mirrors/freetype
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[base] Restore FT_MulFix inlining.

Browse Source 
* include/freetype.h (FT_MulFix): Unconditionally defined.

* src/base/ftcalc.c [FT_MULFIX_ASSEMBLER]: Move code from here...

* include/internal/ftcalc.h [FT_MULFIX_ASSEMBLER]: ... to here,
which conditionally replaces the function with an inline version
through the macro.
This commit is contained in:
Alexei Podtelezhnikov 2014-08-11 23:39:34 -04:00
parent 90be4b6377
commit 4728993fa8
4 changed files with 240 additions and 245 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
ChangeLog
View File

@ -1,3 +1,15 @@
2014-08-11 Alexei Podtelezhnikov <apodtele@gmail.com>
[base] Restore FT_MulFix inlining.
* include/freetype.h (FT_MulFix): Unconditionally defined.
* src/base/ftcalc.c [FT_MULFIX_ASSEMBLER]: Move code from here...
* include/internal/ftcalc.h [FT_MULFIX_ASSEMBLER]: ... to here,
which conditionally replaces the function with an inline version
through the macro.
2014-08-08 Alexei Podtelezhnikov <apodtele@gmail.com>
* src/base/ftbitmap.c (ft_gray_for_premultiplied_srgb_bgra): Refactor.

17
include/freetype.h
View File

@ -3762,12 +3762,6 @@ FT_BEGIN_HEADER
FT_Long c );
/* */
/* The following #if 0 ... #endif is for the documentation formatter, */
/* hiding the internal `FT_MULFIX_INLINED' macro. */
#if 0
/*************************************************************************/
/* */
/* <Function> */
@ -3801,17 +3795,6 @@ FT_BEGIN_HEADER
FT_MulFix( FT_Long a,
FT_Long b );
/* */
#endif
#ifdef FT_MULFIX_INLINED
#define FT_MulFix( a, b ) FT_MULFIX_INLINED( a, b )
#else
FT_EXPORT( FT_Long )
FT_MulFix( FT_Long a,
FT_Long b );
#endif
/*************************************************************************/
/* */

