Skip to content

Latest commit

 

History

History
485 lines (397 loc) · 14.7 KB

usage.md

File metadata and controls

485 lines (397 loc) · 14.7 KB
Raw

Usage

Inherit Xbyak::CodeGenerator class and make the class method.

#include <xbyak/xbyak.h>

struct Code : Xbyak::CodeGenerator {
    Code(int x)
    {
        mov(eax, x);
        ret();
    }
};

Or you can pass the instance of CodeGenerator without inheriting.

void genCode(Xbyak::CodeGenerator& code, int x) {
    using namespace Xbyak::util;
    code.mov(eax, x);
    code.ret();
}

Make an instance of the class and get the function pointer by calling getCode() and call it.

Code c(5);
int (*f)() = c.getCode<int (*)()>();
printf("ret=%d\n", f()); // ret = 5

Syntax

Similar to MASM/NASM syntax with parentheses.

NASM              Xbyak
mov eax, ebx  --> mov(eax, ebx);
inc ecx           inc(ecx);
ret           --> ret();

Addressing

Use qword, dword, word and byte if it is necessary to specify the size of memory, otherwise use ptr.

(ptr|qword|dword|word|byte) [base + index * (1|2|4|8) + displacement]
                            [rip + 32bit disp] ; x64 only

NASM                   Xbyak
mov eax, [ebx+ecx] --> mov(eax, ptr [ebx+ecx]);
mov al, [ebx+ecx]  --> mov(al, ptr [ebx + ecx]);
test byte [esp], 4 --> test(byte [esp], 4);
inc qword [rax]    --> inc(qword [rax]);

Note: qword, ... are member variables, then don't use dword as unsigned int type.

How to use Selector (Segment Register)

mov eax, [fs:eax] --> putSeg(fs);
                      mov(eax, ptr [eax]);
mov ax, cs        --> mov(ax, cs);

Note: Segment class is not derived from Operand.

AVX

vaddps(xmm1, xmm2, xmm3); // xmm1 <- xmm2 + xmm3
vaddps(xmm2, xmm3, ptr [rax]); // use ptr to access memory
vgatherdpd(xmm1, ptr [ebp + 256 + xmm2*4], xmm3);

Note: If XBYAK_ENABLE_OMITTED_OPERAND is defined, then you can use two operand version for backward compatibility. But the newer version will not support it.

vaddps(xmm2, xmm3); // xmm2 <- xmm2 + xmm3

AVX-512

vaddpd zmm2, zmm5, zmm30                --> vaddpd(zmm2, zmm5, zmm30);
vaddpd xmm30, xmm20, [rax]              --> vaddpd(xmm30, xmm20, ptr [rax]);
vaddps xmm30, xmm20, [rax]              --> vaddps(xmm30, xmm20, ptr [rax]);
vaddpd zmm2{k5}, zmm4, zmm2             --> vaddpd(zmm2 | k5, zmm4, zmm2);
vaddpd zmm2{k5}{z}, zmm4, zmm2          --> vaddpd(zmm2 | k5 | T_z, zmm4, zmm2);
vaddpd zmm2{k5}{z}, zmm4, zmm2,{rd-sae} --> vaddpd(zmm2 | k5 | T_z, zmm4, zmm2 | T_rd_sae);
                                            vaddpd(zmm2 | k5 | T_z | T_rd_sae, zmm4, zmm2); // the position of `|` is arbitrary.
vcmppd k4{k3}, zmm1, zmm2, {sae}, 5     --> vcmppd(k4 | k3, zmm1, zmm2 | T_sae, 5);

vaddpd xmm1, xmm2, [rax+256]            --> vaddpd(xmm1, xmm2, ptr [rax+256]);
vaddpd xmm1, xmm2, [rax+256]{1to2}      --> vaddpd(xmm1, xmm2, ptr_b [rax+256]);
vaddpd ymm1, ymm2, [rax+256]{1to4}      --> vaddpd(ymm1, ymm2, ptr_b [rax+256]);
vaddpd zmm1, zmm2, [rax+256]{1to8}      --> vaddpd(zmm1, zmm2, ptr_b [rax+256]);
vaddps zmm1, zmm2, [rax+rcx*8+8]{1to16} --> vaddps(zmm1, zmm2, ptr_b [rax+rcx*8+8]);
vmovsd [rax]{k1}, xmm4                  --> vmovsd(ptr [rax] | k1, xmm4);

