commits
tags
/* Bra.c -- Branch converters for RISC code
2023-04-02 : Igor Pavlov : Public domain */
#include "Precomp.h"
#include "Bra.h"
#include "CpuArch.h"
#include "RotateDefs.h"
#if defined(MY_CPU_SIZEOF_POINTER) \
&& ( MY_CPU_SIZEOF_POINTER == 4 \
|| MY_CPU_SIZEOF_POINTER == 8)
#define BR_CONV_USE_OPT_PC_PTR
#endif
#ifdef BR_CONV_USE_OPT_PC_PTR
#define BR_PC_INIT pc -= (UInt32)(SizeT)p;
#define BR_PC_GET (pc + (UInt32)(SizeT)p)
#else
#define BR_PC_INIT pc += (UInt32)size;
#define BR_PC_GET (pc - (UInt32)(SizeT)(lim - p))
// #define BR_PC_INIT
// #define BR_PC_GET (pc + (UInt32)(SizeT)(p - data))
#endif
#define BR_CONVERT_VAL(v, c) if (encoding) v += c; else v -= c;
// #define BR_CONVERT_VAL(v, c) if (!encoding) c = (UInt32)0 - c; v += c;
#define Z7_BRANCH_CONV(name) z7_BranchConv_ ## name
#define Z7_BRANCH_FUNC_MAIN(name) \
static \
Z7_FORCE_INLINE \
Z7_ATTRIB_NO_VECTOR \
Byte *Z7_BRANCH_CONV(name)(Byte *p, SizeT size, UInt32 pc, int encoding)
#define Z7_BRANCH_FUNC_IMP(name, m, encoding) \
Z7_NO_INLINE \
Z7_ATTRIB_NO_VECTOR \
Byte *m(name)(Byte *data, SizeT size, UInt32 pc) \
{ return Z7_BRANCH_CONV(name)(data, size, pc, encoding); } \
#ifdef Z7_EXTRACT_ONLY
#define Z7_BRANCH_FUNCS_IMP(name) \
Z7_BRANCH_FUNC_IMP(name, Z7_BRANCH_CONV_DEC, 0)
#else
#define Z7_BRANCH_FUNCS_IMP(name) \
Z7_BRANCH_FUNC_IMP(name, Z7_BRANCH_CONV_DEC, 0) \
Z7_BRANCH_FUNC_IMP(name, Z7_BRANCH_CONV_ENC, 1)
#endif
#if defined(__clang__)
#define BR_EXTERNAL_FOR
#define BR_NEXT_ITERATION continue;
#else
#define BR_EXTERNAL_FOR for (;;)
#define BR_NEXT_ITERATION break;
#endif
#if defined(__clang__) && (__clang_major__ >= 8) \
|| defined(__GNUC__) && (__GNUC__ >= 1000) \
// GCC is not good for __builtin_expect() here
/* || defined(_MSC_VER) && (_MSC_VER >= 1920) */
// #define Z7_unlikely [[unlikely]]
// #define Z7_LIKELY(x) (__builtin_expect((x), 1))
#define Z7_UNLIKELY(x) (__builtin_expect((x), 0))
// #define Z7_likely [[likely]]
#else
// #define Z7_LIKELY(x) (x)
#define Z7_UNLIKELY(x) (x)
// #define Z7_likely
#endif
Z7_BRANCH_FUNC_MAIN(ARM64)
{
// Byte *p = data;
const Byte *lim;
const UInt32 flag = (UInt32)1 << (24 - 4);
const UInt32 mask = ((UInt32)1 << 24) - (flag << 1);
size &= ~(SizeT)3;
// if (size == 0) return p;
lim = p + size;
BR_PC_INIT
pc -= 4; // because (p) will point to next instruction
BR_EXTERNAL_FOR
{
// Z7_PRAGMA_OPT_DISABLE_LOOP_UNROLL_VECTORIZE
for (;;)
{
UInt32 v;
if Z7_UNLIKELY(p == lim)
return p;
v = GetUi32a(p);
p += 4;
if Z7_UNLIKELY(((v - 0x94000000) & 0xfc000000) == 0)
{
UInt32 c = BR_PC_GET >> 2;
BR_CONVERT_VAL(v, c)
v &= 0x03ffffff;
v |= 0x94000000;
SetUi32a(p - 4, v)
BR_NEXT_ITERATION
}
// v = rotlFixed(v, 8); v += (flag << 8) - 0x90; if Z7_UNLIKELY((v & ((mask << 8) + 0x9f)) == 0)
v -= 0x90000000; if Z7_UNLIKELY((v & 0x9f000000) == 0)
{
UInt32 z, c;
// v = rotrFixed(v, 8);
v += flag; if Z7_UNLIKELY(v & mask) continue;
z = (v & 0xffffffe0) | (v >> 26);
c = (BR_PC_GET >> (12 - 3)) & ~(UInt32)7;
BR_CONVERT_VAL(z, c)
v &= 0x1f;
v |= 0x90000000;
v |= z << 26;
v |= 0x00ffffe0 & ((z & (((flag << 1) - 1))) - flag);
SetUi32a(p - 4, v)
}
}
}
}
Z7_BRANCH_FUNCS_IMP(ARM64)
Z7_BRANCH_FUNC_MAIN(ARM)
{
// Byte *p = data;
const Byte *lim;
size &= ~(SizeT)3;
lim = p + size;
BR_PC_INIT
/* in ARM: branch offset is relative to the +2 instructions from current instruction.