228
include/internal/ftcalc.h
View File

@ -58,6 +58,234 @@ FT_BEGIN_HEADER
/* */
/*************************************************************************/
#ifndef FT_CONFIG_OPTION_NO_ASSEMBLER
/* Provide assembler fragments for performance-critical functions. */
/* These must be defined `static __inline__' with GCC. */
#if defined( __CC_ARM ) || defined( __ARMCC__ ) /* RVCT */
#define FT_MULFIX_ASSEMBLER FT_MulFix_arm
/* documentation is in freetype.h */
static __inline FT_Int32
FT_MulFix_arm( FT_Int32 a,
FT_Int32 b )
{
register FT_Int32 t, t2;
__asm
{
smull t2, t, b, a /* (lo=t2,hi=t) = a*b */
mov a, t, asr #31 /* a = (hi >> 31) */
add a, a, #0x8000 /* a += 0x8000 */
adds t2, t2, a /* t2 += a */
adc t, t, #0 /* t += carry */
mov a, t2, lsr #16 /* a = t2 >> 16 */
orr a, a, t, lsl #16 /* a |= t << 16 */
}
return a;
}
#endif /* __CC_ARM || __ARMCC__ */
#ifdef __GNUC__
#if defined( __arm__ ) && \
( !defined( __thumb__ ) || defined( __thumb2__ ) ) && \
!( defined( __CC_ARM ) || defined( __ARMCC__ ) )
#define FT_MULFIX_ASSEMBLER FT_MulFix_arm
/* documentation is in freetype.h */
static __inline__ FT_Int32
FT_MulFix_arm( FT_Int32 a,
FT_Int32 b )
{
register FT_Int32 t, t2;
__asm__ __volatile__ (
"smull %1, %2, %4, %3\n\t" /* (lo=%1,hi=%2) = a*b */
"mov %0, %2, asr #31\n\t" /* %0 = (hi >> 31) */
#if defined( __clang__ ) && defined( __thumb2__ )
"add.w %0, %0, #0x8000\n\t" /* %0 += 0x8000 */
#else
"add %0, %0, #0x8000\n\t" /* %0 += 0x8000 */
#endif
"adds %1, %1, %0\n\t" /* %1 += %0 */
"adc %2, %2, #0\n\t" /* %2 += carry */
"mov %0, %1, lsr #16\n\t" /* %0 = %1 >> 16 */
"orr %0, %0, %2, lsl #16\n\t" /* %0 |= %2 << 16 */
: "=r"(a), "=&r"(t2), "=&r"(t)
: "r"(a), "r"(b)
: "cc" );
return a;
}
#endif /* __arm__ && */
/* ( __thumb2__ || !__thumb__ ) && */
/* !( __CC_ARM || __ARMCC__ ) */
#if defined( __i386__ )
#define FT_MULFIX_ASSEMBLER FT_MulFix_i386
/* documentation is in freetype.h */
static __inline__ FT_Int32
FT_MulFix_i386( FT_Int32 a,
FT_Int32 b )
{
register FT_Int32 result;
__asm__ __volatile__ (
"imul %%edx\n"
"movl %%edx, %%ecx\n"
"sarl $31, %%ecx\n"
"addl $0x8000, %%ecx\n"
"addl %%ecx, %%eax\n"
"adcl $0, %%edx\n"
"shrl $16, %%eax\n"
"shll $16, %%edx\n"
"addl %%edx, %%eax\n"
: "=a"(result), "=d"(b)
: "a"(a), "d"(b)
: "%ecx", "cc" );
return result;
}
#endif /* i386 */
#endif /* __GNUC__ */
#ifdef _MSC_VER /* Visual C++ */
#ifdef _M_IX86
#define FT_MULFIX_ASSEMBLER FT_MulFix_i386
/* documentation is in freetype.h */
static __inline FT_Int32
FT_MulFix_i386( FT_Int32 a,
FT_Int32 b )
{
register FT_Int32 result;
__asm
{
mov eax, a
mov edx, b
imul edx
mov ecx, edx
sar ecx, 31
add ecx, 8000h
add eax, ecx
adc edx, 0
shr eax, 16
shl edx, 16
add eax, edx
mov result, eax
}
return result;
}
#endif /* _M_IX86 */
#endif /* _MSC_VER */
#if defined( __GNUC__ ) && defined( __x86_64__ )
#define FT_MULFIX_ASSEMBLER FT_MulFix_x86_64
static __inline__ FT_Int32
FT_MulFix_x86_64( FT_Int32 a,
FT_Int32 b )
{
/* Temporarily disable the warning that C90 doesn't support */
/* `long long'. */
#if ( __GNUC__ > 4 ) || ( ( __GNUC__ == 4 ) && ( __GNUC_MINOR__ >= 6 ) )
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wlong-long"
#endif
#if 1
/* Technically not an assembly fragment, but GCC does a really good */
/* job at inlining it and generating good machine code for it. */
long long ret, tmp;
ret = (long long)a * b;
tmp = ret >> 63;
ret += 0x8000 + tmp;
return (FT_Int32)( ret >> 16 );
#else
/* For some reason, GCC 4.6 on Ubuntu 12.04 generates invalid machine */
/* code from the lines below. The main issue is that `wide_a' is not */
/* properly initialized by sign-extending `a'. Instead, the generated */
/* machine code assumes that the register that contains `a' on input */
/* can be used directly as a 64-bit value, which is wrong most of the */
/* time. */
long long wide_a = (long long)a;
long long wide_b = (long long)b;
long long result;
__asm__ __volatile__ (
"imul %2, %1\n"
"mov %1, %0\n"
"sar $63, %0\n"
"lea 0x8000(%1, %0), %0\n"
"sar $16, %0\n"
: "=&r"(result), "=&r"(wide_a)
: "r"(wide_b)
: "cc" );
return (FT_Int32)result;
#endif
#if ( __GNUC__ > 4 ) || ( ( __GNUC__ == 4 ) && ( __GNUC_MINOR__ >= 6 ) )
#pragma GCC diagnostic pop
#endif
}
#endif /* __GNUC__ && __x86_64__ */
#if defined( __GNUC__ )
#if ( __GNUC__ > 3 ) || ( ( __GNUC__ == 3 ) && ( __GNUC_MINOR__ >= 4 ) )
#if FT_SIZEOF_INT == 4
#define FT_MSB_BUILTIN( x ) ( 31 - __builtin_clz( x ) )
#elif FT_SIZEOF_LONG == 4
#define FT_MSB_BUILTIN( x ) ( 31 - __builtin_clzl( x ) )
#endif
#endif
#endif /* __GNUC__ */
#endif /* !FT_CONFIG_OPTION_NO_ASSEMBLER */
#ifdef FT_CONFIG_OPTION_INLINE_MULFIX
#ifdef FT_MULFIX_ASSEMBLER
#define FT_MulFix( a, b ) FT_MULFIX_ASSEMBLER( a, b )
#endif
#endif
/*************************************************************************/
/* */

