VS2005和VS2010 C++标准库 exp结果不一致

如图，今天在工作的时候，要迁移老代码到新的编译工具上，然后两边跑同样的案例不一致，经排查是在一个只根据参数计算结果的函数里发生了不一致，单独把变量摘出来后，复现和debug的时候一致。怀疑是05的exp标准库为了性能、或者当时指令集有限等原因导致为了性能损失了精度。

 10的汇编代码

exp:
000000005AB2D010  sub         rsp,88h  
000000005AB2D017  movsd       mmword ptr [rsp+30h],xmm0  
000000005AB2D01D  mov         rax,qword ptr [__real_inf (5AB31090h)]  
000000005AB2D024  mov         rdx,qword ptr [rsp+30h]  
000000005AB2D029  and         rax,rdx  
000000005AB2D02C  mov         r9,rdx  
000000005AB2D02F  cmp         rax,qword ptr [__real_inf (5AB31090h)]  
000000005AB2D036  je          __x_is_inf_or_nan (5AB2D260h)  
000000005AB2D03C  and         r9,qword ptr [__exp_mant_mask (5AB31070h)]  
000000005AB2D043  cmp         r9,qword ptr [__real_x_near0_threshold (5AB31080h)]  
000000005AB2D04A  movsd       xmm3,mmword ptr [__real_64_by_log2 (5AB310E0h)]  
000000005AB2D052  jl          __process_result_one (5AB2D180h)  
000000005AB2D058  mulsd       xmm3,xmm0  
000000005AB2D05C  comisd      xmm3,mmword ptr [__real_p65536 (5AB310C0h)]  
000000005AB2D064  ja          __y_is_inf (5AB2D240h)  
000000005AB2D06A  comisd      xmm3,mmword ptr [__real_m68800 (5AB310D0h)]  
000000005AB2D072  jbe         __y_is_zero (5AB2D220h)  
000000005AB2D078  cvtpd2dq    xmm4,xmm3  
000000005AB2D07C  lea         r10,[__two_to_jby64_head_table (5AB31360h)]  
000000005AB2D083  lea         r11,[__two_to_jby64_tail_table (5AB31560h)]  
000000005AB2D08A  cvtdq2pd    xmm1,xmm4  
000000005AB2D08E  movsd       xmm2,mmword ptr [__real_log2_by_64_head (5AB310F0h)]  
000000005AB2D096  mulsd       xmm2,xmm1  
000000005AB2D09A  movd        ecx,xmm4  
000000005AB2D09E  mov         rax,3Fh  
000000005AB2D0A5  and         eax,ecx  
000000005AB2D0A7  subsd       xmm0,xmm2  
000000005AB2D0AB  mulsd       xmm1,mmword ptr [__real_log2_by_64_tail (5AB31100h)]  
000000005AB2D0B3  movsd       xmm2,xmm0  
000000005AB2D0B7  sub         ecx,eax  
000000005AB2D0B9  sar         ecx,6  
000000005AB2D0BC  addsd       xmm2,xmm1  
000000005AB2D0C0  movsd       xmm1,xmm2  
000000005AB2D0C4  movsd       xmm0,mmword ptr [__real_1_by_2 (5AB31150h)]  
000000005AB2D0CC  movsd       xmm3,mmword ptr [__real_1_by_24 (5AB31130h)]  
000000005AB2D0D4  movsd       xmm4,mmword ptr [__real_1_by_720 (5AB31110h)]  
000000005AB2D0DC  mulsd       xmm1,xmm2  
000000005AB2D0E0  mulsd       xmm0,xmm2  
000000005AB2D0E4  mulsd       xmm3,xmm2  
000000005AB2D0E8  mulsd       xmm4,xmm2  
000000005AB2D0EC  movsd       xmm5,xmm1  
000000005AB2D0F0  mulsd       xmm1,xmm2  
000000005AB2D0F4  addsd       xmm0,mmword ptr [__real_one (5AB31040h)]  
000000005AB2D0FC  addsd       xmm3,mmword ptr [__real_1_by_6 (5AB31140h)]  
000000005AB2D104  mulsd       xmm5,xmm1  
