2. PHP 7 release
Nov 12th
* if everything goes as planned

3. http://hhvm.com/blog/9293/lockdown-results-and-hhvm-performance

4. Memory optimization

5. Memory optimization
• Reduce number of allocations

6. Memory optimization
• Reduce number of allocations
Memory

7. Memory optimization
• Reduce number of allocations
• PHP 5 spends 20% of CPU time in the allocator
Memory

8. Memory optimization
• Reduce number of allocations
• Reduce memory usage

9. Memory optimization
• Reduce number of allocations
• Reduce memory usage
• Memory access has high latency

10. Memory optimization
• Reduce number of allocations
• Reduce memory usage
• Memory access has high latency
CPU

11. Memory optimization
• Reduce number of allocations
• Reduce memory usage
• Memory access has high latency
CPU L1D
1ns 32KB

12. Memory optimization
• Reduce number of allocations
• Reduce memory usage
• Memory access has high latency
CPU
4ns
L2
256 KB
L1D
1ns 32KB

13. Memory optimization
• Reduce number of allocations
• Reduce memory usage
• Memory access has high latency
CPU L312ns
a few MB
4ns
L2
256 KB
L1D
1ns 32KB

14. RAM
100ns
Memory optimization
• Reduce number of allocations
• Reduce memory usage
• Memory access has high latency
CPU L312ns
a few MB
4ns
L2
256 KB
L1D
1ns 32KB

15. RAM
100ns
Memory optimization
• Reduce number of allocations
• Reduce memory usage
• Memory access has high latency
• Less data => more fits into the cache
CPU L312ns
a few MB
4ns
L2
256 KB
L1D
1ns 32KB

16. Memory optimization
• Reduce number of allocations
• Reduce memory usage
• Reduce indirection

17. Memory optimization
• Reduce number of allocations
• Reduce memory usage
• Reduce indirection
• Pointer points to pointer pointing to pointer pointing to pointer pointing to
pointer pointing to pointer pointing to what you actually want
What you
have
What you
want

18. Memory optimization
• Reduce number of allocations
• Reduce memory usage
• Reduce indirection
• Pointer points to pointer pointing to pointer pointing to pointer pointing to
pointer pointing to pointer pointing to what you actually want
• Fewer memory accesses
What you
have
What you
want

19. Zvals: PHP 5
• Zval = Type + Value
value
ty
zval

20. Zvals: PHP 5
• Zval = Type + Value
value
ty
zval
$a = 42;Code:

21. Zvals: PHP 5
• Zval = Type + Value
value = int(42)
ty
zval *
zval
$a:
$a = 42;Code:

22. Zvals: PHP 5
• Zval = Type + Value
value (simple):
null, bool, int, float
ty
zval *
zval
$a:
$a = 42;Code:

23. Zvals: PHP 5
• Zval = Type + Value
value (complex)
ty
zval *
zval
$a:
Complex data structure:
string, array, object
$a = [];Code:

24. Zvals: PHP 5
• Zval = Type + Value
value (complex)
ty
zval *
zval
$a:
Complex data structure:
string, array, object
$a = [];
$b = $a;
Code:

25. Zvals: PHP 5
• Zval = Type + Value + Refcount
value (complex)
refcount = 1 ty
zval *
zval
$a:
Complex data structure:
string, array, object
$a = [];
$b = $a;
Code:

26. Zvals: PHP 5
• Zval = Type + Value + Refcount
value (complex)
refcount = 2 ty
zval *
zval
$a:
Complex data structure:
string, array, object
$a = [];
$b = $a;
Code:
zval *$b:

27. Zvals: PHP 7
• Zval = Type + Value
value (complex)
ty
zval *
zval
$a:
$a = [];
$b = $a;
Code:
zval *$b:
Complex data structure:
string, array, object
refcount = 2

28. Zvals: PHP 7
• Zval = Type + Value
zval *
zval
$a:
$a = [];
$b = $a;
Code:
zval *$b:
Complex data structure:
string, array, object
refcount = 2
value (complex)
type_info

29. value (complex)
type_info
value (complex)
type_info
Zvals: PHP 7
• Zval = Type + Value
$a:
Complex data structure:
string, array, object
$b:
refcount = 2
zval
zval

30. PHP 5 PHP 7
value (simple):
null, bool, int, float
refcount ty
gc_buffer
zval * value (simple): …
type_info
1 allocations
1 level of indirection
40 bytes
no allocations
no indirection
16 bytes

31. PHP 5 PHP 7
value (complex)
refcount ty
gc_root
zval *
Complex data structure:
string, array, object
value (complex)
type_info
Complex data structure:
string, array, object
refcount gc_info
2 allocations
2 levels of indirection
40 bytes
1 allocation
1 level of indirection
24 bytes

