Merge remote branch 'origin/gallium-0.2' into gallium-0.2

10 files changed, 1249 insertions, 831 deletions

diff --git a/src/gallium/auxiliary/gallivm/gallivm_builtins.cpp b/src/gallium/auxiliary/gallivm/gallivm_builtins.cpp
index 0fc5c4ec5ca..fcc5c05794a 100644
--- a/src/gallium/auxiliary/gallivm/gallivm_builtins.cpp
+++ b/src/gallium/auxiliary/gallivm/gallivm_builtins.cpp
@@ -1,140 +1,140 @@
 static const unsigned char llvm_builtins_data[] = {
-0x42,0x43,0xc0,0xde,0x21,0x0c,0x00,0x00,0x29,0x02,0x00,0x00,0x01,0x10,0x00,0x00,
+0x42,0x43,0xc0,0xde,0x21,0x0c,0x00,0x00,0x27,0x02,0x00,0x00,0x01,0x10,0x00,0x00,
 0x10,0x00,0x00,0x00,0x07,0x81,0x23,0x91,0x41,0xc8,0x04,0x49,0x06,0x10,0x32,0x39,
 0x92,0x01,0x84,0x0c,0x25,0x05,0x08,0x19,0x1e,0x04,0x8b,0x62,0x80,0x14,0x45,0x02,
 0x42,0x92,0x0b,0x42,0xa4,0x10,0x32,0x14,0x38,0x08,0x18,0x49,0x0a,0x32,0x44,0x24,
 0x48,0x0a,0x90,0x21,0x23,0x44,0x72,0x80,0x8c,0x14,0x21,0x86,0x0a,0x8a,0x0a,0x64,
-0x0c,0x1f,0x00,0x00,0x49,0x18,0x00,0x00,0x02,0x00,0x00,0x00,0x0b,0x04,0x00,0x0c,
-0x00,0x00,0x00,0x00,0x51,0x20,0x00,0x00,0x12,0x00,0x00,0x00,0x32,0x22,0x48,0x09,
-0x20,0x65,0x82,0x84,0x00,0x26,0x45,0x48,0x05,0x09,0x26,0x45,0xc6,0x05,0x42,0x52,
-0x26,0x08,0xae,0x19,0x80,0x61,0x04,0x02,0x98,0x23,0x00,0x83,0x29,0x80,0x21,0x00,
-0xb2,0x73,0x04,0x01,0x51,0x8a,0xf4,0x08,0x92,0xa4,0x39,0x47,0x80,0x50,0x2b,0x03,
-0x00,0xa0,0x08,0x21,0x5c,0x46,0x2b,0x44,0x08,0x21,0xd4,0x40,0x14,0x01,0x80,0x11,
-0x80,0x22,0x88,0x00,0x13,0xa2,0x74,0xb0,0x03,0x3c,0xb0,0x83,0x36,0x80,0x87,0x71,
-0x68,0x03,0x76,0x48,0x07,0x77,0xa8,0x07,0x7c,0x68,0x83,0x73,0x70,0x87,0x7a,0xd8,
-0x70,0x0f,0xe5,0xd0,0x06,0xf0,0xa0,0x07,0x73,0x20,0x07,0x7a,0x30,0x07,0x72,0xa0,
-0x07,0x73,0x20,0x07,0x6d,0x90,0x0e,0x71,0xa0,0x07,0x78,0xa0,0x07,0x78,0xd0,0x06,
-0xe9,0x80,0x07,0x7a,0x80,0x07,0x7a,0x80,0x07,0x6d,0x90,0x0e,0x71,0x60,0x07,0x7a,
-0x10,0x07,0x76,0xa0,0x07,0x71,0x60,0x07,0x6d,0x90,0x0e,0x73,0x20,0x07,0x7a,0x30,
-0x07,0x72,0xa0,0x07,0x73,0x20,0x07,0x6d,0x90,0x0e,0x76,0x40,0x07,0x7a,0x30,0x07,
-0x72,0xa0,0x07,0x76,0x40,0x07,0x6d,0x60,0x0e,0x73,0x20,0x07,0x7a,0x30,0x07,0x72,
-0xa0,0x07,0x73,0x20,0x07,0x6d,0x60,0x0e,0x76,0x40,0x07,0x7a,0x30,0x07,0x72,0xa0,
-0x07,0x76,0x40,0x07,0x6d,0x60,0x0f,0x76,0x40,0x07,0x7a,0x60,0x07,0x74,0xa0,0x07,
-0x76,0x40,0x07,0x6d,0x60,0x0f,0x71,0x20,0x07,0x78,0xa0,0x07,0x71,0x20,0x07,0x78,
-0xa0,0x07,0x71,0x20,0x07,0x78,0xd0,0x06,0xe1,0x00,0x07,0x7a,0x00,0x07,0x7a,0x60,
-0x07,0x74,0xd0,0x06,0xe6,0x80,0x07,0x70,0xa0,0x07,0x71,0x20,0x07,0x78,0xa0,0x07,
-0x71,0x20,0x07,0x78,0xa0,0xf3,0x40,0x88,0x04,0x32,0x32,0x02,0x04,0x20,0x76,0x46,
-0xfc,0x6c,0x48,0x92,0x00,0x40,0x00,0x00,0x00,0x00,0x0c,0x49,0x12,0x20,0x00,0x00,
-0x00,0x00,0x80,0x21,0x89,0x02,0x00,0x01,0x00,0x00,0x00,0x30,0x24,0x59,0x00,0x20,
-0x08,0x00,0x00,0x00,0x86,0x24,0x0a,0x00,0x04,0x00,0x00,0x00,0xc0,0x90,0x84,0x01,
-0x02,0x00,0x00,0x00,0x00,0x18,0x92,0x1c,0x40,0x00,0x00,0x00,0x00,0x00,0x43,0x12,
-0x05,0x00,0x02,0x00,0x00,0x00,0x60,0x48,0x72,0x00,0x01,0x00,0x00,0x00,0x00,0x0c,
-0x49,0x14,0x00,0x08,0x00,0x00,0x00,0x80,0x21,0x49,0x01,0x00,0x41,0x00,0x00,0x00,
-0x90,0x05,0x02,0x00,0x10,0x00,0x00,0x00,0x32,0x1e,0x98,0x10,0x19,0x11,0x4c,0x90,
+0x0c,0x1f,0x00,0x00,0x49,0x18,0x00,0x00,0x03,0x00,0x00,0x00,0x0b,0x84,0xff,0xff,
+0xff,0xff,0x1f,0xc0,0x00,0x00,0x00,0x00,0x51,0x20,0x00,0x00,0x12,0x00,0x00,0x00,
+0x32,0x22,0x48,0x09,0x20,0x65,0x82,0x84,0x00,0x26,0x45,0x48,0x05,0x09,0x26,0x45,
+0xc6,0x05,0x42,0x52,0x26,0x08,0xae,0x19,0x80,0x61,0x04,0x02,0x98,0x23,0x00,0x83,
+0x29,0x80,0x21,0x00,0xb2,0x73,0x04,0x01,0x51,0x8a,0xf4,0x08,0x92,0xa4,0x39,0x47,
+0x80,0x50,0x2b,0x03,0x00,0xa0,0x08,0x21,0x5c,0x46,0x2b,0x44,0x08,0x21,0xd4,0x40,
+0x14,0x01,0x80,0x11,0x80,0x22,0x88,0x00,0x13,0x30,0x7c,0xc0,0x03,0x3b,0xf8,0x05,
+0x3b,0xa0,0x83,0x36,0xa8,0x07,0x77,0x58,0x07,0x77,0x78,0x87,0x7b,0x70,0x87,0x36,
+0x60,0x87,0x74,0x70,0x87,0x7a,0xc0,0x87,0x36,0x38,0x07,0x77,0xa8,0x87,0x0d,0xf7,
+0x50,0x0e,0x6d,0x00,0x0f,0x7a,0x60,0x07,0x74,0xa0,0x07,0x76,0x40,0x07,0x7a,0x60,
+0x07,0x74,0xd0,0x06,0xe9,0x10,0x07,0x7a,0x80,0x07,0x7a,0x80,0x07,0x6d,0x90,0x0e,
+0x78,0xa0,0x07,0x78,0xa0,0x07,0x78,0xd0,0x06,0xe9,0x10,0x07,0x76,0xa0,0x07,0x71,
+0x60,0x07,0x7a,0x10,0x07,0x76,0xd0,0x06,0xe9,0x30,0x07,0x72,0xa0,0x07,0x73,0x20,
+0x07,0x7a,0x30,0x07,0x72,0xd0,0x06,0xe9,0x60,0x07,0x74,0xa0,0x07,0x76,0x40,0x07,
+0x7a,0x60,0x07,0x74,0xd0,0x06,0xe6,0x30,0x07,0x72,0xa0,0x07,0x73,0x20,0x07,0x7a,
+0x30,0x07,0x72,0xd0,0x06,0xe6,0x60,0x07,0x74,0xa0,0x07,0x76,0x40,0x07,0x7a,0x60,
+0x07,0x74,0xd0,0x06,0xf6,0x60,0x07,0x74,0xa0,0x07,0x76,0x40,0x07,0x7a,0x60,0x07,
+0x74,0xd0,0x06,0xf6,0x10,0x07,0x72,0x80,0x07,0x7a,0x10,0x07,0x72,0x80,0x07,0x7a,
+0x10,0x07,0x72,0x80,0x07,0x6d,0x10,0x0e,0x70,0xa0,0x07,0x70,0xa0,0x07,0x76,0x40,
+0x07,0x6d,0x60,0x0e,0x78,0x00,0x07,0x7a,0x10,0x07,0x72,0x80,0x07,0x7a,0x10,0x07,
+0x72,0x80,0x07,0x3a,0x0f,0x84,0x48,0x20,0x23,0x24,0x40,0x00,0x62,0x67,0x88,0x9f,
+0x19,0x92,0x24,0x00,0x10,0x04,0x00,0x00,0x00,0x43,0x92,0x04,0x08,0x00,0x00,0x00,
+0x00,0x60,0x48,0xa2,0x00,0x40,0x10,0x00,0x00,0x00,0x0c,0x49,0x16,0x00,0x08,0x02,
+0x00,0x00,0x80,0x21,0x89,0x02,0x00,0x41,0x00,0x00,0x00,0x30,0x24,0x61,0x80,0x00,
+0x00,0x00,0x00,0x00,0x86,0x24,0x07,0x10,0x00,0x00,0x00,0x00,0xc0,0x90,0x44,0x01,
+0x80,0x20,0x00,0x00,0x00,0x18,0x92,0x1c,0x40,0x00,0x00,0x00,0x00,0x00,0x43,0x12,
+0x05,0x00,0x82,0x00,0x00,0x00,0x60,0x48,0x52,0x00,0x40,0x10,0x00,0x00,0x00,0x64,
+0x81,0x00,0x00,0x00,0x10,0x00,0x00,0x00,0x32,0x1e,0x98,0x10,0x19,0x11,0x4c,0x90,
 0x8c,0x09,0x26,0x47,0xc6,0x04,0x43,0x8a,0x8a,0x59,0x8b,0x43,0x50,0xd2,0x09,0x02,
 0x81,0xd2,0x73,0x50,0xc9,0x0c,0x2a,0x99,0x41,0x25,0x33,0xa8,0x64,0x56,0x28,0x66,
 0x2d,0x0e,0x41,0xcf,0x2a,0x15,0x04,0x4a,0xcf,0x41,0x25,0x33,0xa8,0x64,0x06,0x95,
 0xcc,0xa0,0x92,0x59,0x01,0x00,0x00,0x00,0x53,0x82,0x26,0x0c,0x04,0x00,0x00,0x00,
 0x22,0x00,0x00,0x00,0x13,0x04,0x41,0x2c,0x10,0x00,0x00,0x00,0x05,0x00,0x00,0x00,
 0x04,0xc6,0x08,0x40,0x10,0x04,0xe1,0x70,0x18,0x23,0x00,0x41,0x10,0x84,0xc3,0x60,
-0x04,0x00,0x00,0x00,0x93,0x0c,0xce,0x43,0x4c,0x31,0x3c,0x8e,0x34,0xc9,0x30,0x41,
-0xc2,0x14,0x03,0x34,0x51,0x93,0x0c,0x4d,0x44,0x4c,0x31,0x44,0x8d,0x35,0x56,0x01,
-0x04,0xc3,0x55,0x21,0x16,0x0e,0x04,0x00,0x0f,0x00,0x00,0x00,0x46,0x41,0x08,0xcc,
-0x73,0x9b,0x05,0x21,0x30,0xcf,0x6e,0x18,0x84,0x00,0x2c,0x8b,0x35,0x04,0x80,0x39,
-0x04,0x81,0x5d,0x20,0x80,0x0f,0x0c,0x43,0xe4,0xd3,0x36,0x81,0x04,0x3e,0x30,0x0c,
-0x91,0x4f,0x5b,0x05,0x12,0xf8,0xc0,0x30,0x44,0x7e,0x7d,0x00,0x05,0xd1,0x4c,0x11,
-0x66,0x12,0x83,0xc0,0x3c,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x61,0x20,0x00,0x00,
-0x2a,0x00,0x00,0x00,0x13,0x04,0x43,0x2c,0x10,0x00,0x00,0x00,0x08,0x00,0x00,0x00,
+0x04,0x00,0x00,0x00,0xc3,0x0d,0xce,0x43,0x4c,0x37,0x3c,0x8e,0x34,0xdc,0x30,0x41,
+0xc2,0x74,0x03,0x34,0x51,0xc3,0x0d,0x4d,0x44,0x4c,0x37,0x44,0x8d,0x35,0x56,0x01,
+0x04,0xc3,0x55,0x21,0x16,0x0e,0x04,0x00,0x0f,0x00,0x00,0x00,0xd6,0x10,0x00,0xe6,
+0x10,0x04,0x76,0x81,0x00,0x3e,0x30,0x0c,0x91,0x4f,0x1b,0x05,0x21,0x30,0x8f,0x6d,
+0x13,0x48,0xe0,0x03,0xc3,0x10,0xf9,0xb4,0x55,0x20,0x81,0x0f,0x0c,0x43,0xe4,0xd7,
+0x66,0x41,0x08,0xcc,0xa3,0x1f,0x40,0x41,0x34,0x53,0x84,0x99,0xc4,0x20,0x30,0x8f,
+0x61,0x10,0x02,0xb0,0x2c,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x61,0x20,0x00,0x00,
+0x27,0x00,0x00,0x00,0x13,0x04,0x43,0x2c,0x10,0x00,0x00,0x00,0x08,0x00,0x00,0x00,
 0x24,0x8a,0xa0,0x0c,0x46,0x00,0x4a,0x80,0xc2,0x1c,0x84,0x55,0x55,0xd6,0x1c,0x84,
 0x45,0x51,0x16,0x81,0x19,0x80,0x11,0x80,0x31,0x02,0x10,0x04,0x41,0xfc,0x03,0x00,
-0x63,0x08,0x0d,0x34,0xc9,0x70,0x55,0xc2,0x2c,0x43,0x20,0x60,0x73,0x0c,0xd3,0x15,
