Author: Armin Rigo <[email protected]>
Branch:
Changeset: r92618:120728f3d439
Date: 2017-10-06 02:58 +0200
http://bitbucket.org/pypy/pypy/changeset/120728f3d439/
Log: Optimize the case of siphash computing the hash of a latin1 unicode
string
diff --git a/rpython/rlib/rsiphash.py b/rpython/rlib/rsiphash.py
--- a/rpython/rlib/rsiphash.py
+++ b/rpython/rlib/rsiphash.py
@@ -11,6 +11,7 @@
from rpython.rlib.objectmodel import not_rpython, always_inline
from rpython.rlib.objectmodel import we_are_translated, dont_inline
from rpython.rlib.objectmodel import keepalive_until_here
+from rpython.rlib.objectmodel import specialize
from rpython.rlib import rgc, jit, rposix
from rpython.rlib.rarithmetic import r_uint64, r_uint32, r_uint
from rpython.rlib.rawstorage import misaligned_is_fine
@@ -26,9 +27,11 @@
return x
def _le32toh(x):
return x
+ BIG_ENDIAN = False
else:
_le64toh = rarithmetic.byteswap
_le32toh = rarithmetic.byteswap
+ BIG_ENDIAN = True
def _decode64(s):
return (r_uint64(ord(s[0])) |
@@ -151,23 +154,20 @@
else:
# NOTE: a latin-1 unicode string must have the same hash as the
# corresponding byte string. If the unicode is all within
- # 0-255, then we need to allocate a byte buffer and copy the
- # latin-1 encoding in it manually. Note also that we give a
+ # 0-255, then we call _siphash24() with a special argument that
+ # will make it load only one byte for every unicode char.
+ # Note also that we give a
# different hash result than CPython on ucs4 platforms, for
# unicode strings where CPython uses 2 bytes per character.
+ addr = rstr._get_raw_buf_unicode(rstr.UNICODE, ll_s, 0)
+ SZ = rffi.sizeof(rstr.UNICODE.chars.OF)
for i in range(length):
if ord(ll_s.chars[i]) > 0xFF:
- addr = rstr._get_raw_buf_unicode(rstr.UNICODE, ll_s, 0)
- length *= rffi.sizeof(rstr.UNICODE.chars.OF)
+ length *= SZ
break
else:
- p = lltype.malloc(rffi.CCHARP.TO, length, flavor='raw')
- i = 0
- while i < length:
- p[i] = chr(ord(ll_s.chars[i]))
- i += 1
- x = _siphash24(llmemory.cast_ptr_to_adr(p), length)
- lltype.free(p, flavor='raw')
+ x = _siphash24(addr, length, SZ)
+ keepalive_until_here(ll_s)
return intmask(x)
x = _siphash24(addr, length)
keepalive_until_here(ll_s)
@@ -215,7 +215,8 @@
@rgc.no_collect
-def _siphash24(addr_in, size):
[email protected](2)
+def _siphash24(addr_in, size, SZ=1):
"""Takes an address pointer and a size. Returns the hash as a r_uint64,
which can then be casted to the expected type."""
@@ -227,10 +228,14 @@
v2 = k0 ^ magic2
v3 = k1 ^ magic3
- direct = (misaligned_is_fine or
+ direct = (SZ == 1) and (misaligned_is_fine or
(rffi.cast(lltype.Signed, addr_in) & 7) == 0)
- index = 0
+ if BIG_ENDIAN:
+ index = SZ - 1
+ else:
+ index = 0
if direct:
+ assert SZ == 1
while size >= 8:
mi = llop.raw_load(rffi.ULONGLONG, addr_in, index)
mi = _le64toh(mi)
@@ -242,30 +247,30 @@
else:
while size >= 8:
mi = (
- r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index)) |
- r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index + 1)) << 8 |
- r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index + 2)) << 16 |
- r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index + 3)) << 24 |
- r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index + 4)) << 32 |
- r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index + 5)) << 40 |
- r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index + 6)) << 48 |
- r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index + 7)) << 56
+ r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index)) |
+ r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index + 1*SZ)) << 8 |
+ r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index + 2*SZ)) << 16
|
+ r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index + 3*SZ)) << 24
|
+ r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index + 4*SZ)) << 32
|
+ r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index + 5*SZ)) << 40
|
+ r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index + 6*SZ)) << 48
|
+ r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index + 7*SZ)) << 56
)
size -= 8
- index += 8
+ index += 8*SZ
v3 ^= mi
v0, v1, v2, v3 = _double_round(v0, v1, v2, v3)
v0 ^= mi
t = r_uint64(0)
if size == 7:
- t = r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index + 6)) << 48
+ t = r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index + 6*SZ)) << 48
size = 6
if size == 6:
- t |= r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index + 5)) << 40
+ t |= r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index + 5*SZ)) << 40
size = 5
if size == 5:
- t |= r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index + 4)) << 32
+ t |= r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index + 4*SZ)) << 32
size = 4
if size == 4:
if direct:
@@ -273,13 +278,13 @@
t |= r_uint64(v)
size = 0
else:
- t |= r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index + 3)) << 24
+ t |= r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index + 3*SZ))<<
24
size = 3
if size == 3:
- t |= r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index + 2)) << 16
+ t |= r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index + 2*SZ)) << 16
size = 2
if size == 2:
- t |= r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index + 1)) << 8
+ t |= r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index + 1*SZ)) << 8
size = 1
if size == 1:
t |= r_uint64(llop.raw_load(rffi.UCHAR, addr_in, index))
_______________________________________________
pypy-commit mailing list
[email protected]
https://mail.python.org/mailman/listinfo/pypy-commit