32. val (char *)
len (int)
A B C 0
zval value:
PHP 5

33. zend_string *
refcount gc_info
hash
len (size_t)
A B C 0
val (char *)
len (int)
A B C 0
zval value: zval value:
PHP 5 PHP 7

34. [0]: (empty)
[1]: (empty)
[2]: (empty)
[3]: (empty)
hash()
“xyz”
Arrays: PHP 5

35. [0]: (empty)
[1]:
[2]: (empty)
[3]: (empty)
hash()
“xyz”
data
key “xyz”
Arrays: PHP 5

36. [0]: (empty)
[1]:
[2]: (empty)
[3]:
“foo”
data
key “foo”
hash()
“xyz”
data
key “xyz”
Arrays: PHP 5

37. [0]: (empty)
[1]:
[2]: (empty)
[3]:
“foo”
data
key “foo”
next
data
key “bar”
next = NULL
hash()
“bar”
“xyz”
data
key “xyz”
next = NULL
Arrays: PHP 5

38. [0]: (empty)
[1]:
[2]: (empty)
[3]:
“foo”
hash()
“bar”
“xyz”
Arrays: PHP 5
head
tail
data
key “foo”
next
data
key “bar”
next = NULL
data
key “xyz”
next = NULL

39. [0]: (empty)
[1]:
[2]: (empty)
[3]:
“foo”
hash()
“bar”
“xyz”
Arrays: PHP 7
data
key “foo”
next
data
key “bar”
next = NULL
data
key “xyz”
next = NULL

40. [0]: (empty)
[1]:
[2]: (empty)
[3]:
“foo”
hash()
“bar”
“xyz”
Arrays: PHP 7
data
key “foo”
next
data
key “bar”
next = NULL
data
key “xyz”
next = NULL
Bucket *
uint32_t

41. PHP 5 without zval PHP 7 with zval
Allocations: 2 + 1 per element 2
Size: 72 + 80 per element 56 + 36 per element
Indirections (lookup): 4 2
Arrays: comparison

42. PHP 5 without zval PHP 7 with zval
Allocations: 2 + 1 per element 2
Size: 72 + 80 per element 56 + 36 per element
Indirections (lookup): 4 2
PHP 5 w/ unique zvals
2 + 2 per element
72 + 112 per element
4
Arrays: comparison

43. Immutable arrays
$arrays = [];
for ($i = 0; $i < 1000000; ++$i) {
$arrays[] = [1, 2, 3, 4, 5, 6, 7, 8];
}

44. Immutable arrays
$arrays = [];
for ($i = 0; $i < 1000000; ++$i) {
$arrays[] = [1, 2, 3, 4, 5, 6, 7, 8];
}
PHP 5
Memory usage 1336 MB
Execution time (create) 0.77 s
Execution time (destroy) 1.45 s

45. Immutable arrays
$arrays = [];
for ($i = 0; $i < 1000000; ++$i) {
$arrays[] = [1, 2, 3, 4, 5, 6, 7, 8];
}
PHP 5 PHP 7
Memory usage 1336 MB 391 MB
Execution time (create) 0.77 s 0.29 s
Execution time (destroy) 1.45 s 0.05 s

46. Immutable arrays
$arrays = [];
for ($i = 0; $i < 1000000; ++$i) {
$arrays[] = [1, 2, 3, 4, 5, 6, 7, 8];
}
PHP 5 PHP 7 PHP 7 with opcache
Memory usage 1336 MB 391 MB 32 MB
Execution time (create) 0.77 s 0.29 s 0.03 s
Execution time (destroy) 1.45 s 0.05 s 0.002 s

47. handle (ID)
handlers
zval value
D V ac
object
dtor()
free_storage()
clone()
handlers
refcount
gc_root
ce
properties
properties_table
guards
[0] (stores $prop1)
[1] (stores $prop2)
[2] (stores $prop3)
zval
zval
zval
object store bucket
Zend object
Objects: PHP 5

48. zend_object * refcount type_info
handle
ce
handlers
properties
value
type_info
value
type_info
value
type_info
$prop1 zval
$prop2 zval
$prop3 zval
Zend object
Objects:
PHP 7

49. PHP 5 without zval PHP 7 with zval
Allocations: 2 1
Size: 96 + 8 per element 48 + 16 per element
Indirections (prop val): 4 1
Objects: comparison

50. Integers
• long  zend_long
• 64-bit integers on LLP64 platforms (= Windows)

51. AST
<?php
$a = 42;
$b = 24;
echo $a + $b;

52. AST
<?php
$a = 42;
$b = 24;
echo $a + $b;
<?php T_OPEN_TAG
$a T_VARIABLE
=
42 T_LNUMBER
;
$b T_VARIABLE
=
24 T_LNUMBER
;
echo T_ECHO
$a T_VARIABLE
+
$b T_VARIABLE
;
lex

53. AST
<?php
$a = 42;
$b = 24;
echo $a + $b;
<?php T_OPEN_TAG
$a T_VARIABLE
=
42 T_LNUMBER
;
$b T_VARIABLE
=
24 T_LNUMBER
;
echo T_ECHO
$a T_VARIABLE
+
$b T_VARIABLE
;
ASSIGN $a 42
ASSIGN $b 24
ADD $a $b ~2
ECHO ~2
RETURN 1
lex
parse + compile

54. AST
<?php
$a = 42;
$b = 24;
echo $a + $b;
<?php T_OPEN_TAG
$a T_VARIABLE
=
42 T_LNUMBER
;
$b T_VARIABLE
=
24 T_LNUMBER
;
echo T_ECHO
$a T_VARIABLE
+
$b T_VARIABLE
;
ASSIGN $a 42
ASSIGN $b 24
ADD $a $b ~2
ECHO ~2
RETURN 1
lex parse
compile
stmts
assign
var 42
$a
assign
$b
echo
+
$a $b
var 24

55. AST
<?php
$a = 42;
$b = 24;
echo $a + $b;
<?php T_OPEN_TAG
$a T_VARIABLE
=
42 T_LNUMBER
;
$b T_VARIABLE
=
24 T_LNUMBER
;
echo T_ECHO
$a T_VARIABLE
+
$b T_VARIABLE
;
ASSIGN $a 42
ASSIGN $b 24
ADD $a $b ~0
ECHO ~0
RETURN 1
lex parse
compile
token_get_all()
nikic/php-ast
phpdbg -p
stmts
assign
var 42
$a
assign
$b
echo
+
$a $b
var 24

56. Virtual machine

57. Virtual machine – stack management (PHP 5)
function foo($a, $b) {
return $a + $b;
}
foo(1, 2);

58. Virtual machine – stack management (PHP 5)
function foo($a, $b) {
return $a + $b;
}
foo(1, 2);
SEND_VAL 1
SEND_VAL 2
DO_FCALL foo
RECV $a
RECV $b
ADD $a $b ~0
RETURN ~0
{main}
foo()

59. Virtual machine – stack management (PHP 5)
function foo($a, $b) {
return $a + $b;
}
foo(1, 2);
SEND_VAL 1
SEND_VAL 2
DO_FCALL foo
RECV $a
RECV $b
ADD $a $b ~0
RETURN ~0
{main}
foo()
…
VM stack

60. Virtual machine – stack management (PHP 5)
function foo($a, $b) {
return $a + $b;
}
foo(1, 2);
SEND_VAL 1
SEND_VAL 2
DO_FCALL foo
RECV $a
RECV $b
ADD $a $b ~0
RETURN ~0
{main}
foo()
…
int(1)
int(2)
VM stack
arg[0]
arg[1]

61. Virtual machine – stack management (PHP 5)
function foo($a, $b) {
return $a + $b;
}
foo(1, 2);
SEND_VAL 1
SEND_VAL 2
DO_FCALL foo
RECV $a
RECV $b
ADD $a $b ~0
RETURN ~0
{main}
foo()
…
int(1)
int(2)
execute_data
VM stack
arg[0]
arg[1]
~0
$a
$b

62. Virtual machine – stack management (PHP 5)
function foo($a, $b) {
return $a + $b;
}
foo(1, 2);
SEND_VAL 1
SEND_VAL 2
DO_FCALL foo
RECV $a
RECV $b
ADD $a $b ~0
RETURN ~0
{main}
foo()
…
int(1)
int(2)
execute_data
int(1)
int(2)
VM stack
arg[0]
arg[1]
~0
$a
$b

63. Virtual machine – stack management (PHP 5)
function foo($a, $b) {
return $a + $b;
}
foo(1, 2);
SEND_VAL 1
SEND_VAL 2
DO_FCALL foo
RECV $a
RECV $b
ADD $a $b ~0
RETURN ~0
{main}
foo()
…
int(1)
int(2)
int(3)
execute_data
int(1)
int(2)
VM stack
arg[0]
arg[1]
~0
$a
$b

64. Virtual machine – stack management (PHP 7)
function foo($a, $b) {
return $a + $b;
}
foo(1, 2);
{main}
foo()
…
VM stack
INIT_FCALL foo
SEND_VAL 1
SEND_VAL 2
DO_FCALL
RECV $a
RECV $b
ADD $a $b ~0
RETURN ~0

65. Virtual machine – stack management (PHP 7)
function foo($a, $b) {
return $a + $b;
}
foo(1, 2);
{main}
foo()
…
execute_data
VM stack
~0
$a
$b
INIT_FCALL foo
SEND_VAL 1
SEND_VAL 2
DO_FCALL
RECV $a
RECV $b
ADD $a $b ~0
RETURN ~0

66. Virtual machine – stack management (PHP 7)
function foo($a, $b) {
return $a + $b;
}
foo(1, 2);
{main}
foo()
…
execute_data
int(1)
int(2)
VM stack
~0
$a
$b
INIT_FCALL foo
SEND_VAL 1
SEND_VAL 2
DO_FCALL
RECV $a
RECV $b
ADD $a $b ~0
RETURN ~0

67. Virtual machine – stack management (PHP 7)
function foo($a, $b) {
return $a + $b;
}
foo(1, 2);
{main}
foo()
…
execute_data
int(1)
int(2)
VM stack
~0
$a
$b
INIT_FCALL foo
SEND_VAL 1
SEND_VAL 2
DO_FCALL
RECV $a
RECV $b
ADD $a $b ~0
RETURN ~0

68. Virtual machine – stack management (PHP 7)
function foo($a, $b) {
return $a + $b;
}
foo(1, 2);
{main}
foo()
…
execute_data
int(1)
int(2)
int(3)
VM stack
~0
$a
$b
INIT_FCALL foo
SEND_VAL 1
SEND_VAL 2
DO_FCALL
RECV $a
RECV $b
ADD $a $b ~0
RETURN ~0

69. Virtual machine – inlined internal functions
• Functions with custom opcodes:
• strlen()
• is_*()
• defined()
• call_user_func()
• call_user_func_array()

70. Virtual machine – inlined internal functions
• Functions with custom opcodes:
• strlen()
• is_*()
• defined()
• call_user_func()
• call_user_func_array()
• Only in global scope or for fully qualified calls
namespace foo;
echo strlen($str);

71. Virtual machine – inlined internal functions
• Functions with custom opcodes:
• strlen()
• is_*()
• defined()
• call_user_func()
• call_user_func_array()
• Only in global scope or for fully qualified calls
namespace foo;
echo strlen($str);
Is this strlen() or foostrlen()?

72. Virtual machine – inlined internal functions
• Functions with custom opcodes:
• strlen()
• is_*()
• defined()
• call_user_func()
• call_user_func_array()
• Only in global scope or for fully qualified calls
namespace foo;
echo strlen($str);
This is definitely strlen()!

73. Virtual machine – global registers
zend_execute_data* execute_data;
const zend_op* opline;

74. Virtual machine – global registers
register zend_execute_data* execute_data __asm__("%r14");
register const zend_op* opline __asm__("%r15");

75. Virtual machine – global registers
register zend_execute_data* execute_data __asm__("%r14");
register const zend_op* opline __asm__("%r15");
• Registers reserved in VM code

76. Virtual machine – global registers
register zend_execute_data* execute_data __asm__("%r14");
register const zend_op* opline __asm__("%r15");
• Registers reserved in VM code
• Avoid reloading from executor_globals / execute_data

77. Virtual machine – global registers
register zend_execute_data* execute_data __asm__("%r14");
register const zend_op* opline __asm__("%r15");
• Registers reserved in VM code
• Avoid reloading from executor_globals / execute_data
• Avoid save/restore during calls

78. Opcache

79. Opcache
• File cache

80. Opcache
• File cache
• Alternative (or addition) to SHM

81. Opcache
• File cache
• Alternative (or addition) to SHM
• Shared hosting? PHP restart / cache reset?

82. Opcache
• File cache
• Alternative (or addition) to SHM
• Shared hosting? PHP restart / cache reset?
• Time for 500 sequential WP 4.1 requests
SHM 3 s
File cache 6 s
No cache 24 s

83. Opcache
• Huge Pages

84. Opcache
• Huge Pages
• Code segment of PHP binary remapped into huge pages
• Reduces iTLB misses

85. Opcache
• Huge Pages
• Code segment of PHP binary remapped into huge pages
• Reduces iTLB misses
CPU L2L1I …

86. Opcache
• Huge Pages
• Code segment of PHP binary remapped into huge pages
• Reduces iTLB misses
CPU L2L1I
L1 iTLB 4K
…
L2 TLB

87. Opcache
• Huge Pages
• Code segment of PHP binary remapped into huge pages
• Reduces iTLB misses
CPU L2L1I
L1 iTLB 4K L1 iTLB 2M
…
L2 TLB

88. @nikita_ppv
nikic@php.net
https://joind.in/talk/view/15865