seal 0.1.0

data/mpg123/src/libmpg123/dct64_i486.c

@@ -0,0 +1,342 @@
+/*
+	dct64_i486.c: DCT64, a plain C variant for i486
+
+	copyright 1998-2006 by the mpg123 project - free software under the terms of the LGPL 2.1
+	see COPYING and AUTHORS files in distribution or http://mpg123.org
+	initially written by Fabrice Bellard
+*/
+
+/* Discrete Cosine Tansform (DCT) for subband synthesis.
+ *
+ * This code is optimized for 80486. It should be compiled with gcc
+ * 2.7.2 or higher.
+ *
+ * Note: This code does not give the necessary accuracy. Moreover, no
+ * overflow test are done.
+ *
+ * (c) 1998 Fabrice Bellard.  
+ */
+
+#include "mpg123lib_intern.h"
+
+#define COS_0_0 16403
+#define COS_0_1 16563
+#define COS_0_2 16890
+#define COS_0_3 17401
+#define COS_0_4 18124
+#define COS_0_5 19101
+#define COS_0_6 20398
+#define COS_0_7 22112
+#define COS_0_8 24396
+#define COS_0_9 27503
+#define COS_0_10 31869
+#define COS_0_11 38320
+#define COS_0_12 48633
+#define COS_0_13 67429
+#define COS_0_14 111660
+#define COS_0_15 333906
+#define COS_1_0 16463
+#define COS_1_1 17121
+#define COS_1_2 18577
+#define COS_1_3 21195
+#define COS_1_4 25826
+#define COS_1_5 34756
+#define COS_1_6 56441
+#define COS_1_7 167154
+#define COS_2_0 16704
+#define COS_2_1 19704
+#define COS_2_2 29490
+#define COS_2_3 83981
+#define COS_3_0 17733
+#define COS_3_1 42813
+#define COS_4_0 23170         
+
+#define SETOUT(out,n,expr) out[FIR_BUFFER_SIZE*(n)]=(expr)
+#define MULL(a,b) (((long long)(a)*(long long)(b)) >> 15)
+#define MUL(a,b) \
+(\
+       ((!(b & 0x3F)) ? (((a)*(b >> 6)) >> 9) :\
+       ((!(b & 0x1F)) ? (((a)*(b >> 5)) >> 10) :\
+       ((!(b & 0x0F)) ? (((a)*(b >> 4)) >> 11) :\
+       ((!(b & 0x07)) ? (((a)*(b >> 3)) >> 12) :\
+       ((!(b & 0x03)) ? (((a)*(b >> 2)) >> 13) :\
+       ((!(b & 0x01)) ? (((a)*(b >> 1)) >> 14) :\
+                        (((a)*(b   )) >> 15))))))))
+
+
+void dct64_1_486(int *out0,int *out1,int *b1,int *b2)
+{
+  b1[0x00] = b2[0x00] + b2[0x1F];
+  b1[0x1F] = MUL((b2[0x00] - b2[0x1F]),COS_0_0);
+
+  b1[0x01] = b2[0x01] + b2[0x1E];
+  b1[0x1E] = MUL((b2[0x01] - b2[0x1E]),COS_0_1);
+
+  b1[0x02] = b2[0x02] + b2[0x1D];
+  b1[0x1D] = MUL((b2[0x02] - b2[0x1D]),COS_0_2);
+
+  b1[0x03] = b2[0x03] + b2[0x1C];
+  b1[0x1C] = MUL((b2[0x03] - b2[0x1C]),COS_0_3);
+
+  b1[0x04] = b2[0x04] + b2[0x1B];
+  b1[0x1B] = MUL((b2[0x04] - b2[0x1B]),COS_0_4);
+
+  b1[0x05] = b2[0x05] + b2[0x1A];
+  b1[0x1A] = MUL((b2[0x05] - b2[0x1A]),COS_0_5);
+
+  b1[0x06] = b2[0x06] + b2[0x19];
+  b1[0x19] = MUL((b2[0x06] - b2[0x19]),COS_0_6);
+
+  b1[0x07] = b2[0x07] + b2[0x18];
+  b1[0x18] = MUL((b2[0x07] - b2[0x18]),COS_0_7);
+
+  b1[0x08] = b2[0x08] + b2[0x17];
+  b1[0x17] = MUL((b2[0x08] - b2[0x17]),COS_0_8);
+
+  b1[0x09] = b2[0x09] + b2[0x16];
+  b1[0x16] = MUL((b2[0x09] - b2[0x16]),COS_0_9);
+
+  b1[0x0A] = b2[0x0A] + b2[0x15];
+  b1[0x15] = MUL((b2[0x0A] - b2[0x15]),COS_0_10);
+
+  b1[0x0B] = b2[0x0B] + b2[0x14];
+  b1[0x14] = MUL((b2[0x0B] - b2[0x14]),COS_0_11);
+
+  b1[0x0C] = b2[0x0C] + b2[0x13];
+  b1[0x13] = MUL((b2[0x0C] - b2[0x13]),COS_0_12);
+
+  b1[0x0D] = b2[0x0D] + b2[0x12];
+  b1[0x12] = MULL((b2[0x0D] - b2[0x12]),COS_0_13);
+
+  b1[0x0E] = b2[0x0E] + b2[0x11];
+  b1[0x11] = MULL((b2[0x0E] - b2[0x11]),COS_0_14);
+
+  b1[0x0F] = b2[0x0F] + b2[0x10];
+  b1[0x10] = MULL((b2[0x0F] - b2[0x10]),COS_0_15);
+
+
+  b2[0x00] = b1[0x00] + b1[0x0F]; 
+  b2[0x0F] = MUL((b1[0x00] - b1[0x0F]),COS_1_0);
+  b2[0x01] = b1[0x01] + b1[0x0E]; 
+  b2[0x0E] = MUL((b1[0x01] - b1[0x0E]),COS_1_1);
+  b2[0x02] = b1[0x02] + b1[0x0D]; 
+  b2[0x0D] = MUL((b1[0x02] - b1[0x0D]),COS_1_2);
+  b2[0x03] = b1[0x03] + b1[0x0C]; 
+  b2[0x0C] = MUL((b1[0x03] - b1[0x0C]),COS_1_3);
+  b2[0x04] = b1[0x04] + b1[0x0B]; 
+  b2[0x0B] = MUL((b1[0x04] - b1[0x0B]),COS_1_4);
+  b2[0x05] = b1[0x05] + b1[0x0A]; 
+  b2[0x0A] = MUL((b1[0x05] - b1[0x0A]),COS_1_5);
+  b2[0x06] = b1[0x06] + b1[0x09]; 
+  b2[0x09] = MUL((b1[0x06] - b1[0x09]),COS_1_6);
+  b2[0x07] = b1[0x07] + b1[0x08]; 
+  b2[0x08] = MULL((b1[0x07] - b1[0x08]),COS_1_7);
+
+  b2[0x10] = b1[0x10] + b1[0x1F];
+  b2[0x1F] = MUL((b1[0x1F] - b1[0x10]),COS_1_0);
+  b2[0x11] = b1[0x11] + b1[0x1E];
+  b2[0x1E] = MUL((b1[0x1E] - b1[0x11]),COS_1_1);
+  b2[0x12] = b1[0x12] + b1[0x1D];
+  b2[0x1D] = MUL((b1[0x1D] - b1[0x12]),COS_1_2);
+  b2[0x13] = b1[0x13] + b1[0x1C];
+  b2[0x1C] = MUL((b1[0x1C] - b1[0x13]),COS_1_3);
+  b2[0x14] = b1[0x14] + b1[0x1B];
+  b2[0x1B] = MUL((b1[0x1B] - b1[0x14]),COS_1_4);
+  b2[0x15] = b1[0x15] + b1[0x1A];
+  b2[0x1A] = MUL((b1[0x1A] - b1[0x15]),COS_1_5);
+  b2[0x16] = b1[0x16] + b1[0x19];
+  b2[0x19] = MUL((b1[0x19] - b1[0x16]),COS_1_6);
+  b2[0x17] = b1[0x17] + b1[0x18];
+  b2[0x18] = MULL((b1[0x18] - b1[0x17]),COS_1_7);
+
+
+  b1[0x00] = b2[0x00] + b2[0x07];
+  b1[0x07] = MUL((b2[0x00] - b2[0x07]),COS_2_0);
+  b1[0x01] = b2[0x01] + b2[0x06];
+  b1[0x06] = MUL((b2[0x01] - b2[0x06]),COS_2_1);
+  b1[0x02] = b2[0x02] + b2[0x05];
+  b1[0x05] = MUL((b2[0x02] - b2[0x05]),COS_2_2);
+  b1[0x03] = b2[0x03] + b2[0x04];
+  b1[0x04] = MULL((b2[0x03] - b2[0x04]),COS_2_3);
+
+  b1[0x08] = b2[0x08] + b2[0x0F];
+  b1[0x0F] = MUL((b2[0x0F] - b2[0x08]),COS_2_0);
+  b1[0x09] = b2[0x09] + b2[0x0E];
+  b1[0x0E] = MUL((b2[0x0E] - b2[0x09]),COS_2_1);
+  b1[0x0A] = b2[0x0A] + b2[0x0D];
+  b1[0x0D] = MUL((b2[0x0D] - b2[0x0A]),COS_2_2);
+  b1[0x0B] = b2[0x0B] + b2[0x0C];
+  b1[0x0C] = MULL((b2[0x0C] - b2[0x0B]),COS_2_3);
+
+  b1[0x10] = b2[0x10] + b2[0x17];
+  b1[0x17] = MUL((b2[0x10] - b2[0x17]),COS_2_0);
+  b1[0x11] = b2[0x11] + b2[0x16];
+  b1[0x16] = MUL((b2[0x11] - b2[0x16]),COS_2_1);
+  b1[0x12] = b2[0x12] + b2[0x15];
+  b1[0x15] = MUL((b2[0x12] - b2[0x15]),COS_2_2);
+  b1[0x13] = b2[0x13] + b2[0x14];
+  b1[0x14] = MULL((b2[0x13] - b2[0x14]),COS_2_3);
+
+  b1[0x18] = b2[0x18] + b2[0x1F];
+  b1[0x1F] = MUL((b2[0x1F] - b2[0x18]),COS_2_0);
+  b1[0x19] = b2[0x19] + b2[0x1E];
+  b1[0x1E] = MUL((b2[0x1E] - b2[0x19]),COS_2_1);
+  b1[0x1A] = b2[0x1A] + b2[0x1D];
+  b1[0x1D] = MUL((b2[0x1D] - b2[0x1A]),COS_2_2);
+  b1[0x1B] = b2[0x1B] + b2[0x1C];
+  b1[0x1C] = MULL((b2[0x1C] - b2[0x1B]),COS_2_3);
+
+
+  b2[0x00] = b1[0x00] + b1[0x03];
+  b2[0x03] = MUL((b1[0x00] - b1[0x03]),COS_3_0);
+  b2[0x01] = b1[0x01] + b1[0x02];
+  b2[0x02] = MUL((b1[0x01] - b1[0x02]),COS_3_1);
+
+  b2[0x04] = b1[0x04] + b1[0x07];
+  b2[0x07] = MUL((b1[0x07] - b1[0x04]),COS_3_0);
+  b2[0x05] = b1[0x05] + b1[0x06];
+  b2[0x06] = MUL((b1[0x06] - b1[0x05]),COS_3_1);
+
+  b2[0x08] = b1[0x08] + b1[0x0B];
+  b2[0x0B] = MUL((b1[0x08] - b1[0x0B]),COS_3_0);
+  b2[0x09] = b1[0x09] + b1[0x0A];
+  b2[0x0A] = MUL((b1[0x09] - b1[0x0A]),COS_3_1);
+  
+  b2[0x0C] = b1[0x0C] + b1[0x0F];
+  b2[0x0F] = MUL((b1[0x0F] - b1[0x0C]),COS_3_0);
+  b2[0x0D] = b1[0x0D] + b1[0x0E];
+  b2[0x0E] = MUL((b1[0x0E] - b1[0x0D]),COS_3_1);
+
+  b2[0x10] = b1[0x10] + b1[0x13];
+  b2[0x13] = MUL((b1[0x10] - b1[0x13]),COS_3_0);
+  b2[0x11] = b1[0x11] + b1[0x12];
+  b2[0x12] = MUL((b1[0x11] - b1[0x12]),COS_3_1);
+
+  b2[0x14] = b1[0x14] + b1[0x17];
+  b2[0x17] = MUL((b1[0x17] - b1[0x14]),COS_3_0);
+  b2[0x15] = b1[0x15] + b1[0x16];
+  b2[0x16] = MUL((b1[0x16] - b1[0x15]),COS_3_1);
+
+  b2[0x18] = b1[0x18] + b1[0x1B];
+  b2[0x1B] = MUL((b1[0x18] - b1[0x1B]),COS_3_0);
+  b2[0x19] = b1[0x19] + b1[0x1A];
+  b2[0x1A] = MUL((b1[0x19] - b1[0x1A]),COS_3_1);
+
+  b2[0x1C] = b1[0x1C] + b1[0x1F];
+  b2[0x1F] = MUL((b1[0x1F] - b1[0x1C]),COS_3_0);
+  b2[0x1D] = b1[0x1D] + b1[0x1E];
+  b2[0x1E] = MUL((b1[0x1E] - b1[0x1D]),COS_3_1);
+
+  {
+    int i;
+    for(i=0;i<32;i+=4) {
+      b1[i+0x00] = b2[i+0x00] + b2[i+0x01];
+      b1[i+0x01] = MUL((b2[i+0x00] - b2[i+0x01]),COS_4_0);
+      b1[i+0x02] = b2[i+0x02] + b2[i+0x03];
+      b1[i+0x03] = MUL((b2[i+0x03] - b2[i+0x02]),COS_4_0);
+    }
+  }
+
+  b1[0x02] += b1[0x03];
+  b1[0x06] += b1[0x07];
+  b1[0x04] += b1[0x06];
+  b1[0x06] += b1[0x05];
+  b1[0x05] += b1[0x07];
+
+  b1[0x0A] += b1[0x0B];
+  b1[0x0E] += b1[0x0F];
+  b1[0x0C] += b1[0x0E];
+  b1[0x0E] += b1[0x0D];
+  b1[0x0D] += b1[0x0F];
+
+  b1[0x12] += b1[0x13];
+  b1[0x16] += b1[0x17];
+  b1[0x14] += b1[0x16];
+  b1[0x16] += b1[0x15];
+  b1[0x15] += b1[0x17];
+
+  b1[0x1A] += b1[0x1B];
+  b1[0x1E] += b1[0x1F];
+  b1[0x1C] += b1[0x1E];
+  b1[0x1E] += b1[0x1D];
+  b1[0x1D] += b1[0x1F];
+
+ SETOUT(out0,16,b1[0x00]);
+ SETOUT(out0,12,b1[0x04]);
+ SETOUT(out0, 8,b1[0x02]);
+ SETOUT(out0, 4,b1[0x06]);
+ SETOUT(out0, 0,b1[0x01]);
+ SETOUT(out1, 0,b1[0x01]);
+ SETOUT(out1, 4,b1[0x05]);
+ SETOUT(out1, 8,b1[0x03]);
+ SETOUT(out1,12,b1[0x07]);
+
+ b1[0x08] += b1[0x0C];
+ SETOUT(out0,14,b1[0x08]);
+ b1[0x0C] += b1[0x0a];
+ SETOUT(out0,10,b1[0x0C]);
+ b1[0x0A] += b1[0x0E];
+ SETOUT(out0, 6,b1[0x0A]);
+ b1[0x0E] += b1[0x09];
+ SETOUT(out0, 2,b1[0x0E]);
+ b1[0x09] += b1[0x0D];
+ SETOUT(out1, 2,b1[0x09]);
+ b1[0x0D] += b1[0x0B];
+ SETOUT(out1, 6,b1[0x0D]);
+ b1[0x0B] += b1[0x0F];
+ SETOUT(out1,10,b1[0x0B]);
+ SETOUT(out1,14,b1[0x0F]);
+
+ b1[0x18] += b1[0x1C];
+ SETOUT(out0,15,b1[0x10] + b1[0x18]);
+ SETOUT(out0,13,b1[0x18] + b1[0x14]);
+ b1[0x1C] += b1[0x1a];
+ SETOUT(out0,11,b1[0x14] + b1[0x1C]);
+ SETOUT(out0, 9,b1[0x1C] + b1[0x12]);
+ b1[0x1A] += b1[0x1E];
+ SETOUT(out0, 7,b1[0x12] + b1[0x1A]);
+ SETOUT(out0, 5,b1[0x1A] + b1[0x16]);
+ b1[0x1E] += b1[0x19];
+ SETOUT(out0, 3,b1[0x16] + b1[0x1E]);
+ SETOUT(out0, 1,b1[0x1E] + b1[0x11]);
+ b1[0x19] += b1[0x1D];
+ SETOUT(out1, 1,b1[0x11] + b1[0x19]);
+ SETOUT(out1, 3,b1[0x19] + b1[0x15]);
+ b1[0x1D] += b1[0x1B];
+ SETOUT(out1, 5,b1[0x15] + b1[0x1D]);
+ SETOUT(out1, 7,b1[0x1D] + b1[0x13]);
+ b1[0x1B] += b1[0x1F];
+ SETOUT(out1, 9,b1[0x13] + b1[0x1B]);
+ SETOUT(out1,11,b1[0x1B] + b1[0x17]);
+ SETOUT(out1,13,b1[0x17] + b1[0x1F]);
+ SETOUT(out1,15,b1[0x1F]);
+}
+
+
+/*
+ * the call via dct64 is a trick to force GCC to use
+ * (new) registers for the b1,b2 pointer to the bufs[xx] field
+ */
+void dct64_i486(int *a,int *b,real *samples)
+{
+  int bufs[64];
+  int i;
+
+#ifdef REAL_IS_FIXED  
+#define TOINT(a) ((a) * 32768 / (int)REAL_FACTOR)
+
+  for(i=0;i<32;i++) {
+    bufs[i]=TOINT(samples[i]);
+  }
+#else      
+  int *p = bufs;
+  register double const scale = ((65536.0 * 32) + 1) * 65536.0;
+  
+  for(i=0;i<32;i++) {
+    *((double *) (p++)) = scale + *samples++; /* beware on bufs overrun: 8B store from x87 */
+  }
+#endif
+  
+  dct64_1_486(a,b,bufs+32,bufs);
+}
+

data/mpg123/src/libmpg123/dct64_mmx.S

@@ -0,0 +1,811 @@
+/*
+	dct64_mmx.s: MMX optimized DCT64
+
+	copyright ?-2006 by the mpg123 project - free software under the terms of the LGPL 2.1
+	see COPYING and AUTHORS files in distribution or http://mpg123.org
+	initially written by the mysterious higway (apparently)
+*/
+
+#include "mangle.h"
+
+.text
+
+	ALIGN32
+.globl ASM_NAME(dct64_mmx)
+ASM_NAME(dct64_mmx):
+
+	xorl %ecx,%ecx
+.globl ASM_NAME(dct64_MMX)
+ASM_NAME(dct64_MMX):
+	pushl %ebx
+	pushl %esi
+	pushl %edi
+	subl $256,%esp
+	movl 280(%esp),%eax
+	flds     (%eax)
+	leal 128(%esp),%edx
+	fadds 124(%eax)
+	movl 272(%esp),%esi
+	fstps    (%edx)
+	movl 276(%esp),%edi
+	flds    4(%eax)
+	movl ASM_VALUE(costab_mmxsse),%ebx
+	fadds 120(%eax)
+	orl %ecx,%ecx
+	fstps   4(%edx)
+	flds     (%eax)
+	movl %esp,%ecx
+	fsubs 124(%eax)
+	fmuls    (%ebx)
+	fstps 124(%edx)
+	flds    4(%eax)
+	fsubs 120(%eax)
+	fmuls   4(%ebx)
+	fstps 120(%edx)
+	flds    8(%eax)
+	fadds 116(%eax)
+	fstps   8(%edx)
+	flds   12(%eax)
+	fadds 112(%eax)
+	fstps  12(%edx)
+	flds    8(%eax)
+	fsubs 116(%eax)
+	fmuls   8(%ebx)
+	fstps 116(%edx)
+	flds   12(%eax)
+	fsubs 112(%eax)
+	fmuls  12(%ebx)
+	fstps 112(%edx)
+	flds   16(%eax)
+	fadds 108(%eax)
+	fstps  16(%edx)
+	flds   20(%eax)
+	fadds 104(%eax)
+	fstps  20(%edx)
+	flds   16(%eax)
+	fsubs 108(%eax)
+	fmuls  16(%ebx)
+	fstps 108(%edx)
+	flds   20(%eax)
+	fsubs 104(%eax)
+	fmuls  20(%ebx)
+	fstps 104(%edx)
+	flds   24(%eax)
+	fadds 100(%eax)
+	fstps  24(%edx)
+	flds   28(%eax)
+	fadds  96(%eax)
+	fstps  28(%edx)
+	flds   24(%eax)
+	fsubs 100(%eax)
+	fmuls  24(%ebx)
+	fstps 100(%edx)
+	flds   28(%eax)
+	fsubs  96(%eax)
+	fmuls  28(%ebx)
+	fstps  96(%edx)
+	flds   32(%eax)
+	fadds  92(%eax)
+	fstps  32(%edx)
+	flds   36(%eax)
+	fadds  88(%eax)
+	fstps  36(%edx)
+	flds   32(%eax)
+	fsubs  92(%eax)
+	fmuls  32(%ebx)
+	fstps  92(%edx)
+	flds   36(%eax)
+	fsubs  88(%eax)
+	fmuls  36(%ebx)
+	fstps  88(%edx)
+	flds   40(%eax)
+	fadds  84(%eax)
+	fstps  40(%edx)
+	flds   44(%eax)
+	fadds  80(%eax)
+	fstps  44(%edx)
+	flds   40(%eax)
+	fsubs  84(%eax)
+	fmuls  40(%ebx)
