froydnj/convert-builtins.py

## convert-builtins.py
#!/usr/bin/env python

import re
import sys

# nvptx preprocessor definitions.
SM_80 = "sm_80"
SM_75 = "sm_75|" + SM_80
SM_72 = "sm_72|" + SM_75
SM_70 = "sm_70|" + SM_72
SM_60 = "sm_60|sm_61|sm_62|" + SM_70
PTX70 = "ptx70"
PTX65 = "ptx65|" + PTX70
PTX64 = "ptx64|" + PTX65
PTX63 = "ptx63|" + PTX64
PTX61 = "ptx61|" + PTX63
PTX60 = "ptx60|" + PTX61

preprocessor_features = {
    "PTX60": PTX60,
    "SM": None,
    "60": None,
    "SM_60": SM_60,
    "HVXV65": None,
    "HVXV60": None,
    "HVXV62": None,
    "HVXV66": None,
    "V5": None,
    "V67": None,
    "V66": None,
    "V60": None,
    "V62": None,
    "V65": None,
    "AND\\(SM_70,PTX60\\)": SM_70 + "|" + PTX60,
    "AND\\(SM_70,PTX61\\)": SM_70 + "|" + PTX61,
    "AND\\(SM_70,PTX63\\)": SM_70 + "|" + PTX63,
    "AND\\(SM_72,PTX63\\)": SM_72 + "|" + PTX63,
    "AND\\(SM_75,PTX63\\)": SM_75 + "|" + PTX63,
}

name_re = "(?P<name>[^, ]+)"
type_re = "(?P<type>\"[^\"]*\")"
attr_re = "(?P<attr>\"[^\"]*\")"
# The alternatives here are a hack to catch NVPTX and other preprocessor macro users.
feat_re = "(?P<feat>\"[^\"]*\"|%s)" % '|'.join(preprocessor_features.keys())
head_re = "(?P<head>\"[^\"]+\")"
lang_re = "(?P<lang>[A-Z0-9_]+)"

def regex_for(macro, *args):
    # Anchor so BUILTIN doesn't get confused with TARGET_BUILTIN or similar.
    return re.compile("^%s\\(%s\\)(?P<trailing>.*)" % (macro, " *, *".join(args)))

BUILTIN = regex_for('BUILTIN', name_re, type_re, attr_re)
ATOMIC_BUILTIN = regex_for('ATOMIC_BUILTIN', name_re, type_re, attr_re)
LIBBUILTIN = regex_for('LIBBUILTIN', name_re, type_re, attr_re, head_re, lang_re)
LANGBUILTIN = regex_for('LANGBUILTIN', name_re, type_re, attr_re, lang_re)
TARGET_BUILTIN = regex_for('TARGET_BUILTIN', name_re, type_re, attr_re, feat_re)
TARGET_HEADER_BUILTIN = regex_for('TARGET_HEADER_BUILTIN', name_re, type_re, attr_re, head_re, lang_re, feat_re)

lang_subs = {
    "GNU_LANG": "GnuLang",
    "C_LANG": "CLang",
    "CXX_LANG": "CxxLang",
    "OBJC_LANG": "ObjcLang",
    "MS_LANG": "MsLang",
    "OCLC20_LANG": "OCLC20Lang",
    "OCLC1X_LANG": "OCLC1xLang",
    "OMP_LANG": "OMPLang",
    "ALL_LANGUAGES": "AllLanguages",
    "ALL_MS_LANGUAGES": "AllMSLanguages",
    "ALL_GNU_LANGUAGES": "AllGnuLanguages",
    "ALL_OCLC_LANGUAGES": "AllOCLCLanguages",
}

Builtin = "def {name} : Builtin<{type}, {attr}>;{trailing}"
AtomicBuiltin = "def {name} : Builtin<{type}, {attr}>;{trailing}"
LibBuiltin = "def {name} : LibraryBuiltin<{type}, {attr}, {head}, {lang}>;{trailing}"
LangBuiltin = "def {name} : LangBuiltin<{type}, {attr}, {lang}>;{trailing}"
TargetBuiltin = "def {name} : TargetBuiltin<{type}, {attr}, {feat}>;{trailing}"
TargetHeaderBuiltin = "def {name} : TargetHeaderBuiltin<{type}, {attr}, {head}, {lang}, {feat}>;{trailing}"

builtins = {
    TARGET_BUILTIN: TargetBuiltin,
    TARGET_HEADER_BUILTIN: TargetHeaderBuiltin,
    BUILTIN: Builtin,
    LANGBUILTIN: LangBuiltin,
    LIBBUILTIN: LibBuiltin,
    ATOMIC_BUILTIN: AtomicBuiltin,
}

builtins_file = sys.argv[1]
# Remove "Builtins" prefix and ".def" suffix.
arch = builtins_file[8:-4]
builtins_file = open(builtins_file, 'r')

