diff options
| author | djm <> | 2010-10-01 22:59:01 +0000 |
|---|---|---|
| committer | djm <> | 2010-10-01 22:59:01 +0000 |
| commit | 8922d4bc4a8b8893d72a48deb2cdf58215f98505 (patch) | |
| tree | 939b752540947d33507b3acc48d76a8bfb7c3dc3 /src/lib/libcrypto/x86_64cpuid.pl | |
| parent | 76262f7bf9262f965142b1b2b2105cb279c5c696 (diff) | |
| download | openbsd-8922d4bc4a8b8893d72a48deb2cdf58215f98505.tar.gz openbsd-8922d4bc4a8b8893d72a48deb2cdf58215f98505.tar.bz2 openbsd-8922d4bc4a8b8893d72a48deb2cdf58215f98505.zip | |
resolve conflicts, fix local changes
Diffstat (limited to 'src/lib/libcrypto/x86_64cpuid.pl')
| -rw-r--r-- | src/lib/libcrypto/x86_64cpuid.pl | 256 |
1 files changed, 164 insertions, 92 deletions
diff --git a/src/lib/libcrypto/x86_64cpuid.pl b/src/lib/libcrypto/x86_64cpuid.pl index 8946b464a8..b771a8539d 100644 --- a/src/lib/libcrypto/x86_64cpuid.pl +++ b/src/lib/libcrypto/x86_64cpuid.pl | |||
| @@ -1,110 +1,38 @@ | |||
| 1 | #!/usr/bin/env perl | 1 | #!/usr/bin/env perl |
| 2 | 2 | ||
| 3 | $output=shift; | 3 | $flavour = shift; |
| 4 | $masm=1 if ($output =~ /\.asm/); | 4 | $output = shift; |
| 5 | open STDOUT,">$output" || die "can't open $output: $!"; | 5 | if ($flavour =~ /\./) { $output = $flavour; undef $flavour; } |
| 6 | |||
| 7 | print<<___ if(defined($masm)); | ||
| 8 | _TEXT SEGMENT | ||
| 9 | PUBLIC OPENSSL_rdtsc | ||
| 10 | |||
| 11 | PUBLIC OPENSSL_atomic_add | ||
| 12 | ALIGN 16 | ||
| 13 | OPENSSL_atomic_add PROC | ||
| 14 | mov eax,DWORD PTR[rcx] | ||
| 15 | \$Lspin: lea r8,DWORD PTR[rdx+rax] | ||
| 16 | lock cmpxchg DWORD PTR[rcx],r8d | ||
| 17 | jne \$Lspin | ||
| 18 | mov eax,r8d | ||
| 19 | cdqe | ||
| 20 | ret | ||
| 21 | OPENSSL_atomic_add ENDP | ||
| 22 | |||
| 23 | PUBLIC OPENSSL_wipe_cpu | ||
| 24 | ALIGN 16 | ||
| 25 | OPENSSL_wipe_cpu PROC | ||
| 26 | pxor xmm0,xmm0 | ||
| 27 | pxor xmm1,xmm1 | ||
| 28 | pxor xmm2,xmm2 | ||
| 29 | pxor xmm3,xmm3 | ||
| 30 | pxor xmm4,xmm4 | ||
| 31 | pxor xmm5,xmm5 | ||
| 32 | xor rcx,rcx | ||
| 33 | xor rdx,rdx | ||
| 34 | xor r8,r8 | ||
| 35 | xor r9,r9 | ||
| 36 | xor r10,r10 | ||
| 37 | xor r11,r11 | ||
| 38 | lea rax,QWORD PTR[rsp+8] | ||
| 39 | ret | ||
| 40 | OPENSSL_wipe_cpu ENDP | ||
| 41 | _TEXT ENDS | ||
| 42 | 6 | ||
| 43 | CRT\$XIU SEGMENT | 7 | $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/); |
| 44 | EXTRN OPENSSL_cpuid_setup:PROC | ||
| 45 | DQ OPENSSL_cpuid_setup | ||
| 46 | CRT\$XIU ENDS | ||
| 47 | 8 | ||
| 48 | ___ | 9 | $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1; |
| 49 | print<<___ if(!defined($masm)); | 10 | open STDOUT,"| $^X ${dir}perlasm/x86_64-xlate.pl $flavour $output"; |
| 11 | |||
| 12 | if ($win64) { $arg1="%rcx"; $arg2="%rdx"; } | ||
