approve it.  It also isn't
clear at all what you want to do; piece by piece it is easier to
explain.

> -; Iterators for converting to/from TFmode
> -(define_mode_iterator IFKF [IF KF])

Yes, IFmode and KFmode have almost nothing in common.  Good to see this
go.  It would be even better if we would not use
rs6000_expand_float128_convert when not needed, either, and all this
would be just gone after expand.

> +(define_expand "extendkfif2"
> +  [(set (match_operand:IF 0 "gpc_reg_operand")
> +	(float_extend:IF (match_operand:KF 1 "gpc_reg_operand")))]
> +  "TARGET_FLOAT128_TYPE"
> +{
> +  rs6000_expand_float128_convert (operands[0], operands[1], false);
> +  DONE;
> +})

This does not belong here.

It really shouldn't *exist* at all: this is *not* a float_extend!  It is
not converting to a wider mode (as required!), but not even to a mode of
higher precision: both IFmode and KFmode can represent (finite, normal)
numbers the other can not.

But it certainly does not belong here in the middle of no-op moves.

> +(define_expand "trunckfif2"
> +  [(set (match_operand:IF 0 "gpc_reg_operand")
> +	(float_truncate:IF (match_operand:KF 1 "gpc_reg_operand")))]
> +  "TARGET_FLOAT128_TYPE"
> +{
> +  rs6000_expand_float128_convert (operands[0], operands[1], false);
> +  DONE;
> +})

I also would expect IBM128 instead of just IF.  This would simplify a
lot.  Why do you not use that, is there a reason?

> +;; Convert between KFmode and TFmode when -mabi=ieeelongdouble
> +(define_insn_and_split "*extendkftf2_internal"

Same for IEEE128.  And this isn't a conversion at all (it's a no-op
move), please don't confuse things by saying it is.

> +  [(set_attr "type" "two")
> +   (set_attr "num_insns" "2")])

Btw, that really should never be needed.  insn_type "two" already means
exactly that.

> +(define_insn_and_split "*extendtfif2_internal"
> +  [(set (match_operand:IF 0 "gpc_reg_operand" "=d,&d")
> +	(float_extend:IF
> +	 (match_operand:TF 1 "input_operand" "0,d")))]
> +   "FLOAT128_IBM_P (TFmode)"
> +  "#"
> +  "&& reload_completed"

Why would this ever need reload_completed?  It is a no-op move!


Segher
  

On Tue, Aug 23, 2022 at 04:13:45PM -0500, Segher Boessenkool wrote:
> Please do not send new patches as replies to other patches.

This was sent as a new patch.

> On Thu, Aug 18, 2022 at 05:48:29PM -0400, Michael Meissner wrote:
> > mprove converting between 128-bit modes that use the same format.
>
> You are missing some characters?  But this is an edited version of the
> subject anyway.  Just don't do that (neither the copying or the
> editing), it just confuses things.

That is the first line from the git commit, which git format-patch puts as the
subject.  I accidently deleted a few extra characters when trimming it down (I
remove the From:, etc. lines from the format-patch output).  But I can just
delete this line if desired.

> Please factor this patch into more pieces, pieces that can be reviewed
> more easily, pieces that change one thing only.
> 
> As is you are just rewriting the lot, and it is not an improvement at
> all this way.  No doubt there are many good pieces in it, but mixed with
> a non-trivial amount of bad pieces I cannot approve it.  It also isn't
> clear at all what you want to do; piece by piece it is easier to
> explain.
> 
> > -; Iterators for converting to/from TFmode
> > -(define_mode_iterator IFKF [IF KF])
> 
> Yes, IFmode and KFmode have almost nothing in common.  Good to see this
> go.  It would be even better if we would not use
> rs6000_expand_float128_convert when not needed, either, and all this
> would be just gone after expand.

I took a look at it, and I have a new version that only does the moves that are
NOPs, and it makes sure all of the functions called have the proper names.  I
will post it on Tuesday, as some of the machines that I use for testing are now
down for the US Labor Day weekend (they need to work on power infrastructure to
the lab the machines are in).

