Copyright 1996, 1997, 1999-2005 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 either: * 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. or * the GNU General Public License as published by the Free Software Foundation; either version 2 of the License, or (at your option) any later version. or both in parallel, as here. 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 General Public License for more details. You should have received copies of the GNU General Public License and the GNU Lesser General Public License along with the GNU MP Library. If not, see https://www.gnu.org/licenses/. This directory contains mpn functions optimized for DEC Alpha processors. ALPHA ASSEMBLY RULES AND REGULATIONS The `.prologue N' pseudo op marks the end of instruction that needs special handling by unwinding. It also says whether $27 is really needed for computing the gp. The `.mask M' pseudo op says which registers are saved on the stack, and at what offset in the frame. Cray T3 code is very very different... "$6" / "$f6" etc is the usual syntax for registers, but on Unicos instead "r6" / "f6" is required. We use the "r6" / "f6" forms, and have m4 defines expand them to "$6" or "$f6" where necessary. "0x" introduces a hex constant in gas and DEC as, but on Unicos "^X" is required. The X() macro accommodates this difference. "cvttqc" is required by DEC as, "cvttq/c" is required by Unicos, and gas will accept either. We use cvttqc and have an m4 define expand to cvttq/c where necessary. "not" as an alias for "ornot r31, ..." is available in gas and DEC as, but not the Unicos assembler. The full "ornot" must be used. "unop" is not available in Unicos. We make an m4 define to the usual "ldq_u r31,0(r30)", and in fact use that define on all systems since it comes out the same. "!literal!123" etc explicit relocations as per Tru64 4.0 are apparently not available in older alpha assemblers (including gas prior to 2.12), according to the GCC manual, so the assembler macro forms must be used (eg. ldgp). RELEVANT OPTIMIZATION ISSUES EV4 1. This chip has very limited store bandwidth. The on-chip L1 cache is write- through, and a cache line is transferred from the store buffer to the off- chip L2 in as much 15 cycles on most systems. This delay hurts mpn_add_n, mpn_sub_n, mpn_lshift, and mpn_rshift. 2. Pairing is possible between memory instructions and integer arithmetic instructions. 3. mulq and umulh are documented to have a latency of 23 cycles, but 2 of these cycles are pipelined. Thus, multiply instructions can be issued at a rate of one each 21st cycle. EV5 1. The memory bandwidth of this chip is good, both for loads and stores. The L1 cache can handle two loads or one store per cycle, but two cycles after a store, no ld can issue. 2. mulq has a latency of 12 cycles and an issue rate of 1 each 8th cycle. umulh has a latency of 14 cycles and an issue rate of 1 each 10th cycle. (Note that published documentation gets these numbers slightly wrong.) 3. mpn_add_n. With 4-fold unrolling, we need 37 instructions, whereof 12 are memory operations. This will take at least ceil(37/2) [dual issue] + 1 [taken branch] = 19 cycles We have 12 memory cycles, plus 4 after-store conflict cycles, or 16 data cache cycles, which should be completely hidden in the 19 issue cycles. The computation is inherently serial, with these dependencies: ldq ldq \ /\ (or) addq | |\ / \ | | addq cmpult \ | | cmpult | \ / or I.e., 3 operations are needed between carry-in and carry-out, making 12 cycles the absolute minimum for the 4 limbs. We could replace the `or' with a cmoveq/cmovne, which could issue one cycle earlier that the `or', but that might waste a cycle on EV4. The total depth remain unaffected, since cmov has a latency of 2 cycles. addq / \ addq cmpult | \ cmpult -> cmovne Montgomery has a slightly different way of computing carry that requires one less instruction, but has depth 4 (instead of the current 3). Since the code is currently instruction issue bound, Montgomery's idea should save us 1/2 cycle per limb, or bring us down to a total of 17 cycles or 4.25 cycles/limb. Unfortunately, this method will not be good for the EV6. 4. addmul_1 and friends: We previously had a scheme for splitting the single- limb operand in 21-bits chunks and the multi-limb operand in 32-bit chunks, and then use FP operations for every 2nd multiply, and integer operations for every 2nd multiply. But it seems much better to split the single-limb operand in 16-bit chunks, since we save many integer shifts and adds that way. See powerpc64/README for some more details. EV6 Here we have a really parallel pipeline, capable of issuing up to 4 integer instructions per cycle. In actual practice, it is never possible to sustain more than 3.5 integer insns/cycle due to rename register shortage. One integer multiply instruction can issue each cycle. To get optimal speed, we need to pretend we are vectorizing the code, i.e., minimize the depth of recurrences. There are two dependencies to watch out for. 1) Address arithmetic dependencies, and 2) carry propagation dependencies. We can avoid serializing due to address arithmetic by unrolling loops, so that addresses don't depend heavily on an index variable. Avoiding serializing because of carry propagation is trickier; the ultimate performance of the code will be determined of the number of latency cycles it takes from accepting carry-in to a vector point until we can generate carry-out. Most integer instructions can execute in either the L0, U0, L1, or U1 pipelines. Shifts only execute in U0 and U1, and multiply only in U1. CMOV instructions split into two internal instructions, CMOV1 and CMOV2. CMOV split the mapping process (see pg 2-26 in cmpwrgd.pdf), suggesting the CMOV should always be placed as the last instruction of an aligned 4 instruction block, or perhaps simply avoided. Perhaps the most important issue is the latency between the L0/U0 and L1/U1 clusters; a result obtained on either cluster has an extra cycle of latency for consumers in the opposite cluster. Because of the dynamic nature of the implementation, it is hard to predict where an instruction will execute. REFERENCES "Alpha Architecture Handbook", version 4, Compaq, October 1998, order number EC-QD2KC-TE. "Alpha 21164 Microprocessor Hardware Reference Manual", Compaq, December 1998, order number EC-QP99C-TE. "Alpha 21264/EV67 Microprocessor Hardware Reference Manual", revision 1.4, Compaq, September 2000, order number DS-0028B-TE. "Compiler Writer's Guide for the Alpha 21264", Compaq, June 1999, order number EC-RJ66A-TE. All of the above are available online from http://ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html ftp://ftp.compaq.com/pub/products/alphaCPUdocs "Tru64 Unix Assembly Language Programmer's Guide", Compaq, March 1996, part number AA-PS31D-TE. "Digital UNIX Calling Standard for Alpha Systems", Digital Equipment Corp, March 1996, part number AA-PY8AC-TE. The above are available online, http://h30097.www3.hp.com/docs/pub_page/V40F_DOCS.HTM (Dunno what h30097 means in this URL, but if it moves try searching for "tru64 online documentation" from the main www.hp.com page.) ---------------- Local variables: mode: text fill-column: 79 End: