Re: Current egcs, binutils and kernel (fwd)

Re: Current egcs, binutils and kernel (fwd)

[Geert Uytterhoeven]

---------- Forwarded message ----------
Date: Tue, 20 Apr 1999 13:15:41 +0200
From: Reinhard Nissl <rnissl@gmx.de>
To: Geert Uytterhoeven <Geert.Uytterhoeven@cs.kuleuven.ac.be>
Cc: "linux-apus@sunsite.auc.dk" <linux-apus@sunsite.auc.dk>
Subject: Re: Current egcs, binutils and kernel

Hi,

Geert Uytterhoeven wrote:

> On Wed, 14 Apr 1999, Reinhard Nissl wrote:
> > has anyone had success in compiling (egcs-1.1.2 and binutils-2.9.1.0.23)
> > the current APUS kernel with support for network block devices (nbd.c)?
> >
> > I get an undefined reference to __lshrdi3 from nbd_ioctl(), which looks
> > like a compiler / binutils bug.
>
> Hence a __lshrdi3() routine needs to be added to arch/ppc/kernel/misc.S.

I had a look into misc.S and found similar routines (__ashrdi3) there. Then I
searched in the egcs-1.1.2 sources for files, where such functions are
referenced. I found definitions in egcs-1.1.2/gcc/config/rs6000/rs6000.md but
they are not native ppc assembler instructions. As I'm not that much used to
*.md files and ppc assembly code, I'm currently not able to define the missing
function in misc.S myself.

I checked the kernel source diffs from version 2.2.4 to 2.2.6 for lshrdi3 and
had only success for arch=sparc. So, is there anybody who can add the missing
function to misc.S for arch=ppc?

> Greetings,
>
>                                                 Geert

Bye.
--
Dipl.-Inform. (FH) Reinhard Nissl
mailto:rnissl@gmx.de




[[ This message was sent via the linuxppc-dev mailing list.  Replies are ]]
[[ not  forced  back  to the list, so be sure to Cc linuxppc-dev if your ]]
[[ reply is of general interest. Please check http://lists.linuxppc.org/ ]]
[[ and http://www.linuxppc.org/ for useful information before posting.   ]]

Re: Current egcs, binutils and kernel (fwd)

[Franz Sirl]

At 13:39 20.04.99 , Geert Uytterhoeven wrote:

>---------- Forwarded message ----------
>Date: Tue, 20 Apr 1999 13:15:41 +0200
>From: Reinhard Nissl <rnissl@gmx.de>
>To: Geert Uytterhoeven <Geert.Uytterhoeven@cs.kuleuven.ac.be>
>Cc: "linux-apus@sunsite.auc.dk" <linux-apus@sunsite.auc.dk>
>Subject: Re: Current egcs, binutils and kernel
>
>Hi,
>
>Geert Uytterhoeven wrote:
>
>> On Wed, 14 Apr 1999, Reinhard Nissl wrote:
>> > has anyone had success in compiling (egcs-1.1.2 and binutils-2.9.1.0.23)
>> > the current APUS kernel with support for network block devices (nbd.c)?
>> >
>> > I get an undefined reference to __lshrdi3 from nbd_ioctl(), which looks
>> > like a compiler / binutils bug.
>>
>> Hence a __lshrdi3() routine needs to be added to arch/ppc/kernel/misc.S.
>
>I had a look into misc.S and found similar routines (__ashrdi3) there. Then I
>searched in the egcs-1.1.2 sources for files, where such functions are
>referenced. I found definitions in egcs-1.1.2/gcc/config/rs6000/rs6000.md but
>they are not native ppc assembler instructions. As I'm not that much used to
>*.md files and ppc assembly code, I'm currently not able to define the missing
>function in misc.S myself.
>
>I checked the kernel source diffs from version 2.2.4 to 2.2.6 for lshrdi3 and
>had only success for arch=sparc. So, is there anybody who can add the missing
>function to misc.S for arch=ppc?


objdump -D libgcc.a shows:

_lshrdi3.o:     file format elf32-powerpc

Disassembly of section .text:

