SSE2 instructions generated when /arch:SSE is specified by David York

Visual Studio 2010 incorporates a common technique for converting a memory based 32 bit signed integer to a 32 bit floating point xmm register using a movd and cvtdq2ps instruction pairing.

The problem for us is that cvtdq2ps is an SSE2 instruction and might give the ol' PIII's and Athlon XP's pause when running programs compiled for SSE targets.

Example of incorrect code generation for SSE-only machines:

void test(float *v, int alpha) {

    float f = v[1] * alpha;
    ...

00000    66 0f 6e 44 24 04     movd     xmm0, DWORD PTR _alpha$[esp-4]
00006    f3 0f 10 48 04         movss     xmm1, DWORD PTR [eax+4]
0000b    0f 5b c0             cvtdq2ps xmm0, xmm0
0000e    f3 0f 59 c8         mulss     xmm1, xmm0

Using these compiler options:

D:\Program Files\Microsoft Visual Studio 10.0\VC\bin\CL.exe /c /I"D:\Program Files\Microsoft DirectX SDK (February 2010)\Include"
/D WIN32 /D NDEBUG /D _WINDOWS /D UNICODE /D _WIN32_WINNT=0x0500 /D_VC80_UPGRADE=0x0600 /D _UNICODE /D UNICODE
/Zi /nologo /W2 /WX- /Ox /Ob1 /Oi /Ot /Oy /GL /GF /Gm- /MT /GS- /Gy
/arch:SSE /fp:fast /fp:except- /Zc:wchar_t /Zc:forScope /GR- /FAcs /Gr /TP /analyze- /errorReport:queue
/Fd".\Release/jack.pdb"