#!/usr/bin/perl -w use strict; my $code .= < / Licence: I hereby disclaim the copyright on this code and place it / in the public domain. / .ident "%Z%%M% %I% %E% SMI" / / The following is Marc Bevand's MD5 implementation optimized for / AMD64. It has been lifted intact, except for changing the comment / character, adding comments, and a kludge to generate valid "lea" / instructions that are not generated by the Solaris "as" assembler / (CR 6628627). / / typedef struct { / uint32_t state[4]; /* state (ABCD) */ / uint32_t count[2]; /* number of bits, modulo 2^64 (lsb first) */ / union { / uint8_t buf8[64]; /* undigested input */ / uint32_t buf32[16]; /* realigned input */ / } buf_un; / } MD5_CTX; / / void md5_block_asm_host_order(MD5_CTX *ctx, const void *inpp, / unsigned int input_length_in_blocks); / / Registers used: / rax A r8 old A / rbx B r9 old B / rcx C r10 tmp / rdx D r11 tmp / rsi ptr r12 tmp / rdi end r13 - / rbp - r14 old C / rsp stack r15 old D / EOF sub lea_offset_register_r10d_ebx # Workaround for a Solaris "gas" assembler bug where compiling the source # errors out and does not generate a valid "lea" instruction. Specifically, # lea OFFSET(REGISTER,%r10d),REGISTER # # For Solaris as, "as -a32" must be used to compile this. # For Solaris gas 2.15, this errors out with this message: # Error: `0xf57c0faf(%eax,%r10d)' is not a valid 64 bit base/index expression # This should be fixed in Solaris gas 2.16. # It assembles with the Linux "as --64" gas 2.17 assembler and runs OK. # # For the ONBLD NV tools, the aw wrapper script fails when -a32 is used: # /ws/onnv-tools/onbld/bin/i386/aw -xarch=amd64 -P -a32 -o lea.o lea.s # aw: as->gas mapping failed at or near arg '-a32' # # For more information, see CRs 6644870 and 6628627. # Note2: Solaris "as" uses "/" for comments; Linux "as" uses "#" for comments. { use Switch; my ($offset, $register, $comment) = @_; # Failed "lea" instruction. # This instruction errors out from the Solaris as assembler. # It assembles with the Linux "as --64" assembler and runs OK. $code .= " / lea $offset($register,%r10d),$register" . " $comment\n"; # Workaround #1 (not used) # One workaround is to generate two "add" instructions that are # functionally equivalent to "lea." The problem is this workaround # is about 4.5% slower than a lea, so is not used. #$code .= " add %r10d,$register\n"; #$code .= " add \&0x%0x,$offset\n"; # Workaround #2 (used) # This workaround hand-generates hex machine code for lea. $code .= " / Solaris as assembly bug CR 6628627 errors out for\n"; $code .= " / the above, so we specify the machine code in hex:\n"; $code .= " .byte 0x67,0x42,0x8d / lea offset(reg,%r10d),reg\n"; switch ($register) { case "%eax" { $code .= " .byte 0x84,0x10 / reg=%eax\n"; } case "%ebx" { $code .= " .byte 0x9c,0x13 / reg=%ebx\n"; } case "%ecx" { $code .= " .byte 0x8c,0x11 / reg=%ecx\n"; } case "%edx" { $code .= " .byte 0x94,0x12 / reg=%edx\n"; } else { $code .= "ERROR: unknown register $register\n"; } } $code .