+	fstps  84(%edx)
+	flds   44(%eax)
+	fsubs  80(%eax)
+	fmuls  44(%ebx)
+	fstps  80(%edx)
+	flds   48(%eax)
+	fadds  76(%eax)
+	fstps  48(%edx)
+	flds   52(%eax)
+	fadds  72(%eax)
+	fstps  52(%edx)
+	flds   48(%eax)
+	fsubs  76(%eax)
+	fmuls  48(%ebx)
+	fstps  76(%edx)
+	flds   52(%eax)
+	fsubs  72(%eax)
+	fmuls  52(%ebx)
+	fstps  72(%edx)
+	flds   56(%eax)
+	fadds  68(%eax)
+	fstps  56(%edx)
+	flds   60(%eax)
+	fadds  64(%eax)
+	fstps  60(%edx)
+	flds   56(%eax)
+	fsubs  68(%eax)
+	fmuls  56(%ebx)
+	fstps  68(%edx)
+	flds   60(%eax)
+	fsubs  64(%eax)
+	fmuls  60(%ebx)
+	fstps  64(%edx)
+	
+	flds     (%edx)
+	fadds  60(%edx)
+	fstps    (%ecx)
+	flds    4(%edx)
+	fadds  56(%edx)
+	fstps   4(%ecx)
+	flds     (%edx)
+	fsubs  60(%edx)
+	fmuls  64(%ebx)
+	fstps  60(%ecx)
+	flds    4(%edx)
+	fsubs  56(%edx)
+	fmuls  68(%ebx)
+	fstps  56(%ecx)
+	flds    8(%edx)
+	fadds  52(%edx)
+	fstps   8(%ecx)
+	flds   12(%edx)
+	fadds  48(%edx)
+	fstps  12(%ecx)
+	flds    8(%edx)
+	fsubs  52(%edx)
+	fmuls  72(%ebx)
+	fstps  52(%ecx)
+	flds   12(%edx)
+	fsubs  48(%edx)
+	fmuls  76(%ebx)
+	fstps  48(%ecx)
+	flds   16(%edx)
+	fadds  44(%edx)
+	fstps  16(%ecx)
+	flds   20(%edx)
+	fadds  40(%edx)
+	fstps  20(%ecx)
+	flds   16(%edx)
+	fsubs  44(%edx)
+	fmuls  80(%ebx)
+	fstps  44(%ecx)
+	flds   20(%edx)
+	fsubs  40(%edx)
+	fmuls  84(%ebx)
+	fstps  40(%ecx)
+	flds   24(%edx)
+	fadds  36(%edx)
+	fstps  24(%ecx)
+	flds   28(%edx)
+	fadds  32(%edx)
+	fstps  28(%ecx)
+	flds   24(%edx)
+	fsubs  36(%edx)
+	fmuls  88(%ebx)
+	fstps  36(%ecx)
+	flds   28(%edx)
+	fsubs  32(%edx)
+	fmuls  92(%ebx)
+	fstps  32(%ecx)
+	
+	flds   64(%edx)
+	fadds 124(%edx)
+	fstps  64(%ecx)
+	flds   68(%edx)
+	fadds 120(%edx)
+	fstps  68(%ecx)
+	flds  124(%edx)
+	fsubs  64(%edx)
+	fmuls  64(%ebx)
+	fstps 124(%ecx)
+	flds  120(%edx)
+	fsubs  68(%edx)
+	fmuls  68(%ebx)
+	fstps 120(%ecx)
+	flds   72(%edx)
+	fadds 116(%edx)
+	fstps  72(%ecx)
+	flds   76(%edx)
+	fadds 112(%edx)
+	fstps  76(%ecx)
+	flds  116(%edx)
+	fsubs  72(%edx)
+	fmuls  72(%ebx)
+	fstps 116(%ecx)
+	flds  112(%edx)
+	fsubs  76(%edx)
+	fmuls  76(%ebx)
+	fstps 112(%ecx)
+	flds   80(%edx)
+	fadds 108(%edx)
+	fstps  80(%ecx)
+	flds   84(%edx)
+	fadds 104(%edx)
+	fstps  84(%ecx)
+	flds  108(%edx)
+	fsubs  80(%edx)
+	fmuls  80(%ebx)
+	fstps 108(%ecx)
+	flds  104(%edx)
+	fsubs  84(%edx)
+	fmuls  84(%ebx)
+	fstps 104(%ecx)
+	flds   88(%edx)
+	fadds 100(%edx)
+	fstps  88(%ecx)
+	flds   92(%edx)
+	fadds  96(%edx)
+	fstps  92(%ecx)
+	flds  100(%edx)
+	fsubs  88(%edx)
+	fmuls  88(%ebx)
+	fstps 100(%ecx)
+	flds   96(%edx)
+	fsubs  92(%edx)
+	fmuls  92(%ebx)
+	fstps  96(%ecx)
+	
+	flds     (%ecx)
+	fadds  28(%ecx)
+	fstps    (%edx)
+	flds     (%ecx)
+	fsubs  28(%ecx)
+	fmuls  96(%ebx)
+	fstps  28(%edx)
+	flds    4(%ecx)
+	fadds  24(%ecx)
+	fstps   4(%edx)
+	flds    4(%ecx)
+	fsubs  24(%ecx)
+	fmuls 100(%ebx)
+	fstps  24(%edx)
+	flds    8(%ecx)
+	fadds  20(%ecx)
+	fstps   8(%edx)
+	flds    8(%ecx)
+	fsubs  20(%ecx)
+	fmuls 104(%ebx)
+	fstps  20(%edx)
+	flds   12(%ecx)
+	fadds  16(%ecx)
+	fstps  12(%edx)
+	flds   12(%ecx)
+	fsubs  16(%ecx)
+	fmuls 108(%ebx)
+	fstps  16(%edx)
+	flds   32(%ecx)
+	fadds  60(%ecx)
+	fstps  32(%edx)
+	flds   60(%ecx)
+	fsubs  32(%ecx)
+	fmuls  96(%ebx)
+	fstps  60(%edx)
+	flds   36(%ecx)
+	fadds  56(%ecx)
+	fstps  36(%edx)
+	flds   56(%ecx)
+	fsubs  36(%ecx)
+	fmuls 100(%ebx)
+	fstps  56(%edx)
+	flds   40(%ecx)
+	fadds  52(%ecx)
+	fstps  40(%edx)
+	flds   52(%ecx)
+	fsubs  40(%ecx)
+	fmuls 104(%ebx)
+	fstps  52(%edx)
+	flds   44(%ecx)
+	fadds  48(%ecx)
+	fstps  44(%edx)
+	flds   48(%ecx)
+	fsubs  44(%ecx)
+	fmuls 108(%ebx)
+	fstps  48(%edx)
+	flds   64(%ecx)
+	fadds  92(%ecx)
+	fstps  64(%edx)
+	flds   64(%ecx)
+	fsubs  92(%ecx)
+	fmuls  96(%ebx)
+	fstps  92(%edx)
+	flds   68(%ecx)
+	fadds  88(%ecx)
+	fstps  68(%edx)
+	flds   68(%ecx)
+	fsubs  88(%ecx)
+	fmuls 100(%ebx)
+	fstps  88(%edx)
+	flds   72(%ecx)
+	fadds  84(%ecx)
+	fstps  72(%edx)
+	flds   72(%ecx)
+	fsubs  84(%ecx)
+	fmuls 104(%ebx)
+	fstps  84(%edx)
+	flds   76(%ecx)
+	fadds  80(%ecx)
+	fstps  76(%edx)
+	flds   76(%ecx)
+	fsubs  80(%ecx)
+	fmuls 108(%ebx)
+	fstps  80(%edx)
+	flds   96(%ecx)
+	fadds 124(%ecx)
+	fstps  96(%edx)
+	flds  124(%ecx)
+	fsubs  96(%ecx)
+	fmuls  96(%ebx)
+	fstps 124(%edx)
+	flds  100(%ecx)
+	fadds 120(%ecx)
+	fstps 100(%edx)
+	flds  120(%ecx)
+	fsubs 100(%ecx)
+	fmuls 100(%ebx)
+	fstps 120(%edx)
+	flds  104(%ecx)
+	fadds 116(%ecx)
+	fstps 104(%edx)
+	flds  116(%ecx)
+	fsubs 104(%ecx)
+	fmuls 104(%ebx)
+	fstps 116(%edx)
+	flds  108(%ecx)
+	fadds 112(%ecx)
+	fstps 108(%edx)
+	flds  112(%ecx)
+	fsubs 108(%ecx)
+	fmuls 108(%ebx)
+	fstps 112(%edx)
+	flds     (%edx)
+	fadds  12(%edx)
+	fstps    (%ecx)
+	flds     (%edx)
+	fsubs  12(%edx)
+	fmuls 112(%ebx)
+	fstps  12(%ecx)
+	flds    4(%edx)
+	fadds   8(%edx)
+	fstps   4(%ecx)
+	flds    4(%edx)
+	fsubs   8(%edx)
+	fmuls 116(%ebx)
+	fstps   8(%ecx)
+	flds   16(%edx)
+	fadds  28(%edx)
+	fstps  16(%ecx)
+	flds   28(%edx)
+	fsubs  16(%edx)
+	fmuls 112(%ebx)
+	fstps  28(%ecx)
+	flds   20(%edx)
+	fadds  24(%edx)
+	fstps  20(%ecx)
+	flds   24(%edx)
+	fsubs  20(%edx)
+	fmuls 116(%ebx)
+	fstps  24(%ecx)
+	flds   32(%edx)
+	fadds  44(%edx)
+	fstps  32(%ecx)
+	flds   32(%edx)
+	fsubs  44(%edx)
+	fmuls 112(%ebx)
+	fstps  44(%ecx)
+	flds   36(%edx)
+	fadds  40(%edx)
+	fstps  36(%ecx)
+	flds   36(%edx)
+	fsubs  40(%edx)
+	fmuls 116(%ebx)
+	fstps  40(%ecx)
+	flds   48(%edx)
+	fadds  60(%edx)
+	fstps  48(%ecx)
+	flds   60(%edx)
+	fsubs  48(%edx)
+	fmuls 112(%ebx)
+	fstps  60(%ecx)
+	flds   52(%edx)
+	fadds  56(%edx)
+	fstps  52(%ecx)
+	flds   56(%edx)
+	fsubs  52(%edx)
+	fmuls 116(%ebx)
+	fstps  56(%ecx)
+	flds   64(%edx)
+	fadds  76(%edx)
+	fstps  64(%ecx)
+	flds   64(%edx)
+	fsubs  76(%edx)
+	fmuls 112(%ebx)
+	fstps  76(%ecx)
+	flds   68(%edx)
+	fadds  72(%edx)
+	fstps  68(%ecx)
+	flds   68(%edx)
+	fsubs  72(%edx)
+	fmuls 116(%ebx)
+	fstps  72(%ecx)
+	flds   80(%edx)
+	fadds  92(%edx)
+	fstps  80(%ecx)
+	flds   92(%edx)
+	fsubs  80(%edx)
+	fmuls 112(%ebx)
+	fstps  92(%ecx)
+	flds   84(%edx)
+	fadds  88(%edx)
+	fstps  84(%ecx)
+	flds   88(%edx)
+	fsubs  84(%edx)
+	fmuls 116(%ebx)
+	fstps  88(%ecx)
+	flds   96(%edx)
+	fadds 108(%edx)
+	fstps  96(%ecx)
+	flds   96(%edx)
