Visual Studio调试与发布版本:比较int和float missmatch

Question

看看这个例子:

#include <stdio.h>

int main() {

  int i= 16777217;
  float f = 16777216.0;
  float g = i;

  if( i == f )
    printf("eq\n");
  else
    printf("neq\n");

  if( g == f )
    printf("eq\n");
  else
    printf("neq\n");

  return 0;
}

其中,在发布模式下使用Visual Studio 2010 C++(VS),gcc或g ++(4.9.2)具有输出

eq
eq

这对我来说是合理的:在第一次比较期间,i隐式转换为浮点数,其中尾数中的有效位被截断.因此,两个i和f具有比较平等同位模式.在第二个中if,相同的转换适用但在定义和初始化时已经执行g.

但是,在调试模式下使用VS,结果是

neq
eq

看来,第一次比较期间的隐式转换if(作为C和C++中通常的算术转换的一部分)并未应用.这是真的？是否存在VS机制,可以在比较浮点数和整数(转换为更高精度的int/float)时防止此类误报？根据MSDN VS C++遵循的标准.

我用这个函数检查了位表示.对于所有编译器,它产生

i = 00000001000000000000000000000001
f = 01001011100000000000000000000000
g = 01001011100000000000000000000000

float.h在VS状态,#define FLT_MANT_DIG 24所以描述的截断问题也应该成立.

我在同一台机器(英特尔i5-3570K)上编译了所有内容,但是在虚拟机中安装了VS. 在另一台机器上使用VS进行编译也会打印neq/eq.

编辑:附加汇编代码

differences_debug.asm

; Listing generated by Microsoft (R) Optimizing Compiler Version 16.00.30319.01 

    TITLE   c:\Users\user\documents\visual studio 2010\Projects\differences\differences\differences.cpp
    .686P
    .XMM
    include listing.inc
    .model  flat

INCLUDELIB MSVCRTD
INCLUDELIB OLDNAMES

PUBLIC  ??_C@_04LMPLCMBC@neq?6?$AA@         ; `string'
PUBLIC  ??_C@_03HNJPMNDP@eq?6?$AA@          ; `string'
PUBLIC  __real@4b800000
PUBLIC  _wmain
EXTRN   __imp__printf:PROC
EXTRN   __fltused:DWORD
EXTRN   __RTC_CheckEsp:PROC
EXTRN   __RTC_Shutdown:PROC
EXTRN   __RTC_InitBase:PROC
;   COMDAT ??_C@_04LMPLCMBC@neq?6?$AA@
; File c:\users\user\documents\visual studio 2010\projects\differences\differences\differences.cpp
CONST   SEGMENT
??_C@_04LMPLCMBC@neq?6?$AA@ DB 'neq', 0aH, 00H      ; `string'
CONST   ENDS
;   COMDAT ??_C@_03HNJPMNDP@eq?6?$AA@
CONST   SEGMENT
??_C@_03HNJPMNDP@eq?6?$AA@ DB 'eq', 0aH, 00H        ; `string'
CONST   ENDS
;   COMDAT __real@4b800000
CONST   SEGMENT
__real@4b800000 DD 04b800000r           ; 1.67772e+007
CONST   ENDS
;   COMDAT rtc$TMZ
rtc$TMZ SEGMENT
__RTC_Shutdown.rtc$TMZ DD FLAT:__RTC_Shutdown
rtc$TMZ ENDS
;   COMDAT rtc$IMZ
rtc$IMZ SEGMENT
__RTC_InitBase.rtc$IMZ DD FLAT:__RTC_InitBase
; Function compile flags: /Odtp /RTCsu /ZI
rtc$IMZ ENDS
;   COMDAT _wmain
_TEXT   SEGMENT
_g$ = -32                       ; size = 4
_f$ = -20                       ; size = 4
_i$ = -8                        ; size = 4
_argc$ = 8                      ; size = 4
_argv$ = 12                     ; size = 4
_wmain  PROC                        ; COMDAT
; Line 7
    push    ebp
    mov ebp, esp
    sub esp, 228                ; 000000e4H
    push    ebx
    push    esi
    push    edi
    lea edi, DWORD PTR [ebp-228]
    mov ecx, 57                 ; 00000039H
    mov eax, -858993460             ; ccccccccH
    rep stosd
; Line 8
    mov DWORD PTR _i$[ebp], 16777217        ; 01000001H
; Line 9
    fld DWORD PTR __real@4b800000
    fstp    DWORD PTR _f$[ebp]
; Line 10
    fild    DWORD PTR _i$[ebp]
    fstp    DWORD PTR _g$[ebp]
; Line 13
    fild    DWORD PTR _i$[ebp]
    fld DWORD PTR _f$[ebp]
    fucompp
    fnstsw  ax
    test    ah, 68                  ; 00000044H
    jp  SHORT $LN4@wmain
; Line 14
    mov esi, esp
    push    OFFSET ??_C@_03HNJPMNDP@eq?6?$AA@
    call    DWORD PTR __imp__printf
    add esp, 4
    cmp esi, esp
    call    __RTC_CheckEsp
; Line 15
    jmp SHORT $LN3@wmain
$LN4@wmain:
; Line 16
    mov esi, esp
    push    OFFSET ??_C@_04LMPLCMBC@neq?6?$AA@
    call    DWORD PTR __imp__printf
    add esp, 4
    cmp esi, esp
    call    __RTC_CheckEsp
$LN3@wmain:
; Line 19
    fld DWORD PTR _g$[ebp]
    fld DWORD PTR _f$[ebp]
    fucompp
    fnstsw  ax
    test    ah, 68                  ; 00000044H
    jp  SHORT $LN2@wmain
; Line 20
    mov esi, esp
    push    OFFSET ??_C@_03HNJPMNDP@eq?6?$AA@
    call    DWORD PTR __imp__printf
    add esp, 4
    cmp esi, esp
    call    __RTC_CheckEsp
; Line 21
    jmp SHORT $LN1@wmain
$LN2@wmain:
; Line 22
    mov esi, esp
    push    OFFSET ??_C@_04LMPLCMBC@neq?6?$AA@
    call    DWORD PTR __imp__printf
    add esp, 4
    cmp esi, esp
    call    __RTC_CheckEsp
$LN1@wmain:
; Line 24
    xor eax, eax
; Line 26
    pop edi
    pop esi
    pop ebx
    add esp, 228                ; 000000e4H
    cmp ebp, esp
    call    __RTC_CheckEsp
    mov esp, ebp
    pop ebp
    ret 0
_wmain  ENDP
_TEXT   ENDS
END