= " .long $offset / offset\n"; } # round1_step() does: # dst = x + ((dst + F(x,y,z) + X[k] + T_i) <<< s) # %r10d = X[k_next] # %r11d = z' (copy of z for the next step) # Each round1_step() takes about 5.3 clocks (9 instructions, 1.7 IPC) sub round1_step { my ($pos, $dst, $x, $y, $z, $k_next, $T_i, $s) = @_; $code .= " mov 0*4(%rsi), %r10d /* (NEXT STEP) X[0] */\n" if ($pos == -1); $code .= " mov %edx, %r11d /* (NEXT STEP) z' = %edx */\n" if ($pos == -1); $code .= " xor $y, %r11d /* y ^ ... */\n"; #lea $T_i($dst,%r10d),$dst /* Const + dst + ... */ lea_offset_register_r10d_ebx($T_i, $dst, "/* Const + dst + ... */"); $code .= <$output" or die "can't open $output: $!"; $code .= < ENTRY_NP(md5_block_asm_host_order) push %rbp push %rbx push %r12 push %r13 push %r14 push %r15 / rdi = arg #1 (ctx, MD5_CTX pointer) / rsi = arg #2 (ptr, data pointer) / rdx = arg #3 (nbr, number of 64-byte blocks to process) mov %rdi, %rbp / rbp = ctx shl \$6, %rdx / rdx = nbr in bytes lea (%rsi,%rdx), %rdi / rdi = end mov 0*4(%rbp), %eax / eax = ctx->A mov 1*4(%rbp), %ebx / ebx = ctx->B mov 2*4(%rbp), %ecx / ecx = ctx->C mov 3*4(%rbp), %edx / edx = ctx->D push %rbp / save ctx / end is 'rdi' / ptr is 'rsi' / A is 'eax' / B is 'ebx' / C is 'ecx' / D is 'edx' cmp %rdi, %rsi / cmp end with ptr je 1f / jmp if ptr == end / BEGIN of loop over 64-byte blocks 2: / save old values of A, B, C, D mov %eax, %r8d mov %ebx, %r9d mov %ecx, %r14d mov %edx, %r15d EOF round1_step(-1,'%eax','%ebx','%ecx','%edx', '1','0xd76aa478', '7'); round1_step( 0,'%edx','%eax','%ebx','%ecx', '2','0xe8c7b756','12'); round1_step( 0,'%ecx','%edx','%eax','%ebx', '3','0x242070db','17'); round1_step( 0,'%ebx','%ecx','%edx','%eax', '4','0xc1bdceee','22'); round1_step( 0,'%eax','%ebx','%ecx','%edx', '5','0xf57c0faf', '7'); round1_step( 0,'%edx','%eax','%ebx','%ecx', '6','0x4787c62a','12'); round1_step( 0,'%ecx','%edx','%eax','%ebx', '7','0xa8304613','17'); round1_step( 0,'%ebx','%ecx','%edx','%eax', '8','0xfd469501','22'); round1_step( 0,'%eax','%ebx','%ecx','%edx', '9','0x698098d8', '7'); round1_step( 0,'%edx','%eax','%ebx','%ecx','10','0x8b44f7af','12'); round1_step( 0,'%ecx','%edx','%eax','%ebx','11','0xffff5bb1','17'); round1_step( 0,'%ebx','%ecx','%edx','%eax','12','0x895cd7be','22'); round1_step( 0,'%eax','%ebx','%ecx','%edx','13','0x6b901122', '7'); round1_step( 0,'%edx','%eax','%ebx','%ecx','14','0xfd987193','12'); round1_step( 0,'%ecx','%edx','%eax','%ebx','15','0xa679438e','17'); round1_step( 1,'%ebx','%ecx','%edx','%eax', '0','0x49b40821','22'); round2_step(-1,'%eax','%ebx','%ecx','%edx', '6','0xf61e2562', '5'); round2_step( 0,'%edx','%eax','%ebx','%ecx','11','0xc040b340', '9'); round2_step( 0,'%ecx','%edx','%eax','%ebx', '0','0x265e5a51','14'); round2_step( 0,'%ebx','%ecx','%edx','%eax', '5','0xe9b6c7aa','20'); round2_step( 0,'%eax','%ebx','%ecx','%edx','10','0xd62f105d', '5'); round2_step( 0,'%edx','%eax','%ebx','%ecx','15', '0x2441453', '9'); round2_step( 0,'%ecx','%edx','%eax','%ebx', '4','0xd8a1e681','14'); round2_step( 0,'%ebx','%ecx','%edx','%eax', '9','0xe7d3fbc8','20'); round2_step( 