gcc/gmp/mpn/powerpc64/README

Copyright 1999, 2000, 2001, 2003, 2004, 2005 Free Software Foundation, Inc.

This file is part of the GNU MP Library.

The GNU MP Library is free software; you can redistribute it and/or modify
it under the terms of the GNU Lesser General Public License as published by
the Free Software Foundation; either version 3 of the License, or (at your
option) any later version.

The GNU MP Library is distributed in the hope that it will be useful, but
WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public
License for more details.

You should have received a copy of the GNU Lesser General Public License
along with the GNU MP Library.  If not, see http://www.gnu.org/licenses/.



                    POWERPC-64 MPN SUBROUTINES


This directory contains mpn functions for 64-bit PowerPC chips.


CODE ORGANIZATION

	mpn/powerpc64          mode-neutral code
	mpn/powerpc64/mode32   code for mode32
	mpn/powerpc64/mode64   code for mode64


The mode32 and mode64 sub-directories contain code which is for use in the
respective chip mode, 32 or 64.  The top-level directory is code that's
unaffected by the mode.

The "adde" instruction is the main difference between mode32 and mode64.  It
operates on either on a 32-bit or 64-bit quantity according to the chip mode.
Other instructions have an operand size in their opcode and hence don't vary.



POWER3/PPC630 pipeline information:

Decoding is 4-way + branch and issue is 8-way with some out-of-order
capability.

Functional units:
LS1  - ld/st unit 1
LS2  - ld/st unit 2
FXU1 - integer unit 1, handles any simple integer instruction
FXU2 - integer unit 2, handles any simple integer instruction
FXU3 - integer unit 3, handles integer multiply and divide
FPU1 - floating-point unit 1
FPU2 - floating-point unit 2

Memory:		  Any two memory operations can issue, but memory subsystem
		  can sustain just one store per cycle.  No need for data
		  prefetch; the hardware has very sophisticated prefetch logic.
Simple integer:	  2 operations (such as add, rl*)
Integer multiply: 1 operation every 9th cycle worst case; exact timing depends
		  on 2nd operand's most significant bit position (10 bits per
		  cycle).  Multiply unit is not pipelined, only one multiply
		  operation in progress is allowed.
Integer divide:	  ?
Floating-point:	  Any plain 2 arithmetic instructions (such as fmul, fadd, and
		  fmadd), latency 4 cycles.
Floating-point divide:
		  ?
Floating-point square root:
		  ?

POWER3/PPC630 best possible times for the main loops:
shift:	      1.5 cycles limited by integer unit contention.
	      With 63 special loops, one for each shift count, we could
	      reduce the needed integer instructions to 2, which would
	      reduce the best possible time to 1 cycle.
add/sub:      1.5 cycles, limited by ld/st unit contention.
mul:	      18 cycles (average) unless floating-point operations are used,
	      but that would only help for multiplies of perhaps 10 and more
	      limbs.
addmul/submul:Same situation as for mul.


POWER4/PPC970 and POWER5 pipeline information:

This is a very odd pipeline, it is basically a VLIW masquerading as a plain
architecture.  Its issue rules are not made public, and since it is so weird,
it is very hard to figure out any useful information from experimentation.
An example:

  A well-aligned loop with nop's take 3, 4, 6, 7, ... cycles.
    3 cycles for  0,  1,  2,  3,  4,  5,  6,  7 nop's
    4 cycles for  8,  9, 10, 11, 12, 13, 14, 15 nop's
    6 cycles for 16, 17, 18, 19, 20, 21, 22, 23 nop's
    7 cycles for 24, 25, 26, 27 nop's
    8 cycles for 28, 29, 30, 31 nop's
    ... continues regularly


Functional units:
LS1  - ld/st unit 1
LS2  - ld/st unit 2
FXU1 - integer unit 1, handles any integer instruction
FXU2 - integer unit 2, handles any integer instruction
FPU1 - floating-point unit 1
FPU2 - floating-point unit 2

While this is one integer unit less than POWER3/PPC630, the remaining units
are more powerful; here they handle multiply and divide.

Memory:		  2 ld/st.  Stores go to the L2 cache, which can sustain just
		  one store per cycle.
		  L1 load latency: to gregs 3-4 cycles, to fregs 5-6 cycles.
		  Operations that modify the address register might be split
		  to use also a an integer issue slot.
Simple integer:	  2 operations every cycle, latency 2.
Integer multiply: 2 operations every 6th cycle, latency 7 cycles.
Integer divide:	  ?
Floating-point:	  Any plain 2 arithmetic instructions (such as fmul, fadd, and
		  fmadd), latency 6 cycles.
Floating-point divide:
		  ?
Floating-point square root:
		  ?


IDEAS

*mul_1: Handling one limb using mulld/mulhdu and two limbs using floating-
point operations should give performance of about 20 cycles for 3 limbs, or 7
cycles/limb.

We should probably split the single-limb operand in 32-bit chunks, and the
multi-limb operand in 16-bit chunks, allowing us to accumulate well in fp
registers.

Problem is to get 32-bit or 16-bit words to the fp registers.  Only 64-bit fp
memops copies bits without fiddling with them.  We might therefore need to
load to integer registers with zero extension, store as 64 bits into temp
space, and then load to fp regs.  Alternatively, load directly to fp space
and add well-chosen constants to get cancelation.  (Other part after given by
subsequent subtraction.)

Possible code mix for load-via-intregs variant:

lwz,std,lfd
fmadd,fmadd,fmul,fmul
fctidz,stfd,ld,fctidz,stfd,ld
add,adde
lwz,std,lfd
fmadd,fmadd,fmul,fmul
fctidz,stfd,ld,fctidz,stfd,ld
add,adde
srd,sld,add,adde,add,adde