000000005AB2D108  addsd       xmm4,mmword ptr [__real_1_by_120 (5AB31120h)]  
000000005AB2D110  mulsd       xmm0,xmm2  
000000005AB2D114  mulsd       xmm3,xmm1  
000000005AB2D118  mulsd       xmm4,xmm5  
000000005AB2D11C  xor         r9d,r9d  
000000005AB2D11F  cmp         ecx,dword ptr [__denormal_threshold (5AB31010h)]  
000000005AB2D125  addsd       xmm3,xmm4  
000000005AB2D129  addsd       xmm0,xmm3  
000000005AB2D12D  cmovle      r9d,ecx  
000000005AB2D131  add         rcx,3FFh  
000000005AB2D138  shl         rcx,34h  
000000005AB2D13C  lea         r8,[__two_to_jby64_table (5AB31160h)]  
000000005AB2D143  mulsd       xmm0,mmword ptr [r8+rax*8]  
000000005AB2D149  cmp         rcx,qword ptr [__real_inf (5AB31090h)]  
000000005AB2D150  addsd       xmm0,mmword ptr [r11+rax*8]  
000000005AB2D156  addsd       xmm0,mmword ptr [r10+rax*8]  
000000005AB2D15C  je          __process_almost_inf (5AB2D190h)  
000000005AB2D15E  test        r9d,r9d  
000000005AB2D161  mov         qword ptr [rsp+30h],rcx  
000000005AB2D166  jne         __process_denormal (5AB2D1A0h)  
000000005AB2D168  mulsd       xmm0,mmword ptr [rsp+30h]  
__final_check:
000000005AB2D16E  add         rsp,88h  
000000005AB2D175  ret  
000000005AB2D176  nop         word ptr [rax+rax]  
__process_result_one:
000000005AB2D180  movsd       xmm0,mmword ptr [__real_one (5AB31040h)]  
000000005AB2D188  jmp         __final_check (5AB2D16Eh)  
000000005AB2D18A  nop         word ptr [rax+rax]  
__process_almost_inf:
000000005AB2D190  orpd        xmm0,xmmword ptr [__enable_almost_inf (5AB31020h)]  
000000005AB2D198  jmp         __final_check (5AB2D16Eh)  
000000005AB2D19A  nop         word ptr [rax+rax]  
__process_denormal:
000000005AB2D1A0  mov         ecx,r9d  
000000005AB2D1A3  xor         r11d,r11d  
000000005AB2D1A6  comisd      xmm0,mmword ptr [__real_one (5AB31040h)]  
000000005AB2D1AE  cmovae      r11d,ecx  
000000005AB2D1B2  cmp         r11d,dword ptr [__denormal_threshold (5AB31010h)]  
000000005AB2D1B9  jne         __process_true_denormal (5AB2D1D0h)  
000000005AB2D1BB  mulsd       xmm0,mmword ptr [rsp+30h]  
000000005AB2D1C1  jmp         __final_check (5AB2D16Eh)  
000000005AB2D1C3  nop         word ptr [rax+rax]  
__process_true_denormal:
000000005AB2D1D0  xor         r8,r8  
000000005AB2D1D3  cmp         rdx,qword ptr [__denormal_tiny_threshold (5AB31060h)]  
000000005AB2D1DA  mov         r9,1  
000000005AB2D1E1  jg          __process_denormal_tiny (5AB2D210h)  
000000005AB2D1E3  add         ecx,432h  
000000005AB2D1E9  cmovs       rcx,r8  
000000005AB2D1ED  shl         r9,cl  
000000005AB2D1F0  mov         rcx,r9  
000000005AB2D1F3  mov         qword ptr [rsp+30h],rcx  
000000005AB2D1F8  mulsd       xmm0,mmword ptr [rsp+30h]  
000000005AB2D1FE  jmp         __final_check (5AB2D16Eh)  