+	fsubs 108(%edx)
+	fmuls 112(%ebx)
+	fstps 108(%ecx)
+	flds  100(%edx)
+	fadds 104(%edx)
+	fstps 100(%ecx)
+	flds  100(%edx)
+	fsubs 104(%edx)
+	fmuls 116(%ebx)
+	fstps 104(%ecx)
+	flds  112(%edx)
+	fadds 124(%edx)
+	fstps 112(%ecx)
+	flds  124(%edx)
+	fsubs 112(%edx)
+	fmuls 112(%ebx)
+	fstps 124(%ecx)
+	flds  116(%edx)
+	fadds 120(%edx)
+	fstps 116(%ecx)
+	flds  120(%edx)
+	fsubs 116(%edx)
+	fmuls 116(%ebx)
+	fstps 120(%ecx)
+	
+	flds   32(%ecx)
+	fadds  36(%ecx)
+	fstps  32(%edx)
+	flds   32(%ecx)
+	fsubs  36(%ecx)
+	fmuls 120(%ebx)
+	fstps  36(%edx)
+	flds   44(%ecx)
+	fsubs  40(%ecx)
+	fmuls 120(%ebx)
+	fsts   44(%edx)
+	fadds  40(%ecx)
+	fadds  44(%ecx)
+	fstps  40(%edx)
+	flds   48(%ecx)
+	fsubs  52(%ecx)
+	fmuls 120(%ebx)
+	flds   60(%ecx)
+	fsubs  56(%ecx)
+	fmuls 120(%ebx)
+	fld      %st(0)
+	fadds  56(%ecx)
+	fadds  60(%ecx)
+	fld      %st(0)
+	fadds  48(%ecx)
+	fadds  52(%ecx)
+	fstps  48(%edx)
+	fadd     %st(2)
+	fstps  56(%edx)
+	fsts   60(%edx)
+	faddp    %st(1)
+	fstps  52(%edx)
+	flds   64(%ecx)
+	fadds  68(%ecx)
+	fstps  64(%edx)
+	flds   64(%ecx)
+	fsubs  68(%ecx)
+	fmuls 120(%ebx)
+	fstps  68(%edx)
+	flds   76(%ecx)
+	fsubs  72(%ecx)
+	fmuls 120(%ebx)
+	fsts   76(%edx)
+	fadds  72(%ecx)
+	fadds  76(%ecx)
+	fstps  72(%edx)
+	flds   92(%ecx)
+	fsubs  88(%ecx)
+	fmuls 120(%ebx)
+	fsts   92(%edx)
+	fadds  92(%ecx)
+	fadds  88(%ecx)
+	fld      %st(0)
+	fadds  80(%ecx)
+	fadds  84(%ecx)
+	fstps  80(%edx)
+	flds   80(%ecx)
+	fsubs  84(%ecx)
+	fmuls 120(%ebx)
+	fadd  %st(0), %st(1)
+	fadds 92(%edx)
+	fstps 84(%edx)
+	fstps 88(%edx)
+	flds   96(%ecx)
+	fadds 100(%ecx)
+	fstps  96(%edx)
+	flds   96(%ecx)
+	fsubs 100(%ecx)
+	fmuls 120(%ebx)
+	fstps 100(%edx)
+	flds  108(%ecx)
+	fsubs 104(%ecx)
+	fmuls 120(%ebx)
+	fsts  108(%edx)
+	fadds 104(%ecx)
+	fadds 108(%ecx)
+	fstps 104(%edx)
+	flds  124(%ecx)
+	fsubs 120(%ecx)
+	fmuls 120(%ebx)
+	fsts  124(%edx)
+	fadds 120(%ecx)
+	fadds 124(%ecx)
+	fld      %st(0)
+	fadds 112(%ecx)
+	fadds 116(%ecx)
+	fstps 112(%edx)
+	flds  112(%ecx)
+	fsubs 116(%ecx)
+	fmuls 120(%ebx)
+	fadd  %st(0),%st(1)
+	fadds 124(%edx)
+	fstps 116(%edx)
+	fstps 120(%edx)
+	jnz .L01
+	
+	flds      (%ecx)
+	fadds    4(%ecx)
+	fstps 1024(%esi)
+	flds      (%ecx)
+	fsubs    4(%ecx)
+	fmuls  120(%ebx)
+	fsts      (%esi)
+	fstps     (%edi)
+	flds   12(%ecx)
+	fsubs   8(%ecx)
+	fmuls 120(%ebx)
+	fsts  512(%edi)
+	fadds  12(%ecx)
+	fadds   8(%ecx)
+	fstps 512(%esi)
+	flds   16(%ecx)
+	fsubs  20(%ecx)
+	fmuls 120(%ebx)
+	flds   28(%ecx)
+	fsubs  24(%ecx)
+	fmuls 120(%ebx)
+	fsts  768(%edi)
+	fld      %st(0)
+	fadds  24(%ecx)
+	fadds  28(%ecx)
+	fld      %st(0)
+	fadds  16(%ecx)
+	fadds  20(%ecx)
+	fstps 768(%esi)
+	fadd     %st(2)
+	fstps 256(%esi)
+	faddp    %st(1)
+	fstps 256(%edi)
+	
+	flds   32(%edx)
+	fadds  48(%edx)
+	fstps 896(%esi)
+	flds   48(%edx)
+	fadds  40(%edx)
+	fstps 640(%esi)
+	flds   40(%edx)
+	fadds  56(%edx)
+	fstps 384(%esi)
+	flds   56(%edx)
+	fadds  36(%edx)
+	fstps 128(%esi)
+	flds   36(%edx)
+	fadds  52(%edx)
+	fstps 128(%edi)
+	flds   52(%edx)
+	fadds  44(%edx)
+	fstps 384(%edi)
+	flds   60(%edx)
+	fsts  896(%edi)
