perlasm/x86_64-xlate.pl: recognize DWARF CFI directives.
authorAndy Polyakov <appro@openssl.org>
Mon, 6 Feb 2017 07:58:34 +0000 (08:58 +0100)
committerAndy Polyakov <appro@openssl.org>
Thu, 9 Feb 2017 19:00:33 +0000 (20:00 +0100)
CFI directives annotate instructions that are significant for stack
unwinding procedure. In addition to directives recognized by GNU
assembler this module implements three synthetic ones:

- .cfi_push annotates push instructions in prologue and translates to
  .cfi_adjust_cfa_offset (if needed) and .cfi_offset;
- .cfi_pop annotates pop instructions in epilogue and translates to
  .cfi_adjust_cfs_offset (if needed) and .cfi_restore;
- .cfi_cfa_expression encodes DW_CFA_def_cfa_expression and passes it
  to .cfi_escape as byte vector;

CFA expression syntax is made up mix of DWARF operator suffixes [subset
of] and references to registers with optional bias. Following example
describes offloaded original stack pointer at specific offset from
current stack pointer:

.cfi_cfa_expression %rsp+40,deref,+8

Final +8 has everything to do with the fact that CFA, Canonical Frame
Address, is reference to top of caller's stack, and on x86_64 call to
subroutine pushes 8-byte return address.

Triggered by request from Adam Langley.