| 13 | else { $arg1="%rdi"; $arg2="%rsi"; } | ||
| 14 | print<<___; | ||
| 50 | #include <machine/asm.h> | 15 | #include <machine/asm.h> |
| 16 | .extern OPENSSL_cpuid_setup | ||
| 17 | .section .init | ||
| 18 | call PIC_PLT(OPENSSL_cpuid_setup) | ||
| 51 | 19 | ||
| 52 | .text | 20 | .text |
| 53 | 21 | ||
| 54 | .globl OPENSSL_atomic_add | 22 | .globl OPENSSL_atomic_add |
| 55 | .type OPENSSL_atomic_add,\@function | 23 | .type OPENSSL_atomic_add,\@abi-omnipotent |
| 56 | .align 16 | 24 | .align 16 |
| 57 | OPENSSL_atomic_add: | 25 | OPENSSL_atomic_add: |
| 58 | movl (%rdi),%eax | 26 | movl ($arg1),%eax |
| 59 | .Lspin: leaq (%rsi,%rax),%r8 | 27 | .Lspin: leaq ($arg2,%rax),%r8 |
| 60 | lock; cmpxchgl %r8d,(%rdi) | 28 | .byte 0xf0 # lock |
| 29 | cmpxchgl %r8d,($arg1) | ||
| 61 | jne .Lspin | 30 | jne .Lspin |
| 62 | movl %r8d,%eax | 31 | movl %r8d,%eax |
| 63 | .byte 0x48,0x98 | 32 | .byte 0x48,0x98 # cltq/cdqe |
| 64 | ret | 33 | ret |
| 65 | .size OPENSSL_atomic_add,.-OPENSSL_atomic_add | 34 | .size OPENSSL_atomic_add,.-OPENSSL_atomic_add |
| 66 | 35 | ||
| 67 | .globl OPENSSL_wipe_cpu | ||
| 68 | .type OPENSSL_wipe_cpu,\@function | ||
| 69 | .align 16 | ||
| 70 | OPENSSL_wipe_cpu: | ||
| 71 | pxor %xmm0,%xmm0 | ||
| 72 | pxor %xmm1,%xmm1 | ||
| 73 | pxor %xmm2,%xmm2 | ||
| 74 | pxor %xmm3,%xmm3 | ||
| 75 | pxor %xmm4,%xmm4 | ||
| 76 | pxor %xmm5,%xmm5 | ||
| 77 | pxor %xmm6,%xmm6 | ||
| 78 | pxor %xmm7,%xmm7 | ||
| 79 | pxor %xmm8,%xmm8 | ||
| 80 | pxor %xmm9,%xmm9 | ||
| 81 | pxor %xmm10,%xmm10 | ||
| 82 | pxor %xmm11,%xmm11 | ||
| 83 | pxor %xmm12,%xmm12 | ||
| 84 | pxor %xmm13,%xmm13 | ||
| 85 | pxor %xmm14,%xmm14 | ||
| 86 | pxor %xmm15,%xmm15 | ||
| 87 | xorq %rcx,%rcx | ||
| 88 | xorq %rdx,%rdx | ||
| 89 | xorq %rsi,%rsi | ||
| 90 | xorq %rdi,%rdi | ||
| 91 | xorq %r8,%r8 | ||
| 92 | xorq %r9,%r9 | ||
| 93 | xorq %r10,%r10 | ||
| 94 | xorq %r11,%r11 | ||
| 95 | leaq 8(%rsp),%rax | ||
| 96 | ret | ||
| 97 | .size OPENSSL_wipe_cpu,.-OPENSSL_wipe_cpu | ||
| 98 | |||
| 99 | .section .init | ||
| 100 | call PIC_PLT(OPENSSL_cpuid_setup) | ||
| 101 | |||
| 102 | ___ | ||
| 103 | |||
| 104 | open STDOUT,"| $^X perlasm/x86_64-xlate.pl $output"; | ||
| 105 | print<<___; | ||
| 106 | .text | ||
| 107 | |||
| 108 | .globl OPENSSL_rdtsc | 36 | .globl OPENSSL_rdtsc |
| 109 | .type OPENSSL_rdtsc,\@abi-omnipotent | 37 | .type OPENSSL_rdtsc,\@abi-omnipotent |
| 110 | .align 16 | 38 | .align 16 |
| @@ -123,6 +51,8 @@ OPENSSL_ia32_cpuid: | |||
| 123 | 51 | ||
| 124 | xor %eax,%eax | 52 | xor %eax,%eax |
| 125 | cpuid | 53 | cpuid |
| 54 | mov %eax,%r11d # max value for standard query level | ||
| 55 | |||