00000000 <__lshrdi3>:
   0:   7c a5 2b 79     mr.     r5,r5
   4:   4d 82 00 20     beqlr
   8:   20 05 00 20     subfic  r0,r5,32
   c:   2c 00 00 00     cmpwi   r0,0
  10:   41 81 00 14     bgt     24 <__lshrdi3+0x24>
  14:   7c 00 00 d0     neg     r0,r0
  18:   39 60 00 00     li      r11,0
  1c:   7c 6c 04 30     srw     r12,r3,r0
  20:   48 00 00 14     b       34 <__lshrdi3+0x34>
  24:   7c 89 2c 30     srw     r9,r4,r5
  28:   7c 60 00 30     slw     r0,r3,r0
  2c:   7c 6b 2c 30     srw     r11,r3,r5
  30:   7d 2c 03 78     or      r12,r9,r0
  34:   7d 63 5b 78     mr      r3,r11
  38:   7d 84 63 78     mr      r4,r12
  3c:   4e 80 00 20     blr

__lshrdi3 is a logical right shift (0 is shifted in from the left, whereas 
__ashrdi3 arithmetic right shift shifts in the sign bit) on a 64-bit 
quantitity. You could also use the example code in Appendix E of the PPC601 
manual.

Franz.


[[ This message was sent via the linuxppc-dev mailing list.  Replies are ]]
[[ not  forced  back  to the list, so be sure to Cc linuxppc-dev if your ]]
[[ reply is of general interest. Please check http://lists.linuxppc.org/ ]]
[[ and http://www.linuxppc.org/ for useful information before posting.   ]]

Re: Current egcs, binutils and kernel (fwd)

[David Edelsohn]

	The original POWER architecture with the MQ register allowed a
short sequence to provide this 64-bit functionality.  The PowerPC
instruction set does not provide the necessary building blocks, so GCC
falls back to the C implementation in libgcc2.c producing the assembly
code shown in the later posting.

David

[[ This message was sent via the linuxppc-dev mailing list.  Replies are ]]
[[ not  forced  back  to the list, so be sure to Cc linuxppc-dev if your ]]
[[ reply is of general interest. Please check http://lists.linuxppc.org/ ]]
[[ and http://www.linuxppc.org/ for useful information before posting.   ]]

Re: Current egcs, binutils and kernel (fwd)

[Gabriel Paubert]

On Tue, 20 Apr 1999, Geert Uytterhoeven wrote:

> 
> ---------- Forwarded message ----------
> Date: Tue, 20 Apr 1999 13:15:41 +0200
> From: Reinhard Nissl <rnissl@gmx.de>
> To: Geert Uytterhoeven <Geert.Uytterhoeven@cs.kuleuven.ac.be>
> Cc: "linux-apus@sunsite.auc.dk" <linux-apus@sunsite.auc.dk>
> Subject: Re: Current egcs, binutils and kernel
> 
> Hi,
> 
> Geert Uytterhoeven wrote:
> 
> > On Wed, 14 Apr 1999, Reinhard Nissl wrote:
> > > has anyone had success in compiling (egcs-1.1.2 and binutils-2.9.1.0.23)
> > > the current APUS kernel with support for network block devices (nbd.c)?
> > >
> > > I get an undefined reference to __lshrdi3 from nbd_ioctl(), which looks
> > > like a compiler / binutils bug.
> >
> > Hence a __lshrdi3() routine needs to be added to arch/ppc/kernel/misc.S.
> 
> I had a look into misc.S and found similar routines (__ashrdi3) there. Then I
> searched in the egcs-1.1.2 sources for files, where such functions are
> referenced. I found definitions in egcs-1.1.2/gcc/config/rs6000/rs6000.md but
> they are not native ppc assembler instructions. As I'm not that much used to
> *.md files and ppc assembly code, I'm currently not able to define the missing
> function in misc.S myself.
> 
> I checked the kernel source diffs from version 2.2.4 to 2.2.6 for lshrdi3 and
> had only success for arch=sparc. So, is there anybody who can add the missing
> function to misc.S for arch=ppc?

I'd suggest the following patch. Note that the current versions of the
long long shifts will not work when the shift count is > 32. There is an
appendix in all good PPC manuals on how to do multiple precision shifts
and I've followed it (except for the exact order for better superscalar
issue/execution and completion, all the code should flow perfectly
through 2 pipes) with one exception: the arithmetic right
shift is one instruction longer but is branchless (conditional clear 
of a register using a shift whose amount is computed by an rlwinm
instruction).