Reviewed-by: Rich Salz <rsalz@openssl.org>
crypto/perlasm/x86_64-xlate.pl

index afbe5e141f86835c4d09bbe1524da90855f4a771..09d293a446f22347d110bd8e867bb5faf7a7c1f5 100755 (executable)
@@ -465,6 +465,240 @@ my %globals;
        }
     }
 }
+{ package cfi_directive;
+    # CFI directives annotate instructions that are significant for
+    # stack unwinding procedure compliant with DWARF specification,
+    # see http://dwarfstd.org/. Besides naturally expected for this
+    # script platform-specific filtering function, this module adds
+    # three auxiliary synthetic directives not recognized by [GNU]
+    # assembler:
+    #
+    # - .cfi_push to annotate push instructions in prologue, which
+    #   translates to .cfi_adjust_cfa_offset (if needed) and
+    #   .cfi_offset;
+    # - .cfi_pop to annotate pop instructions in epilogue, which
+    #   translates to .cfi_adjust_cfa_offset (if needed) and
+    #   .cfi_restore;
+    # - [and most notably] .cfi_cfa_expression which encodes
+    #   DW_CFA_def_cfa_expression and passes it to .cfi_escape as
+    #   byte vector;
+    #
+    # CFA expressions were introduced in DWARF specification version
+    # 3 and describe how to deduce CFA, Canonical Frame Address. This
+    # becomes handy if your stack frame is variable and you can't
+    # spare register for [previous] frame pointer. Suggested directive
+    # syntax is made-up mix of DWARF operator suffixes [subset of]
+    # and references to registers with optional bias. Following example
+    # describes offloaded *original* stack pointer at specific offset
+    # from *current* stack pointer:
+    #
+    #   .cfi_cfa_expression     %rsp+40,deref,+8
+    #
+    # Final +8 has everything to do with the fact that CFA is defined
+    # as reference to top of caller's stack, and on x86_64 call to
+    # subroutine pushes 8-byte return address. In other words original
+    # stack pointer upon entry to a subroutine is 8 bytes off from CFA.
+
+    # Below constants are taken from "DWARF Expressions" section of the
+    # DWARF specification, section is numbered 7.7 in versions 3 and 4.
+    my %DW_OP_simple = (       # no-arg operators, mapped directly
+       deref   => 0x06,        dup     => 0x12,
+       drop    => 0x13,        over    => 0x14,
+       pick    => 0x15,        swap    => 0x16,
+       rot     => 0x17,        xderef  => 0x18,
+
+       abs     => 0x19,        and     => 0x1a,
+       div     => 0x1b,        minus   => 0x1c,
+       mod     => 0x1d,        mul     => 0x1e,
+       neg     => 0x1f,        not     => 0x20,
+       or      => 0x21,        plus    => 0x22,
+       shl     => 0x24,        shr     => 0x25,
+       shra    => 0x26,        xor     => 0x27,
+       );
+
+    my %DW_OP_complex = (      # used in specific subroutines
+       constu          => 0x10,        # uleb128
+       consts          => 0x11,        # sleb128
+       plus_uconst     => 0x23,        # uleb128
+       lit0            => 0x30,        # add 0-31 to opcode
+       reg0            => 0x50,        # add 0-31 to opcode
+       breg0           => 0x70,        # add 0-31 to opcole, sleb128
+       regx            => 0x90,        # uleb28
+       fbreg           => 0x91,        # sleb128
+       bregx           => 0x92,        # uleb128, sleb128
+       piece           => 0x93,        # uleb128
+       );
+
+    # Following constants are defined in x86_64 ABI supplement, for
+    # example avaiable at https://www.uclibc.org/docs/psABI-x86_64.pdf,
+    # see section 3.7 "Stack Unwind Algorithm".
+    my %DW_reg_idx = (
+       "%rax"=>0,  "%rdx"=>1,  "%rcx"=>2,  "%rbx"=>3,
+       "%rsi"=>4,  "%rdi"=>5,  "%rbp"=>6,  "%rsp"=>7,
+       "%r8" =>8,  "%r9" =>9,  "%r10"=>10, "%r11"=>11,
+       "%r12"=>12, "%r13"=>13, "%r14"=>14, "%r15"=>15
+       );
+
+    my ($cfa_reg, $cfa_rsp);
+
+    # [us]leb128 format is variable-length integer representation base
+    # 2^128, with most significant bit of each byte being 0 denoting
+    # *last* most significat digit. See "Variable Length Data" in the
+    # DWARF specification, numbered 7.6 at least in versions 3 and 4.
+    sub sleb128 {
+       use integer;    # get right shift extend sign
+
+       my $val = shift;
+       my $sign = ($val < 0) ? -1 : 0;
+       my @ret = ();
+
+       while(1) {
+           push @ret, $val&0x7f;
+
+           # see if remaining bits are same and equal to most
+           # significant bit of the current digit, if so, it's
+           # last digit...
+           last if (($val>>6) == $sign);
+
+           @ret[-1] |= 0x80;
+           $val >>= 7;
+       }
+
+       return @ret;
+    }
+    sub uleb128 {
+       my $val = shift;
+       my @ret = ();
+
+       while(1) {
+           push @ret, $val&0x7f;
+
+           # see if it's last significant digit...
+           last if (($val >>= 7) == 0);
+
+           @ret[-1] |= 0x80;
+       }
+
+       return @ret;
+    }