vcvtpd2dq xmm16, oword [eax+33]         --> vcvtpd2dq(xmm16, xword [eax+33]); // use xword for m128 instead of oword
                                            vcvtpd2dq(xmm16, ptr [eax+33]); // default xword
vcvtpd2dq xmm21, [eax+32]{1to2}         --> vcvtpd2dq(xmm21, ptr_b [eax+32]);
vcvtpd2dq xmm0, yword [eax+33]          --> vcvtpd2dq(xmm0, yword [eax+33]); // use yword for m256
vcvtpd2dq xmm19, [eax+32]{1to4}         --> vcvtpd2dq(xmm19, yword_b [eax+32]); // use yword_b to broadcast

vfpclassps k5{k3}, zword [rax+64], 5    --> vfpclassps(k5|k3, zword [rax+64], 5); // specify m512
vfpclasspd k5{k3}, [rax+64]{1to2}, 5    --> vfpclasspd(k5|k3, xword_b [rax+64], 5); // broadcast 64-bit to 128-bit
vfpclassps k5{k3}, [rax+64]{1to4}, 5    --> vfpclassps(k5|k3, yword_b [rax+64], 5); // broadcast 64-bit to 256-bit

Remark

  • k1, ..., k7 are opmask registers.
    • k0 is dealt as no mask.
    • e.g. vmovaps(zmm0|k0, ptr[rax]); and vmovaps(zmm0|T_z, ptr[rax]); are same to vmovaps(zmm0, ptr[rax]);.
  • use | T_z, | T_sae, | T_rn_sae, | T_rd_sae, | T_ru_sae, | T_rz_sae instead of ,{z}, ,{sae}, ,{rn-sae}, ,{rd-sae}, ,{ru-sae}, ,{rz-sae} respectively.
  • k4 | k3 is different from k3 | k4.
  • use ptr_b for broadcast {1toX}. X is automatically determined.
  • specify xword/yword/zword(_b) for m128/m256/m512 if necessary.

Selecting AVX512-VNNI, AVX-VNNI, AVX-VNNI-INT8, AVX10.2.

Some mnemonics have some types of encodings: VEX, EVEX, AVX10.2. The functions for these mnemonics include an optional parameter as the last argument to specify the encoding. The default behavior depends on the order in which the instruction was introduced (whether VEX, EVEX or AVX10.2 came first), and can be specified using setDefaultEncoding.

vpdpbusd(xm0, xm1, xm2); // default encoding: EVEX (AVX512-VNNI)
vpdpbusd(xm0, xm1, xm2, EvexEncoding); // same as the above
vpdpbusd(xm0, xm1, xm2, VexEncoding); // VEX (AVX-VNNI)
setDefaultEncoding(VexEncoding); // change default encoding
vpdpbusd(xm0, xm1, xm2); // VEX

vmpsadbw(xm1, xm3, xm15, 3); // default encoding: AVX
vmpsadbw(xm1, xm3, xm15, 3, PreAVX10v2Encoding); // same as the above
vmpsadbw(xm1, xm3, xm15, 3, AVX10v2Encoding); // AVX10.2
setDefaultEncodingAVX10(AVX10v2Encoding); // change default encoding
vmpsadbw(xm1, xm3, xm15, 3); // AVX10.2
  • setDefaultEncoding(PreferredEncoding enc = EvexEncoding)
    • Configure encoding for AVX512-VNNI or AVX-VNNI instructions.
  • setDefaultEncodingAVX10(PreferredEncoding enc = PreAVXv2Encoding)
    • Configure encoding for pre-AVX10.2 and AVX10.2 instructions.
setDefaultEncoding EvexEncoding (default) VexEncoding
feature AVX512-VNNI AVX-VNNI
  • Target functions: vpdpbusd, vpdpbusds, vpdpwssd, vpdpwssds
setDefaultEncodingAVX10 PreAVX10v2Encoding (default) AVX10v2Encoding
feature AVX-VNNI-INT8, AVX512-FP16 AVX10.2
  • Target functions: vmpsadbw, vpdpbssd, vpdpbssds, vpdpbsud, vpdpbsuds, vpdpbuud, vpdpbuuds, vpdpwsud vpdpwsuds vpdpwusd vpdpwusds vpdpwuud, vpdpwuuds and vmovd, vmovw with MEM-to-MEM.

Remark

  1. vmovd and vmovw instructions with REG-to-XMM or XMM-to-REG operands are always encoded using AVX10.1. When used with XMM-to-XMM operands, these instructions are always encoded using AVX10.2.

  2. vmovd and vmovw instructions with XMM-to-MEM or MEM-to-XMM operands support multiple encoding formats, including AVX, AVX512F, AVX512-FP16, and AVX10.2.

