First proper commit.

3 files changed, 1106 insertions, 0 deletions

diff --git a/tools/asm_processor/asm_processor.py b/tools/asm_processor/asm_processor.py
new file mode 100644
index 000000000..1ff25db81
--- /dev/null
+++ b/tools/asm_processor/asm_processor.py
@@ -0,0 +1,1050 @@
+#!/usr/bin/env python3
+import argparse
+import tempfile
+import struct
+import copy
+import sys
+import re
+import os
+from collections import namedtuple
+
+MAX_FN_SIZE = 100
+
+EI_NIDENT     = 16
+EI_CLASS      = 4
+EI_DATA       = 5
+EI_VERSION    = 6
+EI_OSABI      = 7
+EI_ABIVERSION = 8
+STN_UNDEF = 0
+
+SHN_UNDEF     = 0
+SHN_ABS       = 0xfff1
+SHN_COMMON    = 0xfff2
+SHN_XINDEX    = 0xffff
+SHN_LORESERVE = 0xff00
+
+STT_NOTYPE  = 0
+STT_OBJECT  = 1
+STT_FUNC    = 2
+STT_SECTION = 3
+STT_FILE    = 4
+STT_COMMON  = 5
+STT_TLS     = 6
+
+STB_LOCAL  = 0
+STB_GLOBAL = 1
+STB_WEAK   = 2
+
+STV_DEFAULT   = 0
+STV_INTERNAL  = 1
+STV_HIDDEN    = 2
+STV_PROTECTED = 3
+
+SHT_NULL          = 0
+SHT_PROGBITS      = 1
+SHT_SYMTAB        = 2
+SHT_STRTAB        = 3
+SHT_RELA          = 4
+SHT_HASH          = 5
+SHT_DYNAMIC       = 6
+SHT_NOTE          = 7
+SHT_NOBITS        = 8
+SHT_REL           = 9
+SHT_SHLIB         = 10
+SHT_DYNSYM        = 11
+SHT_INIT_ARRAY    = 14
+SHT_FINI_ARRAY    = 15
+SHT_PREINIT_ARRAY = 16
+SHT_GROUP         = 17
+SHT_SYMTAB_SHNDX  = 18
+SHT_MIPS_GPTAB    = 0x70000003
+SHT_MIPS_DEBUG    = 0x70000005
+SHT_MIPS_REGINFO  = 0x70000006
+SHT_MIPS_OPTIONS  = 0x7000000d
+
+SHF_WRITE            = 0x1
+SHF_ALLOC            = 0x2
+SHF_EXECINSTR        = 0x4
+SHF_MERGE            = 0x10
+SHF_STRINGS          = 0x20
+SHF_INFO_LINK        = 0x40
+SHF_LINK_ORDER       = 0x80
+SHF_OS_NONCONFORMING = 0x100
+SHF_GROUP            = 0x200
+SHF_TLS              = 0x400
+
+R_MIPS_32   = 2
+R_MIPS_26   = 4
+R_MIPS_HI16 = 5
+R_MIPS_LO16 = 6
+
+
+class ElfHeader:
+    """
+    typedef struct {
+        unsigned char   e_ident[EI_NIDENT];
+        Elf32_Half      e_type;
+        Elf32_Half      e_machine;
+        Elf32_Word      e_version;
+        Elf32_Addr      e_entry;
+        Elf32_Off       e_phoff;
+        Elf32_Off       e_shoff;
+        Elf32_Word      e_flags;
+        Elf32_Half      e_ehsize;
+        Elf32_Half      e_phentsize;
+        Elf32_Half      e_phnum;
+        Elf32_Half      e_shentsize;
+        Elf32_Half      e_shnum;
+        Elf32_Half      e_shstrndx;
+    } Elf32_Ehdr;
+    """
+
+    def __init__(self, data):
+        self.e_ident = data[:EI_NIDENT]
+        self.e_type, self.e_machine, self.e_version, self.e_entry, self.e_phoff, self.e_shoff, self.e_flags, self.e_ehsize, self.e_phentsize, self.e_phnum, self.e_shentsize, self.e_shnum, self.e_shstrndx = struct.unpack('>HHIIIIIHHHHHH', data[EI_NIDENT:])
+        assert self.e_ident[EI_CLASS] == 1 # 32-bit
+        assert self.e_ident[EI_DATA] == 2 # big-endian
+        assert self.e_type == 1 # relocatable
+        assert self.e_machine == 8 # MIPS I Architecture
+        assert self.e_phoff == 0 # no program header
+        assert self.e_shoff != 0 # section header
+        assert self.e_shstrndx != SHN_UNDEF
+
+    def to_bin(self):
+        return self.e_ident + struct.pack('>HHIIIIIHHHHHH', self.e_type,
+                self.e_machine, self.e_version, self.e_entry, self.e_phoff,
+                self.e_shoff, self.e_flags, self.e_ehsize, self.e_phentsize,
+                self.e_phnum, self.e_shentsize, self.e_shnum, self.e_shstrndx)
+
+
+class Symbol:
+    """
+    typedef struct {
+        Elf32_Word      st_name;
+        Elf32_Addr      st_value;
+        Elf32_Word      st_size;
+        unsigned char   st_info;
+        unsigned char   st_other;
+        Elf32_Half      st_shndx;
+    } Elf32_Sym;
+    """
+
+    def __init__(self, data, strtab):
+        self.st_name, self.st_value, self.st_size, st_info, self.st_other, self.st_shndx = struct.unpack('>IIIBBH', data)
+        assert self.st_shndx != SHN_XINDEX, "too many sections (SHN_XINDEX not supported)"
+        self.bind = st_info >> 4
+        self.type = st_info & 15
+        self.name = strtab.lookup_str(self.st_name)
+        self.visibility = self.st_other & 3
+
+    def to_bin(self):
+        st_info = (self.bind << 4) | self.type
+        return struct.pack('>IIIBBH', self.st_name, self.st_value, self.st_size, st_info, self.st_other, self.st_shndx)
+
+
+class Relocation:
+    def __init__(self, data, sh_type):
+        self.sh_type = sh_type
+        if sh_type == SHT_REL:
+            self.r_offset, self.r_info = struct.unpack('>II', data)
+        else:
+            self.r_offset, self.r_info, self.r_addend = struct.unpack('>III', data)
+        self.sym_index = self.r_info >> 8
+        self.rel_type = self.r_info & 0xff
+
+    def to_bin(self):
+        self.r_info = (self.sym_index << 8) | self.rel_type
+        if self.sh_type == SHT_REL:
+            return struct.pack('>II', self.r_offset, self.r_info)
+        else:
+            return struct.pack('>III', self.r_offset, self.r_info, self.r_addend)
+
+
+class Section:
+    """
+    typedef struct {
+        Elf32_Word   sh_name;
+        Elf32_Word   sh_type;
+        Elf32_Word   sh_flags;
+        Elf32_Addr   sh_addr;
+        Elf32_Off    sh_offset;
+        Elf32_Word   sh_size;
+        Elf32_Word   sh_link;
