On 02/04/2013 01:50 PM, Dmitry Stogov wrote:
I can't remember if I did any special benchmarks except for bench.php when
I introduced fast math functions.
That time, I rearranged the code to allow inlining of the most probable
paths and added assembler code to catch overflow (C can't do it in optimal
way). As I remember the bench.php showed some visible improvement.
OK, in that case I would still recommend using fewer hardcoded numbers
and offsets: please refer to the applied patch, it makes things a bit
more robust.
However, as I don't expect any speedup for ARM in this area, I will not
pursue this any further.
Are there any other optimizations involving inline assembler that you
implemented for PHP? I would like to review them to see if I need to add
an ARM version.
Cheers,
Ard.
>From 16f61b05b3911c91b52c5eff66c2274d5a7c817f Mon Sep 17 00:00:00 2001
From: Ard Biesheuvel <ard.biesheu...@linaro.org>
Date: Mon, 4 Feb 2013 17:56:11 +0100
Subject: [PATCH] Remove hardcoded constants and struct offsets from asm
arithmetic
---
Zend/zend_operators.h | 80 ++++++++++++++++++++++++++++++++-----------------
1 file changed, 52 insertions(+), 28 deletions(-)
diff --git a/Zend/zend_operators.h b/Zend/zend_operators.h
index 20a5277..b51a000 100644
--- a/Zend/zend_operators.h
+++ b/Zend/zend_operators.h
@@ -476,6 +476,10 @@ ZEND_API void zend_update_current_locale(void);
#define zend_update_current_locale()
#endif
+/* The offset in bytes between the value and type fields of a zval */
+#define ZVAL_OFFSETOF_TYPE \
+ (__builtin_offsetof(zval,type) - __builtin_offsetof(zval,value))
+
static zend_always_inline int fast_increment_function(zval *op1)
{
if (EXPECTED(Z_TYPE_P(op1) == IS_LONG)) {
@@ -485,20 +489,24 @@ static zend_always_inline int fast_increment_function(zval *op1)
"jno 0f\n\t"
"movl $0x0, (%0)\n\t"
"movl $0x41e00000, 0x4(%0)\n\t"
- "movb $0x2,0xc(%0)\n"
+ "movb %1, %c2(%0)\n"
"0:"
:
- : "r"(op1));
+ : "r"(&op1->value),
+ "n"(IS_DOUBLE),
+ "n"(ZVAL_OFFSETOF_TYPE));
#elif defined(__GNUC__) && defined(__x86_64__)
__asm__(
"incq (%0)\n\t"
"jno 0f\n\t"
"movl $0x0, (%0)\n\t"
"movl $0x43e00000, 0x4(%0)\n\t"
- "movb $0x2,0x14(%0)\n"
+ "movb %1, %c2(%0)\n"
"0:"
:
- : "r"(op1));
+ : "r"(&op1->value),
+ "n"(IS_DOUBLE),
+ "n"(ZVAL_OFFSETOF_TYPE));
#else
if (UNEXPECTED(Z_LVAL_P(op1) == LONG_MAX)) {
/* switch to double */
@@ -522,20 +530,24 @@ static zend_always_inline int fast_decrement_function(zval *op1)
"jno 0f\n\t"
"movl $0x00200000, (%0)\n\t"
"movl $0xc1e00000, 0x4(%0)\n\t"
- "movb $0x2,0xc(%0)\n"
+ "movb %1,%c2(%0)\n"
"0:"
:
- : "r"(op1));
+ : "r"(&op1->value),
+ "n"(IS_DOUBLE),
+ "n"(ZVAL_OFFSETOF_TYPE));
#elif defined(__GNUC__) && defined(__x86_64__)
__asm__(
"decq (%0)\n\t"
"jno 0f\n\t"
"movl $0x00000000, (%0)\n\t"
"movl $0xc3e00000, 0x4(%0)\n\t"
- "movb $0x2,0x14(%0)\n"
+ "movb %1,%c2(%0)\n"
"0:"
:
- : "r"(op1));
