punpcklbw（在MMX/SSE/AVX中交错）有哪些用例？
-
可以使用哪些类别的算法
punpcklbw？ -
特别是，在
punpcklbw xmm0, xmm0做什么？ -
然而，什么是
maskedPow2_Value有用的？maskedValue = 0x101010101010101i64 * *(_QWORD *)&Val; // Val 是 int maskedPow2_Value = 0x101010101010101i64 * maskedValue;
（或mov r9, 101010101010101h; imul rdx, r9;两次）
一个完整的例子（该函数名为 CompressPacket 但它可能会产生误导），作为 IDA 反编译的结果：
void *__cdecl CompressPacket(void *Dst, int Val, size_t Size)
{
__int64 maskedPow2_Value; // rdx
unsigned int v5; // ecx
__int64 *bufferOut; // rcx
size_t size_; // r9
size_t i; // r9
size_t size__; // r9
size_t counter; // r8
size_t j; // r9
void *result; // rax
__m128i v13; // xmm0
__int64 lsb4; // rax
size_t counter1; // r9
size_t k; // r9
size_t lsb4_; // r8
__int64 maskedValue; // rdx
*(_QWORD *)&Val = (unsigned __int8)Val;
maskedValue = 0x101010101010101i64 * *(_QWORD *)&Val;
bufferOut = (__int64 *)((char *)Dst + Size);
result = Dst;
switch ( Size )
{
case 0ui64:
return result;
case 1ui64:
goto LBL_1_F;
case 2ui64:
goto LBL_2_E;
case 3ui64:
goto LBL_3_F;
case 4ui64:
goto LBL_4_C;
case 5ui64:
goto LBL_5_D;
case 6ui64:
goto LBL_6_E;
case 7ui64:
goto LBL_7_F;
case 8ui64:
*(bufferOut - 1) = maskedValue;
return result;
case 9ui64:
*(__int64 *)((char *)bufferOut - 9) = maskedValue;
*((_BYTE *)bufferOut - 1) = maskedValue;
return result;
case 0xAui64:
*(__int64 *)((char *)bufferOut - 10) = maskedValue;
*((_WORD *)bufferOut - 1) = maskedValue;
return result;
case 0xBui64:
*(__int64 *)((char *)bufferOut - 11) = maskedValue;
goto LBL_3_F;
case 0xCui64:
*(__int64 *)((char *)bufferOut - 12) = maskedValue;
LBL_4_C:
*((_DWORD *)bufferOut - 1) = maskedValue;
return result;
case 0xDui64:
*(__int64 *)((char *)bufferOut - 13) = maskedValue;
LBL_5_D:
*(_DWORD *)((char *)bufferOut - 5) = maskedValue;
*((_BYTE *)bufferOut - 1) = maskedValue;
return result;
case 0xEui64:
*(__int64 *)((char *)bufferOut - 14) = maskedValue;
LBL_6_E:
*(_DWORD *)((char *)bufferOut - 6) = maskedValue;
LBL_2_E:
*((_WORD *)bufferOut - 1) = maskedValue;
return result;
case 0xFui64:
*(__int64 *)((char *)bufferOut - 15) = maskedValue;
LBL_7_F:
*(_DWORD *)((char *)bufferOut - 7) = maskedValue;
LBL_3_F:
*(_WORD *)((char *)bufferOut - 3) = maskedValue;
LBL_1_F:
*((_BYTE *)bufferOut - 1) = maskedValue;
return result;
default:
if ( _bittest(dword_7FFFF4B237D8, 1u) )
{
memset(bufferOut, maskedValue, Size);
return Dst;
}
maskedPow2_Value = 0x101010101010101i64 * maskedValue;
if ( !_bittest(dword_7FFFF4B237D8, 2u) )
{
if ( Size >= 0x40 )
{
v5 = -(int)bufferOut & 7;
if ( v5 )
{
Size -= v5;
*(_QWORD *)Dst = maskedPow2_Value;
}
bufferOut = (__int64 *)((char *)Dst + v5);
size_ = Size;
Size &= 0x3Fu;
