[PATCH] use __builtin_convertvector for _mm_cvtepi32_ps to enable constant propagation

[Matthias Kretz]

  Testcase:


    __m128 cvt0() { return _mm_cvtepi32_ps(_mm_set1_epi32(2)); }
    __m128 cvt0(__m128i x) { return _mm_cvtepi32_ps(x); }

  compiles to IR:


    define <4 x float> @_Z4cvt0v() #0 {
    entry:
      ret <4 x float> <float 2.000000e+00, float 2.000000e+00, float 2.000000e+00, float 2.000000e+00>
    }

    define <4 x float> @_Z4cvt0Dv2_x(<2 x i64> %x) #0 {
    entry:
      %0 = bitcast <2 x i64> %x to <4 x i32>
      %conv.i = sitofp <4 x i32> %0 to <4 x float>
      ret <4 x float> %conv.i
    }

  and x86:


    0000000000000000 <cvt0()>:
       0:·       c5 f8 28 05 00 00 00 00 · vmovaps 0x0(%rip),%xmm0        # 8 <cvt0()+0x8>·    4: R_X86_64_PC32·         .LCPI0_0-0x4
       8:·       c3                   ·    retq

    0000000000000010 <cvt0(long long __vector(2))>:
      10:·       c5 f8 5b c0          ·    vcvtdq2ps %xmm0,%xmm0
      14:·       c3                   ·    retq

http://llvm-reviews.chandlerc.com/D1792