Subversion Repositories openrisc_2011-10-31

@section mmo backend

The mmo object format is used exclusively together with Professor

Donald E.@: Knuth's educational 64-bit processor MMIX.  The simulator

@command{mmix} which is available at

@url{http://www-cs-faculty.stanford.edu/~knuth/programs/mmix.tar.gz}

understands this format.  That package also includes a combined

assembler and linker called @command{mmixal}.  The mmo format has

no advantages feature-wise compared to e.g. ELF.  It is a simple

non-relocatable object format with no support for archives or

debugging information, except for symbol value information and

line numbers (which is not yet implemented in BFD).  See

@url{http://www-cs-faculty.stanford.edu/~knuth/mmix.html} for more

information about MMIX.  The ELF format is used for intermediate

object files in the BFD implementation.

@c We want to xref the symbol table node.  A feature in "chew"

@c requires that "commands" do not contain spaces in the

@c arguments.  Hence the hyphen in "Symbol-table".

@menu

* File layout::

* Symbol-table::

* mmo section mapping::

@end menu

@node File layout, Symbol-table, mmo, mmo

@subsection File layout

The mmo file contents is not partitioned into named sections as

with e.g.@: ELF.  Memory areas is formed by specifying the

location of the data that follows.  Only the memory area

@samp{0x0000@dots{}00} to @samp{0x01ff@dots{}ff} is executable, so

it is used for code (and constants) and the area

@samp{0x2000@dots{}00} to @samp{0x20ff@dots{}ff} is used for

writable data.  @xref{mmo section mapping}.

There is provision for specifying ``special data'' of 65536

different types.  We use type 80 (decimal), arbitrarily chosen the

same as the ELF @code{e_machine} number for MMIX, filling it with

section information normally found in ELF objects. @xref{mmo

section mapping}.

Contents is entered as 32-bit words, xor:ed over previous

contents, always zero-initialized.  A word that starts with the

byte @samp{0x98} forms a command called a @samp{lopcode}, where

the next byte distinguished between the thirteen lopcodes.  The

two remaining bytes, called the @samp{Y} and @samp{Z} fields, or

the @samp{YZ} field (a 16-bit big-endian number), are used for

various purposes different for each lopcode.  As documented in

@url{http://www-cs-faculty.stanford.edu/~knuth/mmixal-intro.ps.gz},

the lopcodes are:

@table @code

@item lop_quote

0x98000001.  The next word is contents, regardless of whether it

starts with 0x98 or not.

@item lop_loc

0x9801YYZZ, where @samp{Z} is 1 or 2.  This is a location

directive, setting the location for the next data to the next

32-bit word (for @math{Z = 1}) or 64-bit word (for @math{Z = 2}),

plus @math{Y * 2^56}.  Normally @samp{Y} is 0 for the text segment

and 2 for the data segment.

@item lop_skip

0x9802YYZZ.  Increase the current location by @samp{YZ} bytes.

@item lop_fixo

0x9803YYZZ, where @samp{Z} is 1 or 2.  Store the current location

as 64 bits into the location pointed to by the next 32-bit

(@math{Z = 1}) or 64-bit (@math{Z = 2}) word, plus @math{Y *

2^56}.

@item lop_fixr

0x9804YYZZ.  @samp{YZ} is stored into the current location plus

@math{2 - 4 * YZ}.

@item lop_fixrx

0x980500ZZ.  @samp{Z} is 16 or 24.  A value @samp{L} derived from

the following 32-bit word are used in a manner similar to

@samp{YZ} in lop_fixr: it is xor:ed into the current location

minus @math{4 * L}.  The first byte of the word is 0 or 1.  If it

is 1, then @math{L = (@var{lowest 24 bits of word}) - 2^Z}, if 0,

then @math{L = (@var{lowest 24 bits of word})}.

@item lop_file

0x9806YYZZ.  @samp{Y} is the file number, @samp{Z} is count of

32-bit words.  Set the file number to @samp{Y} and the line

counter to 0.  The next @math{Z * 4} bytes contain the file name,

padded with zeros if the count is not a multiple of four.  The

same @samp{Y} may occur multiple times, but @samp{Z} must be 0 for

all but the first occurrence.

@item lop_line

0x9807YYZZ.  @samp{YZ} is the line number.  Together with

lop_file, it forms the source location for the next 32-bit word.

Note that for each non-lopcode 32-bit word, line numbers are

assumed incremented by one.

@item lop_spec

0x9808YYZZ.  @samp{YZ} is the type number.  Data until the next

lopcode other than lop_quote forms special data of type @samp{YZ}.

@xref{mmo section mapping}.

Other types than 80, (or type 80 with a content that does not

parse) is stored in sections named @code{.MMIX.spec_data.@var{n}}

where @var{n} is the @samp{YZ}-type.  The flags for such a

sections say not to allocate or load the data.  The vma is 0.

