Copyright 1997, 1999, 2000, 2001, 2002 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 2.1 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; see the file COPYING.LIB. If not, write to
the Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
02110-1301, USA.
This directory contains mpn functions for 64-bit V9 SPARC
RELEVANT OPTIMIZATION ISSUES
Notation:
IANY = shift/add/sub/logical/sethi
IADDLOG = add/sub/logical/sethi
MEM = ld*/st*
FA = fadd*/fsub*/f*to*/fmov*
FM = fmul*
UltraSPARC can issue four instructions per cycle, with these restrictions:
* Two IANY instructions, but only one of these may be a shift. If there is a
shift and an IANY instruction, the shift must precede the IANY instruction.
* One FA.
* One FM.
* One branch.
* One MEM.
* IANY/IADDLOG/MEM must be insn 1, 2, or 3 in an issue bundle. Taken branches
should not be in slot 4, since that makes the delay insn come from separate
bundle.
* If two IANY/IADDLOG instructions are to be executed in the same cycle and one
of these is setting the condition codes, that instruction must be the second
one.
To summarize, ignoring branches, these are the bundles that can reach the peak
execution speed:
insn1 iany iany mem iany iany mem iany iany mem
insn2 iaddlog mem iany mem iaddlog iany mem iaddlog iany
insn3 mem iaddlog iaddlog fa fa fa fm fm fm
insn4 fa/fm fa/fm fa/fm fm fm fm fa fa fa
The 64-bit integer multiply instruction mulx takes from 5 cycles to 35 cycles,
depending on the position of the most significant bit of the first source
operand. When used for 32x32->64 multiplication, it needs 20 cycles.
Furthermore, it stalls the processor while executing. We stay away from that
instruction, and instead use floating-point operations.
Floating-point add and multiply units are fully pipelined. The latency for
UltraSPARC-1/2 is 3 cycles and for UltraSPARC-3 it is 4 cycles.
Integer conditional move instructions cannot dual-issue with other integer
instructions. No conditional move can issue 1-5 cycles after a load. (This
might have been fixed for UltraSPARC-3.)
The UltraSPARC-3 pipeline is very simular to he one of UltraSPARC-1/2 , but is
somewhat slower. Branches execute slower, and there may be other new stalls.
But integer multiply doesn't stall the entire CPU and also has a much lower
latency. But it's still not pipelined, and thus useless for our needs.
STATUS
* mpn_lshift, mpn_rshift: The current code runs at 2.0 cycles/limb on
UltraSPARC-1/2 and 2.65 on UltraSPARC-3. For UltraSPARC-1/2, the IEU0
functional unit is saturated with shifts.
* mpn_add_n, mpn_sub_n: The current code runs at 4 cycles/limb on
UltraSPARC-1/2 and 4.5 cycles/limb on UltraSPARC-3. The 4 instruction
recurrency is the speed limiter.
* mpn_addmul_1: The current code runs at 14 cycles/limb asymptotically on
UltraSPARC-1/2 and 17.5 cycles/limb on UltraSPARC-3. On UltraSPARC-1/2, the
code sustains 4 instructions/cycle. It might be possible to invent a better
way of summing the intermediate 49-bit operands, but it is unlikely that it
will save enough instructions to save an entire cycle.
The load-use of the u operand is not enough scheduled for good L2 cache
performance. The UltraSPARC-1/2 L1 cache is direct mapped, and since we use
temporary stack slots that will conflict with the u and r operands, we miss
to L2 very often. The load-use of the std/ldx pairs via the stack are
perhaps over-scheduled.
It would be possible to save two instructions: (1) The mov could be avoided
if the std/ldx were less scheduled. (2) The ldx of the r operand could be
split into two ld instructions, saving the shifts/masks.
It should be possible to reach 14 cycles/limb for UltraSPARC-3 if the fp
operations where rescheduled for this processor's 4-cycle latency.
* mpn_mul_1: The current code is a straightforward edit of the mpn_addmul_1
code. It would be possible to shave one or two cycles from it, with some
labour.
* mpn_submul_1: Simpleminded code just calling mpn_mul_1 + mpn_sub_n. This
means that it runs at 18 cycles/limb on UltraSPARC-1/2 and 23 cycles/limb on
UltraSPARC-3. It would be possible to either match the mpn_addmul_1
performance, or in the worst case use one more instruction group.
* US1/US2 cache conflict resolving. The direct mapped L1 date cache of US1/US2
is a problem for mul_1, addmul_1 (and a prospective submul_1). We should
allocate a larger cache area, and put the stack temp area in a place that
doesn't cause cache conflicts.