+    sub const {
+       my $val = shift;
+
+       if ($val >= 0 && $val < 32) {
+            return ($DW_OP_complex{lit0}+$val);
+       }
+       return ($DW_OP_complex{consts}, sleb128($val));
+    }
+    sub reg {
+       my $val = shift;
+
+       return if ($val !~ m/^(%r\w+)(?:([\+\-])((?:0x)?[0-9a-f]+))?/);
+
+       my $reg = $DW_reg_idx{$1};
+       my $off = eval ("0 $2 $3");
+
+       return (($DW_OP_complex{breg0} + $reg), sleb128($off));
+       # Yes, we use DW_OP_bregX+0 to push register value and not
+       # DW_OP_regX, because latter would require even DW_OP_piece,
+       # which would be a waste under the circumstances. If you have
+       # to use DWP_OP_reg, use "regx:N"...
+    }
+    sub cfa_expression {
+       my $line = shift;
+       my @ret;
+
+       foreach my $token (split(/,\s*/,$line)) {
+           if ($token =~ /^%r/) {
+               push @ret,reg($token);
+           } elsif ($token =~ /(\w+):(\-?(?:0x)?[0-9a-f]+)(U?)/i) {
+               my $i = 1*eval($2);
+               push @ret,$DW_OP_complex{$1}, ($3 ? uleb128($i) : sleb128($i));
+           } elsif (my $i = 1*eval($token) or $token eq "0") {
+               if ($token =~ /^\+/) {
+                   push @ret,$DW_OP_complex{plus_uconst},uleb128($i);
+               } else {
+                   push @ret,const($i);
+               }
+           } else {
+               push @ret,$DW_OP_simple{$token};
+           }
+       }
+
+       # Finally we return DW_CFA_def_cfa_expression, 15, followed by
+       # length of the expression and of course the expression itself.
+       return (15,scalar(@ret),@ret);
+    }
+    sub re {
+       my      ($class, $line) = @_;
+       my      $self = {};
+       my      $ret;
+
+       if ($$line =~ s/^\s*\.cfi_(\w+)\s+//) {
+           bless $self,$class;
+           $ret = $self;
+           undef $self->{value};
+           my $dir = $1;
+
+           SWITCH: for ($dir) {
+           # What is $cfa_rsp? Effectively it's difference between %rsp
+           # value and current CFA, Canonical Frame Address, which is
+           # why it starts with -8. Recall that CFA is top of caller's
+           # stack...
+           /startproc/ && do { ($cfa_reg, $cfa_rsp) = ("%rsp", -8); last; };
+           /endproc/   && do { ($cfa_reg, $cfa_rsp) = ("%rsp",  0); last; };
+           /def_cfa_register/
+                       && do { $cfa_reg = $$line; last; };
+           /def_cfa_offset/
+                       && do { $cfa_rsp = -1*eval($$line) if ($cfa_reg eq "%rsp");
+                               last;
+                             };
+           /adjust_cfa_offset/
+                       && do { $cfa_rsp -= 1*eval($$line) if ($cfa_reg eq "%rsp");
+                               last;
+                             };
+           /def_cfa/   && do { if ($$line =~ /(%r\w+)\s*,\s*(\.+)/) {
+                                   $cfa_reg = $1;
+                                   $cfa_rsp = -1*eval($2) if ($cfa_reg eq "%rsp");
+                               }
+                               last;
+                             };
+           /push/      && do { $dir = undef;
+                               $cfa_rsp -= 8;
+                               if ($cfa_reg eq "%rsp") {
+                                   $self->{value} = ".cfi_adjust_cfa_offset\t8\n";
+                               }
+                               $self->{value} .= ".cfi_offset\t$$line,$cfa_rsp";
+                               last;
+                             };
+           /pop/       && do { $dir = undef;
+                               $cfa_rsp += 8;
+                               if ($cfa_reg eq "%rsp") {
+                                   $self->{value} = ".cfi_adjust_cfa_offset\t-8\n";
+                               }
+                               $self->{value} .= ".cfi_restore\t$$line";
+                               last;
+                             };
+           /cfa_expression/
+                       && do { $dir = undef;
+                               $self->{value} = ".cfi_escape\t" .
+                                       join(",", map(sprintf("0x%02x", $_),
+                                                     cfa_expression($$line)));
+                               last;
+                             };
+           }
+
+           $self->{value} = ".cfi_$dir\t$$line" if ($dir);
+
+           $$line = "";
+       }
+
+       return $ret;
+    }
+    sub out {
+       my $self = shift;
+       return ($elf ? $self->{value} : undef);
+    }
+}
 { package directive;   # pick up directives, which start with .
     sub re {
        my      ($class, $line) = @_;
@@ -472,6 +706,9 @@ my %globals;
        my      $ret;
        my      $dir;
 
+       # chain-call to cfi_directive
+       $ret = cfi_directive->re($line) and return $ret;
+
        if ($$line =~ /^\s*(\.\w+)/) {
            bless $self,$class;
            $dir = $1;