> 
> > +(define_expand "extendkfif2"
> > +  [(set (match_operand:IF 0 "gpc_reg_operand")
> > +	(float_extend:IF (match_operand:KF 1 "gpc_reg_operand")))]
> > +  "TARGET_FLOAT128_TYPE"
> > +{
> > +  rs6000_expand_float128_convert (operands[0], operands[1], false);
> > +  DONE;
> > +})
> 
> This does not belong here.
> 
> It really shouldn't *exist* at all: this is *not* a float_extend!  It is
> not converting to a wider mode (as required!), but not even to a mode of
> higher precision: both IFmode and KFmode can represent (finite, normal)
> numbers the other can not.

We know that TFmode (if -mabi=ieeelongdouble) and KFmode are the same, just
like TFmode (if -mabi=ibmlongdouble) and IFmode are the same.  But RTL does not
know that these modes use the same representation.  So to convert between them,
it needs to use either FLOAT_EXTEND or FLOAT_TRUNCATE, depending on which
precision each of the three modes have (i.e. rs6000-modes.h).  So you need
these conversions in RTL.

Unfortunately, you can't just use SUBREG before register allocation is done.
So I do define_insn_and_split to cover this.

> 
> But it certainly does not belong here in the middle of no-op moves.
> 
> > +(define_expand "trunckfif2"
> > +  [(set (match_operand:IF 0 "gpc_reg_operand")
> > +	(float_truncate:IF (match_operand:KF 1 "gpc_reg_operand")))]
> > +  "TARGET_FLOAT128_TYPE"
> > +{
> > +  rs6000_expand_float128_convert (operands[0], operands[1], false);
> > +  DONE;
> > +})
> 
> I also would expect IBM128 instead of just IF.  This would simplify a
> lot.  Why do you not use that, is there a reason?

If you use IBM128, you then need to create a mode_attr that for a given mode
gives the other mode.  Sure it can be done, but for the insns involved it was
just simpler to duplicate the insns.

So for example, for IBM floating point my current patches are:

	(define_insn_and_split "extendtfif2"
	  [(set (match_operand:IF 0 "gpc_reg_operand" "=wa,wa,r,r")
		(float_extend:IF
		 (match_operand:TF 1 "gpc_reg_operand" "0,wa,0,r")))]
	  "TARGET_HARD_FLOAT && TARGET_IBM128 && FLOAT128_IBM_P (TFmode)"
	  "#"
	  "&& reload_completed"
	  [(set (match_dup 0)
		(match_dup 2))]
	{
	  operands[2] = gen_lowpart (IFmode, operands[1]);
	}
	  [(set_attr "num_insns" "2")
	   (set_attr "length" "8")])

	(define_insn_and_split "extendiftf2"
	  [(set (match_operand:TF 0 "gpc_reg_operand" "=wa,wa,r,r")
		(float_extend:TF
		 (match_operand:IF 1 "gpc_reg_operand" "0,wa,0,r")))]
	  "TARGET_HARD_FLOAT && TARGET_IBM128 && FLOAT128_IBM_P (TFmode)"
	  "#"
	  "&& reload_completed"
	  [(set (match_dup 0)
		(match_dup 2))]
	{
	  operands[2] = gen_lowpart (TFmode, operands[1]);
	}
	  [(set_attr "num_insns" "2")
	   (set_attr "length" "8")])

You could rewrite that as:

	(define_mode_attr IBM128_other [(IF "TF") ("TF" "IF")])

	(define_insn_and_split "extend<mode><IBM128_other>2"
	  [(set (match_operand:IBM128 0 "gpc_reg_operand" "=wa,wa,r,r")
		(float_extend:IBM128
		 (match_operand:IBM128_othr 1 "gpc_reg_operand" "0,wa,0,r")))]
	  "TARGET_HARD_FLOAT && TARGET_IBM128 && FLOAT128_IBM_P (TFmode)"
	  "#"
	  "&& reload_completed"
	  [(set (match_dup 0)
		(match_dup 2))]
	{
	  operands[2] = gen_lowpart (TFmode, operands[1]);
	}
	  [(set_attr "num_insns" "2")
	   (set_attr "length" "8")])

But for the IEEE side, combining the two insns won't work, since going from
TFmode to KFmode will generate a FLOAT_TRUNCATE instead of a FLOAT_EXTEND.