| 126 | xor %eax,%eax | 56 | xor %eax,%eax |
| 127 | cmp \$0x756e6547,%ebx # "Genu" | 57 | cmp \$0x756e6547,%ebx # "Genu" |
| 128 | setne %al | 58 | setne %al |
| @@ -132,8 +62,54 @@ OPENSSL_ia32_cpuid: | |||
| 132 | or %eax,%r9d | 62 | or %eax,%r9d |
| 133 | cmp \$0x6c65746e,%ecx # "ntel" | 63 | cmp \$0x6c65746e,%ecx # "ntel" |
| 134 | setne %al | 64 | setne %al |
| 135 | or %eax,%r9d | 65 | or %eax,%r9d # 0 indicates Intel CPU |
| 66 | jz .Lintel | ||
| 67 | |||
| 68 | cmp \$0x68747541,%ebx # "Auth" | ||
| 69 | setne %al | ||
| 70 | mov %eax,%r10d | ||
| 71 | cmp \$0x69746E65,%edx # "enti" | ||
| 72 | setne %al | ||
| 73 | or %eax,%r10d | ||
| 74 | cmp \$0x444D4163,%ecx # "cAMD" | ||
| 75 | setne %al | ||
| 76 | or %eax,%r10d # 0 indicates AMD CPU | ||
| 77 | jnz .Lintel | ||
| 78 | |||
| 79 | # AMD specific | ||
| 80 | mov \$0x80000000,%eax | ||
| 81 | cpuid | ||
| 82 | cmp \$0x80000008,%eax | ||
| 83 | jb .Lintel | ||
| 84 | |||
| 85 | mov \$0x80000008,%eax | ||
| 86 | cpuid | ||
| 87 | movzb %cl,%r10 # number of cores - 1 | ||
| 88 | inc %r10 # number of cores | ||
| 89 | |||
| 90 | mov \$1,%eax | ||
| 91 | cpuid | ||
| 92 | bt \$28,%edx # test hyper-threading bit | ||
| 93 | jnc .Ldone | ||
| 94 | shr \$16,%ebx # number of logical processors | ||
| 95 | cmp %r10b,%bl | ||
| 96 | ja .Ldone | ||
| 97 | and \$0xefffffff,%edx # ~(1<<28) | ||
| 98 | jmp .Ldone | ||
| 99 | |||
| 100 | .Lintel: | ||
| 101 | cmp \$4,%r11d | ||
| 102 | mov \$-1,%r10d | ||
| 103 | jb .Lnocacheinfo | ||
| 104 | |||
| 105 | mov \$4,%eax | ||
| 106 | mov \$0,%ecx # query L1D | ||
| 107 | cpuid | ||
| 108 | mov %eax,%r10d | ||
| 109 | shr \$14,%r10d | ||
| 110 | and \$0xfff,%r10d # number of cores -1 per L1D | ||
| 136 | 111 | ||
| 112 | .Lnocacheinfo: | ||
| 137 | mov \$1,%eax | 113 | mov \$1,%eax |
| 138 | cpuid | 114 | cpuid |
| 139 | cmp \$0,%r9d | 115 | cmp \$0,%r9d |
| @@ -146,6 +122,11 @@ OPENSSL_ia32_cpuid: | |||
| 146 | .Lnotintel: | 122 | .Lnotintel: |
| 147 | bt \$28,%edx # test hyper-threading bit | 123 | bt \$28,%edx # test hyper-threading bit |
| 148 | jnc .Ldone | 124 | jnc .Ldone |
| 125 | and \$0xefffffff,%edx # ~(1<<28) | ||
| 126 | cmp \$0,%r10d | ||
| 127 | je .Ldone | ||
| 128 | |||
| 129 | or \$0x10000000,%edx # 1<<28 | ||
| 149 | shr \$16,%ebx | 130 | shr \$16,%ebx |
| 150 | cmp \$1,%bl # see if cache is shared | 131 | cmp \$1,%bl # see if cache is shared |
| 151 | ja .Ldone | 132 | ja .Ldone |
| @@ -157,5 +138,96 @@ OPENSSL_ia32_cpuid: | |||
| 157 | or %rcx,%rax | 138 | or %rcx,%rax |
| 158 | ret | 139 | ret |
| 159 | .size OPENSSL_ia32_cpuid,.-OPENSSL_ia32_cpuid | 140 | .size OPENSSL_ia32_cpuid,.-OPENSSL_ia32_cpuid |
| 141 | |||
| 142 | .globl OPENSSL_cleanse | ||