+        Elf32_Word   sh_info;
+        Elf32_Word   sh_addralign;
+        Elf32_Word   sh_entsize;
+    } Elf32_Shdr;
+    """
+
+    def __init__(self, header, data, index):
+        self.sh_name, self.sh_type, self.sh_flags, self.sh_addr, self.sh_offset, self.sh_size, self.sh_link, self.sh_info, self.sh_addralign, self.sh_entsize = struct.unpack('>IIIIIIIIII', header)
+        assert not self.sh_flags & SHF_LINK_ORDER
+        if self.sh_entsize != 0:
+            assert self.sh_size % self.sh_entsize == 0
+        if self.sh_type == SHT_NOBITS:
+            self.data = ''
+        else:
+            self.data = data[self.sh_offset:self.sh_offset + self.sh_size]
+        self.index = index
+        self.relocated_by = []
+
+    @staticmethod
+    def from_parts(sh_name, sh_type, sh_flags, sh_link, sh_info, sh_addralign, sh_entsize, data, index):
+        header = struct.pack('>IIIIIIIIII', sh_name, sh_type, sh_flags, 0, 0, len(data), sh_link, sh_info, sh_addralign, sh_entsize)
+        return Section(header, data, index)
+
+    def lookup_str(self, index):
+        assert self.sh_type == SHT_STRTAB
+        to = self.data.find(b'\0', index)
+        assert to != -1
+        return self.data[index:to].decode('latin1')
+
+    def add_str(self, string):
+        assert self.sh_type == SHT_STRTAB
+        ret = len(self.data)
+        self.data += string.encode('latin1') + b'\0'
+        return ret
+
+    def is_rel(self):
+        return self.sh_type == SHT_REL or self.sh_type == SHT_RELA
+
+    def header_to_bin(self):
+        if self.sh_type != SHT_NOBITS:
+            self.sh_size = len(self.data)
+        return struct.pack('>IIIIIIIIII', self.sh_name, self.sh_type, self.sh_flags, self.sh_addr, self.sh_offset, self.sh_size, self.sh_link, self.sh_info, self.sh_addralign, self.sh_entsize)
+
+    def late_init(self, sections):
+        if self.sh_type == SHT_SYMTAB:
+            self.init_symbols(sections)
+        elif self.is_rel():
+            self.rel_target = sections[self.sh_info]
+            self.rel_target.relocated_by.append(self)
+            self.init_relocs()
+
+    def find_symbol(self, name):
+        assert self.sh_type == SHT_SYMTAB
+        for s in self.symbol_entries:
+            if s.name == name:
+                return (s.st_shndx, s.st_value)
+        return None
+
+    def find_symbol_in_section(self, name, section):
+        pos = self.find_symbol(name)
+        assert pos is not None
+        assert pos[0] == section.index
+        return pos[1]
+
+    def init_symbols(self, sections):
+        assert self.sh_type == SHT_SYMTAB
+        assert self.sh_entsize == 16
+        self.strtab = sections[self.sh_link]
+        entries = []
+        for i in range(0, self.sh_size, self.sh_entsize):
+            entries.append(Symbol(self.data[i:i+self.sh_entsize], self.strtab))
+        self.symbol_entries = entries
+
+    def init_relocs(self):
+        assert self.is_rel()
+        entries = []
+        for i in range(0, self.sh_size, self.sh_entsize):
+            entries.append(Relocation(self.data[i:i+self.sh_entsize], self.sh_type))
+        self.relocations = entries
+
+    def local_symbols(self):
+        assert self.sh_type == SHT_SYMTAB
+        return self.symbol_entries[:self.sh_info]
+
+    def global_symbols(self):
+        assert self.sh_type == SHT_SYMTAB
+        return self.symbol_entries[self.sh_info:]
+
+
+class ElfFile:
+    def __init__(self, data):
+        self.data = data
+        assert data[:4] == b'\x7fELF', "not an ELF file"
+
+        self.elf_header = ElfHeader(data[0:52])
+
+        offset, size = self.elf_header.e_shoff, self.elf_header.e_shentsize
+        null_section = Section(data[offset:offset + size], data, 0)
+        num_sections = self.elf_header.e_shnum or null_section.sh_size
+
+        self.sections = [null_section]
+        for i in range(1, num_sections):
+            ind = offset + i * size
+            self.sections.append(Section(data[ind:ind + size], data, i))
+
+        symtab = None
+        for s in self.sections:
+            if s.sh_type == SHT_SYMTAB:
+                assert not symtab
+                symtab = s
+        assert symtab is not None
+        self.symtab = symtab
+
+        shstr = self.sections[self.elf_header.e_shstrndx]
+        for s in self.sections:
+            s.name = shstr.lookup_str(s.sh_name)
+            s.late_init(self.sections)
+
+    def find_section(self, name):
+        for s in self.sections:
+            if s.name == name:
+                return s
+        return None
+
+    def add_section(self, name, sh_type, sh_flags, sh_link, sh_info, sh_addralign, sh_entsize, data):
+        shstr = self.sections[self.elf_header.e_shstrndx]
+        sh_name = shstr.add_str(name)
+        s = Section.from_parts(sh_name=sh_name, sh_type=sh_type,
+                sh_flags=sh_flags, sh_link=sh_link, sh_info=sh_info,
+                sh_addralign=sh_addralign, sh_entsize=sh_entsize, data=data,
+                index=len(self.sections))
+        self.sections.append(s)
+        s.name = name
+        s.late_init(self.sections)
+        return s
+
+    def drop_irrelevant_sections(self):
+        # We can only drop sections at the end, since otherwise section
+        # references might be wrong. Luckily, these sections typically are.