(p) will point to next instruction */
pc += 8 - 4;
for (;;)
{
for (;;)
{
if Z7_UNLIKELY(p >= lim) { return p; } p += 4; if Z7_UNLIKELY(p[-1] == 0xeb) break;
if Z7_UNLIKELY(p >= lim) { return p; } p += 4; if Z7_UNLIKELY(p[-1] == 0xeb) break;
}
{
UInt32 v = GetUi32a(p - 4);
UInt32 c = BR_PC_GET >> 2;
BR_CONVERT_VAL(v, c)
v &= 0x00ffffff;
v |= 0xeb000000;
SetUi32a(p - 4, v)
}
}
}
Z7_BRANCH_FUNCS_IMP(ARM)
Z7_BRANCH_FUNC_MAIN(PPC)
{
// Byte *p = data;
const Byte *lim;
size &= ~(SizeT)3;
lim = p + size;
BR_PC_INIT
pc -= 4; // because (p) will point to next instruction
for (;;)
{
UInt32 v;
for (;;)
{
if Z7_UNLIKELY(p == lim)
return p;
// v = GetBe32a(p);
v = *(UInt32 *)(void *)p;
p += 4;
// if ((v & 0xfc000003) == 0x48000001) break;
// if ((p[-4] & 0xFC) == 0x48 && (p[-1] & 3) == 1) break;
if Z7_UNLIKELY(
((v - Z7_CONV_BE_TO_NATIVE_CONST32(0x48000001))
& Z7_CONV_BE_TO_NATIVE_CONST32(0xfc000003)) == 0) break;
}
{
v = Z7_CONV_NATIVE_TO_BE_32(v);
{
UInt32 c = BR_PC_GET;
BR_CONVERT_VAL(v, c)
}
v &= 0x03ffffff;
v |= 0x48000000;
SetBe32a(p - 4, v)
}
}
}
Z7_BRANCH_FUNCS_IMP(PPC)
#ifdef Z7_CPU_FAST_ROTATE_SUPPORTED
#define BR_SPARC_USE_ROTATE
#endif
Z7_BRANCH_FUNC_MAIN(SPARC)
{
// Byte *p = data;
const Byte *lim;
const UInt32 flag = (UInt32)1 << 22;
size &= ~(SizeT)3;
lim = p + size;
BR_PC_INIT
pc -= 4; // because (p) will point to next instruction
for (;;)
{
UInt32 v;
for (;;)
{
if Z7_UNLIKELY(p == lim)
return p;
/* // the code without GetBe32a():
{ const UInt32 v = GetUi16a(p) & 0xc0ff; p += 4; if (v == 0x40 || v == 0xc07f) break; }
*/
v = GetBe32a(p);
p += 4;
#ifdef BR_SPARC_USE_ROTATE
v = rotlFixed(v, 2);
v += (flag << 2) - 1;
if Z7_UNLIKELY((v & (3 - (flag << 3))) == 0)
#else
v += (UInt32)5 << 29;
v ^= (UInt32)7 << 29;
v += flag;
if Z7_UNLIKELY((v & (0 - (flag << 1))) == 0)
#endif
break;
}
{
// UInt32 v = GetBe32a(p - 4);
#ifndef BR_SPARC_USE_ROTATE
v <<= 2;
#endif
{
UInt32 c = BR_PC_GET;
BR_CONVERT_VAL(v, c)
}
v &= (flag << 3) - 1;
#ifdef BR_SPARC_USE_ROTATE
v -= (flag << 2) - 1;
v = rotrFixed(v, 2);
#else
v -= (flag << 2);
v >>= 2;
v |= (UInt32)1 << 30;
#endif
SetBe32a(p - 4, v)
}
}
}
Z7_BRANCH_FUNCS_IMP(SPARC)
Z7_BRANCH_FUNC_MAIN(ARMT)
{
// Byte *p = data;
Byte *lim;
size &= ~(SizeT)1;
// if (size == 0) return p;
if (size <= 2) return p;
size -= 2;
lim = p + size;
BR_PC_INIT
/* in ARM: branch offset is relative to the +2 instructions from current instruction.
(p) will point to the +2 instructions from current instruction */
// pc += 4 - 4;
// if (encoding) pc -= 0xf800 << 1; else pc += 0xf800 << 1;
// #define ARMT_TAIL_PROC { goto armt_tail; }
#define ARMT_TAIL_PROC { return p; }
do
{
/* in MSVC 32-bit x86 compilers:
UInt32 version : it loads value from memory with movzx
Byte version : it loads value to 8-bit register (AL/CL)
movzx version is slightly faster in some cpus
*/
unsigned b1;
// Byte / unsigned
b1 = p[1];
// optimized version to reduce one (p >= lim) check:
// unsigned a1 = p[1]; b1 = p[3]; p += 2; if Z7_LIKELY((b1 & (a1 ^ 8)) < 0xf8)
for (;;)
{
unsigned b3; // Byte / UInt32
/* (Byte)(b3) normalization can use low byte computations in MSVC.
