📄 integer_8cpp-source.html
字号:
<a name="l00492"></a>00492 <span class="preprocessor"></span><span class="preprocessor"> #define SquPrologue \</span><a name="l00493"></a>00493 <span class="preprocessor"> AS2( mov eax, A) \</span><a name="l00494"></a>00494 <span class="preprocessor"> AS2( mov ecx, C) \</span><a name="l00495"></a>00495 <span class="preprocessor"> SaveEBX \</span><a name="l00496"></a>00496 <span class="preprocessor"> AS2( lea ebx, s_maskLow16)</span><a name="l00497"></a>00497 <span class="preprocessor"></span><span class="preprocessor"> #define MulPrologue \</span><a name="l00498"></a>00498 <span class="preprocessor"> AS2( mov eax, A) \</span><a name="l00499"></a>00499 <span class="preprocessor"> AS2( mov edi, B) \</span><a name="l00500"></a>00500 <span class="preprocessor"> AS2( mov ecx, C) \</span><a name="l00501"></a>00501 <span class="preprocessor"> SaveEBX \</span><a name="l00502"></a>00502 <span class="preprocessor"> AS2( lea ebx, s_maskLow16)</span><a name="l00503"></a>00503 <span class="preprocessor"></span><span class="preprocessor"> #define TopPrologue \</span><a name="l00504"></a>00504 <span class="preprocessor"> AS2( mov eax, A) \</span><a name="l00505"></a>00505 <span class="preprocessor"> AS2( mov edi, B) \</span><a name="l00506"></a>00506 <span class="preprocessor"> AS2( mov ecx, C) \</span><a name="l00507"></a>00507 <span class="preprocessor"> AS2( mov esi, L) \</span><a name="l00508"></a>00508 <span class="preprocessor"> SaveEBX \</span><a name="l00509"></a>00509 <span class="preprocessor"> AS2( lea ebx, s_maskLow16)</span><a name="l00510"></a>00510 <span class="preprocessor"></span><span class="preprocessor"> #define SquEpilogue RestoreEBX</span><a name="l00511"></a>00511 <span class="preprocessor"></span><span class="preprocessor"> #define MulEpilogue RestoreEBX</span><a name="l00512"></a>00512 <span class="preprocessor"></span><span class="preprocessor"> #define TopEpilogue RestoreEBX</span><a name="l00513"></a>00513 <span class="preprocessor"></span><span class="preprocessor">#endif</span><a name="l00514"></a>00514 <span class="preprocessor"></span><a name="l00515"></a>00515 <span class="preprocessor">#ifdef CRYPTOPP_X64_MASM_AVAILABLE</span><a name="l00516"></a>00516 <span class="preprocessor"></span><span class="keyword">extern</span> <span class="stringliteral">"C"</span> {<a name="l00517"></a>00517 <span class="keywordtype">int</span> Baseline_Add(<span class="keywordtype">size_t</span> N, word *C, <span class="keyword">const</span> word *A, <span class="keyword">const</span> word *B);<a name="l00518"></a>00518 <span class="keywordtype">int</span> Baseline_Sub(<span class="keywordtype">size_t</span> N, word *C, <span class="keyword">const</span> word *A, <span class="keyword">const</span> word *B);<a name="l00519"></a>00519 }<a name="l00520"></a>00520 <span class="preprocessor">#elif defined(CRYPTOPP_X64_ASM_AVAILABLE) && defined(__GNUC__)</span><a name="l00521"></a>00521 <span class="preprocessor"></span><span class="keywordtype">int</span> Baseline_Add(<span class="keywordtype">size_t</span> N, word *C, <span class="keyword">const</span> word *A, <span class="keyword">const</span> word *B)<a name="l00522"></a>00522 {<a name="l00523"></a>00523 word result;<a name="l00524"></a>00524 __asm__ __volatile__<a name="l00525"></a>00525 (<a name="l00526"></a>00526 <span class="stringliteral">".intel_syntax;"</span><a name="l00527"></a>00527 AS1( neg %1)<a name="l00528"></a>00528 ASJ( jz, 1, f)<a name="l00529"></a>00529 AS2( mov %0,[%3+8*%1])<a name="l00530"></a>00530 AS2( <span class="keyword">add</span> %0,[%4+8*%1])<a name="l00531"></a>00531 AS2( mov [%2+8*%1],%0)<a name="l00532"></a>00532 ASL(0)<a name="l00533"></a>00533 AS2( mov %0,[%3+8*%1+8])<a name="l00534"></a>00534 AS2( adc %0,[%4+8*%1+8])<a name="l00535"></a>00535 AS2( mov [%2+8*%1+8],%0)<a name="l00536"></a>00536 AS2( lea %1,[%1+2])<a name="l00537"></a>00537 ASJ( jrcxz, 1, f)<a name="l00538"></a>00538 AS2( mov %0,[%3+8*%1])<a name="l00539"></a>00539 AS2( adc %0,[%4+8*%1])<a name="l00540"></a>00540 AS2( mov [%2+8*%1],%0)<a name="l00541"></a>00541 ASJ( jmp, 0, b)<a name="l00542"></a>00542 ASL(1)<a name="l00543"></a>00543 AS2( mov %0, 0)<a name="l00544"></a>00544 AS2( adc %0, %0)<a name="l00545"></a>00545 <span class="stringliteral">".att_syntax;"</span><a name="l00546"></a>00546 : <span class="stringliteral">"=&r"</span> (result)<a name="l00547"></a>00547 : <span class="stringliteral">"c"</span> (N), <span