+	fadds  44(%edx)
+	fstps 640(%edi)
+	flds   96(%edx)
+	fadds 112(%edx)
+	fld      %st(0)
+	fadds  64(%edx)
+	fstps 960(%esi)
+	fadds  80(%edx)
+	fstps 832(%esi)
+	flds  112(%edx)
+	fadds 104(%edx)
+	fld      %st(0)
+	fadds  80(%edx)
+	fstps 704(%esi)
+	fadds  72(%edx)
+	fstps 576(%esi)
+	flds  104(%edx)
+	fadds 120(%edx)
+	fld      %st(0)
+	fadds  72(%edx)
+	fstps 448(%esi)
+	fadds  88(%edx)
+	fstps 320(%esi)
+	flds  120(%edx)
+	fadds 100(%edx)
+	fld      %st(0)
+	fadds  88(%edx)
+	fstps 192(%esi)
+	fadds  68(%edx)
+	fstps  64(%esi)
+	flds  100(%edx)
+	fadds 116(%edx)
+	fld      %st(0)
+	fadds  68(%edx)
+	fstps  64(%edi)
+	fadds  84(%edx)
+	fstps 192(%edi)
+	flds  116(%edx)
+	fadds 108(%edx)
+	fld      %st(0)
+	fadds  84(%edx)
+	fstps 320(%edi)
+	fadds  76(%edx)
+	fstps 448(%edi)
+	flds  108(%edx)
+	fadds 124(%edx)
+	fld      %st(0)
+	fadds  76(%edx)
+	fstps 576(%edi)
+	fadds  92(%edx)
+	fstps 704(%edi)
+	flds  124(%edx)
+	fsts  960(%edi)
+	fadds  92(%edx)
+	fstps 832(%edi)
+	addl $256,%esp
+	popl %edi
+	popl %esi
+	popl %ebx
+	ret
+.L01:	
+	flds      (%ecx)
+	fadds    4(%ecx)
+	fistps 512(%esi)
+	flds      (%ecx)
+	fsubs    4(%ecx)
+	fmuls  120(%ebx)
+
+	fistps    (%esi)
+
+	flds    12(%ecx)
+	fsubs    8(%ecx)
+	fmuls  120(%ebx)
+	fists  256(%edi)
+	fadds   12(%ecx)
+	fadds    8(%ecx)
+	fistps 256(%esi)
+	flds   16(%ecx)
+	fsubs  20(%ecx)
+	fmuls 120(%ebx)
+	flds   28(%ecx)
+	fsubs  24(%ecx)
+	fmuls 120(%ebx)
+	fists 384(%edi)
+	fld      %st(0)
+	fadds  24(%ecx)
+	fadds  28(%ecx)
+	fld      %st(0)
+	fadds  16(%ecx)
+	fadds  20(%ecx)
+	fistps 384(%esi)
+	fadd     %st(2)
+	fistps 128(%esi)
+	faddp    %st(1)
+	fistps 128(%edi)
+	
+	flds    32(%edx)
+	fadds   48(%edx)
+	fistps 448(%esi)
+	flds   48(%edx)
+	fadds  40(%edx)
+	fistps 320(%esi)
+	flds   40(%edx)
+	fadds  56(%edx)
+	fistps 192(%esi)
+	flds   56(%edx)
+	fadds  36(%edx)
+	fistps 64(%esi)
+	flds   36(%edx)
+	fadds  52(%edx)
+	fistps 64(%edi)
+	flds   52(%edx)
+	fadds  44(%edx)
+	fistps 192(%edi)
+	flds   60(%edx)
+	fists  448(%edi)
+	fadds  44(%edx)
+	fistps 320(%edi)
+	flds   96(%edx)
+	fadds 112(%edx)
+	fld      %st(0)
+	fadds  64(%edx)
+	fistps 480(%esi)
+	fadds  80(%edx)
+	fistps 416(%esi)
+	flds  112(%edx)
+	fadds 104(%edx)
+	fld      %st(0)
+	fadds  80(%edx)
+	fistps 352(%esi)
+	fadds  72(%edx)
+	fistps 288(%esi)
+	flds  104(%edx)
+	fadds 120(%edx)
+	fld      %st(0)
+	fadds  72(%edx)
+	fistps 224(%esi)
+	fadds  88(%edx)
+	fistps 160(%esi)
+	flds  120(%edx)
+	fadds 100(%edx)
+	fld      %st(0)
+	fadds  88(%edx)
+	fistps 96(%esi)
+	fadds  68(%edx)
+	fistps 32(%esi)
+	flds  100(%edx)
+	fadds 116(%edx)
+	fld      %st(0)
+	fadds  68(%edx)
+	fistps 32(%edi)
+	fadds  84(%edx)
+	fistps 96(%edi)
+	flds  116(%edx)
+	fadds 108(%edx)
+	fld      %st(0)
+	fadds  84(%edx)
+	fistps 160(%edi)
+	fadds  76(%edx)
+	fistps 224(%edi)
+	flds  108(%edx)
+	fadds 124(%edx)
+	fld      %st(0)
+	fadds  76(%edx)
+	fistps 288(%edi)
+	fadds  92(%edx)
+	fistps 352(%edi)
+	flds  124(%edx)
+	fists  480(%edi)
+	fadds  92(%edx)
+	fistps 416(%edi)
+	movsw
+	addl $256,%esp
+	popl %edi
+	popl %esi
+	popl %ebx
+	ret
+
+NONEXEC_STACK