228
src/base/ftcalc.c
View File

@ -40,235 +40,7 @@
#include FT_INTERNAL_OBJECTS_H
#ifndef FT_CONFIG_OPTION_NO_ASSEMBLER
/* Provide assembler fragments for performance-critical functions. */
/* These must be defined `static __inline__' with GCC. */
#if defined( __CC_ARM ) || defined( __ARMCC__ ) /* RVCT */
#define FT_MULFIX_ASSEMBLER FT_MulFix_arm
/* documentation is in freetype.h */
static __inline FT_Int32
FT_MulFix_arm( FT_Int32 a,
FT_Int32 b )
{
register FT_Int32 t, t2;
__asm
{
smull t2, t, b, a /* (lo=t2,hi=t) = a*b */
mov a, t, asr #31 /* a = (hi >> 31) */
add a, a, #0x8000 /* a += 0x8000 */
adds t2, t2, a /* t2 += a */
adc t, t, #0 /* t += carry */
mov a, t2, lsr #16 /* a = t2 >> 16 */
orr a, a, t, lsl #16 /* a |= t << 16 */
}
return a;
}
#endif /* __CC_ARM || __ARMCC__ */
#ifdef __GNUC__
#if defined( __arm__ ) && \
( !defined( __thumb__ ) || defined( __thumb2__ ) ) && \
!( defined( __CC_ARM ) || defined( __ARMCC__ ) )
#define FT_MULFIX_ASSEMBLER FT_MulFix_arm
/* documentation is in freetype.h */
static __inline__ FT_Int32
FT_MulFix_arm( FT_Int32 a,
FT_Int32 b )
{
register FT_Int32 t, t2;
__asm__ __volatile__ (
"smull %1, %2, %4, %3\n\t" /* (lo=%1,hi=%2) = a*b */
"mov %0, %2, asr #31\n\t" /* %0 = (hi >> 31) */
#if defined( __clang__ ) && defined( __thumb2__ )
"add.w %0, %0, #0x8000\n\t" /* %0 += 0x8000 */
#else
"add %0, %0, #0x8000\n\t" /* %0 += 0x8000 */
#endif
"adds %1, %1, %0\n\t" /* %1 += %0 */
"adc %2, %2, #0\n\t" /* %2 += carry */
"mov %0, %1, lsr #16\n\t" /* %0 = %1 >> 16 */
"orr %0, %0, %2, lsl #16\n\t" /* %0 |= %2 << 16 */
: "=r"(a), "=&r"(t2), "=&r"(t)
: "r"(a), "r"(b)
: "cc" );
return a;
}
#endif /* __arm__ && */
/* ( __thumb2__ || !__thumb__ ) && */
/* !( __CC_ARM || __ARMCC__ ) */
#if defined( __i386__ )
#define FT_MULFIX_ASSEMBLER FT_MulFix_i386
/* documentation is in freetype.h */
static __inline__ FT_Int32
FT_MulFix_i386( FT_Int32 a,
FT_Int32 b )
{
register FT_Int32 result;
__asm__ __volatile__ (
"imul %%edx\n"
"movl %%edx, %%ecx\n"
"sarl $31, %%ecx\n"
"addl $0x8000, %%ecx\n"
"addl %%ecx, %%eax\n"
"adcl $0, %%edx\n"
"shrl $16, %%eax\n"
"shll $16, %%edx\n"
"addl %%edx, %%eax\n"
: "=a"(result), "=d"(b)
: "a"(a), "d"(b)
: "%ecx", "cc" );
return result;
}
#endif /* i386 */
#endif /* __GNUC__ */
#ifdef _MSC_VER /* Visual C++ */
#ifdef _M_IX86
#define FT_MULFIX_ASSEMBLER FT_MulFix_i386
/* documentation is in freetype.h */
static __inline FT_Int32
FT_MulFix_i386( FT_Int32 a,
FT_Int32 b )
{
register FT_Int32 result;
__asm
{
mov eax, a
mov edx, b
imul edx
mov ecx, edx
sar ecx, 31
add ecx, 8000h
add eax, ecx
adc edx, 0
shr eax, 16
shl edx, 16
add eax, edx
mov result, eax
}
return result;
}
#endif /* _M_IX86 */
#endif /* _MSC_VER */
#if defined( __GNUC__ ) && defined( __x86_64__ )
#define FT_MULFIX_ASSEMBLER FT_MulFix_x86_64
static __inline__ FT_Int32
FT_MulFix_x86_64( FT_Int32 a,
FT_Int32 b )
{
/* Temporarily disable the warning that C90 doesn't support */
/* `long long'. */
#if ( __GNUC__ > 4 ) || ( ( __GNUC__ == 4 ) && ( __GNUC_MINOR__ >= 6 ) )
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wlong-long"
#endif
#if 1
/* Technically not an assembly fragment, but GCC does a really good */
/* job at inlining it and generating good machine code for it. */
long long ret, tmp;
ret = (long long)a * b;
tmp = ret >> 63;
ret += 0x8000 + tmp;
return (FT_Int32)( ret >> 16 );
#else
/* For some reason, GCC 4.6 on Ubuntu 12.04 generates invalid machine */
/* code from the lines below. The main issue is that `wide_a' is not */
/* properly initialized by sign-extending `a'. Instead, the generated */
/* machine code assumes that the register that contains `a' on input */
/* can be used directly as a 64-bit value, which is wrong most of the */
/* time. */
long long wide_a = (long long)a;
long long wide_b = (long long)b;
long long result;
__asm__ __volatile__ (
"imul %2, %1\n"
"mov %1, %0\n"
"sar $63, %0\n"
"lea 0x8000(%1, %0), %0\n"
"sar $16, %0\n"
: "=&r"(result), "=&r"(wide_a)
: "r"(wide_b)
: "cc" );
return (FT_Int32)result;
#endif
#if ( __GNUC__ > 4 ) || ( ( __GNUC__ == 4 ) && ( __GNUC_MINOR__ >= 6 ) )
#pragma GCC diagnostic pop
#endif
}
#endif /* __GNUC__ && __x86_64__ */
#if defined( __GNUC__ )
#if ( __GNUC__ > 3 ) || ( ( __GNUC__ == 3 ) && ( __GNUC_MINOR__ >= 4 ) )
#if FT_SIZEOF_INT == 4
#define FT_MSB_BUILTIN( x ) ( 31 - __builtin_clz( x ) )
#elif FT_SIZEOF_LONG == 4
#define FT_MSB_BUILTIN( x ) ( 31 - __builtin_clzl( x ) )
#endif
#endif
#endif /* __GNUC__ */
#endif /* !FT_CONFIG_OPTION_NO_ASSEMBLER */
#ifdef FT_CONFIG_OPTION_INLINE_MULFIX
#ifdef FT_MULFIX_ASSEMBLER
#define FT_MULFIX_INLINED FT_MULFIX_ASSEMBLER
#endif
#endif
#ifdef FT_MULFIX_INLINED
#undef FT_MulFix
#endif