+        while self.sections[-1].sh_type in [SHT_MIPS_DEBUG, SHT_MIPS_GPTAB]:
+            self.sections.pop()
+
+    def write(self, filename):
+        outfile = open(filename, 'wb')
+        outidx = 0
+        def write_out(data):
+            nonlocal outidx
+            outfile.write(data)
+            outidx += len(data)
+        def pad_out(align):
+            if align and outidx % align:
+                write_out(b'\0' * (align - outidx % align))
+
+        self.elf_header.e_shnum = len(self.sections)
+        write_out(self.elf_header.to_bin())
+
+        for s in self.sections:
+            if s.sh_type != SHT_NOBITS and s.sh_type != SHT_NULL:
+                pad_out(s.sh_addralign)
+                s.sh_offset = outidx
+                write_out(s.data)
+
+        pad_out(4)
+        self.elf_header.e_shoff = outidx
+        for s in self.sections:
+            write_out(s.header_to_bin())
+
+        outfile.seek(0)
+        outfile.write(self.elf_header.to_bin())
+        outfile.close()
+
+
+def is_temp_name(name):
+    return name.startswith('_asmpp_')
+
+
+class Failure(Exception):
+    def __init__(self, message):
+        self.message = message
+
+    def __str__(self):
+        return self.message
+
+
+class GlobalState:
+    def __init__(self, min_instr_count, skip_instr_count):
+        # A value that hopefully never appears as a 32-bit rodata constant (or we
+        # miscompile late rodata). Increases by 1 in each step.
+        self.late_rodata_hex = 0xE0123456
+        self.namectr = 0
+        self.min_instr_count = min_instr_count
+        self.skip_instr_count = skip_instr_count
+
+    def next_late_rodata_hex(self):
+        dummy_bytes = struct.pack('>I', self.late_rodata_hex)
+        if (self.late_rodata_hex & 0xffff) == 0:
+            # Avoid lui
+            self.late_rodata_hex += 1
+        self.late_rodata_hex += 1
+        return dummy_bytes
+
+    def make_name(self, cat):
+        self.namectr += 1
+        return '_asmpp_{}{}'.format(cat, self.namectr)
+
+
+Function = namedtuple('Function', ['text_glabels', 'asm_conts', 'late_rodata_dummy_bytes', 'late_rodata_asm_conts', 'fn_desc', 'data'])
+
+
+class GlobalAsmBlock:
+    def __init__(self, fn_desc):
+        self.fn_desc = fn_desc
+        self.cur_section = '.text'
+        self.asm_conts = []
+        self.late_rodata_asm_conts = []
+        self.late_rodata_alignment = 0
+        self.late_rodata_alignment_from_content = False
+        self.text_glabels = []
+        self.fn_section_sizes = {
+            '.text': 0,
+            '.data': 0,
+            '.bss': 0,
+            '.rodata': 0,
+            '.late_rodata': 0,
+        }
+        self.fn_ins_inds = []
+        self.glued_line = ''
+        self.num_lines = 0
+
+    def fail(self, message, line=None):
+        context = self.fn_desc
+        if line:
+            context += ", at line \"" + line + "\""
+        raise Failure(message + "\nwithin " + context)
+
+    def count_quoted_size(self, line, z, real_line, output_enc):
+        line = line.encode(output_enc).decode('latin1')
+        in_quote = False
+        num_parts = 0
+        ret = 0
+        i = 0
+        digits = "0123456789" # 0-7 would be more sane, but this matches GNU as
+        while i < len(line):
+            c = line[i]
+            i += 1
+            if not in_quote:
+                if c == '"':
+                    in_quote = True
+                    num_parts += 1
+            else:
+                if c == '"':
+                    in_quote = False
+                    continue
+                ret += 1
+                if c != '\\':
+                    continue
+                if i == len(line):
+                    self.fail("backslash at end of line not supported", real_line)
+                c = line[i]
+                i += 1
+                # (if c is in "bfnrtv", we have a real escaped literal)
+                if c == 'x':
+                    # hex literal, consume any number of hex chars, possibly none
+                    while i < len(line) and line[i] in digits + "abcdefABCDEF":
+                        i += 1
+                elif c in digits:
+                    # octal literal, consume up to two more digits
+                    it = 0
+                    while i < len(line) and line[i] in digits and it < 2:
+                        i += 1
+                        it += 1
+
+        if in_quote:
+            self.fail("unterminated string literal", real_line)
+        if num_parts == 0:
+            self.fail(".ascii with no string", real_line)
+        return ret + num_parts if z else ret
+
+
+    def align4(self):
+        while self.fn_section_sizes[self.cur_section] % 4 != 0:
+            self.fn_section_sizes[self.cur_section] += 1
+
+    def add_sized(self, size, line):
+        if self.cur_section in ['.text', '.late_rodata']:
+            if size % 4 != 0:
+                self.fail("size must be a multiple of 4", line)
+        if size < 0:
+            self.fail("size cannot be negative", line)
+        self.fn_section_sizes[self.cur_section] += size
+        if self.cur_section == '.text':
+            if not self.text_glabels:
+                self.fail(".text block without an initial glabel", line)
+            self.fn_ins_inds.append((self.num_lines - 1, size // 4))
+
+    def process_line(self, line, output_enc):
+        self.num_lines += 1
+        if line.endswith('\\'):
+            self.glued_line += line[:-1]
+            return
+        line = self.glued_line + line
+        self.glued_line = ''