Contents of multiple occurrences of special data @var{n} is

concatenated to the data of the previous lop_spec @var{n}s.  The

location in data or code at which the lop_spec occurred is lost.

@item lop_pre

0x980901ZZ.  The first lopcode in a file.  The @samp{Z} field forms the

length of header information in 32-bit words, where the first word

tells the time in seconds since @samp{00:00:00 GMT Jan 1 1970}.

@item lop_post

0x980a00ZZ.  @math{Z > 32}.  This lopcode follows after all

content-generating lopcodes in a program.  The @samp{Z} field

denotes the value of @samp{rG} at the beginning of the program.

The following @math{256 - Z} big-endian 64-bit words are loaded

into global registers @samp{$G} @dots{} @samp{$255}.

@item lop_stab

0x980b0000.  The next-to-last lopcode in a program.  Must follow

immediately after the lop_post lopcode and its data.  After this

lopcode follows all symbols in a compressed format

(@pxref{Symbol-table}).

@item lop_end

0x980cYYZZ.  The last lopcode in a program.  It must follow the

lop_stab lopcode and its data.  The @samp{YZ} field contains the

number of 32-bit words of symbol table information after the

preceding lop_stab lopcode.

@end table

Note that the lopcode "fixups"; @code{lop_fixr}, @code{lop_fixrx} and

@code{lop_fixo} are not generated by BFD, but are handled.  They are

generated by @code{mmixal}.

This trivial one-label, one-instruction file:

@example

 :Main TRAP 1,2,3

@end example

can be represented this way in mmo:

@example

 0x98090101 - lop_pre, one 32-bit word with timestamp.

 <timestamp>

 0x98010002 - lop_loc, text segment, using a 64-bit address.

              Note that mmixal does not emit this for the file above.

 0x00000000 - Address, high 32 bits.

 0x00000000 - Address, low 32 bits.

 0x98060002 - lop_file, 2 32-bit words for file-name.

 0x74657374 - "test"

 0x2e730000 - ".s\0\0"

 0x98070001 - lop_line, line 1.

 0x00010203 - TRAP 1,2,3

 0x980a00ff - lop_post, setting $255 to 0.

 0x00000000

 0x00000000

 0x980b0000 - lop_stab for ":Main" = 0, serial 1.

 0x203a4040   @xref{Symbol-table}.

 0x10404020

 0x4d206120

 0x69016e00

 0x81000000

 0x980c0005 - lop_end; symbol table contained five 32-bit words.

@end example

@node Symbol-table, mmo section mapping, File layout, mmo

@subsection Symbol table format

From mmixal.w (or really, the generated mmixal.tex) in

@url{http://www-cs-faculty.stanford.edu/~knuth/programs/mmix.tar.gz}):

``Symbols are stored and retrieved by means of a @samp{ternary

search trie}, following ideas of Bentley and Sedgewick. (See

ACM--SIAM Symp.@: on Discrete Algorithms @samp{8} (1997), 360--369;

R.@:Sedgewick, @samp{Algorithms in C} (Reading, Mass.@:

Addison--Wesley, 1998), @samp{15.4}.)  Each trie node stores a

character, and there are branches to subtries for the cases where

a given character is less than, equal to, or greater than the

character in the trie.  There also is a pointer to a symbol table

entry if a symbol ends at the current node.''

So it's a tree encoded as a stream of bytes.  The stream of bytes

acts on a single virtual global symbol, adding and removing

characters and signalling complete symbol points.  Here, we read

the stream and create symbols at the completion points.

First, there's a control byte @code{m}.  If any of the listed bits

in @code{m} is nonzero, we execute what stands at the right, in

the listed order:

@example

 (MMO3_LEFT)

 0x40 - Traverse left trie.

        (Read a new command byte and recurse.)

 (MMO3_SYMBITS)

 0x2f - Read the next byte as a character and store it in the

        current character position; increment character position.

        Test the bits of @code{m}:

        (MMO3_WCHAR)

        0x80 - The character is 16-bit (so read another byte,

               merge into current character.

        (MMO3_TYPEBITS)

        0xf  - We have a complete symbol; parse the type, value

               and serial number and do what should be done

               with a symbol.  The type and length information

               is in j = (m & 0xf).

               (MMO3_REGQUAL_BITS)

               j == 0xf: A register variable.  The following

                         byte tells which register.

               j <= 8:   An absolute symbol.  Read j bytes as the

                         big-endian number the symbol equals.

                         A j = 2 with two zero bytes denotes an

                         unknown symbol.

               j > 8:    As with j <= 8, but add (0x20 << 56)

                         to the value in the following j - 8

                         bytes.

