lavr: handle clipping in the float to s32 conversion

We cannot clip to INT_MAX because that value cannot be exactly
represented by a float value and ends up overflowing during conversion
anyway. We need to use a slightly smaller float value, which ends up
with slightly inaccurate results for samples which clip or nearly clip,
but it is close enough. Using doubles as intermediates in the conversion
would be more accurate, but it takes about twice as much time.

Signed-off-by: Luca Barbato <lu_zero@gentoo.org>

libavresample/x86/audio_convert.asm

@@ -28,6 +28,7 @@ SECTION_RODATA 32
 
 pf_s32_inv_scale: times 8 dd 0x30000000
 pf_s32_scale:     times 8 dd 0x4f000000
+pf_s32_clip:      times 8 dd 0x4effffff
 pf_s16_inv_scale: times 4 dd 0x38000000
 pf_s16_scale:     times 4 dd 0x47000000
 pb_shuf_unpack_even:      db -1, -1,  0,  1, -1, -1,  2,  3, -1, -1,  8,  9, -1, -1, 10, 11
@@ -197,17 +198,22 @@ cglobal conv_flt_to_s16, 3,3,5, dst, src, len
 ;------------------------------------------------------------------------------
 
 %macro CONV_FLT_TO_S32 0
-cglobal conv_flt_to_s32, 3,3,5, dst, src, len
+cglobal conv_flt_to_s32, 3,3,6, dst, src, len
     lea     lenq, [lend*4]
     add     srcq, lenq
     add     dstq, lenq
     neg     lenq
     mova      m4, [pf_s32_scale]
+    mova      m5, [pf_s32_clip]
 .loop:
     mulps     m0, m4, [srcq+lenq         ]
     mulps     m1, m4, [srcq+lenq+1*mmsize]
     mulps     m2, m4, [srcq+lenq+2*mmsize]
     mulps     m3, m4, [srcq+lenq+3*mmsize]
+    minps     m0, m0, m5
+    minps     m1, m1, m5
+    minps     m2, m2, m5
+    minps     m3, m3, m5
     cvtps2dq  m0, m0
     cvtps2dq  m1, m1
     cvtps2dq  m2, m2