+0x63,0x08,0x0d,0x34,0xdc,0x70,0x55,0xc2,0x2c,0x43,0x20,0x60,0x73,0x0c,0xd3,0x15,
 0x8d,0x21,0x34,0xd1,0x18,0x42,0xf3,0x8c,0x55,0x00,0x81,0xa0,0x6d,0x73,0x0c,0x19,
-0xe7,0x60,0x87,0x52,0x38,0x10,0x00,0x00,0x13,0x00,0x00,0x00,0x17,0x60,0x20,0xc5,
-0x74,0x10,0x8d,0x65,0x14,0x13,0xf3,0xd4,0xb4,0x6d,0x14,0x13,0xf3,0xd4,0xb8,0x69,
-0x14,0x13,0xf3,0xd4,0xb6,0x75,0x14,0x13,0xf3,0xd4,0xba,0x35,0x0c,0x13,0xf3,0x9c,
-0x80,0xe4,0x36,0x48,0x81,0x10,0xc3,0x4a,0x4c,0x54,0xd4,0x6c,0x8b,0x23,0x28,0x76,
-0x41,0x4c,0xcc,0xa3,0x1b,0x07,0x21,0x00,0xcb,0x72,0x00,0x05,0xd1,0x4c,0x11,0x66,
-0x18,0x83,0xc0,0x3c,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x61,0x20,0x00,0x00,
-0x81,0x00,0x00,0x00,0x13,0x04,0x4d,0x2c,0x10,0x00,0x00,0x00,0x04,0x00,0x00,0x00,
-0x24,0xca,0x60,0x04,0xa0,0x04,0x8a,0x80,0xc2,0x0c,0x00,0x91,0x11,0x00,0x00,0x00,
-0x63,0x08,0x4d,0x64,0x16,0xc1,0x49,0x86,0xab,0x22,0x66,0x19,0x02,0x01,0x1b,0x43,
-0x70,0xa2,0x59,0x82,0x61,0x0c,0xe1,0x89,0x66,0x09,0x86,0x81,0x0a,0x20,0x0b,0x34,
-0x61,0x8e,0x81,0xda,0xa2,0x31,0x84,0x46,0xb2,0x8e,0xe0,0x24,0x83,0x57,0x11,0xb3,
-0x0c,0x44,0xf1,0x8d,0x21,0x38,0xd2,0x2c,0x81,0x31,0x86,0xf0,0x48,0xb3,0x04,0xc6,
-0x40,0x05,0x00,0x06,0x44,0x18,0x14,0x73,0x0c,0x9c,0x18,0x48,0x63,0x08,0xcd,0x64,
-0x64,0x40,0x70,0x92,0xa1,0x0c,0x2a,0x62,0x96,0xe1,0x40,0xcc,0x60,0x0c,0xc1,0x99,
-0x66,0x09,0x92,0x31,0x84,0x67,0x9a,0x25,0x48,0x06,0x2a,0x80,0x33,0x38,0xd0,0x00,
-0x99,0x63,0x18,0x83,0x34,0x98,0xc6,0x10,0x1a,0xc8,0xd6,0x80,0xe0,0x24,0x03,0x1b,
-0x54,0xc4,0x2c,0x83,0xb2,0xb4,0xc1,0x18,0x82,0x03,0xcd,0x12,0x30,0x63,0x08,0x0f,
-0x34,0x4b,0xc0,0x0c,0x54,0x00,0x6e,0xa0,0xbc,0xc1,0x32,0xc7,0xa0,0x06,0x70,0x00,
-0x61,0x1c,0x84,0x03,0x01,0x00,0x00,0x00,0x4e,0x00,0x00,0x00,0x76,0x52,0x4c,0xcc,
-0x73,0xd3,0x24,0x05,0x64,0xec,0xcd,0x8d,0xcc,0xe5,0x87,0x46,0xc6,0x50,0x8a,0x89,
-0x79,0xee,0xdb,0x54,0x8a,0x89,0x79,0xee,0xdd,0x1a,0x88,0x89,0x79,0x68,0x73,0x20,
-0x26,0xe6,0xa9,0xed,0x81,0x98,0x98,0xc7,0x36,0x0b,0x62,0x62,0x9e,0xdb,0x32,0x88,
-0x89,0x79,0x72,0xd3,0x20,0x26,0xe6,0xd9,0x8d,0x83,0x98,0x98,0xa7,0xb7,0x95,0x62,
-0x62,0x9e,0xbb,0x27,0x2d,0x20,0x63,0x6f,0x6e,0x64,0x2e,0x3a,0x34,0x35,0x56,0x62,
-0x08,0x4e,0x53,0xd9,0xba,0xb5,0x14,0x02,0xf3,0xe0,0xf5,0x25,0x2c,0x82,0xd3,0x0c,
-0xbe,0xe0,0x34,0xd3,0x8d,0x9b,0x88,0x21,0x38,0xcd,0x60,0xd7,0x24,0x01,0x63,0xec,
-0xcd,0x8d,0xcc,0x45,0x87,0x44,0x80,0x8c,0xbd,0xb9,0x91,0xb9,0xfc,0xc4,0xd0,0x90,
-0x02,0x8c,0xb1,0x37,0x37,0x32,0x97,0x1f,0x73,0x29,0x26,0xe6,0xc1,0x71,0x7b,0x29,
-0x26,0xe6,0xc1,0x77,0xfb,0x28,0x04,0xe6,0xa9,0x6f,0x52,0x01,0x32,0xf6,0xe6,0x46,
-0xe6,0xa2,0x13,0x73,0x63,0x18,0x83,0xc0,0x3c,0xb6,0x41,0x08,0x4e,0x33,0x58,0x47,
-0x31,0x31,0x4f,0x5d,0x1f,0xc3,0x22,0x38,0xcd,0xe0,0x0b,0x4e,0x33,0xe1,0xbc,0xa5,
-0x18,0x82,0xd3,0x0c,0x77,0x6e,0x20,0xc5,0xc4,0x3c,0xb5,0x4e,0x3a,0x40,0xc6,0xde,
-0xdc,0xc8,0x5c,0x7e,0x64,0x70,0x2c,0xa4,0x98,0x98,0xa7,0xee,0xed,0x82,0x10,0x9c,
-0xa6,0xba,0x81,0x44,0x70,0x9a,0xc1,0x17,0x9c,0x66,0x32,0x93,0x42,0x60,0x1e,0x7b,
-0xb7,0x98,0x62,0x62,0x9e,0xbc,0x36,0x16,0x43,0x70,0x9a,0x0a,0xa7,0x6d,0xa4,0x98,
-0x98,0xc7,0xbe,0x8d,0xa4,0x98,0x98,0xc7,0xce,0x0d,0xc6,0x10,0x9c,0x66,0xc0,0x7b,
-0x12,0x02,0x32,0xf6,0xe6,0x46,0xe6,0xa2,0x33,0x13,0x73,0x06,0x8b,0xe0,0x34,0x83,
-0x2f,0x38,0xcd,0x64,0xd3,0xe6,0x61,0x08,0x4e,0x53,0xd5,0xf6,0x01,0x14,0x44,0x33,
-0x45,0x18,0x00,0x00,0x00,0x00,0x00,0x00,0x61,0x20,0x00,0x00,0x4a,0x00,0x00,0x00,
-0x13,0x04,0x41,0x2c,0x10,0x00,0x00,0x00,0x07,0x00,0x00,0x00,0x24,0xca,0x60,0x04,
-0xa0,0x04,0x8a,0x80,0xc2,0x0c,0x00,0xb9,0x61,0x0c,0x04,0x10,0x1e,0xe1,0x19,0xc6,
-0x40,0x02,0xe1,0x11,0x1e,0x00,0x00,0x00,0x63,0x08,0xcd,0x63,0x15,0xc1,0x31,0x84,
-0x06,0xb2,0x8b,0xe0,0x18,0x42,0x13,0x59,0x46,0x70,0x0c,0xa1,0x71,0x6c,0x23,0x38,
-0x16,0x02,0x04,0xc7,0x64,0x61,0x1a,0x37,0x16,0x01,0x04,0x48,0x35,0xc7,0x20,0x79,
-0xcf,0x58,0x04,0x10,0x20,0xd5,0x1c,0xc3,0x07,0x06,0xd0,0x58,0x04,0x10,0x20,0xd5,
-0x1c,0x43,0x18,0x88,0x41,0x34,0x16,0x01,0x04,0x48,0x35,0xc7,0x30,0x06,0x64,0xe0,
-0x98,0x47,0xd0,0xc0,0x80,0xa0,0x89,0x01,0x41,0x23,0x03,0x82,0x63,0x21,0x40,0x70,
-0x50,0x66,0x70,0x06,0x68,0x90,0x06,0x58,0x06,0xe1,0x40,0x00,0x25,0x00,0x00,0x00,
-0x56,0x52,0x4c,0xcc,0x73,0xd3,0x56,0x41,0x4c,0xcc,0x53,0xdb,0x05,0x31,0x31,0xcf,
-0x6d,0x19,0xc4,0xc4,0x3c,0xba,0x6d,0x10,0x13,0xf3,0xf4,0xd6,0x41,0x08,0xc0,0xb2,
-0x18,0x46,0x21,0x38,0x4d,0x85,0x9b,0x46,0x21,0x38,0x4d,0xb5,0x9b,0x8a,0x21,0x00,
-0xcb,0x82,0xdf,0x66,0x62,0x08,0x4e,0x53,0xdd,0xb7,0x9d,0x18,0x82,0xd3,0x54,0xb7,
-0x6e,0x28,0x86,0xe0,0x34,0xd5,0xdd,0xdb,0x47,0x31,0x31,0x4f,0x9d,0x9b,0x87,0x21,
-0x00,0xcb,0x52,0xdf,0x06,0x62,0x08,0xc0,0xb2,0xd4,0xbc,0x59,0x10,0x82,0xd3,0x54,
-0x96,0x62,0x08,0x4e,0x53,0xe1,0xb6,0x85,0x14,0x13,0xf3,0xd8,0xb4,0x8d,0x14,0x13,
-0xf3,0xd8,0xb9,0x89,0x18,0x02,0xb0,0x2c,0xf6,0x6d,0x24,0x86,0x00,0x2c,0x8b,0xcd,
-0x1b,0x87,0x21,0x38,0x4d,0x55,0xd3,0xd6,0x30,0x54,0xc0,0x72,0x00,0x05,0xd1,0x4c,
-0x11,0x06,0x00,0x00,0x00,0x00,0x00,0x00,0x61,0x20,0x00,0x00,0x19,0x00,0x00,0x00,
-0x13,0x04,0x41,0x2c,0x10,0x00,0x00,0x00,0x03,0x00,0x00,0x00,0x24,0x4a,0x60,0x04,
-0x80,0xc2,0x0c,0x00,0x00,0x00,0x00,0x00,0x63,0x08,0xcd,0x33,0x16,0x01,0x04,0x48,
-0x34,0xc7,0x00,0x49,0xcf,0x58,0x04,0x10,0x28,0xd1,0x1c,0xc3,0x44,0x39,0x58,0x85,
-0x03,0x01,0x00,0x00,0x0a,0x00,0x00,0x00,0x16,0x41,0x4c,0xcc,0x63,0xdb,0x04,0x31,
-0x31,0x4f,0x6e,0x0d,0x43,0x05,0x2c,0x07,0x50,0x10,0xcd,0x14,0x61,0x56,0x41,0x4c,
-0xcc,0xd3,0x1b,0x45,0x21,0x00,0xcb,0xb2,0x9b,0x04,0x21,0x00,0xcb,0x02,0x00,0x00,
-0x00,0x00,0x00,0x00,0x61,0x20,0x00,0x00,0x1b,0x00,0x00,0x00,0x13,0x04,0x41,0x2c,
-0x10,0x00,0x00,0x00,0x03,0x00,0x00,0x00,0x24,0xca,0x60,0x04,0xa0,0x04,0x8a,0x80,
-0xc2,0x0c,0x00,0x00,0x63,0x08,0xcd,0x33,0x16,0x01,0x04,0xca,0x34,0xc7,0x20,0x51,
-0xcf,0x1c,0x43,0x45,0x41,0x73,0x0c,0x16,0x15,0xcd,0x31,0x5c,0x94,0x83,0x58,0x38,
-0x10,0x00,0x00,0x00,0x0b,0x00,0x00,0x00,0x86,0x51,0x4c,0xcc,0x53,0xe7,0x76,0x51,
-0x4c,0xcc,0x53,0xdb,0x36,0x41,0x4c,0xcc,0x63,0x5b,0x05,0x31,0x31,0x8f,0x6e,0x0d,
-0x43,0x05,0x2c,0x66,0x41,0x4c,0xcc,0xd3,0x1f,0x40,0x41,0x34,0x53,0x84,0x19,0x05,
-0x21,0x00,0xcb,0x02,0x00,0x00,0x00,0x00,0x61,0x20,0x00,0x00,0x2f,0x00,0x00,0x00,
-0x13,0x04,0x45,0x2c,0x10,0x00,0x00,0x00,0x03,0x00,0x00,0x00,0x24,0xca,0xa0,0x04,
-0x46,0x00,0x8a,0x80,0xc0,0x08,0x00,0x00,0x63,0x08,0x0d,0x34,0xc9,0x30,0x49,0xc4,
-0x2c,0x03,0x11,0x50,0x63,0x08,0xcd,0x33,0xc9,0x50,0x49,0xc4,0x2c,0x03,0x21,0x58,
-0x63,0x08,0x4d,0x34,0xc9,0x70,0x49,0xc4,0x2c,0x03,0x31,0x60,0x63,0x08,0x8d,0x33,
-0xc9,0x90,0x49,0x84,0x69,0x22,0x70,0xc3,0x27,0x1c,0x08,0x00,0x1a,0x00,0x00,0x00,
-0x96,0x51,0x4c,0xcc,0x53,0xdf,0x66,0x41,0x08,0xcc,0x83,0xdb,0x04,0x31,0x31,0x4f,
-0x6d,0x15,0xc4,0xc4,0x3c,0xb7,0x61,0x10,0x02,0xf3,0xf0,0x47,0x20,0xb9,0x0d,0x52,
-0x20,0xc4,0xb0,0x12,0x13,0x15,0x35,0xdb,0xe2,0x08,0x8a,0x5d,0x10,0x13,0xf3,0xec,
-0x37,0x90,0x2c,0x4e,0xf4,0x47,0x87,0x54,0xd7,0x17,0x70,0x2c,0x4e,0xf4,0x47,0x87,
-0x74,0x02,0xc8,0xe2,0x44,0x7f,0x74,0x48,0xb9,0x69,0x14,0x02,0xf3,0xd4,0xb8,0x6d,
-0x18,0x11,0x31,0x55,0xc0,0x62,0x0d,0x43,0x05,0x2c,0x07,0x50,0x10,0xcd,0x14,0x61,
-0x46,0x31,0x08,0xcc,0x03,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x71,0x20,0x00,0x00,
-0x12,0x00,0x00,0x00,0x66,0x40,0x54,0x82,0x23,0x59,0xc2,0x20,0x09,0x92,0x1d,0x18,
-0x4f,0x84,0x34,0x53,0x61,0x03,0xc4,0xe3,0x58,0x85,0x05,0x14,0xbe,0x34,0x45,0xb5,
-0x21,0x10,0x82,0x23,0x15,0x46,0x30,0x2c,0xc8,0x64,0x02,0x06,0xf0,0x3c,0x91,0x73,
-0x19,0x00,0xe1,0x4b,0x53,0x64,0x0a,0x84,0x84,0x34,0x85,0x31,0x10,0x0a,0xb2,0x3c,
-0x56,0x30,0x08,0xcc,0x63,0x0b,0x44,0x25,0x21,0x0d,0x00,0x00,0x00,0x00,0x00,0x00};
+0xe7,0x60,0x87,0x52,0x38,0x10,0x00,0x00,0x10,0x00,0x00,0x00,0x27,0x50,0x20,0x05,
+0xd1,0x0c,0x17,0x60,0x20,0xc5,0x74,0x10,0x8d,0x65,0x14,0x13,0xf3,0xd4,0xb4,0x6d,
+0x14,0x13,0xf3,0xd4,0xb8,0x69,0x14,0x13,0xf3,0xd4,0xb6,0x75,0x14,0x13,0xf3,0xd4,
+0xba,0x35,0x0c,0x13,0xf3,0xd8,0x05,0x31,0x31,0x8f,0x6e,0x1c,0x84,0x00,0x2c,0xcb,
+0x01,0x14,0x44,0x33,0x45,0x98,0x61,0x0c,0x02,0xf3,0x00,0x00,0x00,0x00,0x00,0x00,
+0x61,0x20,0x00,0x00,0x81,0x00,0x00,0x00,0x13,0x04,0x4d,0x2c,0x10,0x00,0x00,0x00,
+0x04,0x00,0x00,0x00,0x24,0xca,0x60,0x04,0xa0,0x04,0x8a,0x80,0xc2,0x0c,0x00,0x91,
+0x11,0x00,0x00,0x00,0x63,0x08,0x4d,0x64,0x16,0xc1,0xe1,0x86,0xab,0x22,0x66,0x19,
+0x02,0x01,0x1b,0x43,0x70,0xa2,0x59,0x82,0x61,0x0c,0xe1,0x89,0x66,0x09,0x86,0x81,
+0x0a,0x20,0x0b,0x34,0x61,0x8e,0x81,0xda,0xa2,0x31,0x84,0x46,0xb2,0x8e,0xe0,0x70,
+0x83,0x57,0x11,0xb3,0x0c,0x44,0xf1,0x8d,0x21,0x38,0xd2,0x2c,0x81,0x31,0x86,0xf0,
+0x48,0xb3,0x04,0xc6,0x40,0x05,0x00,0x06,0x44,0x18,0x14,0x73,0x0c,0x9c,0x18,0x48,
+0x63,0x08,0xcd,0x64,0x64,0x40,0x70,0xb8,0xa1,0x0c,0x2a,0x62,0x96,0xe1,0x40,0xcc,
+0x60,0x0c,0xc1,0x99,0x66,0x09,0x92,0x31,0x84,0x67,0x9a,0x25,0x48,0x06,0x2a,0x80,
+0x33,0x38,0xd0,0x00,0x99,0x63,0x18,0x83,0x34,0x98,0xc6,0x10,0x1a,0xc8,0xd6,0x80,
+0xe0,0x70,0x03,0x1b,0x54,0xc4,0x2c,0x83,0xb2,0xb4,0xc1,0x18,0x82,0x03,0xcd,0x12,
+0x30,0x63,0x08,0x0f,0x34,0x4b,0xc0,0x0c,0x54,0x00,0x6e,0xa0,0xbc,0xc1,0x32,0xc7,
+0xa0,0x06,0x70,0x00,0x61,0x1c,0x84,0x03,0x01,0x00,0x00,0x00,0x4e,0x00,0x00,0x00,
+0x76,0x52,0x4c,0xcc,0x73,0xd3,0x24,0x05,0x64,0xec,0xcd,0x8d,0xcc,0xe5,0x87,0x46,
+0xc6,0x50,0x8a,0x89,0x79,0xee,0xdb,0x54,0x8a,0x89,0x79,0xee,0xdd,0x1a,0x88,0x89,
+0x79,0x68,0x73,0x20,0x26,0xe6,0xa9,0xed,0x81,0x98,0x98,0xc7,0x36,0x0b,0x62,0x62,
+0x9e,0xdb,0x32,0x88,0x89,0x79,0x72,0xd3,0x20,0x26,0xe6,0xd9,0x8d,0x83,0x98,0x98,
+0xa7,0xb7,0x95,0x62,0x62,0x9e,0xbb,0x27,0x2d,0x20,0x63,0x6f,0x6e,0x64,0x2e,0x3a,
+0x34,0x35,0x56,0x62,0x08,0x4e,0x53,0xd9,0xba,0xb5,0x14,0x02,0xf3,0xe0,0xf5,0x25,
+0x2c,0x82,0xd3,0x0c,0xbe,0xe0,0x34,0xd3,0x8d,0x9b,0x88,0x21,0x38,0xcd,0x60,0xd7,
+0x24,0x01,0x63,0xec,0xcd,0x8d,0xcc,0x45,0x87,0x44,0x80,0x8c,0xbd,0xb9,0x91,0xb9,
+0xfc,0xc4,0xd0,0x90,0x02,0x8c,0xb1,0x37,0x37,0x32,0x97,0x1f,0x73,0x29,0x26,0xe6,
+0xc1,0x71,0x7b,0x29,0x26,0xe6,0xc1,0x77,0xfb,0x28,0x04,0xe6,0xa9,0x6f,0x52,0x01,
+0x32,0xf6,0xe6,0x46,0xe6,0xa2,0x13,0x73,0x63,0x18,0x83,0xc0,0x3c,0xb6,0x41,0x08,
+0x4e,0x33,0x58,0x47,0x31,0x31,0x4f,0x5d,0x1f,0xc3,0x22,0x38,0xcd,0xe0,0x0b,0x4e,
+0x33,0xe1,0xbc,0xa5,0x18,0x82,0xd3,0x0c,0x77,0x6e,0x20,0xc5,0xc4,0x3c,0xb5,0x4e,
+0x3a,0x40,0xc6,0xde,0xdc,0xc8,0x5c,0x7e,0x64,0x70,0x2c,0xa4,0x98,0x98,0xa7,0xee,
+0x6f,0x20,0x11,0x9c,0x66,0xf0,0x05,0xa7,0x99,0xec,0x82,0x10,0x9c,0xa6,0x32,0x93,
+0x42,0x60,0x1e,0x7b,0xb7,0x98,0x62,0x62,0x9e,0xbc,0x36,0x16,0x43,0x70,0x9a,0x0a,
+0xa7,0x6d,0xa4,0x98,0x98,0xc7,0xbe,0x8d,0xa4,0x98,0x98,0xc7,0xce,0x0d,0xc6,0x10,
+0x9c,0x66,0xc0,0x7b,0x12,0x02,0x32,0xf6,0xe6,0x46,0xe6,0xa2,0x33,0x13,0x73,0x06,
+0x8b,0xe0,0x34,0x83,0x2f,0x38,0xcd,0x64,0xd3,0x07,0x50,0x10,0xcd,0x14,0x61,0xe6,
+0x61,0x08,0x4e,0x53,0xd5,0x36,0x00,0x00,0x00,0x00,0x00,0x00,0x61,0x20,0x00,0x00,
+0x4a,0x00,0x00,0x00,0x13,0x04,0x41,0x2c,0x10,0x00,0x00,0x00,0x07,0x00,0x00,0x00,
+0x24,0xca,0x60,0x04,0xa0,0x04,0x8a,0x80,0xc2,0x0c,0x00,0xb9,0x61,0x0c,0x04,0x10,
+0x1e,0xe1,0x19,0xc6,0x40,0x02,0xe1,0x11,0x1e,0x00,0x00,0x00,0x63,0x08,0xcd,0x63,
+0x15,0xc1,0x31,0x84,0x06,0xb2,0x8b,0xe0,0x18,0x42,0x13,0x59,0x46,0x70,0x0c,0xa1,
+0x71,0x6c,0x23,0x38,0x16,0x02,0x04,0xc7,0x64,0x61,0x1a,0x37,0x16,0x01,0x04,0x48,
+0x35,0xc7,0x20,0x79,0xcf,0x58,0x04,0x10,0x20,0xd5,0x1c,0xc3,0x07,0x06,0xd0,0x58,
+0x04,0x10,0x20,0xd5,0x1c,0x43,0x18,0x88,0x41,0x34,0x16,0x01,0x04,0x48,0x35,0xc7,
+0x30,0x06,0x64,0xe0,0x98,0x47,0xd0,0xc0,0x80,0xa0,0x89,0x01,0x41,0x23,0x03,0x82,
+0x63,0x21,0x40,0x70,0x50,0x66,0x70,0x06,0x68,0x90,0x06,0x58,0x06,0xe1,0x40,0x00,
+0x25,0x00,0x00,0x00,0x56,0x52,0x4c,0xcc,0x73,0xd3,0x56,0x41,0x4c,0xcc,0x53,0xdb,
+0x05,0x31,0x31,0xcf,0x6d,0x19,0xc4,0xc4,0x3c,0xba,0x6d,0x10,0x13,0xf3,0xf4,0xd6,
+0x41,0x08,0xc0,0xb2,0x18,0x46,0x21,0x38,0x4d,0x85,0x9b,0x46,0x21,0x38,0x4d,0xb5,
+0x9b,0x8a,0x21,0x00,0xcb,0x82,0xdf,0x66,0x62,0x08,0x4e,0x53,0xdd,0xb7,0x9d,0x18,
+0x82,0xd3,0x54,0xb7,0x6e,0x28,0x86,0xe0,0x34,0xd5,0xdd,0xdb,0x47,0x31,0x31,0x4f,
+0x9d,0x9b,0x87,0x21,0x00,0xcb,0x52,0xdf,0x06,0x62,0x08,0xc0,0xb2,0xd4,0xbc,0x59,
+0x10,0x82,0xd3,0x54,0x96,0x62,0x08,0x4e,0x53,0xe1,0xb6,0x85,0x14,0x13,0xf3,0xd8,
+0xb4,0x8d,0x14,0x13,0xf3,0xd8,0xb9,0x89,0x18,0x02,0xb0,0x2c,0xf6,0x6d,0x24,0x86,
+0x00,0x2c,0x8b,0xcd,0x1b,0x87,0x21,0x38,0x4d,0x55,0xd3,0xd6,0x30,0x54,0xc0,0x72,
+0x00,0x05,0xd1,0x4c,0x11,0x06,0x00,0x00,0x00,0x00,0x00,0x00,0x61,0x20,0x00,0x00,
+0x19,0x00,0x00,0x00,0x13,0x04,0x41,0x2c,0x10,0x00,0x00,0x00,0x03,0x00,0x00,0x00,
+0x24,0x4a,0x60,0x04,0x80,0xc2,0x0c,0x00,0x00,0x00,0x00,0x00,0x63,0x08,0xcd,0x33,
+0x16,0x01,0x04,0x48,0x34,0xc7,0x00,0x49,0xcf,0x58,0x04,0x10,0x28,0xd1,0x1c,0xc3,
+0x44,0x39,0x58,0x85,0x03,0x01,0x00,0x00,0x0a,0x00,0x00,0x00,0x26,0x41,0x08,0xc0,
+0xb2,0x18,0x45,0x21,0x00,0xcb,0xb2,0x5b,0x04,0x31,0x31,0x8f,0x6d,0x13,0xc4,0xc4,
+0x3c,0xb9,0x35,0x0c,0x15,0xb0,0x58,0x05,0x31,0x31,0x4f,0x7f,0x00,0x05,0xd1,0x4c,
+0x11,0x06,0x00,0x00,0x00,0x00,0x00,0x00,0x61,0x20,0x00,0x00,0x1b,0x00,0x00,0x00,
+0x13,0x04,0x41,0x2c,0x10,0x00,0x00,0x00,0x03,0x00,0x00,0x00,0x24,0xca,0x60,0x04,
+0xa0,0x04,0x8a,0x80,0xc2,0x0c,0x00,0x00,0x63,0x08,0xcd,0x33,0x16,0x01,0x04,0xca,
+0x34,0xc7,0x20,0x51,0xcf,0x1c,0x43,0x45,0x41,0x73,0x0c,0x16,0x15,0xcd,0x31,0x5c,
+0x94,0x83,0x58,0x38,0x10,0x00,0x00,0x00,0x0b,0x00,0x00,0x00,0x76,0x51,0x4c,0xcc,
+0x53,0xdb,0x86,0x51,0x4c,0xcc,0x53,0xe7,0x36,0x41,0x4c,0xcc,0x63,0x5b,0x05,0x31,
+0x31,0x8f,0x6e,0x16,0xc4,0xc4,0x3c,0xbd,0x51,0x10,0x02,0xb0,0x2c,0xd6,0x30,0x54,
+0xc0,0x72,0x00,0x05,0xd1,0x4c,0x11,0x06,0x00,0x00,0x00,0x00,0x61,0x20,0x00,0x00,
+0x2c,0x00,0x00,0x00,0x13,0x04,0x45,0x2c,0x10,0x00,0x00,0x00,0x03,0x00,0x00,0x00,
+0x24,0xca,0xa0,0x04,0x46,0x00,0x8a,0x80,0xc0,0x08,0x00,0x00,0x63,0x08,0x0d,0x34,
+0xdc,0x30,0x49,0xc4,0x2c,0x03,0x11,0x50,0x63,0x08,0xcd,0x33,0xdc,0x50,0x49,0xc4,
+0x2c,0x03,0x21,0x58,0x63,0x08,0x4d,0x34,0xdc,0x70,0x49,0xc4,0x2c,0x03,0x31,0x60,
+0x63,0x08,0x8d,0x33,0xdc,0x90,0x49,0x84,0x69,0x22,0x70,0xc3,0x27,0x1c,0x08,0x00,
+0x17,0x00,0x00,0x00,0x96,0x51,0x4c,0xcc,0x53,0xdf,0x66,0x41,0x08,0xcc,0x83,0xdb,
+0x04,0x31,0x31,0x4f,0x6d,0x15,0xc4,0xc4,0x3c,0xb7,0x61,0x10,0x02,0xf3,0xf0,0x76,
+0x41,0x4c,0xcc,0xb3,0x1f,0x81,0x11,0x11,0x13,0x15,0x35,0x37,0x90,0x2c,0x4e,0xf4,
+0x47,0x87,0x54,0xd7,0x17,0x70,0x2c,0x4e,0xf4,0x47,0x87,0x74,0x02,0xc8,0xe2,0x44,
+0x7f,0x74,0x48,0xb9,0x69,0x14,0x02,0xf3,0xd4,0xb8,0x6d,0x18,0x11,0x31,0x55,0xc0,
+0x62,0x0d,0x43,0x05,0x2c,0x07,0x50,0x10,0xcd,0x14,0x61,0x46,0x31,0x08,0xcc,0x03,
+0x00,0x00,0x00,0x00,0x71,0x20,0x00,0x00,0x12,0x00,0x00,0x00,0x66,0x40,0x54,0x82,
+0x23,0x19,0xc3,0xa0,0x20,0x8b,0x1d,0x18,0x4f,0x84,0x34,0x53,0x61,0x03,0xc4,0xe3,
+0x58,0x85,0x05,0x14,0xbe,0x34,0x45,0xb5,0x21,0x10,0x82,0x23,0x15,0x46,0x30,0x2c,
+0xc8,0x64,0x02,0x06,0xf0,0x3c,0x91,0x73,0x19,0x00,0xe1,0x4b,0x53,0x64,0x0a,0x84,
+0x84,0x34,0x85,0x25,0x0c,0x92,0x20,0x59,0xc1,0x20,0x30,0x8f,0x2d,0x10,0x95,0x84,
+0x34,0x00,0x00,0x00,0x00,0x00,0x00,0x00};
diff --git a/src/gallium/auxiliary/gallivm/gallivm_cpu.cpp b/src/gallium/auxiliary/gallivm/gallivm_cpu.cpp
index e64bfb1c6cb..3a2f2878a30 100644
--- a/src/gallium/auxiliary/gallivm/gallivm_cpu.cpp
+++ b/src/gallium/auxiliary/gallivm/gallivm_cpu.cpp
@@ -46,6 +46,7 @@
 #include "tgsi/tgsi_dump.h"
 
 #include "util/u_memory.h"
+#include "util/u_math.h"
 
 #include <llvm/Module.h>
 #include <llvm/CallingConv.h>
@@ -157,8 +158,8 @@ void gallivm_cpu_jit_compile(struct gallivm_cpu_engine *cpu, struct gallivm_prog
    llvm::ExistingModuleProvider *mp = new llvm::ExistingModuleProvider(mod);
    llvm::ExecutionEngine *ee = cpu->engine;
    assert(ee);
-   /*FIXME : remove */
-   ee->DisableLazyCompilation();
+   /*FIXME : why was this disabled ? we need it for pow/sqrt/... */
+   ee->DisableLazyCompilation(false);
    ee->addModuleProvider(mp);
 
    llvm::Function *func = func_for_shader(prog);
@@ -201,7 +202,6 @@ int gallivm_cpu_vs_exec(struct gallivm_prog *prog,
    unsigned int i, j;
    unsigned slot;
    vertex_shader_runner runner = reinterpret_cast<vertex_shader_runner>(prog->function);
-
    assert(runner);
 
    for (i = 0; i < count; i += MAX_TGSI_VERTICES) {
diff --git a/src/gallium/auxiliary/gallivm/instructions.cpp b/src/gallium/auxiliary/gallivm/instructions.cpp
index a82dc30306d..599975d5add 100644
--- a/src/gallium/auxiliary/gallivm/instructions.cpp
+++ b/src/gallium/auxiliary/gallivm/instructions.cpp
@@ -83,6 +83,7 @@ Instructions::Instructions(llvm::Module *mod, llvm::Function *func, llvm::BasicB
    m_llvmPow   = 0;
    m_llvmFloor = 0;
    m_llvmFlog  = 0;
+   m_llvmFexp  = 0;
    m_llvmLit  = 0;
    m_fmtPtr = 0;
 
@@ -92,194 +93,271 @@ Instructions::Instructions(llvm::Module *mod, llvm::Function *func, llvm::BasicB
    m_mod = ParseBitcodeFile(buffer);
 }
 
+llvm::BasicBlock * Instructions::currentBlock() const
+{
+   return m_builder.GetInsertBlock();
+}
+
+llvm::Value * Instructions::abs(llvm::Value *in)
+{
+   std::vector<llvm::Value*> vec = extractVector(in);
+   Value *xabs  = callFAbs(vec[0]);
+   Value *yabs  = callFAbs(vec[1]);
+   Value *zabs  = callFAbs(vec[2]);
+   Value *wabs  = callFAbs(vec[3]);
+   return vectorFromVals(xabs, yabs, zabs, wabs);
+}
+
 llvm::Value * Instructions::add(llvm::Value *in1, llvm::Value *in2)
 {
    return m_builder.CreateAdd(in1, in2, name("add"));
 }
 
-llvm::Value * Instructions::madd(llvm::Value *in1, llvm::Value *in2,
-                                 llvm::Value *in3)
+llvm::Value * Instructions::arl(llvm::Value *in)
 {
-   Value *mulRes = mul(in1, in2);
-   return add(mulRes, in3);
+   return floor(in);
 }
- 
-llvm::Value * Instructions::mul(llvm::Value *in1, llvm::Value *in2)
+
+void Instructions::beginLoop()
 {
-   return m_builder.CreateMul(in1, in2, name("mul"));
+   BasicBlock *begin = BasicBlock::Create(name("loop"), m_func,0);
+   BasicBlock *end = BasicBlock::Create(name("endloop"), m_func,0);
+
+   m_builder.CreateBr(begin);
+   Loop loop;
+   loop.begin = begin;
+   loop.end   = end;
+   m_builder.SetInsertPoint(begin);
+   m_loopStack.push(loop);
 }
 
-const char * Instructions::name(const char *prefix)
+void Instructions::bgnSub(unsigned label)
 {
-   ++m_idx;
-   snprintf(m_name, 32, "%s%d", prefix, m_idx);
-   return m_name;
+   llvm::Function *func = findFunction(label);
+
+   Function::arg_iterator args = func->arg_begin();
+   Value *ptr_INPUT = args++;
+   ptr_INPUT->setName("INPUT");
+   m_storage->pushArguments(ptr_INPUT);
+
+   llvm::BasicBlock *entry = BasicBlock::Create("entry", func, 0);
+
+   m_func = func;
+   m_builder.SetInsertPoint(entry);
 }
 
-llvm::Value * Instructions::dp3(llvm::Value *in1, llvm::Value *in2)
+void Instructions::brk()
 {
-   Value *mulRes = mul(in1, in2);
-   Value *x = m_builder.CreateExtractElement(mulRes,
-                                                          m_storage->constantInt(0),
-                                                          name("extractx"));
-   Value *y = m_builder.CreateExtractElement(mulRes,
-                                                          m_storage->constantInt(1),
-                                                          name("extracty"));
-   Value *z = m_builder.CreateExtractElement(mulRes,
-                                                          m_storage->constantInt(2),
-                                                          name("extractz"));
-   Value *xy = m_builder.CreateAdd(x, y,name("xy"));
-   Value *dot3 = m_builder.CreateAdd(xy, z, name("dot3"));
-   return vectorFromVals(dot3, dot3, dot3, dot3);
+   assert(!m_loopStack.empty());
+   BasicBlock *unr = BasicBlock::Create(name("unreachable"), m_func,0);
+   m_builder.CreateBr(m_loopStack.top().end);
+   m_builder.SetInsertPoint(unr);
 }
 
-llvm::Value *Instructions::callFSqrt(llvm::Value *val)
+void Instructions::cal(int label, llvm::Value *input)
 {
-   if (!m_llvmFSqrt) {
-      // predeclare the intrinsic
-      std::vector<const Type*> fsqrtArgs;
-      fsqrtArgs.push_back(Type::FloatTy);
-      PAListPtr fsqrtPal;
-      FunctionType* fsqrtType = FunctionType::get(
-         /*Result=*/Type::FloatTy,
-         /*Params=*/fsqrtArgs,
-         /*isVarArg=*/false);
-      m_llvmFSqrt = Function::Create(
-         /*Type=*/fsqrtType,
-         /*Linkage=*/GlobalValue::ExternalLinkage,
-         /*Name=*/"llvm.sqrt.f32", m_mod);
-      m_llvmFSqrt->setCallingConv(CallingConv::C);
-      m_llvmFSqrt->setParamAttrs(fsqrtPal);
-   }
-   CallInst *call = m_builder.CreateCall(m_llvmFSqrt, val,
-                                         name("sqrt"));
-   call->setCallingConv(CallingConv::C);
-   call->setTailCall(false);
-   return call;
+   std::vector<Value*> params;
+   params.push_back(input);
+   llvm::Function *func = findFunction(label);
+
+   m_builder.CreateCall(func, params.begin(), params.end());
 }
 
-llvm::Value * Instructions::rsq(llvm::Value *in1)
+llvm::Value * Instructions::ceil(llvm::Value *in)
 {
-   Value *x = m_builder.CreateExtractElement(in1,
-                                             m_storage->constantInt(0),
-                                             name("extractx"));
-   Value *abs  = callFAbs(x);
-   Value *sqrt = callFSqrt(abs);
-
-   Value *rsqrt = m_builder.CreateFDiv(ConstantFP::get(APFloat(1.f)),
-                                       sqrt,
-                                       name("rsqrt"));
-   return vectorFromVals(rsqrt, rsqrt, rsqrt, rsqrt);
+   std::vector<llvm::Value*> vec = extractVector(in);
+   return vectorFromVals(callCeil(vec[0]), callCeil(vec[1]),
+                         callCeil(vec[2]), callCeil(vec[3]));
 }
 
-llvm::Value * Instructions::vectorFromVals(llvm::Value *x, llvm::Value *y,
-                                           llvm::Value *z, llvm::Value *w)
+llvm::Value * Instructions::clamp(llvm::Value *in1)
 {
-   Constant *const_vec = Constant::getNullValue(m_floatVecType);
-   Value *res = m_builder.CreateInsertElement(const_vec, x,
-                                              m_storage->constantInt(0),
-                                              name("vecx"));
-   res = m_builder.CreateInsertElement(res, y, m_storage->constantInt(1),
-                               name("vecxy"));
-   res = m_builder.CreateInsertElement(res, z, m_storage->constantInt(2),
-                               name("vecxyz"));
-   if (w)
-      res = m_builder.CreateInsertElement(res, w, m_storage->constantInt(3),
-                                          name("vecxyzw"));
-   return res;
+   llvm::Value *zero = constVector(0.0f, 0.0f, 0.0f, 0.0f);
+   llvm::Value *one = constVector(1.0f, 1.0f, 1.0f, 1.0f);
+   return min( max(zero, in1), one);
 }
 
-llvm::Value *Instructions::callFAbs(llvm::Value *val)
+llvm::Value * Instructions::cmp(llvm::Value *in1, llvm::Value *in2, llvm::Value *in3)
 {
-   if (!m_llvmFAbs) {
-      // predeclare the intrinsic
-      std::vector<const Type*> fabsArgs;
-      fabsArgs.push_back(Type::FloatTy);
-      PAListPtr fabsPal;
-      FunctionType* fabsType = FunctionType::get(
-         /*Result=*/Type::FloatTy,
-         /*Params=*/fabsArgs,
-         /*isVarArg=*/false);
-      m_llvmFAbs = Function::Create(
-         /*Type=*/fabsType,
-         /*Linkage=*/GlobalValue::ExternalLinkage,
-         /*Name=*/"fabs", m_mod);
-      m_llvmFAbs->setCallingConv(CallingConv::C);
-      m_llvmFAbs->setParamAttrs(fabsPal);
-   }
-   CallInst *call = m_builder.CreateCall(m_llvmFAbs, val,
-                                         name("fabs"));
-   call->setCallingConv(CallingConv::C);
+   llvm::Function *func = m_mod->getFunction("cmp");
+   assert(func);
+
+   std::vector<Value*> params;
+   params.push_back(in1);
+   params.push_back(in2);
+   params.push_back(in3);
+   CallInst *call = m_builder.CreateCall(func, params.begin(), params.end(), name("cmpres"));
    call->setTailCall(false);
    return call;
 }
 
-llvm::Value * Instructions::lit(llvm::Value *in)
+llvm::Value * Instructions::cnd(llvm::Value *in1, llvm::Value *in2, llvm::Value *in3)
 {
-   if (!m_llvmLit) {
-      m_llvmLit = m_mod->getFunction("lit");
-   }
-   CallInst *call = m_builder.CreateCall(m_llvmLit, in, name("litres"));
-   call->setCallingConv(CallingConv::C);
-   call->setTailCall(false);
-   return call;
+   std::vector<llvm::Value*> vec1 = extractVector(in1);
+   std::vector<llvm::Value*> vec2 = extractVector(in2);
+   std::vector<llvm::Value*> vec3 = extractVector(in3);
+   Constant *half = ConstantFP::get(APFloat(0.5f));
+
+   Value *xcmp  = m_builder.CreateFCmpOGT(vec1[0], half, name("xcmp"));
+   Value *selx = m_builder.CreateSelect(xcmp, vec2[0], vec3[0],
+                                        name("selx"));
+
+   Value *ycmp  = m_builder.CreateFCmpOGT(vec1[1], half, name("ycmp"));
+   Value *sely = m_builder.CreateSelect(ycmp, vec2[1], vec3[1],
+                                        name("sely"));
+
+   Value *zcmp  = m_builder.CreateFCmpOGT(vec1[2], half, name("zcmp"));
+   Value *selz = m_builder.CreateSelect(zcmp, vec2[2], vec3[2],
+                                        name("selz"));
+
+   Value *wcmp  = m_builder.CreateFCmpOGT(vec1[3], half, name("wcmp"));
+   Value *selw = m_builder.CreateSelect(wcmp, vec2[3], vec3[3],
+                                        name("selw"));
+
+   return vectorFromVals(selx, sely, selz, selw);
 }
 
-llvm::Value * Instructions::sub(llvm::Value *in1, llvm::Value *in2)
+llvm::Value * Instructions::cnd0(llvm::Value *in1, llvm::Value *in2, llvm::Value *in3)
 {
-   Value *res = m_builder.CreateSub(in1, in2, name("sub"));
-   return res;
+   std::vector<llvm::Value*> vec1 = extractVector(in1);
+   std::vector<llvm::Value*> vec2 = extractVector(in2);
+   std::vector<llvm::Value*> vec3 = extractVector(in3);
+   Constant *zero = Constant::getNullValue(Type::FloatTy);
+
+   Value *xcmp  = m_builder.CreateFCmpOGE(vec1[0], zero, name("xcmp"));
+   Value *selx = m_builder.CreateSelect(xcmp, vec2[0], vec3[0],
+                                        name("selx"));
+
+   Value *ycmp  = m_builder.CreateFCmpOGE(vec1[1], zero, name("ycmp"));
+   Value *sely = m_builder.CreateSelect(ycmp, vec2[1], vec3[1],
+                                        name("sely"));
+
+   Value *zcmp  = m_builder.CreateFCmpOGE(vec1[2], zero, name("zcmp"));
+   Value *selz = m_builder.CreateSelect(zcmp, vec2[2], vec3[2],
+                                        name("selz"));
+
+   Value *wcmp  = m_builder.CreateFCmpOGE(vec1[3], zero, name("wcmp"));
+   Value *selw = m_builder.CreateSelect(wcmp, vec2[3], vec3[3],
+                                        name("selw"));
+
+   return vectorFromVals(selx, sely, selz, selw);
 }
 
-llvm::Value * Instructions::callPow(llvm::Value *val1, llvm::Value *val2)
+llvm::Value * Instructions::cos(llvm::Value *in)
 {
-   if (!m_llvmPow) {
-      // predeclare the intrinsic
-      std::vector<const Type*> powArgs;
-      powArgs.push_back(Type::FloatTy);
-      powArgs.push_back(Type::FloatTy);
-      PAListPtr powPal;
-      FunctionType* powType = FunctionType::get(
-         /*Result=*/Type::FloatTy,
-         /*Params=*/powArgs,
-         /*isVarArg=*/false);
-      m_llvmPow = Function::Create(
-         /*Type=*/powType,
-         /*Linkage=*/GlobalValue::ExternalLinkage,
-         /*Name=*/"llvm.pow.f32", m_mod);
-      m_llvmPow->setCallingConv(CallingConv::C);
-      m_llvmPow->setParamAttrs(powPal);
-   }
-   std::vector<Value*> params;
-   params.push_back(val1);
-   params.push_back(val2);
-   CallInst *call = m_builder.CreateCall(m_llvmPow, params.begin(), params.end(),
-                                         name("pow"));
-   call->setCallingConv(CallingConv::C);
+#if 0
+   llvm::Function *func = m_mod->getFunction("vcos");
+   assert(func);
+
+   CallInst *call = m_builder.CreateCall(func, in, name("cosres"));
    call->setTailCall(false);
    return call;
+#else
+   std::vector<llvm::Value*> elems = extractVector(in);
+   Function *func = m_mod->getFunction("cosf");
+   assert(func);
+   CallInst *cos = m_builder.CreateCall(func, elems[0], name("cosres"));
+   cos->setCallingConv(CallingConv::C);
+   cos->setTailCall(true);
+   return vectorFromVals(cos, cos, cos, cos);
+#endif
 }
 
-llvm::Value * Instructions::pow(llvm::Value *in1, llvm::Value *in2)
+llvm::Value * Instructions::cross(llvm::Value *in1, llvm::Value *in2)
 {
    Value *x1 = m_builder.CreateExtractElement(in1,
                                               m_storage->constantInt(0),
                                               name("x1"));
+   Value *y1 = m_builder.CreateExtractElement(in1,
+                                              m_storage->constantInt(1),
+                                              name("y1"));
+   Value *z1 = m_builder.CreateExtractElement(in1,
+                                              m_storage->constantInt(2),
+                                              name("z1"));
+
    Value *x2 = m_builder.CreateExtractElement(in2,
                                               m_storage->constantInt(0),
                                               name("x2"));
-   llvm::Value *val = callPow(x1, x2);
-   return vectorFromVals(val, val, val, val);
+   Value *y2 = m_builder.CreateExtractElement(in2,
+                                              m_storage->constantInt(1),
+                                              name("y2"));
+   Value *z2 = m_builder.CreateExtractElement(in2,
+                                              m_storage->constantInt(2),
+                                              name("z2"));
+   Value *y1z2 = mul(y1, z2);
+   Value *z1y2 = mul(z1, y2);
+
+   Value *z1x2 = mul(z1, x2);
+   Value *x1z2 = mul(x1, z2);
+
+   Value *x1y2 = mul(x1, y2);
+   Value *y1x2 = mul(y1, x2);
+
+   return vectorFromVals(sub(y1z2, z1y2), sub(z1x2, x1z2), sub(x1y2, y1x2));
 }
 
-llvm::Value * Instructions::rcp(llvm::Value *in1)
+llvm::Value * Instructions::ddx(llvm::Value *in)
 {
-   Value *x1 = m_builder.CreateExtractElement(in1,
-                                              m_storage->constantInt(0),
-                                              name("x1"));
-   Value *res = m_builder.CreateFDiv(ConstantFP::get(APFloat(1.f)),
-                                     x1, name("rcp"));
-   return vectorFromVals(res, res, res, res);
+   // FIXME
+   assert(0);
+}
+
+llvm::Value * Instructions::ddy(llvm::Value *in)
+{
+   // FIXME
+   assert(0);
+}
+
+llvm::Value * Instructions::div(llvm::Value *in1, llvm::Value *in2)
+{
+   return m_builder.CreateFDiv(in1, in2, name("div"));
+}
+
+llvm::Value * Instructions::dot2add(llvm::Value *in1, llvm::Value *in2, llvm::Value *in3)
+{
+   Value *mulRes = mul(in1, in2);
+   Value *x = m_builder.CreateExtractElement(mulRes,
+                                                          m_storage->constantInt(0),
+                                                          name("extractx"));
+   Value *y = m_builder.CreateExtractElement(mulRes,
+                                                          m_storage->constantInt(1),
+                                                          name("extracty"));
+   Value *z = m_builder.CreateExtractElement(in3,
+                                                          m_storage->constantInt(2),
+                                                          name("extractz"));
+   Value *xy = m_builder.CreateAdd(x, y,name("xy"));
+   Value *dot2add = m_builder.CreateAdd(xy, z, name("dot2add"));
+   return vectorFromVals(dot2add, dot2add, dot2add, dot2add);
+}
+
+llvm::Value * Instructions::dp2(llvm::Value *in1, llvm::Value *in2)
+{
+   Value *mulRes = mul(in1, in2);
+   Value *x = m_builder.CreateExtractElement(mulRes,
+                                                          m_storage->constantInt(0),
+                                                          name("extractx"));
+   Value *y = m_builder.CreateExtractElement(mulRes,
+                                                          m_storage->constantInt(1),
+                                                          name("extracty"));
+   Value *xy = m_builder.CreateAdd(x, y,name("xy"));
+   return vectorFromVals(xy, xy, xy, xy);
+}
+
+llvm::Value * Instructions::dp3(llvm::Value *in1, llvm::Value *in2)
+{
+   Value *mulRes = mul(in1, in2);
+   Value *x = m_builder.CreateExtractElement(mulRes,
+                                                          m_storage->constantInt(0),
+                                                          name("extractx"));
+   Value *y = m_builder.CreateExtractElement(mulRes,
+                                                          m_storage->constantInt(1),
+                                                          name("extracty"));
+   Value *z = m_builder.CreateExtractElement(mulRes,
+                                                          m_storage->constantInt(2),
+                                                          name("extractz"));
+   Value *xy = m_builder.CreateAdd(x, y,name("xy"));
+   Value *dot3 = m_builder.CreateAdd(xy, z, name("dot3"));
+   return vectorFromVals(dot3, dot3, dot3, dot3);
 }
 
 llvm::Value * Instructions::dp4(llvm::Value *in1, llvm::Value *in2)
@@ -321,6 +399,53 @@ llvm::Value * Instructions::dst(llvm::Value *in1, llvm::Value *in2)
                          ry, z, w);
 }
 
+void Instructions::elseop()
+{
+   assert(!m_ifStack.empty());
+   BasicBlock *ifend = BasicBlock::Create(name("ifend"), m_func,0);
+   m_builder.CreateBr(ifend);
+   m_builder.SetInsertPoint(m_ifStack.top());
+   currentBlock()->setName(name("ifelse"));
+   m_ifStack.pop();
+   m_ifStack.push(ifend);
+}
+
+void Instructions::endif()
+{
+   assert(!m_ifStack.empty());
+   m_builder.CreateBr(m_ifStack.top());
+   m_builder.SetInsertPoint(m_ifStack.top());
+   m_ifStack.pop();
+}
+
+void Instructions::endLoop()
+{
+   assert(!m_loopStack.empty());
+   Loop loop = m_loopStack.top();
+   m_builder.CreateBr(loop.begin);
+   loop.end->moveAfter(currentBlock());
+   m_builder.SetInsertPoint(loop.end);
+   m_loopStack.pop();
+}
+
+void Instructions::end()
+{
+   m_builder.CreateRetVoid();
+}
+
+void Instructions::endSub()
+{
+   m_func = 0;
+   m_builder.SetInsertPoint(0);
+}
+
+llvm::Value * Instructions::exp(llvm::Value *in)
+{
+   std::vector<llvm::Value*> vec = extractVector(in);
+   return vectorFromVals(callFExp(vec[0]), callFExp(vec[1]),
+                             callFExp(vec[2]), callFExp(vec[3]));
+}
+
 llvm::Value * Instructions::ex2(llvm::Value *in)
 {
    llvm::Value *val = callPow(ConstantFP::get(APFloat(2.f)),
@@ -330,31 +455,6 @@ llvm::Value * Instructions::ex2(llvm::Value *in)
    return vectorFromVals(val, val, val, val);
 }
 
-llvm::Value * Instructions::callFloor(llvm::Value *val)
-{
-   if (!m_llvmFloor) {
-      // predeclare the intrinsic
-      std::vector<const Type*> floorArgs;
-      floorArgs.push_back(Type::FloatTy);
-      PAListPtr floorPal;
-      FunctionType* floorType = FunctionType::get(
-         /*Result=*/Type::FloatTy,
-         /*Params=*/floorArgs,
-         /*isVarArg=*/false);
-      m_llvmFloor = Function::Create(
-         /*Type=*/floorType,
-         /*Linkage=*/GlobalValue::ExternalLinkage,
-         /*Name=*/"floorf", m_mod);
-      m_llvmFloor->setCallingConv(CallingConv::C);
-      m_llvmFloor->setParamAttrs(floorPal);
-   }
-   CallInst *call =  m_builder.CreateCall(m_llvmFloor, val,
-                                          name("floorf"));
-   call->setCallingConv(CallingConv::C);
-   call->setTailCall(false);
-   return call;
-}
-
 llvm::Value * Instructions::floor(llvm::Value *in)
 {
    std::vector<llvm::Value*> vec = extractVector(in);
@@ -362,42 +462,52 @@ llvm::Value * Instructions::floor(llvm::Value *in)
                          callFloor(vec[2]), callFloor(vec[3]));
 }
 
-llvm::Value * Instructions::arl(llvm::Value *in)
-{
-   return floor(in);
-}
-
 llvm::Value * Instructions::frc(llvm::Value *in)
 {
    llvm::Value *flr = floor(in);
    return sub(in, flr);
 }
 
-llvm::Value * Instructions::callFLog(llvm::Value *val)
+void Instructions::ifop(llvm::Value *in)
 {
-   if (!m_llvmFlog) {
-      // predeclare the intrinsic
-      std::vector<const Type*> flogArgs;
-      flogArgs.push_back(Type::FloatTy);
-      PAListPtr flogPal;
-      FunctionType* flogType = FunctionType::get(
-         /*Result=*/Type::FloatTy,
-         /*Params=*/flogArgs,
-         /*isVarArg=*/false);
-      m_llvmFlog = Function::Create(
-         /*Type=*/flogType,
-         /*Linkage=*/GlobalValue::ExternalLinkage,
-         /*Name=*/"logf", m_mod);
-      m_llvmFlog->setCallingConv(CallingConv::C);
-      m_llvmFlog->setParamAttrs(flogPal);
-   }
-   CallInst *call = m_builder.CreateCall(m_llvmFlog, val,
-                                         name("logf"));
-   call->setCallingConv(CallingConv::C);
+   BasicBlock *ifthen = BasicBlock::Create(name("ifthen"), m_func,0);
+   BasicBlock *ifend = BasicBlock::Create(name("ifthenend"), m_func,0);
+
+   //BasicBlock *yblock = new BasicBlock(name("yblock"), m_func,0);
+   //BasicBlock *zblock = new BasicBlock(name("zblock"), m_func,0);
+   //BasicBlock *wblock = new BasicBlock(name("wblock"), m_func,0);
+
+   Constant *float0 = Constant::getNullValue(Type::FloatTy);
+
+   Value *x = m_builder.CreateExtractElement(in, m_storage->constantInt(0),
+                                             name("extractx"));
+   Value *xcmp = m_builder.CreateFCmpUNE(x, float0, name("xcmp"));
+   m_builder.CreateCondBr(xcmp, ifthen, ifend);
+   //m_builder.SetInsertPoint(yblock);
+
+   m_builder.SetInsertPoint(ifthen);
+   m_ifStack.push(ifend);
+}
+
+llvm::Value * Instructions::kil(llvm::Value *in)
+{
+   llvm::Function *func = m_mod->getFunction("kil");
+   assert(func);
+
+   CallInst *call = m_builder.CreateCall(func, in, name("kilpres"));
    call->setTailCall(false);
    return call;
 }
 
+llvm::Value * Instructions::lerp(llvm::Value *in1, llvm::Value *in2,
+                                 llvm::Value *in3)
+{
+   llvm::Value *m = mul(in1, in2);
+   llvm::Value *vec1 = constVector(1.f, 1.f, 1.f, 1.f);
+   llvm::Value *s = sub(vec1, in1);
+   return add(m, mul(s, in3));
+}
+
 llvm::Value * Instructions::lg2(llvm::Value *in)
 {
    std::vector<llvm::Value*> vec = extractVector(in);
@@ -407,142 +517,176 @@ llvm::Value * Instructions::lg2(llvm::Value *in)
                              callFLog(vec[2]), callFLog(vec[3])), const_vec);
 }
 
-llvm::Value * Instructions::min(llvm::Value *in1, llvm::Value *in2)
+llvm::Value * Instructions::lit(llvm::Value *in)
+{
+   if (!m_llvmLit) {
+      m_llvmLit = m_mod->getFunction("lit");
+   }
+   CallInst *call = m_builder.CreateCall(m_llvmLit, in, name("litres"));
+   call->setCallingConv(CallingConv::C);
+   call->setTailCall(false);
+   return call;
+}
+
+llvm::Value * Instructions::log(llvm::Value *in)
+{
+   std::vector<llvm::Value*> vec = extractVector(in);
+   return vectorFromVals(callFLog(vec[0]), callFLog(vec[1]),
+                             callFLog(vec[2]), callFLog(vec[3]));
+}
+
+llvm::Value * Instructions::madd(llvm::Value *in1, llvm::Value *in2,
+                                 llvm::Value *in3)
+{
+   Value *mulRes = mul(in1, in2);
+   return add(mulRes, in3);
+}
+
+llvm::Value * Instructions::max(llvm::Value *in1, llvm::Value *in2)
 {
    std::vector<llvm::Value*> vec1 = extractVector(in1);
    std::vector<llvm::Value*> vec2 = extractVector(in2);
 
-   Value *xcmp  = m_builder.CreateFCmpOLT(vec1[0], vec2[0], name("xcmp"));
+   Value *xcmp  = m_builder.CreateFCmpOGT(vec1[0], vec2[0],
+                                          name("xcmp"));
    Value *selx = m_builder.CreateSelect(xcmp, vec1[0], vec2[0],
                                         name("selx"));
 
-   Value *ycmp  = m_builder.CreateFCmpOLT(vec1[1], vec2[1], name("ycmp"));
+   Value *ycmp  = m_builder.CreateFCmpOGT(vec1[1], vec2[1],
+                                          name("ycmp"));
    Value *sely = m_builder.CreateSelect(ycmp, vec1[1], vec2[1],
                                         name("sely"));
 
-   Value *zcmp  = m_builder.CreateFCmpOLT(vec1[2], vec2[2], name("zcmp"));
+   Value *zcmp  = m_builder.CreateFCmpOGT(vec1[2], vec2[2],
+                                          name("zcmp"));
    Value *selz = m_builder.CreateSelect(zcmp, vec1[2], vec2[2],
                                         name("selz"));
 
-   Value *wcmp  = m_builder.CreateFCmpOLT(vec1[3], vec2[3], name("wcmp"));
+   Value *wcmp  = m_builder.CreateFCmpOGT(vec1[3], vec2[3],
+                                          name("wcmp"));
    Value *selw = m_builder.CreateSelect(wcmp, vec1[3], vec2[3],
                                         name("selw"));
 
    return vectorFromVals(selx, sely, selz, selw);
 }
 
-llvm::Value * Instructions::max(llvm::Value *in1, llvm::Value *in2)
+llvm::Value * Instructions::min(llvm::Value *in1, llvm::Value *in2)
 {
    std::vector<llvm::Value*> vec1 = extractVector(in1);
    std::vector<llvm::Value*> vec2 = extractVector(in2);
 
-   Value *xcmp  = m_builder.CreateFCmpOGT(vec1[0], vec2[0],
-                                          name("xcmp"));
+   Value *xcmp  = m_builder.CreateFCmpOLT(vec1[0], vec2[0], name("xcmp"));
    Value *selx = m_builder.CreateSelect(xcmp, vec1[0], vec2[0],
                                         name("selx"));
 
-   Value *ycmp  = m_builder.CreateFCmpOGT(vec1[1], vec2[1],
-                                          name("ycmp"));
+   Value *ycmp  = m_builder.CreateFCmpOLT(vec1[1], vec2[1], name("ycmp"));
    Value *sely = m_builder.CreateSelect(ycmp, vec1[1], vec2[1],
                                         name("sely"));
 
-   Value *zcmp  = m_builder.CreateFCmpOGT(vec1[2], vec2[2],
-                                          name("zcmp"));
+   Value *zcmp  = m_builder.CreateFCmpOLT(vec1[2], vec2[2], name("zcmp"));
    Value *selz = m_builder.CreateSelect(zcmp, vec1[2], vec2[2],
                                         name("selz"));
 
-   Value *wcmp  = m_builder.CreateFCmpOGT(vec1[3], vec2[3],
-                                          name("wcmp"));
+   Value *wcmp  = m_builder.CreateFCmpOLT(vec1[3], vec2[3], name("wcmp"));
    Value *selw = m_builder.CreateSelect(wcmp, vec1[3], vec2[3],
                                         name("selw"));
 
    return vectorFromVals(selx, sely, selz, selw);
 }
 
-void Instructions::printVector(llvm::Value *val)
+llvm::Value * Instructions::mul(llvm::Value *in1, llvm::Value *in2)
 {
-   static const char *frmt = "Vector is [%f, %f, %f, %f]\x0A";
+   return m_builder.CreateMul(in1, in2, name("mul"));
+}
 
-   if (!m_fmtPtr) {
-      Constant *format = ConstantArray::get(frmt, true);
-      ArrayType *arrayTy = ArrayType::get(IntegerType::get(8), strlen(frmt) + 1);
-      GlobalVariable* globalFormat = new GlobalVariable(
-         /*Type=*/arrayTy,
-         /*isConstant=*/true,
-         /*Linkage=*/GlobalValue::InternalLinkage,
-         /*Initializer=*/0, // has initializer, specified below
-         /*Name=*/name(".str"),
-         m_mod);
-      globalFormat->setInitializer(format);
+llvm::Value * Instructions::neg(llvm::Value *in)
+{
+   Value *neg = m_builder.CreateNeg(in, name("neg"));
+   return neg;
+}
 
-      Constant* const_int0 = Constant::getNullValue(IntegerType::get(32));
-      std::vector<Constant*> const_ptr_21_indices;
-      const_ptr_21_indices.push_back(const_int0);
-      const_ptr_21_indices.push_back(const_int0);
-      m_fmtPtr = ConstantExpr::getGetElementPtr(globalFormat,
-                                                &const_ptr_21_indices[0], const_ptr_21_indices.size());
-   }
+llvm::Value * Instructions::nrm(llvm::Value *in)
+{
+   llvm::Value *v = rsq(in);
+   return mul(v, in);
+}
 
-   Function *func_printf = m_mod->getFunction("printf");
-   if (!func_printf)
-      func_printf = declarePrintf();
-   assert(func_printf);
-   std::vector<llvm::Value*> vec = extractVector(val);
-   Value *dx = m_builder.CreateFPExt(vec[0], Type::DoubleTy, name("dx"));
-   Value *dy = m_builder.CreateFPExt(vec[1], Type::DoubleTy, name("dy"));
-   Value *dz = m_builder.CreateFPExt(vec[2], Type::DoubleTy, name("dz"));
-   Value *dw = m_builder.CreateFPExt(vec[3], Type::DoubleTy, name("dw"));
-   std::vector<Value*> params;
-   params.push_back(m_fmtPtr);
-   params.push_back(dx);
-   params.push_back(dy);
-   params.push_back(dz);
-   params.push_back(dw);
-   CallInst *call = m_builder.CreateCall(func_printf, params.begin(), params.end(),
-                                         name("printf"));
-   call->setCallingConv(CallingConv::C);
-   call->setTailCall(true);
+llvm::Value * Instructions::pow(llvm::Value *in1, llvm::Value *in2)
+{
+   Value *x1 = m_builder.CreateExtractElement(in1,
+                                              m_storage->constantInt(0),
+                                              name("x1"));
+   Value *x2 = m_builder.CreateExtractElement(in2,
+                                              m_storage->constantInt(0),
+                                              name("x2"));
+   llvm::Value *val = callPow(x1, x2);
+   return vectorFromVals(val, val, val, val);
 }
 
-llvm::Function * Instructions::declarePrintf()
+llvm::Value * Instructions::rcp(llvm::Value *in1)
 {
-   std::vector<const Type*> args;
-   PAListPtr params;
-   FunctionType* funcTy = FunctionType::get(
-      /*Result=*/IntegerType::get(32),
-      /*Params=*/args,
-      /*isVarArg=*/true);
-   Function* func_printf = Function::Create(
-      /*Type=*/funcTy,
-      /*Linkage=*/GlobalValue::ExternalLinkage,
-      /*Name=*/"printf", m_mod);
-   func_printf->setCallingConv(CallingConv::C);
-   func_printf->setParamAttrs(params);
-   return func_printf;
+   Value *x1 = m_builder.CreateExtractElement(in1,
+                                              m_storage->constantInt(0),
+                                              name("x1"));
+   Value *res = m_builder.CreateFDiv(ConstantFP::get(APFloat(1.f)),
+                                     x1, name("rcp"));
+   return vectorFromVals(res, res, res, res);
+}
+
+llvm::Value * Instructions::rsq(llvm::Value *in1)
+{
+   Value *x = m_builder.CreateExtractElement(in1,
+                                             m_storage->constantInt(0),
+                                             name("extractx"));
+   Value *abs  = callFAbs(x);
+   Value *sqrt = callFSqrt(abs);
+
+   Value *rsqrt = m_builder.CreateFDiv(ConstantFP::get(APFloat(1.f)),
+                                       sqrt,
+                                       name("rsqrt"));
+   return vectorFromVals(rsqrt, rsqrt, rsqrt, rsqrt);
 }
 
+llvm::Value * Instructions::scs(llvm::Value *in)
+{
+   llvm::Function *func = m_mod->getFunction("scs");
+   assert(func);
 
-llvm::Value * Instructions::sgt(llvm::Value *in1, llvm::Value *in2)
+   CallInst *call = m_builder.CreateCall(func, in, name("scsres"));
+   call->setTailCall(false);
+   return call;
+}
+
+llvm::Value * Instructions::seq(llvm::Value *in1, llvm::Value *in2)
 {
    Constant *const1f = ConstantFP::get(APFloat(1.000000e+00f));
    Constant *const0f = Constant::getNullValue(Type::FloatTy);
 
    std::vector<llvm::Value*> vec1 = extractVector(in1);
    std::vector<llvm::Value*> vec2 = extractVector(in2);
-   Value *xcmp = m_builder.CreateFCmpOGT(vec1[0], vec2[0], name("xcmp"));
+
+   Value *xcmp = m_builder.CreateFCmpOEQ(vec1[0], vec2[0], name("xcmp"));
    Value *x = m_builder.CreateSelect(xcmp, const1f, const0f, name("xsel"));
 
-   Value *ycmp = m_builder.CreateFCmpOGT(vec1[1], vec2[1], name("ycmp"));
+   Value *ycmp = m_builder.CreateFCmpOEQ(vec1[1], vec2[1], name("ycmp"));
    Value *y = m_builder.CreateSelect(ycmp, const1f, const0f, name("ysel"));
 
-   Value *zcmp = m_builder.CreateFCmpOGT(vec1[2], vec2[2], name("zcmp"));
+   Value *zcmp = m_builder.CreateFCmpOEQ(vec1[2], vec2[2], name("zcmp"));
    Value *z = m_builder.CreateSelect(zcmp, const1f, const0f, name("zsel"));
 
-   Value *wcmp = m_builder.CreateFCmpOGT(vec1[3], vec2[3], name("wcmp"));
+   Value *wcmp = m_builder.CreateFCmpOEQ(vec1[3], vec2[3], name("wcmp"));
    Value *w = m_builder.CreateSelect(wcmp, const1f, const0f, name("wsel"));
 
    return vectorFromVals(x, y, z, w);
 }
+
+llvm::Value * Instructions::sfl(llvm::Value *in1, llvm::Value *in2)
+{
+   Constant *const0f = Constant::getNullValue(Type::FloatTy);
+
+   return vectorFromVals(const0f, const0f, const0f, const0f);
+}
+
 llvm::Value * Instructions::sge(llvm::Value *in1, llvm::Value *in2)
 {
    Constant *const1f = ConstantFP::get(APFloat(1.000000e+00f));
@@ -566,157 +710,118 @@ llvm::Value * Instructions::sge(llvm::Value *in1, llvm::Value *in2)
    return vectorFromVals(x, y, z, w);
 }
 
-
-llvm::Value * Instructions::slt(llvm::Value *in1, llvm::Value *in2)
+llvm::Value * Instructions::sgt(llvm::Value *in1, llvm::Value *in2)
 {
    Constant *const1f = ConstantFP::get(APFloat(1.000000e+00f));
    Constant *const0f = Constant::getNullValue(Type::FloatTy);
 
    std::vector<llvm::Value*> vec1 = extractVector(in1);
    std::vector<llvm::Value*> vec2 = extractVector(in2);
-
-   Value *xcmp = m_builder.CreateFCmpOLT(vec1[0], vec2[0], name("xcmp"));
+   Value *xcmp = m_builder.CreateFCmpOGT(vec1[0], vec2[0], name("xcmp"));
    Value *x = m_builder.CreateSelect(xcmp, const1f, const0f, name("xsel"));
 
-   Value *ycmp = m_builder.CreateFCmpOLT(vec1[1], vec2[1], name("ycmp"));
+   Value *ycmp = m_builder.CreateFCmpOGT(vec1[1], vec2[1], name("ycmp"));
    Value *y = m_builder.CreateSelect(ycmp, const1f, const0f, name("ysel"));
 
-   Value *zcmp = m_builder.CreateFCmpOLT(vec1[2], vec2[2], name("zcmp"));
+   Value *zcmp = m_builder.CreateFCmpOGT(vec1[2], vec2[2], name("zcmp"));
    Value *z = m_builder.CreateSelect(zcmp, const1f, const0f, name("zsel"));
 
-   Value *wcmp = m_builder.CreateFCmpOLT(vec1[3], vec2[3], name("wcmp"));
+   Value *wcmp = m_builder.CreateFCmpOGT(vec1[3], vec2[3], name("wcmp"));
    Value *w = m_builder.CreateSelect(wcmp, const1f, const0f, name("wsel"));
 
    return vectorFromVals(x, y, z, w);
 }
 
-llvm::Value * Instructions::cross(llvm::Value *in1, llvm::Value *in2)
+llvm::Value * Instructions::sin(llvm::Value *in)
 {
-   Value *x1 = m_builder.CreateExtractElement(in1,
-                                              m_storage->constantInt(0),
-                                              name("x1"));
-   Value *y1 = m_builder.CreateExtractElement(in1,
-                                              m_storage->constantInt(1),
-                                              name("y1"));
-   Value *z1 = m_builder.CreateExtractElement(in1,
-                                              m_storage->constantInt(2),
-                                              name("z1"));
+   llvm::Function *func = m_mod->getFunction("vsin");
+   assert(func);
 
-   Value *x2 = m_builder.CreateExtractElement(in2,
-                                              m_storage->constantInt(0),
-                                              name("x2"));
-   Value *y2 = m_builder.CreateExtractElement(in2,
-                                              m_storage->constantInt(1),
-                                              name("y2"));
-   Value *z2 = m_builder.CreateExtractElement(in2,
-                                              m_storage->constantInt(2),
-                                              name("z2"));
-   Value *y1z2 = mul(y1, z2);
-   Value *z1y2 = mul(z1, y2);
+   CallInst *call = m_builder.CreateCall(func, in, name("sinres"));
+   call->setTailCall(false);
+   return call;
+}
 
-   Value *z1x2 = mul(z1, x2);
-   Value *x1z2 = mul(x1, z2);
+llvm::Value * Instructions::sle(llvm::Value *in1, llvm::Value *in2)
+{
+   Constant *const1f = ConstantFP::get(APFloat(1.000000e+00f));
+   Constant *const0f = Constant::getNullValue(Type::FloatTy);
 
-   Value *x1y2 = mul(x1, y2);
-   Value *y1x2 = mul(y1, x2);
+   std::vector<llvm::Value*> vec1 = extractVector(in1);
+   std::vector<llvm::Value*> vec2 = extractVector(in2);
 
-   return vectorFromVals(sub(y1z2, z1y2), sub(z1x2, x1z2), sub(x1y2, y1x2));
-}
+   Value *xcmp = m_builder.CreateFCmpOLE(vec1[0], vec2[0], name("xcmp"));
+   Value *x = m_builder.CreateSelect(xcmp, const1f, const0f, name("xsel"));
 
+   Value *ycmp = m_builder.CreateFCmpOLE(vec1[1], vec2[1], name("ycmp"));
+   Value *y = m_builder.CreateSelect(ycmp, const1f, const0f, name("ysel"));
 
-llvm::Value * Instructions::abs(llvm::Value *in)
-{
-   std::vector<llvm::Value*> vec = extractVector(in);
-   Value *xabs  = callFAbs(vec[0]);
-   Value *yabs  = callFAbs(vec[1]);
-   Value *zabs  = callFAbs(vec[2]);
-   Value *wabs  = callFAbs(vec[3]);
-   return vectorFromVals(xabs, yabs, zabs, wabs);
+   Value *zcmp = m_builder.CreateFCmpOLE(vec1[2], vec2[2], name("zcmp"));
+   Value *z = m_builder.CreateSelect(zcmp, const1f, const0f, name("zsel"));
+
+   Value *wcmp = m_builder.CreateFCmpOLE(vec1[3], vec2[3], name("wcmp"));
+   Value *w = m_builder.CreateSelect(wcmp, const1f, const0f, name("wsel"));
+
+   return vectorFromVals(x, y, z, w);
 }
 
-void Instructions::ifop(llvm::Value *in)
+llvm::Value * Instructions::slt(llvm::Value *in1, llvm::Value *in2)
 {
-   BasicBlock *ifthen = BasicBlock::Create(name("ifthen"), m_func,0);
-   BasicBlock *ifend = BasicBlock::Create(name("ifthenend"), m_func,0);
+   Constant *const1f = ConstantFP::get(APFloat(1.000000e+00f));
+   Constant *const0f = Constant::getNullValue(Type::FloatTy);
 
-   //BasicBlock *yblock = new BasicBlock(name("yblock"), m_func,0);
-   //BasicBlock *zblock = new BasicBlock(name("zblock"), m_func,0);
-   //BasicBlock *wblock = new BasicBlock(name("wblock"), m_func,0);
+   std::vector<llvm::Value*> vec1 = extractVector(in1);
+   std::vector<llvm::Value*> vec2 = extractVector(in2);
 
-   Constant *float0 = Constant::getNullValue(Type::FloatTy);
+   Value *xcmp = m_builder.CreateFCmpOLT(vec1[0], vec2[0], name("xcmp"));
+   Value *x = m_builder.CreateSelect(xcmp, const1f, const0f, name("xsel"));
 
-   Value *x = m_builder.CreateExtractElement(in, m_storage->constantInt(0),
-                                             name("extractx"));
-   Value *xcmp = m_builder.CreateFCmpUNE(x, float0, name("xcmp"));
-   m_builder.CreateCondBr(xcmp, ifthen, ifend);
-   //m_builder.SetInsertPoint(yblock);
+   Value *ycmp = m_builder.CreateFCmpOLT(vec1[1], vec2[1], name("ycmp"));
+   Value *y = m_builder.CreateSelect(ycmp, const1f, const0f, name("ysel"));
 
-   m_builder.SetInsertPoint(ifthen);
-   m_ifStack.push(ifend);
-}
+   Value *zcmp = m_builder.CreateFCmpOLT(vec1[2], vec2[2], name("zcmp"));
+   Value *z = m_builder.CreateSelect(zcmp, const1f, const0f, name("zsel"));
 
-llvm::BasicBlock * Instructions::currentBlock() const
-{
-   return m_builder.GetInsertBlock();
-}
+   Value *wcmp = m_builder.CreateFCmpOLT(vec1[3], vec2[3], name("wcmp"));
+   Value *w = m_builder.CreateSelect(wcmp, const1f, const0f, name("wsel"));
 
-void Instructions::elseop()
-{
-   assert(!m_ifStack.empty());
-   BasicBlock *ifend = BasicBlock::Create(name("ifend"), m_func,0);
-   m_builder.CreateBr(ifend);
-   m_builder.SetInsertPoint(m_ifStack.top());
-   currentBlock()->setName(name("ifelse"));
-   m_ifStack.pop();
-   m_ifStack.push(ifend);
+   return vectorFromVals(x, y, z, w);
 }
 
-void Instructions::endif()
+llvm::Value * Instructions::sne(llvm::Value *in1, llvm::Value *in2)
 {
-   assert(!m_ifStack.empty());
-   m_builder.CreateBr(m_ifStack.top());
-   m_builder.SetInsertPoint(m_ifStack.top());
-   m_ifStack.pop();
-}
+   Constant *const1f = ConstantFP::get(APFloat(1.000000e+00f));
+   Constant *const0f = Constant::getNullValue(Type::FloatTy);
 
-llvm::Value * Instructions::lerp(llvm::Value *in1, llvm::Value *in2,
-                                 llvm::Value *in3)
-{
-   llvm::Value *m = mul(in1, in2);
-   llvm::Value *vec1 = constVector(1.f, 1.f, 1.f, 1.f);
-   llvm::Value *s = sub(vec1, in1);
-   return add(m, mul(s, in3));
-}
+   std::vector<llvm::Value*> vec1 = extractVector(in1);
+   std::vector<llvm::Value*> vec2 = extractVector(in2);
 
-void Instructions::beginLoop()
-{
-   BasicBlock *begin = BasicBlock::Create(name("loop"), m_func,0);
-   BasicBlock *end = BasicBlock::Create(name("endloop"), m_func,0);
+   Value *xcmp = m_builder.CreateFCmpONE(vec1[0], vec2[0], name("xcmp"));
+   Value *x = m_builder.CreateSelect(xcmp, const1f, const0f, name("xsel"));
 