               Then comes the serial number, as a variant of

               uleb128, but better named ubeb128:

               Read bytes and shift the previous value left 7

               (multiply by 128).  Add in the new byte, repeat

               until a byte has bit 7 set.  The serial number

               is the computed value minus 128.

        (MMO3_MIDDLE)

        0x20 - Traverse middle trie.  (Read a new command byte

               and recurse.)  Decrement character position.

 (MMO3_RIGHT)

 0x10 - Traverse right trie.  (Read a new command byte and

        recurse.)

@end example

Let's look again at the @code{lop_stab} for the trivial file

(@pxref{File layout}).

@example

 0x980b0000 - lop_stab for ":Main" = 0, serial 1.

 0x203a4040

 0x10404020

 0x4d206120

 0x69016e00

 0x81000000

@end example

This forms the trivial trie (note that the path between ``:'' and

``M'' is redundant):

@example

 203a     ":"

 40       /

 40      /

 10      \

 40      /

 40     /

 204d  "M"

 2061  "a"

 2069  "i"

 016e  "n" is the last character in a full symbol, and

       with a value represented in one byte.

 00    The value is 0.

 81    The serial number is 1.

@end example

@node mmo section mapping, , Symbol-table, mmo

@subsection mmo section mapping

The implementation in BFD uses special data type 80 (decimal) to

encapsulate and describe named sections, containing e.g.@: debug

information.  If needed, any datum in the encapsulation will be

quoted using lop_quote.  First comes a 32-bit word holding the

number of 32-bit words containing the zero-terminated zero-padded

segment name.  After the name there's a 32-bit word holding flags

describing the section type.  Then comes a 64-bit big-endian word

with the section length (in bytes), then another with the section

start address.  Depending on the type of section, the contents

might follow, zero-padded to 32-bit boundary.  For a loadable

section (such as data or code), the contents might follow at some

later point, not necessarily immediately, as a lop_loc with the

same start address as in the section description, followed by the

contents.  This in effect forms a descriptor that must be emitted

before the actual contents.  Sections described this way must not

overlap.

For areas that don't have such descriptors, synthetic sections are

formed by BFD.  Consecutive contents in the two memory areas

@samp{0x0000@dots{}00} to @samp{0x01ff@dots{}ff} and

@samp{0x2000@dots{}00} to @samp{0x20ff@dots{}ff} are entered in

sections named @code{.text} and @code{.data} respectively.  If an area

is not otherwise described, but would together with a neighboring

lower area be less than @samp{0x40000000} bytes long, it is joined

with the lower area and the gap is zero-filled.  For other cases,

a new section is formed, named @code{.MMIX.sec.@var{n}}.  Here,

@var{n} is a number, a running count through the mmo file,

starting at 0.

A loadable section specified as:

@example

 .section secname,"ax"

 TETRA 1,2,3,4,-1,-2009

 BYTE 80

@end example

and linked to address @samp{0x4}, is represented by the sequence:

@example

 0x98080050 - lop_spec 80

 0x00000002 - two 32-bit words for the section name

 0x7365636e - "secn"

 0x616d6500 - "ame\0"

 0x00000033 - flags CODE, READONLY, LOAD, ALLOC

 0x00000000 - high 32 bits of section length

 0x0000001c - section length is 28 bytes; 6 * 4 + 1 + alignment to 32 bits

 0x00000000 - high 32 bits of section address

 0x00000004 - section address is 4

 0x98010002 - 64 bits with address of following data

 0x00000000 - high 32 bits of address

 0x00000004 - low 32 bits: data starts at address 4

 0x00000001 - 1

 0x00000002 - 2

 0x00000003 - 3

 0x00000004 - 4

 0xffffffff - -1

 0xfffff827 - -2009

 0x50000000 - 80 as a byte, padded with zeros.

@end example

Note that the lop_spec wrapping does not include the section

contents.  Compare this to a non-loaded section specified as:

@example

 .section thirdsec

 TETRA 200001,100002

 BYTE 38,40

@end example

This, when linked to address @samp{0x200000000000001c}, is

represented by:

@example

 0x98080050 - lop_spec 80

 0x00000002 - two 32-bit words for the section name

 0x7365636e - "thir"

 0x616d6500 - "dsec"

 0x00000010 - flag READONLY

 0x00000000 - high 32 bits of section length

 0x0000000c - section length is 12 bytes; 2 * 4 + 2 + alignment to 32 bits

 0x20000000 - high 32 bits of address

 0x0000001c - low 32 bits of address 0x200000000000001c

 0x00030d41 - 200001

 0x000186a2 - 100002

 0x26280000 - 38, 40 as bytes, padded with zeros

@end example

For the latter example, the section contents must not be

loaded in memory, and is therefore specified as part of the

special data.  The address is usually unimportant but might

provide information for e.g.@: the DWARF 2 debugging format.