mirror of
https://github.com/LadybirdBrowser/ladybird.git
synced 2024-11-21 23:20:20 +00:00
AK: Reorder AK/Math after the removal of i686
Without i686 we can remove some implementations for sqrt and round_to.
This commit is contained in:
Hendiadyoin1
Jelle Raaijmakers
parent
ce0f41b9fb
commit
1bdc4e6b29
Notes:
sideshowbarker
2024-07-17 09:56:35 +09:00
Author: https://github.com/Hendiadyoin1 Commit: https://github.com/SerenityOS/serenity/commit/1bdc4e6b29 Pull-request: https://github.com/SerenityOS/serenity/pull/16705 Reviewed-by: https://github.com/gmta ✅
1 changed files with 119 additions and 123 deletions
226
AK/Math.h
226
AK/Math.h
|
|
@ -123,6 +123,20 @@ constexpr T sqrt(T x)
|
|||
CONSTEXPR_STATE(sqrt, x);
|
||||
|
||||
#if ARCH(X86_64)
|
||||
if constexpr (IsSame<T, float>) {
|
||||
float res;
|
||||
asm("sqrtss %1, %0"
|
||||
: "=x"(res)
|
||||
: "x"(x));
|
||||
return res;
|
||||
}
|
||||
if constexpr (IsSame<T, double>) {
|
||||
double res;
|
||||
asm("sqrtsd %1, %0"
|
||||
: "=x"(res)
|
||||
: "x"(x));
|
||||
return res;
|
||||
}
|
||||
T res;
|
||||
asm("fsqrt"
|
||||
: "=t"(res)
|
||||
|
|
@ -140,45 +154,8 @@ constexpr T rsqrt(T x)
|
|||
{
|
||||
#if ARCH(AARCH64)
|
||||
AARCH64_INSTRUCTION(frsqrte, x);
|
||||
#endif
|
||||
return (T)1. / sqrt(x);
|
||||
}
|
||||
|
||||
#if ARCH(x86_64)
|
||||
template<>
|
||||
constexpr float sqrt(float x)
|
||||
{
|
||||
if (is_constant_evaluated())
|
||||
return __builtin_sqrtf(x);
|
||||
|
||||
float res;
|
||||
asm("sqrtss %1, %0"
|
||||
: "=x"(res)
|
||||
: "x"(x));
|
||||
return res;
|
||||
}
|
||||
|
||||
# ifdef __SSE2__
|
||||
template<>
|
||||
constexpr double sqrt(double x)
|
||||
{
|
||||
if (is_constant_evaluated())
|
||||
return __builtin_sqrt(x);
|
||||
|
||||
double res;
|
||||
asm("sqrtsd %1, %0"
|
||||
: "=x"(res)
|
||||
: "x"(x));
|
||||
return res;
|
||||
}
|
||||
# endif
|
||||
|
||||
template<>
|
||||
constexpr float rsqrt(float x)
|
||||
{
|
||||
if (is_constant_evaluated())
|
||||
return 1.f / __builtin_sqrtf(x);
|
||||
|
||||
#elif ARCH(X86_64)
|
||||
if constexpr (IsSame<T, float>) {
|
||||
float res;
|
||||
asm("rsqrtss %1, %0"
|
||||
: "=x"(res)
|
||||
|
|
@ -186,6 +163,8 @@ constexpr float rsqrt(float x)
|
|||
return res;
|
||||
}
|
||||
#endif
|
||||
return (T)1. / sqrt(x);
|
||||
}
|
||||
|
||||
template<FloatingPoint T>
|
||||
constexpr T cbrt(T x)
|
||||
|
|
@ -583,9 +562,12 @@ using Hyperbolic::sinh;
|
|||
using Hyperbolic::tanh;
|
||||
|
||||
template<Integral I, FloatingPoint P>
|
||||
ALWAYS_INLINE I round_to(P value)
|
||||
{
|
||||
ALWAYS_INLINE I round_to(P value);
|
||||
|
||||
#if ARCH(X86_64)
|
||||
template<Integral I>
|
||||
ALWAYS_INLINE I round_to(long double value)
|
||||
{
|
||||
// Note: fistps outputs into a signed integer location (i16, i32, i64),
|
||||
// so lets be nice and tell the compiler that.