You could then use a code iterator for both FLOAT_TRUNCATE and FLOAT_EXTEND.
Recognizing both forms might protect us in case some day somebody reorders the
precision within rs6000-modes.h.  But that is a lot of work to do for minimal
gain.

But I'm not convinced that it improves things.

> > +;; Convert between KFmode and TFmode when -mabi=ieeelongdouble
> > +(define_insn_and_split "*extendkftf2_internal"
> 
> Same for IEEE128.  And this isn't a conversion at all (it's a no-op
> move), please don't confuse things by saying it is.

That is how RTL goes between modes.  Without a whole lot of changes to the
machine independent portion of the compiler, I don't see anyway of avoiding the
converts.

You might say well always just use one mode for IEEE and one mode for IBM.
I've tried, and I couldn't get it to work.  Besides if we do this, we break
people using __attribute__((mode(...))) which in turn is used by glibc and
libstdc++ to get _Complex forms of the __float128 type.

> 
> > +  [(set_attr "type" "two")
> > +   (set_attr "num_insns" "2")])
> 
> Btw, that really should never be needed.  insn_type "two" already means
> exactly that.
> 
> > +(define_insn_and_split "*extendtfif2_internal"
> > +  [(set (match_operand:IF 0 "gpc_reg_operand" "=d,&d")
> > +	(float_extend:IF
> > +	 (match_operand:TF 1 "input_operand" "0,d")))]
> > +   "FLOAT128_IBM_P (TFmode)"
> > +  "#"
> > +  "&& reload_completed"
> 
> Why would this ever need reload_completed?  It is a no-op move!

Various predicates do not allow SUBREG's of these types before register
allocation.
  

On Fri, Sep 02, 2022 at 01:51:33PM -0400, Michael Meissner wrote:
> On Tue, Aug 23, 2022 at 04:13:45PM -0500, Segher Boessenkool wrote:
> > Please do not send new patches as replies to other patches.
> 
> This was sent as a new patch.

It probably was the partially copied to the mail body subject line that
confused me.  Sorry.

> > On Thu, Aug 18, 2022 at 05:48:29PM -0400, Michael Meissner wrote:
> > > mprove converting between 128-bit modes that use the same format.
> >
> > You are missing some characters?  But this is an edited version of the
> > subject anyway.  Just don't do that (neither the copying or the
> > editing), it just confuses things.
> 
> That is the first line from the git commit, which git format-patch puts as the
> subject.  I accidently deleted a few extra characters when trimming it down (I
> remove the From:, etc. lines from the format-patch output).  But I can just
> delete this line if desired.

Don't copy that line in the first place?  Don't copy *anything* in fact,
the way you copied stuff in that other patch (which is very hard to
review, so it will have to wait a bit more) makes it impossible to
apply the patch (which whould be just git-am).

> > Please factor this patch into more pieces, pieces that can be reviewed
> > more easily, pieces that change one thing only.

Please do this.  It is the biggest problem I have with most of your
patches: you seem to save up development of a week, and then send it out
as big omnibus patch an hour or two before my weekend.  This is not
ideal.

> > > +(define_expand "extendkfif2"
> > > +  [(set (match_operand:IF 0 "gpc_reg_operand")
> > > +	(float_extend:IF (match_operand:KF 1 "gpc_reg_operand")))]
> > > +  "TARGET_FLOAT128_TYPE"
> > > +{
> > > +  rs6000_expand_float128_convert (operands[0], operands[1], false);
> > > +  DONE;
> > > +})
> > 
> > This does not belong here.
> > 
> > It really shouldn't *exist* at all: this is *not* a float_extend!  It is
> > not converting to a wider mode (as required!), but not even to a mode of
> > higher precision: both IFmode and KFmode can represent (finite, normal)
> > numbers the other can not.
> 
> We know that TFmode (if -mabi=ieeelongdouble) and KFmode are the same, just
> like TFmode (if -mabi=ibmlongdouble) and IFmode are the same.  But RTL does not
> know that these modes use the same representation.  So to convert between them,
> it needs to use either FLOAT_EXTEND or FLOAT_TRUNCATE, depending on which
> precision each of the three modes have (i.e. rs6000-modes.h).  So you need
> these conversions in RTL.