+ : "r"(&op1->value),
+ "n"(IS_DOUBLE),
+ "n"(ZVAL_OFFSETOF_TYPE));
#else
if (UNEXPECTED(Z_LVAL_P(op1) == LONG_MIN)) {
/* switch to double */
@@ -560,19 +572,22 @@ static zend_always_inline int fast_add_function(zval *result, zval *op1, zval *o
"addl (%2), %%eax\n\t"
"jo 0f\n\t"
"movl %%eax, (%0)\n\t"
- "movb $0x1,0xc(%0)\n\t"
+ "movb %3, %c5(%0)\n\t"
"jmp 1f\n"
"0:\n\t"
"fildl (%1)\n\t"
"fildl (%2)\n\t"
"faddp %%st, %%st(1)\n\t"
- "movb $0x2,0xc(%0)\n\t"
+ "movb %4, %c5(%0)\n\t"
"fstpl (%0)\n"
"1:"
:
- : "r"(result),
- "r"(op1),
- "r"(op2)
+ : "r"(&result->value),
+ "r"(&op1->value),
+ "r"(&op2->value),
+ "n"(IS_LONG),
+ "n"(IS_DOUBLE),
+ "n"(ZVAL_OFFSETOF_TYPE)
: "eax");
#elif defined(__GNUC__) && defined(__x86_64__)
__asm__(
@@ -580,19 +595,22 @@ static zend_always_inline int fast_add_function(zval *result, zval *op1, zval *o
"addq (%2), %%rax\n\t"
"jo 0f\n\t"
"movq %%rax, (%0)\n\t"
- "movb $0x1,0x14(%0)\n\t"
+ "movb %3, %c5(%0)\n\t"
"jmp 1f\n"
"0:\n\t"
"fildq (%1)\n\t"
"fildq (%2)\n\t"
"faddp %%st, %%st(1)\n\t"
- "movb $0x2,0x14(%0)\n\t"
+ "movb %4, %c5(%0)\n\t"
"fstpl (%0)\n"
"1:"
:
- : "r"(result),
- "r"(op1),
- "r"(op2)
+ : "r"(&result->value),
+ "r"(&op1->value),
+ "r"(&op2->value),
+ "n"(IS_LONG),
+ "n"(IS_DOUBLE),
+ "n"(ZVAL_OFFSETOF_TYPE)
: "rax");
#else
Z_LVAL_P(result) = Z_LVAL_P(op1) + Z_LVAL_P(op2);
@@ -635,7 +653,7 @@ static zend_always_inline int fast_sub_function(zval *result, zval *op1, zval *o
"subl (%2), %%eax\n\t"
"jo 0f\n\t"
"movl %%eax, (%0)\n\t"
- "movb $0x1,0xc(%0)\n\t"
+ "movb %3, %c5(%0)\n\t"
"jmp 1f\n"
"0:\n\t"
"fildl (%2)\n\t"
@@ -645,13 +663,16 @@ static zend_always_inline int fast_sub_function(zval *result, zval *op1, zval *o
#else
"fsubp %%st, %%st(1)\n\t"
#endif
- "movb $0x2,0xc(%0)\n\t"
+ "movb %4, %c5(%0)\n\t"
"fstpl (%0)\n"
"1:"
:
- : "r"(result),
- "r"(op1),
- "r"(op2)
+ : "r"(&result->value),
+ "r"(&op1->value),
+ "r"(&op2->value),
+ "n"(IS_LONG),
+ "n"(IS_DOUBLE),
+ "n"(ZVAL_OFFSETOF_TYPE)
: "eax");
#elif defined(__GNUC__) && defined(__x86_64__)
__asm__(
@@ -659,7 +680,7 @@ static zend_always_inline int fast_sub_function(zval *result, zval *op1, zval *o
"subq (%2), %%rax\n\t"
"jo 0f\n\t"
"movq %%rax, (%0)\n\t"
- "movb $0x1,0x14(%0)\n\t"
+ "movb %3, %c5(%0)\n\t"
"jmp 1f\n"
"0:\n\t"
"fildq (%2)\n\t"
@@ -669,13 +690,16 @@ static zend_always_inline int fast_sub_function(zval *result, zval *op1, zval *o
#else
"fsubp %%st, %%st(1)\n\t"
#endif
- "movb $0x2,0x14(%0)\n\t"
+ "movb %4, %c5(%0)\n\t"
"fstpl (%0)\n"
"1:"
:
- : "r"(result),
- "r"(op1),
- "r"(op2)
+ : "r"(&result->value),
+ "r"(&op1->value),
+ "r"(&op2->value),
+ "n"(IS_LONG),
+ "n"(IS_DOUBLE),
+ "n"(ZVAL_OFFSETOF_TYPE)
: "rax");
#else
Z_LVAL_P(result) = Z_LVAL_P(op1) - Z_LVAL_P(op2);
--
1.7.10.4
--
PHP Internals - PHP Runtime Development Mailing List
To unsubscribe, visit: http://www.php.net/unsub.php