|
||||
Conditional<sizeof(I) >= sizeof(i16), MakeSigned<I>, i16> ret;
|
||||
|
|
@ -606,115 +588,129 @@ ALWAYS_INLINE I round_to(P value)
|
|||
: "st");
|
||||
}
|
||||
return static_cast<I>(ret);
|
||||
}
|
||||
|
||||
template<Integral I>
|
||||
ALWAYS_INLINE I round_to(float value)
|
||||
{
|
||||
// FIXME: round_to<u64> might will cause issues, aka the indefinite value being set,
|
||||
// if the value surpasses the i64 limit, even if the result could fit into an u64
|
||||
// To solve this we would either need to detect that value or do a range check and
|
||||
// then do a more specialized conversion, which might include a division (which is expensive)
|
||||
if constexpr (sizeof(I) == sizeof(i64) || IsSame<I, u32>) {
|
||||
i64 ret;
|
||||
asm("cvtss2si %1, %0"
|
||||
: "=r"(ret)
|
||||
: "xm"(value));
|
||||
return static_cast<I>(ret);
|
||||
}
|
||||
i32 ret;
|
||||
asm("cvtss2si %1, %0"
|
||||
: "=r"(ret)
|
||||
: "xm"(value));
|
||||
return static_cast<I>(ret);
|
||||
}
|
||||
|
||||
template<Integral I>
|
||||
ALWAYS_INLINE I round_to(double value)
|
||||
{
|
||||
// FIXME: round_to<u64> might will cause issues, aka the indefinite value being set,
|
||||
// if the value surpasses the i64 limit, even if the result could fit into an u64
|
||||
// To solve this we would either need to detect that value or do a range check and
|
||||
// then do a more specialized conversion, which might include a division (which is expensive)
|
||||
if constexpr (sizeof(I) == sizeof(i64) || IsSame<I, u32>) {
|
||||
i64 ret;
|
||||
asm("cvtsd2si %1, %0"
|
||||
: "=r"(ret)
|
||||
: "xm"(value));
|
||||
return static_cast<I>(ret);
|
||||
}
|
||||
i32 ret;
|
||||
asm("cvtsd2si %1, %0"
|
||||
: "=r"(ret)
|
||||
: "xm"(value));
|
||||
return static_cast<I>(ret);
|
||||
}
|
||||
|
||||
#elif ARCH(AARCH64)
|
||||
if constexpr (IsSigned<I>) {
|
||||
if constexpr (sizeof(I) <= sizeof(i32)) {
|
||||
template<Signed I>
|
||||
ALWAYS_INLINE I round_to(float value)
|
||||
{
|
||||
if constexpr (sizeof(I) <= sizeof(u32)) {
|
||||
i32 res;
|
||||
if constexpr (IsSame<P, float>) {
|
||||
asm("fcvtns %w0, %s1"
|
||||
: "=r"(res)
|
||||
: "w"(value));
|
||||
} else if constexpr (IsSame<P, double>) {
|
||||
asm("fcvtns %w0, %d1"
|
||||
: "=r"(res)
|
||||
: "w"(value));
|
||||
} else if constexpr (IsSame<P, long double>) {
|
||||
TODO();
|
||||
}
|
||||
return static_cast<I>(res);
|
||||
}
|
||||
// either long or long long aka i64
|
||||
i64 res;
|
||||
if constexpr (IsSame<P, float>) {
|
||||
asm("fcvtns %0, %s1"
|
||||
: "=r"(res)
|
||||
: "w"(value));
|
||||
} else if constexpr (IsSame<P, double>) {
|
||||
return static_cast<I>(res);
|
||||
}
|
||||
|
||||
template<Signed I>
|
||||
ALWAYS_INLINE I round_to(double value)
|
||||
{
|
||||
if constexpr (sizeof(I) <= sizeof(u32)) {
|
||||
i32 res;
|
||||
asm("fcvtns %w0, %d1"
|
||||
: "=r"(res)
|
||||
: "w"(value));
|
||||
return static_cast<I>(res);
|
||||
}
|
||||
i64 res;
|
||||
asm("fcvtns %0, %d1"
|
||||
: "=r"(res)
|
||||
: "w"(value));
|
||||