0,'%eax','%ebx','%ecx','%edx','14','0x21e1cde6', '5'); round2_step( 0,'%edx','%eax','%ebx','%ecx', '3','0xc33707d6', '9'); round2_step( 0,'%ecx','%edx','%eax','%ebx', '8','0xf4d50d87','14'); round2_step( 0,'%ebx','%ecx','%edx','%eax','13','0x455a14ed','20'); round2_step( 0,'%eax','%ebx','%ecx','%edx', '2','0xa9e3e905', '5'); round2_step( 0,'%edx','%eax','%ebx','%ecx', '7','0xfcefa3f8', '9'); round2_step( 0,'%ecx','%edx','%eax','%ebx','12','0x676f02d9','14'); round2_step( 1,'%ebx','%ecx','%edx','%eax', '0','0x8d2a4c8a','20'); round3_step(-1,'%eax','%ebx','%ecx','%edx', '8','0xfffa3942', '4'); round3_step( 0,'%edx','%eax','%ebx','%ecx','11','0x8771f681','11'); round3_step( 0,'%ecx','%edx','%eax','%ebx','14','0x6d9d6122','16'); round3_step( 0,'%ebx','%ecx','%edx','%eax', '1','0xfde5380c','23'); round3_step( 0,'%eax','%ebx','%ecx','%edx', '4','0xa4beea44', '4'); round3_step( 0,'%edx','%eax','%ebx','%ecx', '7','0x4bdecfa9','11'); round3_step( 0,'%ecx','%edx','%eax','%ebx','10','0xf6bb4b60','16'); round3_step( 0,'%ebx','%ecx','%edx','%eax','13','0xbebfbc70','23'); round3_step( 0,'%eax','%ebx','%ecx','%edx', '0','0x289b7ec6', '4'); round3_step( 0,'%edx','%eax','%ebx','%ecx', '3','0xeaa127fa','11'); round3_step( 0,'%ecx','%edx','%eax','%ebx', '6','0xd4ef3085','16'); round3_step( 0,'%ebx','%ecx','%edx','%eax', '9', '0x4881d05','23'); round3_step( 0,'%eax','%ebx','%ecx','%edx','12','0xd9d4d039', '4'); round3_step( 0,'%edx','%eax','%ebx','%ecx','15','0xe6db99e5','11'); round3_step( 0,'%ecx','%edx','%eax','%ebx', '2','0x1fa27cf8','16'); round3_step( 1,'%ebx','%ecx','%edx','%eax', '0','0xc4ac5665','23'); round4_step(-1,'%eax','%ebx','%ecx','%edx', '7','0xf4292244', '6'); round4_step( 0,'%edx','%eax','%ebx','%ecx','14','0x432aff97','10'); round4_step( 0,'%ecx','%edx','%eax','%ebx', '5','0xab9423a7','15'); round4_step( 0,'%ebx','%ecx','%edx','%eax','12','0xfc93a039','21'); round4_step( 0,'%eax','%ebx','%ecx','%edx', '3','0x655b59c3', '6'); round4_step( 0,'%edx','%eax','%ebx','%ecx','10','0x8f0ccc92','10'); round4_step( 0,'%ecx','%edx','%eax','%ebx', '1','0xffeff47d','15'); round4_step( 0,'%ebx','%ecx','%edx','%eax', '8','0x85845dd1','21'); round4_step( 0,'%eax','%ebx','%ecx','%edx','15','0x6fa87e4f', '6'); round4_step( 0,'%edx','%eax','%ebx','%ecx', '6','0xfe2ce6e0','10'); round4_step( 0,'%ecx','%edx','%eax','%ebx','13','0xa3014314','15'); round4_step( 0,'%ebx','%ecx','%edx','%eax', '4','0x4e0811a1','21'); round4_step( 0,'%eax','%ebx','%ecx','%edx','11','0xf7537e82', '6'); round4_step( 0,'%edx','%eax','%ebx','%ecx', '2','0xbd3af235','10'); round4_step( 0,'%ecx','%edx','%eax','%ebx', '9','0x2ad7d2bb','15'); round4_step( 1,'%ebx','%ecx','%edx','%eax', '0','0xeb86d391','21'); $code .= <A = A mov %ebx, 1*4(%rbp) / ctx->B = B mov %ecx, 2*4(%rbp) / ctx->C = C mov %edx, 3*4(%rbp) / ctx->D = D pop %r15 pop %r14 pop %r13 pop %r12 pop %rbx pop %rbp ret SET_SIZE(md5_block_asm_host_order) #else /* LINTED */ /* Nothing to be linted in this file--it's pure assembly source. */ #endif /* !lint && !__lint */ EOF print $code;