I've also fixed a few other oddities in the code: 

- atomic_dec_and_test uses cntlzw the way God intended to evaluate 
`(x==0) ? 1 : 0' without any branch

- the abs function is also branchless now (it would nevertheless be better
to use the __builtin_abs function of GCC)

	Greetings,
	Gabriel.

--- linux-2.2.6/arch/ppc/kernel/misc.S	Thu Mar 11 05:30:32 1999
+++ linux/arch/ppc/kernel/misc.S	Tue Apr 20 20:14:03 1999
@@ -228,10 +228,8 @@
 	subi	r5,r5,1		/* Perform 'add' operation */
 	stwcx.	r5,0,r3		/* Update with new value */
 	bne-	10b		/* Retry if "reservation" (i.e. lock) lost */
-	cmpi	0,r5,0		/* Return 'true' IFF 0 */
-	li	r3,1
-	beqlr
-	li	r3,0
+	cntlzw	r3,r5		/* Return 'true' IFF 0 */
+	srwi	r3,r3,5		/* But do it the clever way */
 	blr
 _GLOBAL(atomic_clear_mask)
 10:	lwarx	r5,0,r4
@@ -355,38 +353,59 @@
 	blr	
 
 /*
- * Extended precision shifts
+ * Extended precision shifts.
+ * 
+ * Updated to be valid for shift counts from 0 to 63 inclusive.
+ * -- Gabriel 
  *
  * R3/R4 has 64 bit value
  * R5    has shift count
  * result in R3/R4
  *
- *  ashrdi3:     XXXYYY/ZZZAAA -> SSSXXX/YYYZZZ
- *  ashldi3:     XXXYYY/ZZZAAA -> YYYZZZ/AAA000
+ *  ashrdi3: arithmetic right shift (sign propagation)     
+ *  lslhdi3: logical right shift	
+ *  ashldi3: left shift
  */
 _GLOBAL(__ashrdi3)
-	li	r6,32
-	sub	r6,r6,r5
-	slw	r7,r3,r6	/* isolate YYY */
-	srw	r4,r4,r5	/* isolate ZZZ */
-	or	r4,r4,r7	/* YYYZZZ */
-	sraw	r3,r3,r5	/* SSSXXX */
+	subfic	r6,r5,32	
+	srw	r4,r4,r5	# LSW = count > 31 ? 0 : LSW >> count
+	addi	r7,r5,32	# could be xori, or addi with -32
+	slw	r6,r3,r6	# t1 = count > 31 ? 0 :	MSW << (32-count)
+	rlwinm	r8,r7,0,32	# t3 = (count < 32) ? 32 : 0
+	sraw	r7,r3,r7	# t2 = MSW >> (count-32)
+	or	r4,r4,r6	# LSW |= t1
+	slw	r7,r7,r8	# t2 = (count < 32) ? 0 : t2
+	sraw	r3,r3,r5	# MSW = MSW >> count
+	or	r4,r4,r7	# LSW |= t2
 	blr
-	
+
 _GLOBAL(__ashldi3)
-	li	r6,32
-	sub	r6,r6,r5
-	srw	r7,r4,r6	/* isolate ZZZ */
-	slw	r4,r4,r5	/* AAA000 */
-	slw	r3,r3,r5	/* YYY--- */
-	or	r3,r3,r7	/* YYYZZZ */
+	subfic	r6,r5,32	
+	slw	r3,r3,r5	# MSW = count > 31 ? 0 : MSW << count
+	addi	r7,r5,32	# could be xori, or addi with -32
+	srw	r6,r4,r6	# t1 = count > 31 ? 0 :	LSW >> (32-count)
+	slw	r7,r4,r7	# t2 = count < 32 ? 0 :	LSW << (count-32)
+	or	r3,r3,r6	# MSW |= t1
+	slw	r4,r4,r5	# LSW = LSW << count
+	or	r3,r3,r7	# MSW |= t2
+	blr
+
+_GLOBAL(__lshrdi3)
+	subfic	r6,r5,32	
+	srw	r4,r4,r5	# LSW = count > 31 ? 0 : LSW >> count
+	addi	r7,r5,32	# could be xori, or addi with -32
+	slw	r6,r3,r6	# t1 = count > 31 ? 0 :	MSW << (32-count)
+	srw	r7,r3,r7	# t2 = count < 32 ? 0 :	MSW >> (count-32)
+	or	r4,r4,r6	# LSW |= t1
+	srw	r3,r3,r5	# MSW = MSW >> count
+	or	r4,r4,r7	# LSW |= t2 
 	blr
 