# Try to handle converting the generic builtins just like architecture ones.
print_target_arch_wrap = arch != ""

have_seen_first_def = False
processing_atomic = False

for line in builtins_file.readlines():
    line = line.rstrip()

    # Skip preprocessor macros.
    if line.startswith('#'):
        continue

    # Two steps to avoid calling re.match twice.
    matched = [(re.match(regex, line), regex, tablegen)
               for (regex, tablegen) in builtins.items()]
    matched = [x for x in matched if x[0]]
    if not matched:
        print line
        continue

    (m, regex, tablegen) = matched[0]
    fields = m.groupdict()
    if regex is ATOMIC_BUILTIN:
        if not processing_atomic:
            processing_atomic = True
            print 'let Atomic = 1 in {'
    else:
        if processing_atomic:
            processing_atomic = False
            print '}'

    # Translate language enums to tablegen definitions.
    if 'lang' in fields:
        fields['lang'] = lang_subs[fields['lang']]
    # We should really be rewriting 'feat' according to preprocessor_features,
    # above, but that's not complete yet -- and we probably really want to
    # translate the preprocessor macros into tablegen definitions anyway.  Just
    # turn things into strings if they're not already strings.
    if 'feat' in fields and not fields['feat'].startswith('"'):
        fields['feat'] = '"%s"' % fields['feat']

    def_line = tablegen.format(**fields)
    if processing_atomic:
        def_line = '  ' + def_line

    if not have_seen_first_def:
        print 'include "BuiltinTableGenDefs.inc"'
        print ''
        if print_target_arch_wrap:
            print 'let TargetArch = "%s" in {' % arch
        have_seen_first_def = True

    print def_line

if have_seen_first_def:
    if print_target_arch_wrap:
        print '}'
	#!/usr/bin/env python

	import re
	import sys

	# nvptx preprocessor definitions.
	SM_80 = "sm_80"
	SM_75 = "sm_75\|" + SM_80
	SM_72 = "sm_72\|" + SM_75
	SM_70 = "sm_70\|" + SM_72
	SM_60 = "sm_60\|sm_61\|sm_62\|" + SM_70
	PTX70 = "ptx70"
	PTX65 = "ptx65\|" + PTX70
	PTX64 = "ptx64\|" + PTX65
	PTX63 = "ptx63\|" + PTX64
	PTX61 = "ptx61\|" + PTX63
	PTX60 = "ptx60\|" + PTX61

	preprocessor_features = {
	"PTX60": PTX60,
	"SM": None,
	"60": None,
	"SM_60": SM_60,
	"HVXV65": None,
	"HVXV60": None,
	"HVXV62": None,
	"HVXV66": None,
	"V5": None,
	"V67": None,
	"V66": None,
	"V60": None,
	"V62": None,
	"V65": None,
	"AND\\(SM_70,PTX60\\)": SM_70 + "\|" + PTX60,
	"AND\\(SM_70,PTX61\\)": SM_70 + "\|" + PTX61,
	"AND\\(SM_70,PTX63\\)": SM_70 + "\|" + PTX63,
	"AND\\(SM_72,PTX63\\)": SM_72 + "\|" + PTX63,
	"AND\\(SM_75,PTX63\\)": SM_75 + "\|" + PTX63,
	}

	name_re = "(?P<name>[^, ]+)"
	type_re = "(?P<type>\"[^\"]*\")"
	attr_re = "(?P<attr>\"[^\"]*\")"
	# The alternatives here are a hack to catch NVPTX and other preprocessor macro users.
	feat_re = "(?P<feat>\"[^\"]*\"\|%s)" % '\|'.join(preprocessor_features.keys())
	head_re = "(?P<head>\"[^\"]+\")"
	lang_re = "(?P<lang>[A-Z0-9_]+)"

	def regex_for(macro, *args):
	# Anchor so BUILTIN doesn't get confused with TARGET_BUILTIN or similar.
	return re.compile("^%s\\(%s\\)(?P<trailing>.)" % (macro, " , *".join(args)))