You have "extends" also when it is to a lower precision!

Also, let me say this again: this needs to be solved.  We cannot make
real progress as long as we keep pretending every pait of floating point
modes can be ordered.

> Unfortunately, you can't just use SUBREG before register allocation is done.

Why not?  It isn't valid (in all cases) *after* RA, but before is fine.

But you do not want a subreg, you need a conversion.

> > But it certainly does not belong here in the middle of no-op moves.

This is still very true.

> So for example, for IBM floating point my current patches are:

> But for the IEEE side, combining the two insns won't work, since going from
> TFmode to KFmode will generate a FLOAT_TRUNCATE instead of a FLOAT_EXTEND.

Yes.  Please just fix the code depending on fundamentally wrong and
unworkable assumptions, instead of adding perpetually worse workarounds
that make things more rickety all the time.

When you added this IF/KF/TF ordering stuff first, I was amazed that it
worked as well as it did.  And this was perhaps the best we could do for
GCC 10, sure.  But the problem should have been fixed since, there is no
sane way forward without doing that.

> > > +;; Convert between KFmode and TFmode when -mabi=ieeelongdouble
> > > +(define_insn_and_split "*extendkftf2_internal"
> > 
> > Same for IEEE128.  And this isn't a conversion at all (it's a no-op
> > move), please don't confuse things by saying it is.
> 
> That is how RTL goes between modes.  Without a whole lot of changes to the
> machine independent portion of the compiler, I don't see anyway of avoiding the
> converts.

It is not a conversion, not in RTL either.  It changes mode, it doesn't
change values.  That is not the same thing.  It's a subreg (one of the
five overloaded meanings of subregs, but :-) )

> > > +(define_insn_and_split "*extendtfif2_internal"
> > > +  [(set (match_operand:IF 0 "gpc_reg_operand" "=d,&d")
> > > +	(float_extend:IF
> > > +	 (match_operand:TF 1 "input_operand" "0,d")))]
> > > +   "FLOAT128_IBM_P (TFmode)"
> > > +  "#"
> > > +  "&& reload_completed"
> > 
> > Why would this ever need reload_completed?  It is a no-op move!
> 
> Various predicates do not allow SUBREG's of these types before register
> allocation.

Modes, not types, and I do not understand what you mean.  Please show an
example?

Either way that is not a reason to do fully the wrong thing.  It might
make doing the right thing slightly more work, sure.


Segher
  

On Tue, Sep 06, 2022 at 05:22:11PM -0500, Segher Boessenkool wrote:
> Please do this.  It is the biggest problem I have with most of your
> patches: you seem to save up development of a week, and then send it out
> as big omnibus patch an hour or two before my weekend.  This is not
> ideal.

This is always going to be the case.  As I'm developing the larger patches,
there are usually at least 3 smaller problems wanting to get out.  I don't know
what these things are until I run into them.

With the long patch review cycle, I have to be working n patches ahead of what
is submitted just to keep busy.

I try my best to segreate the patches into smaller chunks, but at the end of
the day there are going to various changes to get to the larger goal.  I really
don't see any way around that.

Things don't always come in 3 line changes.  A lot of times I make the 3 line
change, and then I have to make several other changes to compensate for it
elsewhere.  If I submit smaller patches, things will break unless all of the
patches in the chain are committed.  And invariably you will then ask why this
particular change is needed, when it is needed by the next patch or the patch
after that.