for ( i = size_ >> 6; i; *(bufferOut - 1) = maskedPow2_Value )
{
*bufferOut = maskedPow2_Value;
bufferOut[1] = maskedPow2_Value;
bufferOut[2] = maskedPow2_Value;
bufferOut += 8;
*(bufferOut - 5) = maskedPow2_Value;
*(bufferOut - 4) = maskedPow2_Value;
--i;
*(bufferOut - 3) = maskedPow2_Value;
*(bufferOut - 2) = maskedPow2_Value;
}
}
size__ = Size;
counter = Size & 7;
for ( j = size__ >> 3; j; --j )
*bufferOut++ = maskedPow2_Value;
for ( ; counter; --counter )
{
*(_BYTE *)bufferOut = maskedPow2_Value;
bufferOut = (__int64 *)((char *)bufferOut + 1);
}
return Dst;
}
v13 = _mm_unpacklo_epi8((__m128i)(unsigned __int64)maskedPow2_Value, (__m128i)(unsigned __int64)maskedPow2_Value);
if ( ((unsigned __int8)bufferOut & 0xF) != 0 )
{
*(__m128i *)bufferOut = v13;
lsb4 = (unsigned __int8)bufferOut & 0xF;
bufferOut = (__int64 *)((char *)bufferOut - lsb4 + 16);
Size = lsb4 + Size - 16;
}
counter1 = Size >> 7;
if ( Size >> 7 )
{
do
{
*(__m128i *)bufferOut = v13;
*((__m128i *)bufferOut + 1) = v13;
bufferOut += 16;
*((__m128i *)bufferOut - 6) = v13;
*((__m128i *)bufferOut - 5) = v13;
--counter1;
*((__m128i *)bufferOut - 4) = v13;
*((__m128i *)bufferOut - 3) = v13;
*((__m128i *)bufferOut - 2) = v13;
*((__m128i *)bufferOut - 1) = v13;
}
while ( counter1 );
Size &= 0x7Fu;
}
for ( k = Size >> 4; k; --k )
{
*(__m128i *)bufferOut = v13;
bufferOut += 2;
}
lsb4_ = Size & 0xF;
if ( lsb4_ )
*(__m128i *)((char *)bufferOut + lsb4_ - 16) = v13;
return Dst;
}
}
以及 IDA 的反汇编：
.text:00007FFFF4AF6440 ; void *__cdecl CompressPacket(void *Dst, int Val, size_t Size)
.text:00007FFFF4AF6440 CompressPacket proc near ; CODE XREF: j_memset?j
.text:00007FFFF4AF6440 ; Concurrency::details::ResourceManager::CreateAllocatedNodeData(void)+49?p ...
.text:00007FFFF4AF6440 mov r11, rcx
.text:00007FFFF4AF6443 movzx edx, dl ; Move with Zero-Extend
.text:00007FFFF4AF6446 cmp r8, 10h ; switch 16 cases
.text:00007FFFF4AF644A jb SetBytes15 ; Jump if Below (CF=1)
.text:00007FFFF4AF6450
.text:00007FFFF4AF6450 def_7FFFF4AF65D2: ; jumptable 00007FFFF4AF65D2 default case
.text:00007FFFF4AF6450 bt cs:dword_7FFFF4B237D8, 1
.text:00007FFFF4AF6458 jnb short mset05 ; Jump if Not Below (CF=0)
.text:00007FFFF4AF645A push rdi
.text:00007FFFF4AF645B mov rdi, rcx
.text:00007FFFF4AF645E mov eax, edx
.text:00007FFFF4AF6460 mov rcx, r8
.text:00007FFFF4AF6463 rep stosb ; Store String
.text:00007FFFF4AF6465 pop rdi
.text:00007FFFF4AF6466 jmp short mset60 ; Jump
.text:00007FFFF4AF6468 ; ---------------------------------------------------------------------------
.text:00007FFFF4AF6468
.text:00007FFFF4AF6468 mset05: ; CODE XREF: CompressPacket+18?j