+
+        real_line = line
+        line = re.sub(r'/\*.*?\*/', '', line)
+        line = re.sub(r'#.*', '', line)
+        line = line.strip()
+        line = re.sub(r'^[a-zA-Z0-9_]+:\s*', '', line)
+        changed_section = False
+        emitting_double = False
+        if line.startswith('glabel ') and self.cur_section == '.text':
+            self.text_glabels.append(line.split()[1])
+        if not line:
+            pass # empty line
+        elif line.startswith('glabel ') or (' ' not in line and line.endswith(':')):
+            pass # label
+        elif line.startswith('.section') or line in ['.text', '.data', '.rdata', '.rodata', '.bss', '.late_rodata']:
+            # section change
+            self.cur_section = '.rodata' if line == '.rdata' else line.split(',')[0].split()[-1]
+            if self.cur_section not in ['.data', '.text', '.rodata', '.late_rodata', '.bss']:
+                self.fail("unrecognized .section directive", real_line)
+            changed_section = True
+        elif line.startswith('.late_rodata_alignment'):
+            if self.cur_section != '.late_rodata':
+                self.fail(".late_rodata_alignment must occur within .late_rodata section", real_line)
+            value = int(line.split()[1])
+            if value not in [4, 8]:
+                self.fail(".late_rodata_alignment argument must be 4 or 8", real_line)
+            if self.late_rodata_alignment and self.late_rodata_alignment != value:
+                self.fail(".late_rodata_alignment alignment assumption conflicts with earlier .double directive. Make sure to provide explicit alignment padding.")
+            self.late_rodata_alignment = value
+            changed_section = True
+        elif line.startswith('.incbin'):
+            self.add_sized(int(line.split(',')[-1].strip(), 0), real_line)
+        elif line.startswith('.word') or line.startswith('.float'):
+            self.align4()
+            self.add_sized(4 * len(line.split(',')), real_line)
+        elif line.startswith('.double'):
+            self.align4()
+            if self.cur_section == '.late_rodata':
+                align8 = self.fn_section_sizes[self.cur_section] % 8
+                # Automatically set late_rodata_alignment, so the generated C code uses doubles.
+                # This gives us correct alignment for the transferred doubles even when the
+                # late_rodata_alignment is wrong, e.g. for non-matching compilation.
+                if not self.late_rodata_alignment:
+                    self.late_rodata_alignment = 8 - align8
+                    self.late_rodata_alignment_from_content = True
+                elif self.late_rodata_alignment != 8 - align8:
+                    if self.late_rodata_alignment_from_content:
+                        self.fail("found two .double directives with different start addresses mod 8. Make sure to provide explicit alignment padding.", real_line)
+                    else:
+                        self.fail(".double at address that is not 0 mod 8 (based on .late_rodata_alignment assumption). Make sure to provide explicit alignment padding.", real_line)
+            self.add_sized(8 * len(line.split(',')), real_line)
+            emitting_double = True
+        elif line.startswith('.space'):
+            self.add_sized(int(line.split()[1], 0), real_line)
+        elif line.startswith('.balign') or line.startswith('.align'):
+            align = int(line.split()[1])
+            if align != 4:
+                self.fail("only .balign 4 is supported", real_line)
+            self.align4()
+        elif line.startswith('.asci'):
+            z = (line.startswith('.asciz') or line.startswith('.asciiz'))
+            self.add_sized(self.count_quoted_size(line, z, real_line, output_enc), real_line)
+        elif line.startswith('.'):
+            # .macro, ...
+            self.fail("asm directive not supported", real_line)
+        else:
+            # Unfortunately, macros are hard to support for .rodata --
+            # we don't know how how space they will expand to before
+            # running the assembler, but we need that information to
+            # construct the C code. So if we need that we'll either
+            # need to run the assembler twice (at least in some rare
+            # cases), or change how this program is invoked.
+            # Similarly, we can't currently deal with pseudo-instructions
+            # that expand to several real instructions.
+            if self.cur_section != '.text':
+                self.fail("instruction or macro call in non-.text section? not supported", real_line)
+            self.add_sized(4, real_line)
+        if self.cur_section == '.late_rodata':
+            if not changed_section:
+                if emitting_double:
+                    self.late_rodata_asm_conts.append(".align 0")
+                self.late_rodata_asm_conts.append(real_line)
+                if emitting_double:
+                    self.late_rodata_asm_conts.append(".align 2")
+        else:
+            self.asm_conts.append(real_line)
+
+    def finish(self, state):
+        src = [''] * (self.num_lines + 1)
+        late_rodata_dummy_bytes = []
+        late_rodata_fn_output = []
+
+        if self.fn_section_sizes['.late_rodata'] > 0:
+            # Generate late rodata by emitting unique float constants.
+            # This requires 3 instructions for each 4 bytes of rodata.
+            # If we know alignment, we can use doubles, which give 3
+            # instructions for 8 bytes of rodata.
+            size = self.fn_section_sizes['.late_rodata'] // 4
+            skip_next = False
+            for i in range(size):
+                if skip_next:
+                    skip_next = False
+                    continue
+                dummy_bytes = state.next_late_rodata_hex()
+                late_rodata_dummy_bytes.append(dummy_bytes)
+                if self.late_rodata_alignment == 4 * ((i + 1) % 2 + 1) and i + 1 < size:
+                    dummy_bytes2 = state.next_late_rodata_hex()
+                    late_rodata_dummy_bytes.append(dummy_bytes2)
+                    fval, = struct.unpack('>d', dummy_bytes + dummy_bytes2)
+                    late_rodata_fn_output.append('*(volatile double*)0 = {};'.format(fval))
+                    skip_next = True
+                else:
+                    fval, = struct.unpack('>f', dummy_bytes)
+                    late_rodata_fn_output.append('*(volatile float*)0 = {}f;'.format(fval))
+                late_rodata_fn_output.append('')
+                late_rodata_fn_output.append('')
+
+        text_name = None
+        if self.fn_section_sizes['.text'] > 0 or late_rodata_fn_output:
+            text_name = state.make_name('func')
+            src[0] = 'void {}(void) {{'.format(text_name)
+            src[self.num_lines] = '}'
+            instr_count = self.fn_section_sizes['.text'] // 4
+            if instr_count < state.min_instr_count:
+                self.fail("too short .text block")
+            tot_emitted = 0
+            tot_skipped = 0
+            fn_emitted = 0
+            fn_skipped = 0
+            rodata_stack = late_rodata_fn_output[::-1]
+            for (line, count) in self.fn_ins_inds:
+                for _ in range(count):
+                    if (fn_emitted > MAX_FN_SIZE and instr_count - tot_emitted > state.min_instr_count and
+                            (not rodata_stack or rodata_stack[-1])):
+                        # Don't let functions become too large. When a function reaches 284
+                        # instructions, and -O2 -framepointer flags are passed, the IRIX
+                        # compiler decides it is a great idea to start optimizing more.
+                        fn_emitted = 0
+                        fn_skipped = 0
+                        src[line] += ' }} void {}(void) {{ '.format(state.make_name('large_func'))
+                    if fn_skipped < state.skip_instr_count:
+                        fn_skipped += 1
+                        tot_skipped += 1
+                    elif rodata_stack:
+                        src[line] += rodata_stack.pop()
+                    else:
+                        src[line] += '*(volatile int*)0 = 0;'
+                    tot_emitted += 1
+                    fn_emitted += 1
+            if rodata_stack:
+                size = len(late_rodata_fn_output) // 3
+                available = instr_count - tot_skipped
+                self.fail(
+                    "late rodata to text ratio is too high: {} / {} must be <= 1/3\n"
+                    "add .late_rodata_alignment (4|8) to the .late_rodata "
+                    "block to double the allowed ratio."
+                        .format(size, available))
+
+        rodata_name = None
+        if self.fn_section_sizes['.rodata'] > 0:
+            rodata_name = state.make_name('rodata')
+            src[self.num_lines] += ' const char {}[{}] = {{1}};'.format(rodata_name, self.fn_section_sizes['.rodata'])
+
+        data_name = None
+        if self.fn_section_sizes['.data'] > 0:
+            data_name = state.make_name('data')
+            src[self.num_lines] += ' char {}[{}] = {{1}};'.format(data_name, self.fn_section_sizes['.data'])
+
+        bss_name = None
+        if self.fn_section_sizes['.bss'] > 0:
+            bss_name = state.make_name('bss')
+            src[self.num_lines] += ' char {}[{}];'.format(bss_name, self.fn_section_sizes['.bss'])
+
+        fn = Function(
+                text_glabels=self.text_glabels,
+                asm_conts=self.asm_conts,
+                late_rodata_dummy_bytes=late_rodata_dummy_bytes,
+                late_rodata_asm_conts=self.late_rodata_asm_conts,
+                fn_desc=self.fn_desc,
+                data={
+                    '.text': (text_name, self.fn_section_sizes['.text']),
+                    '.data': (data_name, self.fn_section_sizes['.data']),
+                    '.rodata': (rodata_name, self.fn_section_sizes['.rodata']),
+                    '.bss': (bss_name, self.fn_section_sizes['.bss']),
+                })
+        return src, fn
+
+def parse_source(f, opt, framepointer, input_enc, output_enc, print_source=None):
+    if opt in ['O2', 'O1']:
+        if framepointer:
+            min_instr_count = 6
+            skip_instr_count = 5
+        else:
+            min_instr_count = 2
+            skip_instr_count = 1
+    elif opt == 'g':
+        if framepointer:
+            min_instr_count = 7
+            skip_instr_count = 7
+        else:
+            min_instr_count = 4
+            skip_instr_count = 4
+    else:
+        if opt != 'g3':
+            raise Failure("must pass one of -g, -O1, -O2, -O2 -g3")
+        if framepointer:
+            min_instr_count = 4
+            skip_instr_count = 4
+        else:
+            min_instr_count = 2
+            skip_instr_count = 2
+
+    state = GlobalState(min_instr_count, skip_instr_count)
+
+    global_asm = None
+    asm_functions = []
+    output_lines = []
+
+    for line_no, raw_line in enumerate(f, 1):
+        raw_line = raw_line.rstrip()
+        line = raw_line.lstrip()
+
+        # Print exactly one output line per source line, to make compiler
+        # errors have correct line numbers. These will be overridden with
+        # reasonable content further down.
+        output_lines.append('')
+
+        if global_asm is not None:
+            if line.startswith(')'):
+                src, fn = global_asm.finish(state)
+                for i, line2 in enumerate(src):
+                    output_lines[start_index + i] = line2
+                asm_functions.append(fn)
+                global_asm = None
+            else:
+                global_asm.process_line(raw_line, output_enc)
+        else:
+            if line in ['GLOBAL_ASM(', '#pragma GLOBAL_ASM(']:
+                global_asm = GlobalAsmBlock("GLOBAL_ASM block at line " + str(line_no))
+                start_index = len(output_lines)
+            elif ((line.startswith('GLOBAL_ASM("') or line.startswith('#pragma GLOBAL_ASM("'))
+                    and line.endswith('")')):
+                fname = line[line.index('(') + 2 : -2]
+                global_asm = GlobalAsmBlock(fname)
+                with open(fname, encoding=input_enc) as f:
+                    for line2 in f:
+                        global_asm.process_line(line2.rstrip(), output_enc)
+                src, fn = global_asm.finish(state)
+                output_lines[-1] = ''.join(src)
+                asm_functions.append(fn)
+                global_asm = None
+            else:
+                output_lines[-1] = raw_line
+
+    if print_source:
+        for line in output_lines:
+            print_source.write(line.encode(output_enc) + b'\n')
+        print_source.flush()
+        if print_source != sys.stdout.buffer:
+            print_source.close()
+
+    return asm_functions
+
+def fixup_objfile(objfile_name, functions, asm_prelude, assembler, output_enc):
+    SECTIONS = ['.data', '.text', '.rodata', '.bss']
+
+    with open(objfile_name, 'rb') as f:
+        objfile = ElfFile(f.read())
+
+    prev_locs = {
+        '.text': 0,
+        '.data': 0,
+        '.rodata': 0,
+        '.bss': 0,
+    }
+    to_copy = {
+        '.text': [],
+        '.data': [],
+        '.rodata': [],
+        '.bss': [],
+    }
+    asm = []
+    late_rodata_dummy_bytes = []
+    late_rodata_asm = []
+    late_rodata_source_name_start = None
+    late_rodata_source_name_end = None
+
+    # Generate an assembly file with all the assembly we need to fill in. For
+    # simplicity we pad with nops/.space so that addresses match exactly, so we
+    # don't have to fix up relocations/symbol references.
+    all_text_glabels = set()
+    for function in functions:
+        ifdefed = False
+        for sectype, (temp_name, size) in function.data.items():
+            if temp_name is None:
+                continue
+            assert size > 0
+            loc = objfile.symtab.find_symbol(temp_name)
+            if loc is None:
+                ifdefed = True
+                break
+            loc = loc[1]
+            prev_loc = prev_locs[sectype]
+            assert loc >= prev_loc, sectype
+            if loc != prev_loc:
+                asm.append('.section ' + sectype)
+                if sectype == '.text':
+                    for i in range((loc - prev_loc) // 4):
+                        asm.append('nop')
+                else:
+                    asm.append('.space {}'.format(loc - prev_loc))
+            to_copy[sectype].append((loc, size, temp_name, function.fn_desc))
+            prev_locs[sectype] = loc + size
+        if not ifdefed:
+            all_text_glabels.update(function.text_glabels)
+            late_rodata_dummy_bytes.append(function.late_rodata_dummy_bytes)
+            late_rodata_asm.append(function.late_rodata_asm_conts)
+            for sectype, (temp_name, size) in function.data.items():
+                if temp_name is not None:
+                    asm.append('.section ' + sectype)
+                    asm.append('glabel ' + temp_name + '_asm_start')
+            asm.append('.text')
+            for line in function.asm_conts:
+                asm.append(line)
+            for sectype, (temp_name, size) in function.data.items():
+                if temp_name is not None:
+                    asm.append('.section ' + sectype)
+                    asm.append('glabel ' + temp_name + '_asm_end')
+    if any(late_rodata_asm):
+        late_rodata_source_name_start = '_asmpp_late_rodata_start'
+        late_rodata_source_name_end = '_asmpp_late_rodata_end'
+        asm.append('.rdata')
+        asm.append('glabel {}'.format(late_rodata_source_name_start))
+        for conts in late_rodata_asm:
+            asm.extend(conts)
+        asm.append('glabel {}'.format(late_rodata_source_name_end))
+
+    o_file = tempfile.NamedTemporaryFile(prefix='asm-processor', suffix='.o', delete=False)
+    o_name = o_file.name
+    o_file.close()
+    s_file = tempfile.NamedTemporaryFile(prefix='asm-processor', suffix='.s', delete=False)
+    s_name = s_file.name
+    try:
+        s_file.write(asm_prelude + b'\n')
+        for line in asm:
+            s_file.write(line.encode(output_enc) + b'\n')
+        s_file.close()
+        ret = os.system(assembler + " " + s_name + " -o " + o_name)
+        if ret != 0:
+            raise Failure("failed to assemble")
+        with open(o_name, 'rb') as f:
+            asm_objfile = ElfFile(f.read())
+
+        # Remove some clutter from objdump output
+        objfile.drop_irrelevant_sections()
+
+        # Unify reginfo sections
+        target_reginfo = objfile.find_section('.reginfo')
+        source_reginfo_data = list(asm_objfile.find_section('.reginfo').data)
+        data = list(target_reginfo.data)
+        for i in range(20):
+            data[i] |= source_reginfo_data[i]
+        target_reginfo.data = bytes(data)
+
+        # Move over section contents
+        modified_text_positions = set()
+        last_rodata_pos = 0
+        for sectype in SECTIONS:
+            if not to_copy[sectype]:
+                continue
+            source = asm_objfile.find_section(sectype)
+            assert source is not None, "didn't find source section: " + sectype
+            for (pos, count, temp_name, fn_desc) in to_copy[sectype]:
+                loc1 = asm_objfile.symtab.find_symbol_in_section(temp_name + '_asm_start', source)
+                loc2 = asm_objfile.symtab.find_symbol_in_section(temp_name + '_asm_end', source)
+                assert loc1 == pos, "assembly and C files don't line up for section " + sectype + ", " + fn_desc
+                if loc2 - loc1 != count:
+                    raise Failure("incorrectly computed size for section " + sectype + ", " + fn_desc + ". If using .double, make sure to provide explicit alignment padding.")
+            if sectype == '.bss':
+                continue
+            target = objfile.find_section(sectype)
+            assert target is not None, "missing target section of type " + sectype
+            data = list(target.data)
+            for (pos, count, _, _) in to_copy[sectype]:
+                data[pos:pos + count] = source.data[pos:pos + count]
+                if sectype == '.text':
+                    assert count % 4 == 0
+                    assert pos % 4 == 0
+                    for i in range(count // 4):
+                        modified_text_positions.add(pos + 4 * i)
+                elif sectype == '.rodata':
+                    last_rodata_pos = pos + count
+            target.data = bytes(data)
+
+        # Move over late rodata. This is heuristic, sadly, since I can't think
+        # of another way of doing it.
+        moved_late_rodata = {}
+        if any(late_rodata_dummy_bytes):
+            source = asm_objfile.find_section('.rodata')
+            target = objfile.find_section('.rodata')
+            source_pos = asm_objfile.symtab.find_symbol_in_section(late_rodata_source_name_start, source)
+            source_end = asm_objfile.symtab.find_symbol_in_section(late_rodata_source_name_end, source)
+            if source_end - source_pos != sum(map(len, late_rodata_dummy_bytes)) * 4:
+                raise Failure("computed wrong size of .late_rodata")
+            new_data = list(target.data)
+            for dummy_bytes_list in late_rodata_dummy_bytes:
+                for index, dummy_bytes in enumerate(dummy_bytes_list):
+                    pos = target.data.index(dummy_bytes, last_rodata_pos)
+                    if target.data.find(dummy_bytes, pos + 4) != -1:
+                        raise Failure("multiple occurrences of late_rodata hex magic. Change asm-processor to use something better than 0xE0123456!")
+                    if index == 0 and len(dummy_bytes_list) > 1 and target.data[pos+4:pos+8] == b'\0\0\0\0':
+                        # Ugly hack to handle double alignment for non-matching builds.
+                        # We were told by .late_rodata_alignment (or deduced from a .double)
+                        # that a function's late_rodata started out 4 (mod 8), and emitted
+                        # a float and then a double. But it was actually 0 (mod 8), so our
+                        # double was moved by 4 bytes. To make them adjacent to keep jump
+                        # tables correct, move the float by 4 bytes as well.
+                        new_data[pos:pos+4] = b'\0\0\0\0'
+                        pos += 4
+                    new_data[pos:pos+4] = source.data[source_pos:source_pos+4]
+                    moved_late_rodata[source_pos] = pos
+                    last_rodata_pos = pos + 4
+                    source_pos += 4
+            target.data = bytes(new_data)
+
+        # Merge strtab data.
+        strtab_adj = len(objfile.symtab.strtab.data)
+        objfile.symtab.strtab.data += asm_objfile.symtab.strtab.data
+
+        # Find relocated symbols
+        relocated_symbols = set()
+        for sectype in SECTIONS:
+            for obj in [asm_objfile, objfile]:
+                sec = obj.find_section(sectype)
+                if sec is None:
+                    continue
+                for reltab in sec.relocated_by:
+                    for rel in reltab.relocations:
+                        relocated_symbols.add(obj.symtab.symbol_entries[rel.sym_index])
+
+        # Move over symbols, deleting the temporary function labels.
+        # Sometimes this naive procedure results in duplicate symbols, or UNDEF
+        # symbols that are also defined the same .o file. Hopefully that's fine.
+        # Skip over local symbols that aren't used relocated against, to avoid
+        # conflicts.
+        new_local_syms = [s for s in objfile.symtab.local_symbols() if not is_temp_name(s.name)]
+        new_global_syms = [s for s in objfile.symtab.global_symbols() if not is_temp_name(s.name)]
+        for i, s in enumerate(asm_objfile.symtab.symbol_entries):
+            is_local = (i < asm_objfile.symtab.sh_info)
+            if is_local and s not in relocated_symbols:
+                continue
+            if is_temp_name(s.name):
+                continue
+            if s.st_shndx not in [SHN_UNDEF, SHN_ABS]:
+                section_name = asm_objfile.sections[s.st_shndx].name
+                if section_name not in SECTIONS:
+                    raise Failure("generated assembly .o must only have symbols for .text, .data, .rodata, ABS and UNDEF, but found " + section_name)
+                s.st_shndx = objfile.find_section(section_name).index
+                # glabel's aren't marked as functions, making objdump output confusing. Fix that.
+                if s.name in all_text_glabels:
+                    s.type = STT_FUNC
+                if objfile.sections[s.st_shndx].name == '.rodata' and s.st_value in moved_late_rodata:
+                    s.st_value = moved_late_rodata[s.st_value]
+            s.st_name += strtab_adj
+            if is_local:
+                new_local_syms.append(s)
+            else:
+                new_global_syms.append(s)
+        new_syms = new_local_syms + new_global_syms
+        for i, s in enumerate(new_syms):
+            s.new_index = i
+        objfile.symtab.data = b''.join(s.to_bin() for s in new_syms)
+        objfile.symtab.sh_info = len(new_local_syms)
+
+        # Move over relocations
+        for sectype in SECTIONS:
+            source = asm_objfile.find_section(sectype)
+            target = objfile.find_section(sectype)
+
+            if target is not None:
+                # fixup relocation symbol indices, since we butchered them above
+                for reltab in target.relocated_by:
+                    nrels = []
+                    for rel in reltab.relocations:
+                        if sectype == '.text' and rel.r_offset in modified_text_positions:
+                            # don't include relocations for late_rodata dummy code
+                            continue
+                        # hopefully we don't have relocations for local or
+                        # temporary symbols, so new_index exists
+                        rel.sym_index = objfile.symtab.symbol_entries[rel.sym_index].new_index
+                        nrels.append(rel)
+                    reltab.relocations = nrels
+                    reltab.data = b''.join(rel.to_bin() for rel in nrels)
+
+            if not source:
+                continue
+
+            target_reltab = objfile.find_section('.rel' + sectype)
+            target_reltaba = objfile.find_section('.rela' + sectype)
+            for reltab in source.relocated_by:
+                for rel in reltab.relocations:
+                    rel.sym_index = asm_objfile.symtab.symbol_entries[rel.sym_index].new_index
+                    if sectype == '.rodata' and rel.r_offset in moved_late_rodata:
+                        rel.r_offset = moved_late_rodata[rel.r_offset]
+                new_data = b''.join(rel.to_bin() for rel in reltab.relocations)
+                if reltab.sh_type == SHT_REL:
+                    if not target_reltab:
+                        target_reltab = objfile.add_section('.rel' + sectype,
+                                sh_type=SHT_REL, sh_flags=0,
+                                sh_link=objfile.symtab.index, sh_info=target.index,
+                                sh_addralign=4, sh_entsize=8, data=b'')
+                    target_reltab.data += new_data
+                else:
+                    if not target_reltaba:
+                        target_reltaba = objfile.add_section('.rela' + sectype,
+                                sh_type=SHT_RELA, sh_flags=0,
+                                sh_link=objfile.symtab.index, sh_info=target.index,
+                                sh_addralign=4, sh_entsize=12, data=b'')
+                    target_reltaba.data += new_data
+
+        objfile.write(objfile_name)
+    finally:
+        s_file.close()
+        os.remove(s_name)
+        try:
+            os.remove(o_name)
+        except:
+            pass
+
+def run_wrapped(argv, outfile):
+    parser = argparse.ArgumentParser(description="Pre-process .c files and post-process .o files to enable embedding assembly into C.")