000000005AB2D203  nop         word ptr [rax+rax]  
__process_denormal_tiny:
000000005AB2D210  movsd       xmm0,mmword ptr [__real_smallest_denormal (5AB31050h)]  
000000005AB2D218  jmp         __final_check (5AB2D16Eh)  
000000005AB2D21D  nop         dword ptr [rax]  
__y_is_zero:
000000005AB2D220  movsd       xmm1,mmword ptr [__real_zero (5AB31030h)]  
000000005AB2D228  movd        xmm0,rdx  
000000005AB2D22D  mov         r8d,dword ptr [__flag_y_zero (5AB31004h)]  
000000005AB2D234  call        _exp_special (5AAB46B0h)  
000000005AB2D239  jmp         __finish (5AB2D2A0h)  
000000005AB2D23E  xchg        ax,ax  
__y_is_inf:
000000005AB2D240  movsd       xmm1,mmword ptr [__real_inf (5AB31090h)]  
000000005AB2D248  movd        xmm0,rdx  
000000005AB2D24D  mov         r8d,dword ptr [__flag_y_inf (5AB31008h)]  
000000005AB2D254  call        _exp_special (5AAB46B0h)  
000000005AB2D259  jmp         __finish (5AB2D2A0h)  
000000005AB2D25B  nop         dword ptr [rax+rax]  
__x_is_inf_or_nan:
000000005AB2D260  cmp         rdx,qword ptr [__real_inf (5AB31090h)]  
000000005AB2D267  je          __finish (5AB2D2A0h)  
000000005AB2D269  cmp         rdx,qword ptr [__real_ninf (5AB310A0h)]  
000000005AB2D270  je          __process_zero (5AB2D290h)  
000000005AB2D272  or          rdx,qword ptr [__real_qnanbit (5AB310B0h)]  
000000005AB2D279  movd        xmm1,rdx  
000000005AB2D27E  mov         r8d,dword ptr [__flag_x_nan (5AB31000h)]  
000000005AB2D285  call        _exp_special (5AAB46B0h)  
000000005AB2D28A  jmp         __finish (5AB2D2A0h)  
000000005AB2D28C  nop         dword ptr [rax]  
__process_zero:
000000005AB2D290  movsd       xmm0,mmword ptr [__real_zero (5AB31030h)]  
000000005AB2D298  jmp         __final_check (5AB2D16Eh)  
000000005AB2D29D  nop         dword ptr [rax]  
__finish:
000000005AB2D2A0  add         rsp,88h  
000000005AB2D2A7  ret  

05的汇编

exp:
00000001400010BA  jmp         qword ptr [__imp_exp (14000A518h)] 
_RTC_InitBase:
00000001400010C0  sub         rsp,38h 
00000001400010C4  cmp         byte ptr [init (1400081A0h)],0 
00000001400010CB  jne         _RTC_InitBase+36h (1400010F6h) 
00000001400010CD  mov         r9d,1 
00000001400010D3  xor         r8d,r8d 
00000001400010D6  xor         edx,edx 
00000001400010D8  xor         ecx,ecx 
00000001400010DA  mov         byte ptr [init (1400081A0h)],1 
00000001400010E1  mov         dword ptr [rsp+20h],0 
00000001400010E9  call        _CRT_RTC_INITW (140001658h) 
00000001400010EE  mov         rcx,rax 
00000001400010F1  call        _RTC_SetErrorFuncW (140001620h) 
00000001400010F6  add         rsp,38h 
00000001400010FA  ret

很明显的差异，我估计10的精度应该更高很多

经过性能测试，05的只有10的一半性能，

100w次，05需要0.007秒~0.008,10需要0.002~0.003

由于还在上班，暂时没法上高精度计时，但是多次重复差不多都是这么个值，还是10又快又准。