.text:00007FFFF4AF6468 mov r9, 101010101010101h
.text:00007FFFF4AF6472 imul rdx, r9 ; Signed Multiply
.text:00007FFFF4AF6476 bt cs:dword_7FFFF4B237D8, 2 ; Bit Test
.text:00007FFFF4AF647E jb msetxmm10 ; Jump if Below (CF=1)
.text:00007FFFF4AF6484 cmp r8, 40h ; '@' ; Compare Two Operands
.text:00007FFFF4AF6488 jb short mset20 ; Jump if Below (CF=1)
.text:00007FFFF4AF648A neg rcx ; Two's Complement Negation
.text:00007FFFF4AF648D and ecx, 7 ; Logical AND
.text:00007FFFF4AF6490 jz short mset10 ; Jump if Zero (ZF=1)
.text:00007FFFF4AF6492 sub r8, rcx ; Integer Subtraction
.text:00007FFFF4AF6495 mov [r11], rdx
.text:00007FFFF4AF6498
.text:00007FFFF4AF6498 mset10: ; CODE XREF: CompressPacket+50?j
.text:00007FFFF4AF6498 add rcx, r11 ; Add
.text:00007FFFF4AF649B mov r9, r8
.text:00007FFFF4AF649E and r8, 3Fh ; Logical AND
.text:00007FFFF4AF64A2 shr r9, 6 ; Shift Logical Right
.text:00007FFFF4AF64A6 jnz short mset80 ; Jump if Not Zero (ZF=0)
.text:00007FFFF4AF64A8
.text:00007FFFF4AF64A8 mset20: ; CODE XREF: CompressPacket+48?j
.text:00007FFFF4AF64A8 ; CompressPacket+CF?j
.text:00007FFFF4AF64A8 mov r9, r8
.text:00007FFFF4AF64AB and r8, 7 ; Logical AND
.text:00007FFFF4AF64AF shr r9, 3 ; Shift Logical Right
.text:00007FFFF4AF64B3 jz short mset40 ; Jump if Zero (ZF=1)
.text:00007FFFF4AF64B5 db 66h, 66h
.text:00007FFFF4AF64B5 xchg ax, ax ; Exchange Register/Memory with Register
.text:00007FFFF4AF64B9 nop ; No Operation
.text:00007FFFF4AF64BA
.text:00007FFFF4AF64BA mset30: ; CODE XREF: CompressPacket+84?j
.text:00007FFFF4AF64BA mov [rcx], rdx
.text:00007FFFF4AF64BD add rcx, 8 ; Add
.text:00007FFFF4AF64C1 dec r9 ; Decrement by 1
.text:00007FFFF4AF64C4 jnz short mset30 ; Jump if Not Zero (ZF=0)
.text:00007FFFF4AF64C6
.text:00007FFFF4AF64C6 mset40: ; CODE XREF: CompressPacket+73?j
.text:00007FFFF4AF64C6 test r8, r8 ; Logical Compare
.text:00007FFFF4AF64C9 jz short mset60 ; Jump if Zero (ZF=1)
.text:00007FFFF4AF64CB
.text:00007FFFF4AF64CB mset50: ; CODE XREF: CompressPacket+93?j
.text:00007FFFF4AF64CB mov [rcx], dl
.text:00007FFFF4AF64CD inc rcx ; Increment by 1
.text:00007FFFF4AF64D0 dec r8 ; Decrement by 1
.text:00007FFFF4AF64D3 jnz short mset50 ; Jump if Not Zero (ZF=0)
.text:00007FFFF4AF64D5
.text:00007FFFF4AF64D5 mset60: ; CODE XREF: CompressPacket+26?j
.text:00007FFFF4AF64D5 ; CompressPacket+89?j
.text:00007FFFF4AF64D5 mov rax, r11
.text:00007FFFF4AF64D8 retn ; Return Near from Procedure
.text:00007FFFF4AF64D8 ; ---------------------------------------------------------------------------
.text:00007FFFF4AF64D9 db 0Fh, 1Fh, 80h, 4 dup(0)
.text:00007FFFF4AF64E0 db 3 dup(66h), 90h
.text:00007FFFF4AF64E4 db 2 dup(66h), 90h