} else if constexpr (IsSame<P, long double>) {
|
||||
TODO();
|
||||
}
|
||||
return static_cast<I>(res);
|
||||
}
|
||||
|
||||
if constexpr (sizeof(I) <= sizeof(u32)) {
|
||||
template<Unsigned U>
|
||||
ALWAYS_INLINE U round_to(float value)
|
||||
{
|
||||
if constexpr (sizeof(U) <= sizeof(u32)) {
|
||||
u32 res;
|
||||
if constexpr (IsSame<P, float>) {
|
||||
asm("fcvtnu %w0, %s1"
|
||||
: "=r"(res)
|
||||
: "w"(value));
|
||||
} else if constexpr (IsSame<P, double>) {
|
||||
asm("fcvtnu %w0, %d1"
|
||||
: "=r"(res)
|
||||
: "w"(value));
|
||||
} else if constexpr (IsSame<P, long double>) {
|
||||
TODO();
|
||||
return static_cast<U>(res);
|
||||
}
|
||||
return static_cast<I>(res);
|
||||
}
|
||||
|
||||
// either unsigned long or unsigned long long aka u64
|
||||
u64 res;
|
||||
if constexpr (IsSame<P, float>) {
|
||||
i64 res;
|
||||
asm("fcvtnu %0, %s1"
|
||||
: "=r"(res)
|
||||
: "w"(value));
|
||||
} else if constexpr (IsSame<P, double>) {
|
||||
asm("fcvtnu %0, %d1"
|
||||
return static_cast<U>(res);
|
||||
}
|
||||
|
||||
template<Unsigned U>
|
||||
ALWAYS_INLINE U round_to(double value)
|
||||
{
|
||||
if constexpr (sizeof(U) <= sizeof(u32)) {
|
||||
u32 res;
|
||||
asm("fcvtns %w0, %d1"
|
||||
: "=r"(res)
|
||||
: "w"(value));
|
||||
} else if constexpr (IsSame<P, long double>) {
|
||||
TODO();
|
||||
return static_cast<U>(res);
|
||||
}
|
||||
return static_cast<I>(res);
|
||||
i64 res;
|
||||
asm("fcvtns %0, %d1"
|
||||
: "=r"(res)
|
||||
: "w"(value));
|
||||
return static_cast<U>(res);
|
||||
}
|
||||
|
||||
#else
|
||||
template<Integral I, FloatingPoint P>
|
||||
ALWAYS_INLINE I round_to(P value)
|
||||
{
|
||||
if constexpr (IsSame<P, long double>)
|
||||
return static_cast<I>(__builtin_llrintl(value));
|
||||
if constexpr (IsSame<P, double>)
|
||||
return static_cast<I>(__builtin_llrint(value));
|
||||
if constexpr (IsSame<P, float>)
|
||||
return static_cast<I>(__builtin_llrintf(value));
|
||||
#endif
|
||||
}
|
||||
|
||||
#if ARCH(x86_64)
|
||||
template<Integral I>
|
||||
ALWAYS_INLINE I round_to(float value)
|
||||
{
|
||||
if constexpr (sizeof(I) == sizeof(i64)) {
|
||||
// Note: Outputting into 64-bit registers or memory locations requires the
|
||||
// REX prefix, so we have to fall back to long doubles on platforms
|
||||
i64 ret;
|
||||
asm("cvtss2si %1, %0"
|
||||
: "=r"(ret)
|
||||
: "xm"(value));
|
||||
return static_cast<I>(ret);
|
||||
}
|
||||
i32 ret;
|
||||
asm("cvtss2si %1, %0"
|
||||
: "=r"(ret)
|
||||
: "xm"(value));
|
||||
return static_cast<I>(ret);
|
||||
}
|
||||
#endif
|
||||
#ifdef __SSE2__
|
||||
template<Integral I>
|
||||
ALWAYS_INLINE I round_to(double value)
|
||||
{
|
||||
if constexpr (sizeof(I) == sizeof(i64)) {
|
||||
i64 ret;
|
||||
asm("cvtsd2si %1, %0"
|
||||
: "=r"(ret)
|
||||
: "xm"(value));
|
||||
return static_cast<I>(ret);
|
||||
}
|
||||
i32 ret;
|
||||
asm("cvtsd2si %1, %0"
|
||||
: "=r"(ret)
|
||||
: "xm"(value));
|
||||
return static_cast<I>(ret);
|
||||
}
|
||||
#endif
|
||||
|
||||
|
|
|
|||
Loading…
Reference in a new issue