> > > > +(define_expand "extendkfif2"
> > > > +  [(set (match_operand:IF 0 "gpc_reg_operand")
> > > > +	(float_extend:IF (match_operand:KF 1 "gpc_reg_operand")))]
> > > > +  "TARGET_FLOAT128_TYPE"
> > > > +{
> > > > +  rs6000_expand_float128_convert (operands[0], operands[1], false);
> > > > +  DONE;
> > > > +})
> > > 
> > > This does not belong here.
> > > 
> > > It really shouldn't *exist* at all: this is *not* a float_extend!  It is
> > > not converting to a wider mode (as required!), but not even to a mode of
> > > higher precision: both IFmode and KFmode can represent (finite, normal)
> > > numbers the other can not.
> > 
> > We know that TFmode (if -mabi=ieeelongdouble) and KFmode are the same, just
> > like TFmode (if -mabi=ibmlongdouble) and IFmode are the same.  But RTL does not
> > know that these modes use the same representation.  So to convert between them,
> > it needs to use either FLOAT_EXTEND or FLOAT_TRUNCATE, depending on which
> > precision each of the three modes have (i.e. rs6000-modes.h).  So you need
> > these conversions in RTL.
> 
> You have "extends" also when it is to a lower precision!
>
> Also, let me say this again: this needs to be solved.  We cannot make
> real progress as long as we keep pretending every pait of floating point
> modes can be ordered.

I have no plans to solve this.  I view it as an unsolvable problem that will
invariably lead to lots and lots of changes that will need buy in from the
other GCC developers.  If you can make the changes, fine.  But I view it as
work with what you have rather than trying to change the whole floating point
infrastructure within GCC.

> > Unfortunately, you can't just use SUBREG before register allocation is done.
> 
> Why not?  It isn't valid (in all cases) *after* RA, but before is fine.

I haven't tracked it down.  But I tend to think it is a whack-a-mole problem
where there will be 50+ small changes that are needed to get this to work.  And
then we get back to the issue that you need all of these changes.

> But you do not want a subreg, you need a conversion.
> 
> > > But it certainly does not belong here in the middle of no-op moves.
> 
> This is still very true.
> 
> > So for example, for IBM floating point my current patches are:
> 
> > But for the IEEE side, combining the two insns won't work, since going from
> > TFmode to KFmode will generate a FLOAT_TRUNCATE instead of a FLOAT_EXTEND.
> 
> Yes.  Please just fix the code depending on fundamentally wrong and
> unworkable assumptions, instead of adding perpetually worse workarounds
> that make things more rickety all the time.
> 
> When you added this IF/KF/TF ordering stuff first, I was amazed that it
> worked as well as it did.  And this was perhaps the best we could do for
> GCC 10, sure.  But the problem should have been fixed since, there is no
> sane way forward without doing that.

When I first looked at it, the number of hidden assumptions that I discovered
just in the changes I made was high.  I believe fundamentally that there will
be a lot more changes to do what you want.  And as I said, this will need a lot
of buy-in from other developers.
  

On Wed, Sep 07, 2022 at 04:25:49PM -0400, Michael Meissner wrote:
> On Tue, Sep 06, 2022 at 05:22:11PM -0500, Segher Boessenkool wrote:
> > Please do this.  It is the biggest problem I have with most of your
> > patches: you seem to save up development of a week, and then send it out
> > as big omnibus patch an hour or two before my weekend.  This is not
> > ideal.
> 
> This is always going to be the case.  As I'm developing the larger patches,
> there are usually at least 3 smaller problems wanting to get out.  I don't know
> what these things are until I run into them.

This is the only large problem.  Humongous, if you want.  It has been
_the_ problem since the beginning of the IEEE QP work.  None of this is
news to anyone.

It makes all other problems tens of times as hard to solve.

The other thing that would have made development way way way less work
in total, and much simpler, is supporting this on more systems.
Everything with VMX for example (with the exact same ABI as we have
now!), or just a la "nof" ("-msoft-float"), which is easy to so as well.
You end up testing more cases, "having" to test more cases, *but* there
are way fewer exceptions.  Anything downstream from us (essentially
everything, we are the C and C++ compilers after all) has to do a lot
of unnecessary work as well, because of this.

But the fundamental problem that IEEE QP and double-double can not be
ordered causes so many problems it is just not funny.  The *workarounds*
for it each cause a cascade of *other problems*, and there is no end in
sight.


Segher
  

I submitted a new patch that rewrites what this patch was trying to do.  I
didn't see the original version I submitted on September 8th, so I just
reposted it.

https://gcc.gnu.org/pipermail/gcc-patches/2022-September/601504.html