$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
open STDOUT,"| $^X ${dir}perlasm/x86_64-xlate.pl $flavour $output";
-if ($win64) { $arg1="%rcx"; $arg2="%rdx"; }
-else { $arg1="%rdi"; $arg2="%rsi"; }
+($arg1,$arg2,$arg3,$arg4)=$win64?("%rcx","%rdx","%r8", "%r9") : # Win64 order
+ ("%rdi","%rsi","%rdx","%rcx"); # Unix order
+
print<<___;
.extern OPENSSL_cpuid_setup
.section .init
call OPENSSL_cpuid_setup
-___
-
-if ($flavour =~ /elf/) {
- print ".section .note.GNU-stack,\"\",\@progbits\n"
-}
-
-print<<___;
.text
.globl OPENSSL_atomic_add
.type OPENSSL_ia32_cpuid,\@abi-omnipotent
.align 16
OPENSSL_ia32_cpuid:
- mov %rbx,%r8
+ mov %rbx,%r8 # save %rbx
xor %eax,%eax
cpuid
# AMD specific
mov \$0x80000000,%eax
cpuid
- cmp \$0x80000008,%eax
+ cmp \$0x80000001,%eax
+ jb .Lintel
+ mov %eax,%r10d
+ mov \$0x80000001,%eax
+ cpuid
+ or %ecx,%r9d
+ and \$0x00000801,%r9d # isolate AMD XOP bit, 1<<11
+
+ cmp \$0x80000008,%r10d
jb .Lintel
mov \$0x80000008,%eax
mov \$1,%eax
cpuid
bt \$28,%edx # test hyper-threading bit
- jnc .Ldone
+ jnc .Lgeneric
shr \$16,%ebx # number of logical processors
cmp %r10b,%bl
- ja .Ldone
+ ja .Lgeneric
and \$0xefffffff,%edx # ~(1<<28)
- jmp .Ldone
+ jmp .Lgeneric
.Lintel:
cmp \$4,%r11d
or \$0x40000000,%edx # use reserved bit to skip unrolled loop
.Lnotintel:
bt \$28,%edx # test hyper-threading bit
- jnc .Ldone
+ jnc .Lgeneric
and \$0xefffffff,%edx # ~(1<<28)
cmp \$0,%r10d
- je .Ldone
+ je .Lgeneric
or \$0x10000000,%edx # 1<<28
shr \$16,%ebx
cmp \$1,%bl # see if cache is shared
- ja .Ldone
+ ja .Lgeneric
and \$0xefffffff,%edx # ~(1<<28)
-.Ldone:
+.Lgeneric:
+ and \$0x00000800,%r9d # isolate AMD XOP flag
+ and \$0xfffff7ff,%ecx
+ or %r9d,%ecx # merge AMD XOP flag
+
shl \$32,%rcx
- mov %edx,%eax
- mov %r8,%rbx
- or %rcx,%rax
+ mov %edx,%ebx
+ or %rcx,%rbx # compose capability vector in %rbx
+ bt \$27+32,%rcx # check OSXSAVE bit
+ jnc .Lclear_avx
+ xor %ecx,%ecx # XCR0
+ .byte 0x0f,0x01,0xd0 # xgetbv
+ and \$6,%eax # isolate XMM and YMM state support
+ cmp \$6,%eax
+ je .Ldone
+.Lclear_avx:
+ mov \$0xefffe7ff,%eax # ~(1<<28|1<<12|1<<11)
+ shl \$32,%rax
+ and %rax,%rbx # clear AVX, FMA and AMD XOP bits
+.Ldone:
+ mov %rbx,%rax
+ mov %r8,%rbx # restore %rbx
ret
.size OPENSSL_ia32_cpuid,.-OPENSSL_ia32_cpuid
ret
.size OPENSSL_wipe_cpu,.-OPENSSL_wipe_cpu
___
+{
+my $out="%r10";
+my $cnt="%rcx";
+my $max="%r11";
+my $lasttick="%r8d";
+my $lastdiff="%r9d";
+my $redzone=win64?8:-8;
+
+print<<___;
+.globl OPENSSL_instrument_bus
+.type OPENSSL_instrument_bus,\@abi-omnipotent
+.align 16
+OPENSSL_instrument_bus:
+ mov $arg1,$out # tribute to Win64
+ mov $arg2,$cnt
+ mov $arg2,$max
+
+ rdtsc # collect 1st tick
+ mov %eax,$lasttick # lasttick = tick
+ mov \$0,$lastdiff # lastdiff = 0
+ clflush ($out)
+ .byte 0xf0 # lock
+ add $lastdiff,($out)
+ jmp .Loop
+.align 16
+.Loop: rdtsc
+ mov %eax,%edx
+ sub $lasttick,%eax
+ mov %edx,$lasttick
+ mov %eax,$lastdiff
+ clflush ($out)
+ .byte 0xf0 # lock
+ add %eax,($out)
+ lea 4($out),$out
+ sub \$1,$cnt
+ jnz .Loop
+
+ mov $max,%rax
+ ret
+.size OPENSSL_instrument_bus,.-OPENSSL_instrument_bus
+
+.globl OPENSSL_instrument_bus2
+.type OPENSSL_instrument_bus2,\@abi-omnipotent
+.align 16
+OPENSSL_instrument_bus2:
+ mov $arg1,$out # tribute to Win64
+ mov $arg2,$cnt
+ mov $arg3,$max
+ mov $cnt,$redzone(%rsp)
+
+ rdtsc # collect 1st tick
+ mov %eax,$lasttick # lasttick = tick
+ mov \$0,$lastdiff # lastdiff = 0
+
+ clflush ($out)
+ .byte 0xf0 # lock
+ add $lastdiff,($out)
+
+ rdtsc # collect 1st diff
+ mov %eax,%edx
+ sub $lasttick,%eax # diff
+ mov %edx,$lasttick # lasttick = tick
+ mov %eax,$lastdiff # lastdiff = diff
+.Loop2:
+ clflush ($out)
+ .byte 0xf0 # lock
+ add %eax,($out) # accumulate diff
+
+ sub \$1,$max
+ jz .Ldone2
+
+ rdtsc
+ mov %eax,%edx
+ sub $lasttick,%eax # diff
+ mov %edx,$lasttick # lasttick = tick
+ cmp $lastdiff,%eax
+ mov %eax,$lastdiff # lastdiff = diff
+ mov \$0,%edx
+ setne %dl
+ sub %rdx,$cnt # conditional --$cnt
+ lea ($out,%rdx,4),$out # conditional ++$out
+ jnz .Loop2
+
+.Ldone2:
+ mov $redzone(%rsp),%rax
+ sub $cnt,%rax
+ ret
+.size OPENSSL_instrument_bus2,.-OPENSSL_instrument_bus2
+___
+}
close STDOUT; # flush