	BUILTIN = regex_for('BUILTIN', name_re, type_re, attr_re)
	ATOMIC_BUILTIN = regex_for('ATOMIC_BUILTIN', name_re, type_re, attr_re)
	LIBBUILTIN = regex_for('LIBBUILTIN', name_re, type_re, attr_re, head_re, lang_re)
	LANGBUILTIN = regex_for('LANGBUILTIN', name_re, type_re, attr_re, lang_re)
	TARGET_BUILTIN = regex_for('TARGET_BUILTIN', name_re, type_re, attr_re, feat_re)
	TARGET_HEADER_BUILTIN = regex_for('TARGET_HEADER_BUILTIN', name_re, type_re, attr_re, head_re, lang_re, feat_re)

	lang_subs = {
	"GNU_LANG": "GnuLang",
	"C_LANG": "CLang",
	"CXX_LANG": "CxxLang",
	"OBJC_LANG": "ObjcLang",
	"MS_LANG": "MsLang",
	"OCLC20_LANG": "OCLC20Lang",
	"OCLC1X_LANG": "OCLC1xLang",
	"OMP_LANG": "OMPLang",
	"ALL_LANGUAGES": "AllLanguages",
	"ALL_MS_LANGUAGES": "AllMSLanguages",
	"ALL_GNU_LANGUAGES": "AllGnuLanguages",
	"ALL_OCLC_LANGUAGES": "AllOCLCLanguages",
	}

	Builtin = "def {name} : Builtin<{type}, {attr}>;{trailing}"
	AtomicBuiltin = "def {name} : Builtin<{type}, {attr}>;{trailing}"
	LibBuiltin = "def {name} : LibraryBuiltin<{type}, {attr}, {head}, {lang}>;{trailing}"
	LangBuiltin = "def {name} : LangBuiltin<{type}, {attr}, {lang}>;{trailing}"
	TargetBuiltin = "def {name} : TargetBuiltin<{type}, {attr}, {feat}>;{trailing}"
	TargetHeaderBuiltin = "def {name} : TargetHeaderBuiltin<{type}, {attr}, {head}, {lang}, {feat}>;{trailing}"

	builtins = {
	TARGET_BUILTIN: TargetBuiltin,
	TARGET_HEADER_BUILTIN: TargetHeaderBuiltin,
	BUILTIN: Builtin,
	LANGBUILTIN: LangBuiltin,
	LIBBUILTIN: LibBuiltin,
	ATOMIC_BUILTIN: AtomicBuiltin,
	}

	builtins_file = sys.argv[1]
	# Remove "Builtins" prefix and ".def" suffix.
	arch = builtins_file[8:-4]
	builtins_file = open(builtins_file, 'r')

	# Try to handle converting the generic builtins just like architecture ones.
	print_target_arch_wrap = arch != ""

	have_seen_first_def = False
	processing_atomic = False

	for line in builtins_file.readlines():
	line = line.rstrip()

	# Skip preprocessor macros.
	if line.startswith('#'):
	continue

	# Two steps to avoid calling re.match twice.
	matched = [(re.match(regex, line), regex, tablegen)
	for (regex, tablegen) in builtins.items()]
	matched = [x for x in matched if x[0]]
	if not matched:
	print line
	continue

	(m, regex, tablegen) = matched[0]
	fields = m.groupdict()
	if regex is ATOMIC_BUILTIN:
	if not processing_atomic:
	processing_atomic = True
	print 'let Atomic = 1 in {'
	else:
	if processing_atomic:
	processing_atomic = False
	print '}'

	# Translate language enums to tablegen definitions.
	if 'lang' in fields:
	fields['lang'] = lang_subs[fields['lang']]
	# We should really be rewriting 'feat' according to preprocessor_features,
	# above, but that's not complete yet -- and we probably really want to
	# translate the preprocessor macros into tablegen definitions anyway. Just
	# turn things into strings if they're not already strings.
	if 'feat' in fields and not fields['feat'].startswith('"'):
	fields['feat'] = '"%s"' % fields['feat']

	def_line = tablegen.format(**fields)
	if processing_atomic:
	def_line = ' ' + def_line

	if not have_seen_first_def:
	print 'include "BuiltinTableGenDefs.inc"'
	print ''
	if print_target_arch_wrap:
	print 'let TargetArch = "%s" in {' % arch
	have_seen_first_def = True

	print def_line

	if have_seen_first_def:
	if print_target_arch_wrap:
	print '}'