| 143 | .type OPENSSL_cleanse,\@abi-omnipotent | ||
| 144 | .align 16 | ||
| 145 | OPENSSL_cleanse: | ||
| 146 | xor %rax,%rax | ||
| 147 | cmp \$15,$arg2 | ||
| 148 | jae .Lot | ||
| 149 | cmp \$0,$arg2 | ||
| 150 | je .Lret | ||
| 151 | .Little: | ||
| 152 | mov %al,($arg1) | ||
| 153 | sub \$1,$arg2 | ||
| 154 | lea 1($arg1),$arg1 | ||
| 155 | jnz .Little | ||
| 156 | .Lret: | ||
| 157 | ret | ||
| 158 | .align 16 | ||
| 159 | .Lot: | ||
| 160 | test \$7,$arg1 | ||
| 161 | jz .Laligned | ||
| 162 | mov %al,($arg1) | ||
| 163 | lea -1($arg2),$arg2 | ||
| 164 | lea 1($arg1),$arg1 | ||
| 165 | jmp .Lot | ||
| 166 | .Laligned: | ||
| 167 | mov %rax,($arg1) | ||
| 168 | lea -8($arg2),$arg2 | ||
| 169 | test \$-8,$arg2 | ||
| 170 | lea 8($arg1),$arg1 | ||
| 171 | jnz .Laligned | ||
| 172 | cmp \$0,$arg2 | ||
| 173 | jne .Little | ||
| 174 | ret | ||
| 175 | .size OPENSSL_cleanse,.-OPENSSL_cleanse | ||
| 176 | ___ | ||
| 177 | |||
| 178 | print<<___ if (!$win64); | ||
| 179 | .globl OPENSSL_wipe_cpu | ||
| 180 | .type OPENSSL_wipe_cpu,\@abi-omnipotent | ||
| 181 | .align 16 | ||
| 182 | OPENSSL_wipe_cpu: | ||
| 183 | pxor %xmm0,%xmm0 | ||
| 184 | pxor %xmm1,%xmm1 | ||
| 185 | pxor %xmm2,%xmm2 | ||
| 186 | pxor %xmm3,%xmm3 | ||
| 187 | pxor %xmm4,%xmm4 | ||
| 188 | pxor %xmm5,%xmm5 | ||
| 189 | pxor %xmm6,%xmm6 | ||
| 190 | pxor %xmm7,%xmm7 | ||
| 191 | pxor %xmm8,%xmm8 | ||
| 192 | pxor %xmm9,%xmm9 | ||
| 193 | pxor %xmm10,%xmm10 | ||
| 194 | pxor %xmm11,%xmm11 | ||
| 195 | pxor %xmm12,%xmm12 | ||
| 196 | pxor %xmm13,%xmm13 | ||
| 197 | pxor %xmm14,%xmm14 | ||
| 198 | pxor %xmm15,%xmm15 | ||
| 199 | xorq %rcx,%rcx | ||
| 200 | xorq %rdx,%rdx | ||
| 201 | xorq %rsi,%rsi | ||
| 202 | xorq %rdi,%rdi | ||
| 203 | xorq %r8,%r8 | ||
| 204 | xorq %r9,%r9 | ||
| 205 | xorq %r10,%r10 | ||
| 206 | xorq %r11,%r11 | ||
| 207 | leaq 8(%rsp),%rax | ||
| 208 | ret | ||
| 209 | .size OPENSSL_wipe_cpu,.-OPENSSL_wipe_cpu | ||
| 160 | ___ | 210 | ___ |
| 211 | print<<___ if ($win64); | ||
| 212 | .globl OPENSSL_wipe_cpu | ||
| 213 | .type OPENSSL_wipe_cpu,\@abi-omnipotent | ||
| 214 | .align 16 | ||
| 215 | OPENSSL_wipe_cpu: | ||
| 216 | pxor %xmm0,%xmm0 | ||
| 217 | pxor %xmm1,%xmm1 | ||
| 218 | pxor %xmm2,%xmm2 | ||
| 219 | pxor %xmm3,%xmm3 | ||
| 220 | pxor %xmm4,%xmm4 | ||
| 221 | pxor %xmm5,%xmm5 | ||
| 222 | xorq %rcx,%rcx | ||
| 223 | xorq %rdx,%rdx | ||
| 224 | xorq %r8,%r8 | ||
| 225 | xorq %r9,%r9 | ||
| 226 | xorq %r10,%r10 | ||
| 227 | xorq %r11,%r11 | ||
| 228 | leaq 8(%rsp),%rax | ||
| 229 | ret | ||
| 230 | .size OPENSSL_wipe_cpu,.-OPENSSL_wipe_cpu | ||
| 231 | ___ | ||
| 232 | |||
| 161 | close STDOUT; # flush | 233 | close STDOUT; # flush |
