1 First up, let me say I don't like writing in assembler. It is not portable, 2 dependant on the particular CPU architecture release and is generally a pig 3 to debug and get right. Having said that, the x86 architecture is probably 4 the most important for speed due to number of boxes and since 5 it appears to be the worst architecture to to get 6 good C compilers for. So due to this, I have lowered myself to do 7 assembler for the inner DES routines in libdes :-). 8 9 The file to implement in assembler is des_enc.c. Replace the following 10 4 functions 11 des_encrypt1(DES_LONG data[2],des_key_schedule ks, int encrypt); 12 des_encrypt2(DES_LONG data[2],des_key_schedule ks, int encrypt); 13 des_encrypt3(DES_LONG data[2],des_key_schedule ks1,ks2,ks3); 14 des_decrypt3(DES_LONG data[2],des_key_schedule ks1,ks2,ks3); 15 16 They encrypt/decrypt the 64 bits held in 'data' using 17 the 'ks' key schedules. The only difference between the 4 functions is that 18 des_encrypt2() does not perform IP() or FP() on the data (this is an 19 optimization for when doing triple DES and des_encrypt3() and des_decrypt3() 20 perform triple des. The triple DES routines are in here because it does 21 make a big difference to have them located near the des_encrypt2 function 22 at link time.. 23 24 Now as we all know, there are lots of different operating systems running on 25 x86 boxes, and unfortunately they normally try to make sure their assembler 26 formating is not the same as the other peoples. 27 The 4 main formats I know of are 28 Microsoft Windows 95/Windows NT 29 Elf Includes Linux and FreeBSD(?). 30 a.out The older Linux. 31 Solaris Same as Elf but different comments :-(. 32 33 Now I was not overly keen to write 4 different copies of the same code, 34 so I wrote a few perl routines to output the correct assembler, given 35 a target assembler type. This code is ugly and is just a hack. 36 The libraries are x86unix.pl and x86ms.pl. 37 des586.pl, des686.pl and des-som[23].pl are the programs to actually 38 generate the assembler. 39 40 So to generate elf assembler 41 perl des-som3.pl elf >dx86-elf.s 42 For Windows 95/NT 43 perl des-som2.pl win32 >win32.asm 44 45 [ update 4 Jan 1996 ] 46 I have added another way to do things. 47 perl des-som3.pl cpp >dx86-cpp.s 48 generates a file that will be included by dx86unix.cpp when it is compiled. 49 To build for elf, a.out, solaris, bsdi etc, 50 cc -E -DELF asm/dx86unix.cpp | as -o asm/dx86-elf.o 51 cc -E -DSOL asm/dx86unix.cpp | as -o asm/dx86-sol.o 52 cc -E -DOUT asm/dx86unix.cpp | as -o asm/dx86-out.o 53 cc -E -DBSDI asm/dx86unix.cpp | as -o asm/dx86bsdi.o 54 This was done to cut down the number of files in the distribution. 55 56 Now the ugly part. I acquired my copy of Intels 57 "Optimization's For Intel's 32-Bit Processors" and found a few interesting 58 things. First, the aim of the exersize is to 'extract' one byte at a time 59 from a word and do an array lookup. This involves getting the byte from 60 the 4 locations in the word and moving it to a new word and doing the lookup. 61 The most obvious way to do this is 62 xor eax, eax # clear word 63 movb al, cl # get low byte 64 xor edi DWORD PTR 0x100+des_SP[eax] # xor in word 65 movb al, ch # get next byte 66 xor edi DWORD PTR 0x300+des_SP[eax] # xor in word 67 shr ecx 16 68 which seems ok. For the pentium, this system appears to be the best. 69 One has to do instruction interleaving to keep both functional units 70 operating, but it is basically very efficient. 71 72 Now the crunch. When a full register is used after a partial write, eg. 73 mov al, cl 74 xor edi, DWORD PTR 0x100+des_SP[eax] 75 386 - 1 cycle stall 76 486 - 1 cycle stall 77 586 - 0 cycle stall 78 686 - at least 7 cycle stall (page 22 of the above mentioned document). 79 80 So the technique that produces the best results on a pentium, according to 81 the documentation, will produce hideous results on a pentium pro. 82 83 To get around this, des686.pl will generate code that is not as fast on 84 a pentium, should be very good on a pentium pro. 85 mov eax, ecx # copy word 86 shr ecx, 8 # line up next byte 87 and eax, 0fch # mask byte 88 xor edi DWORD PTR 0x100+des_SP[eax] # xor in array lookup 89 mov eax, ecx # get word 90 shr ecx 8 # line up next byte 91 and eax, 0fch # mask byte 92 xor edi DWORD PTR 0x300+des_SP[eax] # xor in array lookup 93 94 Due to the execution units in the pentium, this actually works quite well. 95 For a pentium pro it should be very good. This is the type of output 96 Visual C++ generates. 97 98 There is a third option. instead of using 99 mov al, ch 100 which is bad on the pentium pro, one may be able to use 101 movzx eax, ch 102 which may not incur the partial write penalty. On the pentium, 103 this instruction takes 4 cycles so is not worth using but on the 104 pentium pro it appears it may be worth while. I need access to one to 105 experiment :-). 106 107 eric (20 Oct 1996) 108 109 22 Nov 1996 - I have asked people to run the 2 different version on pentium 110 pros and it appears that the intel documentation is wrong. The 111 mov al,bh is still faster on a pentium pro, so just use the des586.pl 112 install des686.pl 113 114 3 Dec 1996 - I added des_encrypt3/des_decrypt3 because I have moved these 115 functions into des_enc.c because it does make a massive performance 116 difference on some boxes to have the functions code located close to 117 the des_encrypt2() function. 118 119 9 Jan 1997 - des-som2.pl is now the correct perl script to use for 120 pentiums. It contains an inner loop from 121 Svend Olaf Mikkelsen <svolaf (a] inet.uni-c.dk> which does raw ecb DES calls at 122 273,000 per second. He had a previous version at 250,000 and the best 123 I was able to get was 203,000. The content has not changed, this is all 124 due to instruction sequencing (and actual instructions choice) which is able 125 to keep both functional units of the pentium going. 126 We may have lost the ugly register usage restrictions when x86 went 32 bit 127 but for the pentium it has been replaced by evil instruction ordering tricks. 128 129 13 Jan 1997 - des-som3.pl, more optimizations from Svend Olaf. 130 raw DES at 281,000 per second on a pentium 100. 131 132