Initially, I tried implementing setDefaultEncodingAVX10 using EvexEncoding (resp. VexEncoding) instead of AVX10v2Encoding (resp. EvexEncoding). However, I abandoned this approach after discovering the complexity of the encoding requirements of vmovd and vmovw.

APX

Advanced Performance Extensions (APX) Architecture Specification

  • Support 64-bit 16 additional GPRs (general-purpose registers) r16, ..., r31
    • 32-bit regs are r16d, ..., r31d
    • 16-bit regs are r16w, ..., r31w
    • 8-bit regs are r16b, ..., r31b
    • add(r20, r21);
    • lea(r30, ptr[r29+r31]);
  • Support three-operand instruction
    • add(r20, r21, r23);
    • add(r20, ptr[rax + rcx * 8 + 0x1234], r23);
  • Support T_nf for NF=1 (status flags update suppression)
    • add(r20|T_nf, r21, r23); // Set EVEX.NF=1
  • Support T_zu for NF=ZU (zero upper) for imul and setcc
    • imul(ax|T_zu, cx, 0x1234); // Set ND=ZU
    • imul(ax|T_zu|T_nf, cx, 0x1234); // Set ND=ZU and EVEX.NF=1
    • setb(r31b|T_zu); // same as set(r31b); movzx(r31, r31b);
    • See sample/zero_upper.cpp

ccmpSCC and ctestSCC

  • ccmpSCC(op1, op2, dfv = 0); // eflags = eflags == SCC ? cmp(op1, op2) : dfv
  • ctestSCC(op1, op2, dfv = 0); // eflags = eflags == SCC ? test(op1, op2) : dfv
  • SCC means source condition code such as z, a, gt.
  • See sample/ccmp.cpp
  • Specify the union of T_of(=8), T_sf(=4), T_zf(=2), or T_cf(=1) for dfv.

Label

Two kinds of Label are supported. (String literal and Label class).

String literal

L("L1");
  jmp("L1");

  jmp("L2");
  ...
  a few mnemonics (8-bit displacement jmp)
  ...
L("L2");

  jmp("L3", T_NEAR);
  ...
  a lot of mnemonics (32-bit displacement jmp)
  ...
L("L3");
  • Call hasUndefinedLabel() to verify your code has no undefined label.
  • you can use a label for immediate value of mov like as mov(eax, "L2").

Support @@, @f, @b like MASM

L("@@"); // <A>
  jmp("@b"); // jmp to <A>
  jmp("@f"); // jmp to <B>
L("@@"); // <B>
  jmp("@b"); // jmp to <B>
  mov(eax, "@b");
  jmp(eax); // jmp to <B>

Local label

Label symbols beginning with a period between inLocalLabel() and outLocalLabel() are treated as a local label. inLocalLabel() and outLocalLabel() can be nested.

void func1()
{
    inLocalLabel();
  L(".lp"); // <A> ; local label
    ...
    jmp(".lp"); // jmp to <A>
  L("aaa"); // global label <C>
    outLocalLabel();

    inLocalLabel();
  L(".lp"); // <B> ; local label
    func1();
    jmp(".lp"); // jmp to <B>
    inLocalLabel();
    jmp("aaa"); // jmp to <C>
}

short and long jump

Xbyak deals with jump mnemonics of an undefined label as short jump if no type is specified. So if the size between jmp and label is larger than 127 byte, then xbyak will cause an error.

jmp("short-jmp"); // short jmp
// small code
L("short-jmp");

jmp("long-jmp");
// long code
L("long-jmp"); // throw exception

Then specify T_NEAR for jmp.

jmp("long-jmp", T_NEAR); // long jmp
// long code
L("long-jmp");

Or call setDefaultJmpNEAR(true); once, then the default type is set to T_NEAR.

jmp("long-jmp"); // long jmp
// long code
L("long-jmp");

Label class

L() and jxx() support Label class.

  Xbyak::Label label1, label2;
L(label1);
  ...
  jmp(label1);
  ...
  jmp(label2);
  ...
L(label2);

Use putL for jmp table

    Label labelTbl, L0, L1, L2;
    mov(rax, labelTbl);
    // rdx is an index of jump table
    jmp(ptr [rax + rdx * sizeof(void*)]);
L(labelTbl);
    putL(L0);
    putL(L1);
    putL(L2);
L(L0);
    ....
L(L1);
    ....

assignL(dstLabel, srcLabel) binds dstLabel with srcLabel.

  Label label2;
  Label label1 = L(); // make label1 ; same to Label label1; L(label1);
  ...
  jmp(label2); // label2 is not determined here
  ...
  assignL(label2, label1); // label2 <- label1

The jmp in the above code jumps to label1 assigned by assignL.