 _GLOBAL(abs)
-	cmpi	0,r3,0
-	bge	10f
-	neg	r3,r3
-10:	blr
+	srawi	r4,r3,31
+	xor	r3,r3,r4
+	sub	r3,r3,r4
+	blr
 
 _GLOBAL(_get_SP)
 	mr	r3,r1		/* Close enough */



[[ This message was sent via the linuxppc-dev mailing list.  Replies are ]]
[[ not  forced  back  to the list, so be sure to Cc linuxppc-dev if your ]]
[[ reply is of general interest. Please check http://lists.linuxppc.org/ ]]
[[ and http://www.linuxppc.org/ for useful information before posting.   ]]

Re: Current egcs, binutils and kernel (fwd)

[Gary Thomas]

I looked at the code in "misc.S" and it appears to me [by cursory glance]
that it already is "__lshrdi3" - i.e. the code that is there is doing a
logical shift and not an arithmetic one!

It seems to me that you could just add "__lshrdi3" as an alternate entry 
point.

I need to investigate the existing code and see if maybe my impression
is just "day's end..."

On 20-Apr-99 Geert Uytterhoeven wrote:
> 
> ---------- Forwarded message ----------
> Date: Tue, 20 Apr 1999 13:15:41 +0200
> From: Reinhard Nissl <rnissl@gmx.de>
> To: Geert Uytterhoeven <Geert.Uytterhoeven@cs.kuleuven.ac.be>
> Cc: "linux-apus@sunsite.auc.dk" <linux-apus@sunsite.auc.dk>
> Subject: Re: Current egcs, binutils and kernel
> 
> Hi,
> 
> Geert Uytterhoeven wrote:
> 
>> On Wed, 14 Apr 1999, Reinhard Nissl wrote:
>> > has anyone had success in compiling (egcs-1.1.2 and binutils-2.9.1.0.23)
>> > the current APUS kernel with support for network block devices (nbd.c)?
>> >
>> > I get an undefined reference to __lshrdi3 from nbd_ioctl(), which looks
>> > like a compiler / binutils bug.
>>
>> Hence a __lshrdi3() routine needs to be added to arch/ppc/kernel/misc.S.
> 
> I had a look into misc.S and found similar routines (__ashrdi3) there. Then I
> searched in the egcs-1.1.2 sources for files, where such functions are
> referenced. I found definitions in egcs-1.1.2/gcc/config/rs6000/rs6000.md but
> they are not native ppc assembler instructions. As I'm not that much used to
> *.md files and ppc assembly code, I'm currently not able to define the missing
> function in misc.S myself.
> 
> I checked the kernel source diffs from version 2.2.4 to 2.2.6 for lshrdi3 and
> had only success for arch=sparc. So, is there anybody who can add the missing
> function to misc.S for arch=ppc?
> 
>> Greetings,
>>
>>                                                 Geert
> 
> Bye.
> --
> Dipl.-Inform. (FH) Reinhard Nissl
> mailto:rnissl@gmx.de
> 
> 
> 

[[ This message was sent via the linuxppc-dev mailing list.  Replies are ]]
[[ not  forced  back  to the list, so be sure to Cc linuxppc-dev if your ]]
[[ reply is of general interest. Please check http://lists.linuxppc.org/ ]]
[[ and http://www.linuxppc.org/ for useful information before posting.   ]]

Re: Current egcs, binutils and kernel (fwd)

[Gabriel Paubert]

On Tue, 20 Apr 1999, David Edelsohn wrote:

> 
> 	The original POWER architecture with the MQ register allowed a
> short sequence to provide this 64-bit functionality.  The PowerPC
> instruction set does not provide the necessary building blocks, so GCC
> falls back to the C implementation in libgcc2.c producing the assembly
> code shown in the later posting.

Yes, but the kernel is not linked with libgcc2. Anyway I've posted a patch
to correctly implement this. And then I've noticed just after that people
do more work than necessary in the cvt_df and cvt_fd routines: floating
point loads and stores never affect or depend on the FPSCR contents, so
there is no point at all in manipulating it (except adding bloat for the
sake of adding bloat).

	Gabriel.


[[ This message was sent via the linuxppc-dev mailing list.  Replies are ]]
[[ not  forced  back  to the list, so be sure to Cc linuxppc-dev if your ]]
[[ reply is of general interest. Please check http://lists.linuxppc.org/ ]]
[[ and http://www.linuxppc.org/ for useful information before posting.   ]]

Re: Current egcs, binutils and kernel (fwd)

[David Edelsohn]

>>>>> Gabriel Paubert writes:

David> The original POWER architecture with the MQ register allowed a
David> short sequence to provide this 64-bit functionality.  The PowerPC
David> instruction set does not provide the necessary building blocks, so GCC
David> falls back to the C implementation in libgcc2.c producing the assembly
David> code shown in the later posting.

Gabriel> Yes, but the kernel is not linked with libgcc2. Anyway I've posted a patch
Gabriel> to correctly implement this. And then I've noticed just after that people
Gabriel> do more work than necessary in the cvt_df and cvt_fd routines: floating
Gabriel> point loads and stores never affect or depend on the FPSCR contents, so
Gabriel> there is no point at all in manipulating it (except adding bloat for the
Gabriel> sake of adding bloat).

	The purpose of my earlier message was to explain what was
occurring in rs6000.md and why GCC was calling the function instead of
inlining the code.  I thought that some people might actually want to
understand *WHY* the function was required and why GCC was behaving as
designed. 

David

[[ This message was sent via the linuxppc-dev mailing list.  Replies are ]]
[[ not  forced  back  to the list, so be sure to Cc linuxppc-dev if your ]]
[[ reply is of general interest. Please check http://lists.linuxppc.org/ ]]
[[ and http://www.linuxppc.org/ for useful information before posting.   ]]

Re: Current egcs, binutils and kernel (fwd)

[Gabriel Paubert]

On Tue, 20 Apr 1999, David Edelsohn wrote:

> 	The purpose of my earlier message was to explain what was
> occurring in rs6000.md and why GCC was calling the function instead of
> inlining the code.  I thought that some people might actually want to
> understand *WHY* the function was required and why GCC was behaving as
> designed. 

Ok, calm down. But I still don't know why GCC does not expand this inline. 
It's 8 or 10 instructions, without branches so it's not much bigger than a
call for which you have to set up 3 parameter registers and which may
destroy many more registers according to the ABI conventions.
Furthermore if only one of the 2 result registers is needed, it may be
simplified quite significantly (i'm not sure it's easy to do in current
GCC to split a variable using 2 registers to remove dead code).

Maybe it's time to add the alternatives in rs6000.md, enabled with a
-minline-longlongshifts option. But I don't yet know enough of GCC to 
implement it.

	Gabriel. 


[[ This message was sent via the linuxppc-dev mailing list.  Replies are ]]
[[ not  forced  back  to the list, so be sure to Cc linuxppc-dev if your ]]
[[ reply is of general interest. Please check http://lists.linuxppc.org/ ]]
[[ and http://www.linuxppc.org/ for useful information before posting.   ]]

Re: Current egcs, binutils and kernel (fwd)

[Paul Mackerras]

Gabriel Paubert <paubert@iram.es> wrote:

> to correctly implement this. And then I've noticed just after that people
> do more work than necessary in the cvt_df and cvt_fd routines: floating
> point loads and stores never affect or depend on the FPSCR contents, so

Arrghl, you're right.  I had thought you could get overflows or other
errors being signalled, that's why I had the fpscr load/save in there,
but it appears from the manual that I was wrong.

Paul.

[[ This message was sent via the linuxppc-dev mailing list.  Replies are ]]
[[ not  forced  back  to the list, so be sure to Cc linuxppc-dev if your ]]
[[ reply is of general interest. Please check http://lists.linuxppc.org/ ]]
[[ and http://www.linuxppc.org/ for useful information before posting.   ]]