+    parser.add_argument('filename', help="path to .c code")
+    parser.add_argument('--post-process', dest='objfile', help="path to .o file to post-process")
+    parser.add_argument('--assembler', dest='assembler', help="assembler command (e.g. \"mips-linux-gnu-as -march=vr4300 -mabi=32\")")
+    parser.add_argument('--asm-prelude', dest='asm_prelude', help="path to a file containing a prelude to the assembly file (with .set and .macro directives, e.g.)")
+    parser.add_argument('--input-enc', default='latin1', help="Input encoding (default: latin1)")
+    parser.add_argument('--output-enc', default='latin1', help="Output encoding (default: latin1)")
+    parser.add_argument('-framepointer', dest='framepointer', action='store_true')
+    parser.add_argument('-g3', dest='g3', action='store_true')
+    group = parser.add_mutually_exclusive_group(required=True)
+    group.add_argument('-O1', dest='opt', action='store_const', const='O1')
+    group.add_argument('-O2', dest='opt', action='store_const', const='O2')
+    group.add_argument('-g', dest='opt', action='store_const', const='g')
+    args = parser.parse_args(argv)
+    opt = args.opt
+    if args.g3:
+        if opt != 'O2':
+            raise Failure("-g3 is only supported together with -O2")
+        opt = 'g3'
+
+    if args.objfile is None:
+        with open(args.filename, encoding=args.input_enc) as f:
+            parse_source(f, opt=opt, framepointer=args.framepointer, input_enc=args.input_enc, output_enc=args.output_enc, print_source=outfile)
+    else:
+        if args.assembler is None:
+            raise Failure("must pass assembler command")
+        with open(args.filename, encoding=args.input_enc) as f:
+            functions = parse_source(f, opt=opt, framepointer=args.framepointer, input_enc=args.input_enc, output_enc=args.output_enc)
+        if not functions:
+            return
+        asm_prelude = b''
+        if args.asm_prelude:
+            with open(args.asm_prelude, 'rb') as f:
+                asm_prelude = f.read()
+        fixup_objfile(args.objfile, functions, asm_prelude, args.assembler, args.output_enc)
+
+def run(argv, outfile=sys.stdout.buffer):
+    try:
+        run_wrapped(argv, outfile)
+    except Failure as e:
+        print("Error:", e, file=sys.stderr)
+        sys.exit(1)
+
+if __name__ == "__main__":
+    run(sys.argv[1:])
diff --git a/tools/asm_processor/build.py b/tools/asm_processor/build.py
new file mode 100644
index 000000000..b30f80104
--- /dev/null
+++ b/tools/asm_processor/build.py
@@ -0,0 +1,51 @@
+#!/usr/bin/env python3
+import sys
+import os
+import shlex
+import subprocess
+import tempfile
+import asm_processor
+
+dir_path = os.path.dirname(os.path.realpath(__file__))
+prelude = os.path.join(dir_path, "prelude.inc")
+
+all_args = sys.argv[1:]
+sep1 = all_args.index('--')
+sep2 = all_args.index('--', sep1+1)
+
+compiler = all_args[:sep1]
+
+assembler = all_args[sep1+1:sep2]
+assembler_sh = ' '.join(shlex.quote(x) for x in assembler)
+
+compile_args = all_args[sep2+1:]
+in_file = compile_args[-1]
+out_ind = compile_args.index('-o')
+out_file = compile_args[out_ind + 1]
+del compile_args[-1]
+del compile_args[out_ind + 1]
+del compile_args[out_ind]
+
+in_dir = os.path.split(os.path.realpath(in_file))[0]
+opt_flags = [x for x in compile_args if x in ['-g3', '-g', '-O1', '-O2', '-framepointer']]
+
+preprocessed_file = tempfile.NamedTemporaryFile(prefix='preprocessed', suffix='.c', delete=False)
+
+try:
+    asmproc_flags = opt_flags + [in_file, '--input-enc', 'utf-8', '--output-enc', 'euc-jp']
+    compile_cmdline = compiler + compile_args + ['-I', in_dir, '-o', out_file, preprocessed_file.name]
+
+    asm_processor.run(asmproc_flags, outfile=preprocessed_file)
+    try:
+        subprocess.check_call(compile_cmdline)
+    except subprocess.CalledProcessError as e:
+        print("Failed to compile file " + in_file + ". Command line:")
+        print()
+        print(' '.join(shlex.quote(x) for x in compile_cmdline))
+        print()
+        # To keep the preprocessed file:
+        # os._exit(1)
+
+    asm_processor.run(asmproc_flags + ['--post-process', out_file, '--assembler', assembler_sh, '--asm-prelude', prelude])
+finally:
+    os.remove(preprocessed_file.name)
diff --git a/tools/asm_processor/prelude.inc b/tools/asm_processor/prelude.inc
new file mode 100644
index 000000000..1ec8c6036
--- /dev/null
+++ b/tools/asm_processor/prelude.inc
@@ -0,0 +1,5 @@
+.set noat
+.set noreorder
+.set gp=64
+.include "macro.inc"
+