Note:

  • srcLabel must be used in L().
  • dstLabel must not be used in L().

Label::getAddress() returns the address specified by the label instance and 0 if not specified.

// not AutoGrow mode
Label  label;
assert(label.getAddress() == 0);
L(label);
assert(label.getAddress() == getCurr());

Rip ; relative addressing

Label label;
mov(eax, ptr [rip + label]); // eax = 4
...

L(label);
dd(4);
int x;
...
  mov(eax, ptr[rip + &x]); // throw exception if the difference between &x and current position is larger than 2GiB

Far jump

Use word|dword|qword instead of ptr to specify the address size.

32 bit mode

jmp(word[eax], T_FAR);  // jmp m16:16(FF /5)
jmp(dword[eax], T_FAR); // jmp m16:32(FF /5)

64 bit mode

jmp(word[rax], T_FAR);  // jmp m16:16(FF /5)
jmp(dword[rax], T_FAR); // jmp m16:32(FF /5)
jmp(qword[rax], T_FAR); // jmp m16:64(REX.W FF /5)

The same applies to call.

Code size

The default max code size is 4096 bytes. Specify the size in constructor of CodeGenerator() if necessary.

class Quantize : public Xbyak::CodeGenerator {
public:
  Quantize()
    : CodeGenerator(8192)
  {
  }
  ...
};

User allocated memory

You can make jit code on prepared memory.

Call setProtectModeRE yourself to change memory mode if using the prepared memory.

uint8_t alignas(4096) buf[8192]; // C++11 or later

struct Code : Xbyak::CodeGenerator {
    Code() : Xbyak::CodeGenerator(sizeof(buf), buf)
    {
        mov(rax, 123);
        ret();
    }
};

int main()
{
    Code c;
    c.setProtectModeRE(); // set memory to Read/Exec
    printf("%d\n", c.getCode<int(*)()>()());
}

Note: See ../sample/test0.cpp.

AutoGrow

The memory region for jit is automatically extended if necessary when AutoGrow is specified in a constructor of CodeGenerator.

Call ready() or readyRE() before calling getCode() to fix jump address.

struct Code : Xbyak::CodeGenerator {
  Code()
    : Xbyak::CodeGenerator(<default memory size>, Xbyak::AutoGrow)
  {
     ...
  }
};
Code c;
// generate code for jit
c.ready(); // mode = Read/Write/Exec

Note:

  • Don't use the address returned by getCurr() before calling ready() because it may be invalid address.

Read/Exec mode

Xbyak set Read/Write/Exec mode to memory to run jit code. If you want to use Read/Exec mode for security, then specify DontSetProtectRWE for CodeGenerator and call setProtectModeRE() after generating jit code.

struct Code : Xbyak::CodeGenerator {
    Code()
        : Xbyak::CodeGenerator(4096, Xbyak::DontSetProtectRWE)
    {
        mov(eax, 123);
        ret();
    }
};

Code c;
c.setProtectModeRE();
...

Call readyRE() instead of ready() when using AutoGrow mode. See protect-re.cpp.

Exception-less mode

If XBYAK_NO_EXCEPTION is defined, then gcc/clang can compile xbyak with -fno-exceptions. In stead of throwing an exception, Xbyak::GetError() returns non-zero value (e.g. ERR_BAD_ADDRESSING) if there is something wrong. The status will not be changed automatically, then you should reset it by Xbyak::ClearError(). CodeGenerator::reset() calls ClearError().

Macro

  • XBYAK32 is defined on 32bit.
  • XBYAK64 is defined on 64bit.
  • XBYAK64_WIN is defined on 64bit Windows(VC).
  • XBYAK64_GCC is defined on 64bit gcc, cygwin.
  • define XBYAK_USE_OP_NAMES on gcc with -fno-operator-names if you want to use and(), ....
  • define XBYAK_ENABLE_OMITTED_OPERAND if you use omitted destination such as vaddps(xmm2, xmm3);(deprecated in the future).
  • define XBYAK_UNDEF_JNL if Bessel function jnl is defined as macro.
  • define XBYAK_NO_EXCEPTION for a compiler option -fno-exceptions.
  • define XBYAK_USE_MEMFD on Linux then /proc/self/maps shows the area used by xbyak.
  • define XBYAK_OLD_DISP_CHECK if the old disp check is necessary (deprecated in the future).

Sample

  • test0.cpp ; tiny sample (x86, x64)
  • quantize.cpp ; JIT optimized quantization by fast division (x86 only)
  • calc.cpp ; assemble and estimate a given polynomial (x86, x64)
  • bf.cpp ; JIT brainfuck (x86, x64)