It gives smaller code, and no loss of speed in some compilers/cpus.
But new MSVC 32-bit x86 compilers use more slow load
from memory to low byte register in that case.
So we try to use full 32-bit computations for faster code.
*/
// if (p >= lim) { ARMT_TAIL_PROC } b3 = b1 + 8; b1 = p[3]; p += 2; if ((b3 & b1) >= 0xf8) break;
if Z7_UNLIKELY(p >= lim) { ARMT_TAIL_PROC } b3 = p[3]; p += 2; if Z7_UNLIKELY((b3 & (b1 ^ 8)) >= 0xf8) break;
if Z7_UNLIKELY(p >= lim) { ARMT_TAIL_PROC } b1 = p[3]; p += 2; if Z7_UNLIKELY((b1 & (b3 ^ 8)) >= 0xf8) break;
}
{
/* we can adjust pc for (0xf800) to rid of (& 0x7FF) operation.
But gcc/clang for arm64 can use bfi instruction for full code here */
UInt32 v =
((UInt32)GetUi16a(p - 2) << 11) |
((UInt32)GetUi16a(p) & 0x7FF);
/*
UInt32 v =
((UInt32)p[1 - 2] << 19)
+ (((UInt32)p[1] & 0x7) << 8)
+ (((UInt32)p[-2] << 11))
+ (p[0]);
*/
p += 2;
{
UInt32 c = BR_PC_GET >> 1;
BR_CONVERT_VAL(v, c)
}
SetUi16a(p - 4, (UInt16)(((v >> 11) & 0x7ff) | 0xf000))
SetUi16a(p - 2, (UInt16)(v | 0xf800))
/*
p[-4] = (Byte)(v >> 11);
p[-3] = (Byte)(0xf0 | ((v >> 19) & 0x7));
p[-2] = (Byte)v;
p[-1] = (Byte)(0xf8 | (v >> 8));
*/
}
}
while (p < lim);
return p;
// armt_tail:
// if ((Byte)((lim[1] & 0xf8)) != 0xf0) { lim += 2; } return lim;
// return (Byte *)(lim + ((Byte)((lim[1] ^ 0xf0) & 0xf8) == 0 ? 0 : 2));
// return (Byte *)(lim + (((lim[1] ^ ~0xfu) & ~7u) == 0 ? 0 : 2));
// return (Byte *)(lim + 2 - (((((unsigned)lim[1] ^ 8) + 8) >> 7) & 2));
}
Z7_BRANCH_FUNCS_IMP(ARMT)
// #define BR_IA64_NO_INLINE
Z7_BRANCH_FUNC_MAIN(IA64)
{
// Byte *p = data;
const Byte *lim;
size &= ~(SizeT)15;
lim = p + size;
pc -= 1 << 4;
pc >>= 4 - 1;
// pc -= 1 << 1;
for (;;)
{
unsigned m;
for (;;)
{
if Z7_UNLIKELY(p == lim)
return p;
m = (unsigned)((UInt32)0x334b0000 >> (*p & 0x1e));
p += 16;
pc += 1 << 1;
if (m &= 3)
break;
}
{
p += (ptrdiff_t)m * 5 - 20; // negative value is expected here.
do
{
const UInt32 t =
#if defined(MY_CPU_X86_OR_AMD64)
// we use 32-bit load here to reduce code size on x86:
GetUi32(p);
#else
GetUi16(p);
#endif
UInt32 z = GetUi32(p + 1) >> m;
p += 5;
if (((t >> m) & (0x70 << 1)) == 0
&& ((z - (0x5000000 << 1)) & (0xf000000 << 1)) == 0)
{
UInt32 v = (UInt32)((0x8fffff << 1) | 1) & z;
z ^= v;
#ifdef BR_IA64_NO_INLINE
v |= (v & ((UInt32)1 << (23 + 1))) >> 3;
{
UInt32 c = pc;
BR_CONVERT_VAL(v, c)
}
v &= (0x1fffff << 1) | 1;
#else
{
if (encoding)
{
// pc &= ~(0xc00000 << 1); // we just need to clear at least 2 bits
pc &= (0x1fffff << 1) | 1;
v += pc;
}
else
{
// pc |= 0xc00000 << 1; // we need to set at least 2 bits
pc |= ~(UInt32)((0x1fffff << 1) | 1);
v -= pc;
}
}
v &= ~(UInt32)(0x600000 << 1);
#endif
v += (0x700000 << 1);
v &= (0x8fffff << 1) | 1;
z |= v;
z <<= m;
SetUi32(p + 1 - 5, z)
}
m++;
}
while (m &= 3); // while (m < 4);
}
}
}
Z7_BRANCH_FUNCS_IMP(IA64)