.text:00007FFFF4AF64E7 ; ---------------------------------------------------------------------------
.text:00007FFFF4AF64E7
.text:00007FFFF4AF64E7 mset80: ; CODE XREF: CompressPacket+66?j
.text:00007FFFF4AF64E7 ; CompressPacket+CD?j
.text:00007FFFF4AF64E7 mov [rcx], rdx
.text:00007FFFF4AF64EA mov [rcx+8], rdx
.text:00007FFFF4AF64EE mov [rcx+10h], rdx
.text:00007FFFF4AF64F2 add rcx, 40h ; '@' ; Add
.text:00007FFFF4AF64F6 mov [rcx-28h], rdx
.text:00007FFFF4AF64FA mov [rcx-20h], rdx
.text:00007FFFF4AF64FE dec r9 ; Decrement by 1
.text:00007FFFF4AF6501 mov [rcx-18h], rdx
.text:00007FFFF4AF6505 mov [rcx-10h], rdx
.text:00007FFFF4AF6509 mov [rcx-8], rdx
.text:00007FFFF4AF650D jnz short mset80 ; Jump if Not Zero (ZF=0)
.text:00007FFFF4AF650F jmp short mset20 ; Jump
.text:00007FFFF4AF650F ; ---------------------------------------------------------------------------
.text:00007FFFF4AF6511 align 20h
.text:00007FFFF4AF6520
.text:00007FFFF4AF6520 msetxmm10: ; CODE XREF: CompressPacket+3E?j
.text:00007FFFF4AF6520 movq xmm0, rdx ; Move 64 bits
.text:00007FFFF4AF6525 punpcklbw xmm0, xmm0 ; Unpack Low Packed Data (Byte->Word)
.text:00007FFFF4AF6529 test cl, 0Fh ; Logical Compare
.text:00007FFFF4AF652C jz short msetxmm20 ; Jump if Zero (ZF=1)
.text:00007FFFF4AF652E movups xmmword ptr [rcx], xmm0 ; Move Unaligned Four Packed Single-FP
.text:00007FFFF4AF6531 mov rax, rcx
.text:00007FFFF4AF6534 and rax, 0Fh ; Logical AND
.text:00007FFFF4AF6538 add rcx, 10h ; Add
.text:00007FFFF4AF653C sub rcx, rax ; Integer Subtraction
.text:00007FFFF4AF653F lea r8, [rax+r8-10h] ; Load Effective Address
.text:00007FFFF4AF6544
.text:00007FFFF4AF6544 msetxmm20: ; CODE XREF: CompressPacket+EC?j
.text:00007FFFF4AF6544 mov r9, r8
.text:00007FFFF4AF6547 shr r9, 7 ; Shift Logical Right
.text:00007FFFF4AF654B jz short msetxmm40 ; Jump if Zero (ZF=1)
.text:00007FFFF4AF654D jmp short msetxmm30 ; Jump
.text:00007FFFF4AF654D ; ---------------------------------------------------------------------------
.text:00007FFFF4AF654F align 10h
.text:00007FFFF4AF6550
.text:00007FFFF4AF6550 msetxmm30: ; CODE XREF: CompressPacket+10D?j
.text:00007FFFF4AF6550 ; CompressPacket+139?j
.text:00007FFFF4AF6550 movaps xmmword ptr [rcx], xmm0 ; Move Aligned Four Packed Single-FP
.text:00007FFFF4AF6553 movaps xmmword ptr [rcx+10h], xmm0 ; Move Aligned Four Packed Single-FP
.text:00007FFFF4AF6557 add rcx, 80h ; '€' ; Add
.text:00007FFFF4AF655E movaps xmmword ptr [rcx-60h], xmm0 ; Move Aligned Four Packed Single-FP
.text:00007FFFF4AF6562 movaps xmmword ptr [rcx-50h], xmm0 ; Move Aligned Four Packed Single-FP
.text:00007FFFF4AF6566 dec r9 ; Decrement by 1
.text:00007FFFF4AF6569 movaps xmmword ptr [rcx-40h], xmm0 ; Move Aligned Four Packed Single-FP
.text:00007FFFF4AF656D movaps xmmword ptr [rcx-30h], xmm0 ; Move Aligned Four Packed Single-FP
.text:00007FFFF4AF6571 movaps xmmword ptr [rcx-20h], xmm0 ; Move Aligned Four Packed Single-FP
.text:00007FFFF4AF6575 movaps xmmword ptr [rcx-10h], xmm0 ; Move Aligned Four Packed Single-FP
.text:00007FFFF4AF6579 jnz short msetxmm30 ; Jump if Not Zero (ZF=0)
.text:00007FFFF4AF657B and r8, 7Fh ; Logical AND
.text:00007FFFF4AF657F
.text:00007FFFF4AF657F msetxmm40: ; CODE XREF: CompressPacket+10B?j
.text:00007FFFF4AF657F mov r9, r8
.text:00007FFFF4AF6582 shr r9, 4 ; Shift Logical Right
.text:00007FFFF4AF6586 jz short msetxmm60 ; Jump if Zero (ZF=1)
.text:00007FFFF4AF6588 nop dword ptr [rax+rax+00000000h] ; No Operation
.text:00007FFFF4AF6590
.text:00007FFFF4AF6590 msetxmm50: ; CODE XREF: CompressPacket+15A?j
.text:00007FFFF4AF6590 movaps xmmword ptr [rcx], xmm0 ; Move Aligned Four Packed Single-FP
.text:00007FFFF4AF6593 add rcx, 10h ; Add
.text:00007FFFF4AF6597 dec r9 ; Decrement by 1
.text:00007FFFF4AF659A jnz short msetxmm50 ; Jump if Not Zero (ZF=0)
.text:00007FFFF4AF659C
.text:00007FFFF4AF659C msetxmm60: ; CODE XREF: CompressPacket+146?j
.text:00007FFFF4AF659C and r8, 0Fh ; Logical AND
.text:00007FFFF4AF65A0 jz short msetxmm70 ; Jump if Zero (ZF=1)
.text:00007FFFF4AF65A2 movups xmmword ptr [r8+rcx-10h], xmm0 ; Move Unaligned Four Packed Single-FP
.text:00007FFFF4AF65A8
.text:00007FFFF4AF65A8 msetxmm70: ; CODE XREF: CompressPacket+160?j
.text:00007FFFF4AF65A8 mov rax, r11
.text:00007FFFF4AF65AB retn ; Return Near from Procedure
.text:00007FFFF4AF65AC ; ---------------------------------------------------------------------------
.text:00007FFFF4AF65AC
.text:00007FFFF4AF65AC SetBytes15: ; CODE XREF: CompressPacket+A?j
.text:00007FFFF4AF65AC mov r9, 101010101010101h
.text:00007FFFF4AF65B6 imul rdx, r9 ; Signed Multiply
.text:00007FFFF4AF65BA lea r9, cs:7FFFF4AB0000h ; Load Effective Address
.text:00007FFFF4AF65C1 mov eax, ds:(jpt_7FFFF4AF65D2 - 7FFFF4AB0000h)[r9+r8*4]
.text:00007FFFF4AF65C9 add r9, rax ; Add
.text:00007FFFF4AF65CC add rcx, r8 ; Add
.text:00007FFFF4AF65CF mov rax, r11
.text:00007FFFF4AF65D2 jmp r9 ; switch jump
.text:00007FFFF4AF65D2 ; ---------------------------------------------------------------------------
.text:00007FFFF4AF65D5 jpt_7FFFF4AF65D2 dd offset msetTab00 - 7FFFF4AB0000h
.text:00007FFFF4AF65D5 ; DATA XREF: CompressPacket+181?r
.text:00007FFFF4AF65D5 dd offset msetTab01 - 7FFFF4AB0000h ; jump table for switch statement
.text:00007FFFF4AF65D5 dd offset msetTab02 - 7FFFF4AB0000h
.text:00007FFFF4AF65D5 dd offset msetTab03 - 7FFFF4AB0000h
.text:00007FFFF4AF65D5 dd offset msetTab04 - 7FFFF4AB0000h
.text:00007FFFF4AF65D5 dd offset msetTab05 - 7FFFF4AB0000h
.text:00007FFFF4AF65D5 dd offset msetTab06 - 7FFFF4AB0000h
.text:00007FFFF4AF65D5 dd offset msetTab07 - 7FFFF4AB0000h
.text:00007FFFF4AF65D5 dd offset msetTab08 - 7FFFF4AB0000h
.text:00007FFFF4AF65D5 dd offset msetTab09 - 7FFFF4AB0000h
.text:00007FFFF4AF65D5 dd offset msetTab10 - 7FFFF4AB0000h
.text:00007FFFF4AF65D5 dd offset msetTab11 - 7FFFF4AB0000h
.text:00007FFFF4AF65D5 dd offset msetTab12 - 7FFFF4AB0000h
.text:00007FFFF4AF65D5 dd offset msetTab13 - 7FFFF4AB0000h
.text:00007FFFF4AF65D5 dd offset msetTab14 - 7FFFF4AB0000h
.text:00007FFFF4AF65D5 dd offset msetTab15 - 7FFFF4AB0000h
.text:00007FFFF4AF6615 align 20h
.text:00007FFFF4AF6620
.text:00007FFFF4AF6620 msetTab15: ; CODE XREF: CompressPacket+192?j
.text:00007FFFF4AF6620 ; DATA XREF: CompressPacket:jpt_7FFFF4AF65D2?o
.text:00007FFFF4AF6620 mov [rcx-0Fh], rdx ; jumptable 00007FFFF4AF65D2 case 15
.text:00007FFFF4AF6624
.text:00007FFFF4AF6624 msetTab07: ; CODE XREF: CompressPacket+192?j
.text:00007FFFF4AF6624 ; DATA XREF: CompressPacket:jpt_7FFFF4AF65D2?o
.text:00007FFFF4AF6624 mov [rcx-7], edx ; jumptable 00007FFFF4AF65D2 case 7
.text:00007FFFF4AF6627
.text:00007FFFF4AF6627 msetTab03: ; CODE XREF: CompressPacket+192?j
.text:00007FFFF4AF6627 ; CompressPacket+1F3?j
.text:00007FFFF4AF6627 ; DATA XREF: ...
.text:00007FFFF4AF6627 mov [rcx-3], dx ; jumptable 00007FFFF4AF65D2 case 3
.text:00007FFFF4AF662B
.text:00007FFFF4AF662B
回答
一个常见的用例是用零解包以将 8 位数字扩展到 16 位（带有零扩展），如 SSE4.1 pmovzxbw。或者特别是解压缩 16 字节寄存器的低半部分和高半部分，以获得两个 8x 16 位元素的向量。 这是“解包”名称有意义的唯一用例，packuswb它的反义词是将 2 个寄存器组合为 1。（或packsswb用于符号饱和。）
“解包”这个名字很奇怪；它只是一个从两个寄存器中交错元素的洗牌。ARM NEON 有一个类似的 shuffle，其助记符是 "zip"。
根据你的情况，这是部分广播字节到XMM寄存器，因为memset的一部分。 即它是什么的一部分_mm_set_epi8(x)。
乘以0x0101010101010101在 64 位整数中重复一个字节 8 次。这让您可以使用标量整数存储来存储奇数 8 个字节（不是 16 的倍数），就像mov [r11], rdx存储一样。
将此 8 字节广播作为输入（via movaq），只需要一次 SIMD shuffle。复制低位 8punpcklqdq是我的选择，因为 8 字节粒度混洗在像 Core 2 这样的老式 CPU 上效率更高。但是将字节相互交错是等效的，因为无论如何它们都是相同的，从而产生 XMM 寄存器包含相同字节的 16 个副本。
事实上，SSE2 可以用一条指令广播一个双字：pshufd xmm0, xmm0, 0，所以如果不是想要一个 8 字节的标量，它可以只使用imul edx, r9d, 0x01010101.
使用 8 字节mov和 16 字节movups存储实现 memset当然需要将其作为输入，如果它使用该策略而不是该rep stosb策略。
使用 SSSE3，您可以直接广播一个字节，其中一个pshufb带有全零向量（无需先乘法）为目标的每个元素选择源的第 0 个元素。或者使用 AVX2 vpbroadcastb。跳过整数乘法步骤就可以了；您可以使用movq [mem], xmm0来自 xmm0 而不是来自 RDX 的 8 字节存储。
使用 xmm 寄存器底部的一个字节和其他元素中的垃圾（即，如果您没有使用imul），2x punpcklbw+pshufd可以仅使用 SSE2 进行广播。或者当然punpcklbw xmm0,xmm0/punpcklwd xmm0,xmm0作为前2次洗牌。或punpcklbw xmm0,xmm0/ pshuflw xmm0,xmm0, 0/ punpcklqdq xmm0,xmm0。