differences_release.asm

; Listing generated by Microsoft (R) Optimizing Compiler Version 16.00.30319.01 

    TITLE   c:\Users\user\documents\visual studio 2010\Projects\differences\differences\differences.cpp
    .686P
    .XMM
    include listing.inc
    .model  flat

INCLUDELIB OLDNAMES

PUBLIC  ??_C@_03HNJPMNDP@eq?6?$AA@          ; `string'
PUBLIC  ??_C@_04LMPLCMBC@neq?6?$AA@         ; `string'
EXTRN   @__security_check_cookie@4:PROC
EXTRN   __imp__printf:PROC
;   COMDAT ??_C@_04LMPLCMBC@neq?6?$AA@
CONST   SEGMENT
??_C@_04LMPLCMBC@neq?6?$AA@ DB 'neq', 0aH, 00H      ; `string'
CONST   ENDS
;   COMDAT ??_C@_03HNJPMNDP@eq?6?$AA@
CONST   SEGMENT
??_C@_03HNJPMNDP@eq?6?$AA@ DB 'eq', 0aH, 00H        ; `string'
CONST   ENDS
PUBLIC  _wmain
EXTRN   __fltused:DWORD
; Function compile flags: /Ogtp
;   COMDAT _wmain
_TEXT   SEGMENT
_argc$ = 8                      ; size = 4
_argv$ = 12                     ; size = 4
_wmain  PROC                        ; COMDAT
; File c:\users\user\documents\visual studio 2010\projects\differences\differences\differences.cpp
; Line 7
    push    esi
; Line 14
    mov esi, DWORD PTR __imp__printf
    push    OFFSET ??_C@_03HNJPMNDP@eq?6?$AA@
    call    esi
; Line 20
    push    OFFSET ??_C@_03HNJPMNDP@eq?6?$AA@
    call    esi
    add esp, 8
; Line 24
    xor eax, eax
    pop esi
; Line 26
    ret 0
_wmain  ENDP
_TEXT   ENDS
END

Answer 1

如果我们解除ASM的发布:

; Line 14
    push    OFFSET ??_C@_03HNJPMNDP@eq?6?$AA@
    call    DWORD PTR __imp__printf
    add esp, 4
; Line 18
    xor eax, eax
; Line 20
    ret 0

它只是打印eq和退出,这表明浮动比较刚刚完全优化.对于调试程序集,我们看到它使用fld和fild说明:

; Line 9
    fld DWORD PTR __real@4b800000
    fstp    DWORD PTR _f$[ebp]
; Line 10
    fild    DWORD PTR _i$[ebp]
    fstp    DWORD PTR _g$[ebp]
; Line 13
    fild    DWORD PTR _i$[ebp]

这些是IA32Visual Studio 2010中使用的默认体系结构的说明.我怀疑您使用/arch:SSE2它会得到不同的结果.

Hans Passant的评论基本上证实了我刚才所说的话.