class="stringliteral">"r"</span> (C+N), <span class="stringliteral">"r"</span> (A+N), <span class="stringliteral">"r"</span> (B+N)<a name="l00548"></a>00548 : <span class="stringliteral">"memory"</span>, <span class="stringliteral">"cc"</span><a name="l00549"></a>00549 );<a name="l00550"></a>00550 <span class="keywordflow">return</span> (<span class="keywordtype">int</span>)result;<a name="l00551"></a>00551 }<a name="l00552"></a>00552 <a name="l00553"></a>00553 <span class="keywordtype">int</span> Baseline_Sub(<span class="keywordtype">size_t</span> N, word *C, <span class="keyword">const</span> word *A, <span class="keyword">const</span> word *B)<a name="l00554"></a>00554 {<a name="l00555"></a>00555 word result;<a name="l00556"></a>00556 __asm__ __volatile__<a name="l00557"></a>00557 (<a name="l00558"></a>00558 <span class="stringliteral">".intel_syntax;"</span><a name="l00559"></a>00559 AS1( neg %1)<a name="l00560"></a>00560 ASJ( jz, 1, f)<a name="l00561"></a>00561 AS2( mov %0,[%3+8*%1])<a name="l00562"></a>00562 AS2( sub %0,[%4+8*%1])<a name="l00563"></a>00563 AS2( mov [%2+8*%1],%0)<a name="l00564"></a>00564 ASL(0)<a name="l00565"></a>00565 AS2( mov %0,[%3+8*%1+8])<a name="l00566"></a>00566 AS2( sbb %0,[%4+8*%1+8])<a name="l00567"></a>00567 AS2( mov [%2+8*%1+8],%0)<a name="l00568"></a>00568 AS2( lea %1,[%1+2])<a name="l00569"></a>00569 ASJ( jrcxz, 1, f)<a name="l00570"></a>00570 AS2( mov %0,[%3+8*%1])<a name="l00571"></a>00571 AS2( sbb %0,[%4+8*%1])<a name="l00572"></a>00572 AS2( mov [%2+8*%1],%0)<a name="l00573"></a>00573 ASJ( jmp, 0, b)<a name="l00574"></a>00574 ASL(1)<a name="l00575"></a>00575 AS2( mov %0, 0)<a name="l00576"></a>00576 AS2( adc %0, %0)<a name="l00577"></a>00577 <span class="stringliteral">".att_syntax;"</span><a name="l00578"></a>00578 : <span class="stringliteral">"=&r"</span> (result)<a name="l00579"></a>00579 : <span class="stringliteral">"c"</span> (N), <span class="stringliteral">"r"</span> (C+N), <span class="stringliteral">"r"</span> (A+N), <span class="stringliteral">"r"</span> (B+N)<a name="l00580"></a>00580 : <span class="stringliteral">"memory"</span>, <span class="stringliteral">"cc"</span><a name="l00581"></a>00581 );<a name="l00582"></a>00582 <span class="keywordflow">return</span> (<span class="keywordtype">int</span>)result;<a name="l00583"></a>00583 }<a name="l00584"></a>00584 <span class="preprocessor">#elif defined(CRYPTOPP_X86_ASM_AVAILABLE) && CRYPTOPP_BOOL_X86</span><a name="l00585"></a>00585 <span class="preprocessor"></span>CRYPTOPP_NAKED <span class="keywordtype">int</span> CRYPTOPP_FASTCALL Baseline_Add(<span class="keywordtype">size_t</span> N, word *C, <span class="keyword">const</span> word *A, <span class="keyword">const</span> word *B)<a name="l00586"></a>00586 {<a name="l00587"></a>00587 AddPrologue<a name="l00588"></a>00588 <a name="l00589"></a>00589 <span class="comment">// now: eax = A, edi = B, edx = C, ecx = N</span><a name="l00590"></a>00590 AS2( lea eax, [eax+4*ecx])<a name="l00591"></a>00591 AS2( lea edi, [edi+4*ecx])<a name="l00592"></a>00592 AS2( lea edx, [edx+4*ecx])<a name="l00593"></a>00593 <a name="l00594"></a>00594 AS1( neg ecx) <span class="comment">// ecx is negative index</span><a name="l00595"></a>00595 AS2( test ecx, 2) <span class="comment">// this clears carry flag</span><a name="l00596"></a>00596 ASJ( jz, 0, f)<a name="l00597"></a>00597 AS2( sub ecx, 2)<a name="l00598"></a>00598 ASJ( jmp, 1, f)<a name="l00599"></a>00599 <a name="l00600"></a>00600 ASL(0)<a name="l00601"></a>00601 ASJ( jecxz, 2, f) <span class="comment">// loop until ecx overflows and becomes zero</span><a name="l00602"></a>00602 AS2( mov esi,[eax+4*ecx])<a name="l00603"></a>00603 AS2( adc esi,[edi+4*ecx])<a name="l00604"></a>00604 AS2( mov [edx+4*ecx],esi)<a name="l00605"></a>00605 AS2( mov esi,[eax+4*ecx+4])<a name="l00606"></a>00606 AS2( adc esi,[edi+4*ecx+4])<a name="l00607"></a>00607 AS2( mov [edx+4*ecx+4],esi)<a name="l00608"></a>00608 ASL(1)<a name="l00609"></a>00609 AS2( mov esi,[eax+4*ecx+8])<a name="l00610"></a>00610 AS2( adc esi,[edi+4*ecx+8])<a name="l00611"></a>00611 AS2( mov [edx+4*ecx+8],esi)<a name="l00612"></a>00612 AS2( mov esi,[eax+4*ecx+12])<a name="l00613"></a>00613 AS2( adc esi,[edi+4*ecx+12])<a name="l00614"></a>00614 AS2( mov [edx+4*ecx+12],esi)<a name="l00615"></a>00615 <a name="l00616"></a>00616 AS2( lea ecx,[ecx+4]) <span class="comment">// advance index, avoid inc which causes slowdown on Intel Core 2</span><a na
⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -