Skip to content

Commit

Permalink
gcc's version of inline assembly for the maths helpers
Browse files Browse the repository at this point in the history
  • Loading branch information
@PoroCYon
PoroCYon committed Sep 9, 2020
1 parent 4d7493e commit df0aac6
Showing 1 changed file with 192 additions and 6 deletions.
198 changes: 192 additions & 6 deletions WaveSabreCore/src/Helpers.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -4,8 +4,17 @@
#define _USE_MATH_DEFINES
#include <math.h>

// TODO: make assembly equivalent for Windows x64 (use intrinsic ?)
// ^--- you probably only need to change esp to rsp? -poro

#if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__GNUC__)
#define ASM_MATH_AVAILABLE (1)
#else /* nor MSVC nor GCC/clang */
#define ASM_MATH_AVAILABLE (0)
#endif

#if ASM_MATH_AVAILABLE == 1
#if defined(_MSC_VER) && defined(_M_IX86)
// TODO: make assembly equivalent for x64 (use intrinsic ?)
static __declspec(naked) double __vectorcall fpuPow(double x, double y)
{
__asm
Expand Down Expand Up @@ -53,7 +62,87 @@ static __declspec(naked) double __vectorcall fpuPow(double x, double y)
ret
}
}
#elif defined(__GNUC__)
#if defined(__x86_64__) || defined(__i386__)
__attribute__((__naked__,__noinline__)) static double fpuPow(double x, double y)
{
// i386 Linux ABI: pass thru the stack, return in st(0)
// x86_64 SysV ABI: pass/return thru xmm0/1
asm volatile(
#ifdef __x86_64__
"subq $8, %%rsp\n"
#else
"movsd 4(%%esp), %%xmm0\n"
"movsd 12(%%esp), %%xmm1\n"
"subl $8, %%esp\n"
#endif
"xorpd %%xmm2, %%xmm2\n"
"comisd %%xmm2, %%xmm1\n"
"jne 1f\n"

"fld1\n"
"jmp 3f\n"

"1:\n"
"comisd %%xmm2, %%xmm0\n"
"jne 2f\n"

"fldz\n"
"jmp 3f\n"

"2:\n"
#ifdef __x86_64__
"movsd %%xmm1, (%%rsp)\n"
"fldl (%%rsp)\n"
"movsd %%xmm0, (%%rsp)\n"
"fldl (%%rsp)\n"
#else
"movsd %%xmm1, (%%esp)\n"
"fldl (%%esp)\n"
"movsd %%xmm0, (%%esp)\n"
"fldl (%%esp)\n"
#endif

"fyl2x\n"
"fld %%st(0)\n"
"frndint\n"
"fsub %%st(0), %%st(1)\n"
"fxch %%st(1)\n"
"fchs\n"
"f2xm1\n"
"fld1\n"
"faddp %%st(0), %%st(1)\n"
"fscale\n"
"fstp %%st(1)\n"

"3:\n"
#ifdef __x86_64__
"fstpl (%%rsp)\n"
"movsd (%%rsp), %%xmm0\n"
"addq $8, %%rsp\n"
#else
"addl $8, %%esp\n"
#endif
"ret\n"
:// no output
:// no input
:"xmm2" // clobbered
);
}
#else
// __builtin_pow only supports integer exponents... so if the exponent
// is an integer, use __builtin_pow, using some preprocessor magic
#define fpuPow(x, y) \
((__builtin_constant_p(y) && ((y) == (int)(y))) \
? __builtin_pow(x, y) \
: pow(x, y)) \

#endif
#else
#error "Unsupported compiler."
#endif /* compiler */

#if defined(_MSC_VER) && defined(_M_IX86)
static __declspec(naked) float __vectorcall fpuPowF(float x, float y)
{
__asm
Expand Down Expand Up @@ -101,7 +190,87 @@ static __declspec(naked) float __vectorcall fpuPowF(float x, float y)
ret
}
}
#elif defined(__GNUC__)
#if defined(__x86_64__) || defined(__i386__)
__attribute__((__naked__,__noinline__)) static float fpuPowF(float x, float y)
{
// i386 Linux ABI: pass thru the stack, return in st(0)
// x86_64 SysV ABI: pass/return thru xmm0/1
asm volatile(
#ifdef __x86_64__
"subq $8, %%rsp\n"
#else
"movss 4(%%esp), %%xmm0\n"
"movss 8(%%esp), %%xmm1\n"
"subl $8, %%esp\n"
#endif
"xorps %%xmm2, %%xmm2\n"
"comiss %%xmm2, %%xmm1\n"
"jne 1f\n"

"fld1\n"
"jmp 3f\n"

"1:\n"
"comiss %%xmm2, %%xmm0\n"
"jne 2f\n"

"fldz\n"
"jmp 3f\n"

"2:\n"
#ifdef __x86_64__
"movss %%xmm1, (%%rsp)\n"
"flds (%%rsp)\n"
"movss %%xmm0, (%%rsp)\n"
"flds (%%rsp)\n"
#else
"movss %%xmm1, (%%esp)\n"
"flds (%%esp)\n"
"movss %%xmm0, (%%esp)\n"
"flds (%%esp)\n"
#endif

"fyl2x\n"
"fld %%st(0)\n"
"frndint\n"
"fsub %%st(0), %%st(1)\n"
"fxch %%st(1)\n"
"fchs\n"
"f2xm1\n"
"fld1\n"
"faddp %%st(0), %%st(1)\n"
"fscale\n"
"fstp %%st(1)\n"

"3:\n"
#ifdef __x86_64__
"fstps (%%rsp)\n"
"movss (%%rsp), %%xmm0\n"
"addq $8, %%rsp\n"
#else
"addl $8, %%esp\n"
#endif
"ret\n"
:// no output
:// no input
:"xmm2" // clobbered
);
}
#else
// __builtin_powf only supports integer exponents... so if the exponent
// is an integer, use __builtin_powf, using some preprocessor magic
#define fpuPowF(x, y) \
((__builtin_constant_p(y) && ((y) == (int)(y))) \
? __builtin_powf(x, y) \
: powf(x, y)) \

#endif
#else
#error "Unsupported compiler."
#endif /* compiler */

#if defined(_MSC_VER) && defined(_M_IX86)
static __declspec(naked) double __vectorcall fpuCos(double x)
{
__asm
Expand All @@ -119,7 +288,24 @@ static __declspec(naked) double __vectorcall fpuCos(double x)
ret
}
}
#endif // defined(_MSC_VER) && defined(_M_IX86)
#elif defined(__GNUC__)
#if defined(__x86_64__) || defined(__i386__)
__attribute__((__always_inline__)) inline static double fpuCos(double x)
{
// not writing the *entire* function body in assembly actually helps
// gcc and clang with inlining and LTO
// ... except trying this with fpuPow/F somehow got botched, so those I
// wrote as pure assembly
asm volatile("fcos\n":"+t"(x)::);
return x;
}
#else /* x86_64 */
#define fpuCos(x) __builtin_cos(x)
#endif /* GNUC, platform */
#else
#error "Unsupported compiler."
#endif /* compiler */
#endif // ASM_MATH_AVAILABLE == 1

namespace WaveSabreCore
{
Expand All @@ -138,7 +324,7 @@ namespace WaveSabreCore
for (int i = 0; i < fastCosTabSize + 1; i++)
{
double phase = double(i) * ((M_PI * 2) / fastCosTabSize);
#if defined(_MSC_VER) && defined(_M_IX86)
#if ASM_MATH_AVAILABLE == 1
fastCosTab[i] = fpuCos(phase);
#else
fastCosTab[i] = cos(phase);
Expand All @@ -153,7 +339,7 @@ namespace WaveSabreCore

double Helpers::Pow(double x, double y)
{
#if defined(_MSC_VER) && defined(_M_IX86)
#if ASM_MATH_AVAILABLE == 1
return fpuPow(x, y);
#else
return pow(x, y);
Expand All @@ -162,7 +348,7 @@ namespace WaveSabreCore

float Helpers::PowF(float x, float y)
{
#if defined(_MSC_VER) && defined(_M_IX86)
#if ASM_MATH_AVAILABLE == 1
return fpuPowF(x, y);
#else
return powf(x, y);
Expand Down Expand Up @@ -365,7 +551,7 @@ namespace WaveSabreCore
{
return (Spread)(int)(param * 2.0f);
}

float Helpers::SpreadToParam(Spread spread)
{
return (float)spread / 2.0f;
Expand Down

0 comments on commit df0aac6

Please sign in to comment.