-   m_builder.CreateBr(begin);
-   Loop loop;
-   loop.begin = begin;
-   loop.end   = end;
-   m_builder.SetInsertPoint(begin);
-   m_loopStack.push(loop);
+   Value *ycmp = m_builder.CreateFCmpONE(vec1[1], vec2[1], name("ycmp"));
+   Value *y = m_builder.CreateSelect(ycmp, const1f, const0f, name("ysel"));
+
+   Value *zcmp = m_builder.CreateFCmpONE(vec1[2], vec2[2], name("zcmp"));
+   Value *z = m_builder.CreateSelect(zcmp, const1f, const0f, name("zsel"));
+
+   Value *wcmp = m_builder.CreateFCmpONE(vec1[3], vec2[3], name("wcmp"));
+   Value *w = m_builder.CreateSelect(wcmp, const1f, const0f, name("wsel"));
+
+   return vectorFromVals(x, y, z, w);
 }
 
-void Instructions::endLoop()
+llvm::Value * Instructions::str(llvm::Value *in1, llvm::Value *in2)
 {
-   assert(!m_loopStack.empty());
-   Loop loop = m_loopStack.top();
-   m_builder.CreateBr(loop.begin);
-   loop.end->moveAfter(currentBlock());
-   m_builder.SetInsertPoint(loop.end);
-   m_loopStack.pop();
+   Constant *const1f = ConstantFP::get(APFloat(1.000000e+00f));
+
+   return vectorFromVals(const1f, const1f, const1f, const1f);
 }
 
-void Instructions::brk()
+llvm::Value * Instructions::sub(llvm::Value *in1, llvm::Value *in2)
 {
-   assert(!m_loopStack.empty());
-   BasicBlock *unr = BasicBlock::Create(name("unreachable"), m_func,0);
-   m_builder.CreateBr(m_loopStack.top().end);
-   m_builder.SetInsertPoint(unr);
+   Value *res = m_builder.CreateSub(in1, in2, name("sub"));
+   return res;
 }
 
 llvm::Value * Instructions::trunc(llvm::Value *in)
@@ -741,18 +846,298 @@ llvm::Value * Instructions::trunc(llvm::Value *in)
    return vectorFromVals(fx, fy, fz, fw);
 }
 
-void Instructions::end()
+llvm::Value * Instructions::x2d(llvm::Value *in1, llvm::Value *in2, llvm::Value *in3)
 {
-   m_builder.CreateRetVoid();
+   std::vector<llvm::Value*> vec1 = extractVector(in1);
+   std::vector<llvm::Value*> vec2 = extractVector(in2);
+   std::vector<llvm::Value*> vec3 = extractVector(in3);
+
+   Value *x2x3 = m_builder.CreateMul( vec2[0], vec3[0], name("x2x3"));
+   Value *y2y3 = m_builder.CreateMul( vec2[1], vec3[1], name("y2y3"));
+   Value *x1px2x3 = m_builder.CreateAdd (vec1[0], x2x3, name("x1 + x2x3"));
+   Value *x1px2x3py2y3 = m_builder.CreateAdd (x1px2x3, y2y3, name("x1 + x2x3 + y2y3"));
+
+   Value *x2z3 = m_builder.CreateMul( vec2[0], vec3[2], name("x2z3"));
+   Value *y2w3 = m_builder.CreateMul( vec2[1], vec3[3], name("y2w3"));
+   Value *y1px2z3 = m_builder.CreateAdd (vec1[1], x2z3, name("y1 + x2z3"));
+   Value *y1px2z3py2w3 = m_builder.CreateAdd (y1px2z3, y2w3, name("y1 + x2z3 + y2w3"));
+
+   return vectorFromVals(x1px2x3py2y3, y1px2z3py2w3, x1px2x3py2y3, y1px2z3py2w3);
 }
 
-void Instructions::cal(int label, llvm::Value *input)
+void Instructions::printVector(llvm::Value *val)
 {
+   static const char *frmt = "Vector is [%f, %f, %f, %f]\x0A";
+
+   if (!m_fmtPtr) {
+      Constant *format = ConstantArray::get(frmt, true);
+      ArrayType *arrayTy = ArrayType::get(IntegerType::get(8), strlen(frmt) + 1);
+      GlobalVariable* globalFormat = new GlobalVariable(
+         /*Type=*/arrayTy,
+         /*isConstant=*/true,
+         /*Linkage=*/GlobalValue::InternalLinkage,
+         /*Initializer=*/0, // has initializer, specified below
+         /*Name=*/name(".str"),
+         m_mod);
+      globalFormat->setInitializer(format);
+
+      Constant* const_int0 = Constant::getNullValue(IntegerType::get(32));
+      std::vector<Constant*> const_ptr_21_indices;
+      const_ptr_21_indices.push_back(const_int0);
+      const_ptr_21_indices.push_back(const_int0);
+      m_fmtPtr = ConstantExpr::getGetElementPtr(globalFormat,
+                                                &const_ptr_21_indices[0], const_ptr_21_indices.size());
+   }
+
+   Function *func_printf = m_mod->getFunction("printf");
+   if (!func_printf)
+      func_printf = declarePrintf();
+   assert(func_printf);
+   std::vector<llvm::Value*> vec = extractVector(val);
+   Value *dx = m_builder.CreateFPExt(vec[0], Type::DoubleTy, name("dx"));
+   Value *dy = m_builder.CreateFPExt(vec[1], Type::DoubleTy, name("dy"));
+   Value *dz = m_builder.CreateFPExt(vec[2], Type::DoubleTy, name("dz"));
+   Value *dw = m_builder.CreateFPExt(vec[3], Type::DoubleTy, name("dw"));
    std::vector<Value*> params;
-   params.push_back(input);
-   llvm::Function *func = findFunction(label);
+   params.push_back(m_fmtPtr);
+   params.push_back(dx);
+   params.push_back(dy);
+   params.push_back(dz);
+   params.push_back(dw);
+   CallInst *call = m_builder.CreateCall(func_printf, params.begin(), params.end(),
+                                         name("printf"));
+   call->setCallingConv(CallingConv::C);
+   call->setTailCall(true);
+}
 
-   m_builder.CreateCall(func, params.begin(), params.end());
+const char * Instructions::name(const char *prefix)
+{
+   ++m_idx;
+   snprintf(m_name, 32, "%s%d", prefix, m_idx);
+   return m_name;
+}
+
+llvm::Value * Instructions::callCeil(llvm::Value *val)
+{
+   if (!m_llvmCeil) {
+      // predeclare the intrinsic
+      std::vector<const Type*> ceilArgs;
+      ceilArgs.push_back(Type::FloatTy);
+      AttrListPtr ceilPal;
+      FunctionType* ceilType = FunctionType::get(
+         /*Result=*/Type::FloatTy,
+         /*Params=*/ceilArgs,
+         /*isVarArg=*/false);
+      m_llvmCeil = Function::Create(
+         /*Type=*/ceilType,
+         /*Linkage=*/GlobalValue::ExternalLinkage,
+         /*Name=*/"ceilf", m_mod);
+      m_llvmCeil->setCallingConv(CallingConv::C);
+      m_llvmCeil->setAttributes(ceilPal);
+   }
+   CallInst *call =  m_builder.CreateCall(m_llvmCeil, val,
+                                          name("ceilf"));
+   call->setCallingConv(CallingConv::C);
+   call->setTailCall(false);
+   return call;
+}
+
+llvm::Value *Instructions::callFAbs(llvm::Value *val)
+{
+   if (!m_llvmFAbs) {
+      // predeclare the intrinsic
+      std::vector<const Type*> fabsArgs;
+      fabsArgs.push_back(Type::FloatTy);
+      AttrListPtr fabsPal;
+      FunctionType* fabsType = FunctionType::get(
+         /*Result=*/Type::FloatTy,
+         /*Params=*/fabsArgs,
+         /*isVarArg=*/false);
+      m_llvmFAbs = Function::Create(
+         /*Type=*/fabsType,
+         /*Linkage=*/GlobalValue::ExternalLinkage,
+         /*Name=*/"fabs", m_mod);
+      m_llvmFAbs->setCallingConv(CallingConv::C);
+      m_llvmFAbs->setAttributes(fabsPal);
+   }
+   CallInst *call = m_builder.CreateCall(m_llvmFAbs, val,
+                                         name("fabs"));
+   call->setCallingConv(CallingConv::C);
+   call->setTailCall(false);
+   return call;
+}
+
+llvm::Value * Instructions::callFExp(llvm::Value *val)
+{
+   if (!m_llvmFexp) {
+      // predeclare the intrinsic
+      std::vector<const Type*> fexpArgs;
+      fexpArgs.push_back(Type::FloatTy);
+      AttrListPtr fexpPal;
+      FunctionType* fexpType = FunctionType::get(
+         /*Result=*/Type::FloatTy,
+         /*Params=*/fexpArgs,
+         /*isVarArg=*/false);
+      m_llvmFexp = Function::Create(
+         /*Type=*/fexpType,
+         /*Linkage=*/GlobalValue::ExternalLinkage,
+         /*Name=*/"expf", m_mod);
+      m_llvmFexp->setCallingConv(CallingConv::C);
+      m_llvmFexp->setAttributes(fexpPal);
+   }
+   CallInst *call = m_builder.CreateCall(m_llvmFexp, val,
+                                         name("expf"));
+   call->setCallingConv(CallingConv::C);
+   call->setTailCall(false);
+   return call;
+}
+
+llvm::Value * Instructions::callFLog(llvm::Value *val)
+{
+   if (!m_llvmFlog) {
+      // predeclare the intrinsic
+      std::vector<const Type*> flogArgs;
+      flogArgs.push_back(Type::FloatTy);
+      AttrListPtr flogPal;
+      FunctionType* flogType = FunctionType::get(
+         /*Result=*/Type::FloatTy,
+         /*Params=*/flogArgs,
+         /*isVarArg=*/false);
+      m_llvmFlog = Function::Create(
+         /*Type=*/flogType,
+         /*Linkage=*/GlobalValue::ExternalLinkage,
+         /*Name=*/"logf", m_mod);
+      m_llvmFlog->setCallingConv(CallingConv::C);
+      m_llvmFlog->setAttributes(flogPal);
+   }
+   CallInst *call = m_builder.CreateCall(m_llvmFlog, val,
+                                         name("logf"));
+   call->setCallingConv(CallingConv::C);
+   call->setTailCall(false);
+   return call;
+}
+
+llvm::Value * Instructions::callFloor(llvm::Value *val)
+{
+   if (!m_llvmFloor) {
+      // predeclare the intrinsic
+      std::vector<const Type*> floorArgs;
+      floorArgs.push_back(Type::FloatTy);
+      AttrListPtr floorPal;
+      FunctionType* floorType = FunctionType::get(
+         /*Result=*/Type::FloatTy,
+         /*Params=*/floorArgs,
+         /*isVarArg=*/false);
+      m_llvmFloor = Function::Create(
+         /*Type=*/floorType,
+         /*Linkage=*/GlobalValue::ExternalLinkage,
+         /*Name=*/"floorf", m_mod);
+      m_llvmFloor->setCallingConv(CallingConv::C);
+      m_llvmFloor->setAttributes(floorPal);
+   }
+   CallInst *call =  m_builder.CreateCall(m_llvmFloor, val,
+                                          name("floorf"));
+   call->setCallingConv(CallingConv::C);
+   call->setTailCall(false);
+   return call;
+}
+
+llvm::Value *Instructions::callFSqrt(llvm::Value *val)
+{
+   if (!m_llvmFSqrt) {
+      // predeclare the intrinsic
+      std::vector<const Type*> fsqrtArgs;
+      fsqrtArgs.push_back(Type::FloatTy);
+      AttrListPtr fsqrtPal;
+      FunctionType* fsqrtType = FunctionType::get(
+         /*Result=*/Type::FloatTy,
+         /*Params=*/fsqrtArgs,
+         /*isVarArg=*/false);
+      m_llvmFSqrt = Function::Create(
+         /*Type=*/fsqrtType,
+         /*Linkage=*/GlobalValue::ExternalLinkage,
+         /*Name=*/"llvm.sqrt.f32", m_mod);
+      m_llvmFSqrt->setCallingConv(CallingConv::C);
+      m_llvmFSqrt->setAttributes(fsqrtPal);
+   }
+   CallInst *call = m_builder.CreateCall(m_llvmFSqrt, val,
+                                         name("sqrt"));
+   call->setCallingConv(CallingConv::C);
+   call->setTailCall(false);
+   return call;
+}
+
+llvm::Value * Instructions::callPow(llvm::Value *val1, llvm::Value *val2)
+{
+   if (!m_llvmPow) {
+      // predeclare the intrinsic
+      std::vector<const Type*> powArgs;
+      powArgs.push_back(Type::FloatTy);
+      powArgs.push_back(Type::FloatTy);
+      AttrListPtr powPal;
+      FunctionType* powType = FunctionType::get(
+         /*Result=*/Type::FloatTy,
+         /*Params=*/powArgs,
+         /*isVarArg=*/false);
+      m_llvmPow = Function::Create(
+         /*Type=*/powType,
+         /*Linkage=*/GlobalValue::ExternalLinkage,
+         /*Name=*/"llvm.pow.f32", m_mod);
+      m_llvmPow->setCallingConv(CallingConv::C);
+      m_llvmPow->setAttributes(powPal);
+   }
+   std::vector<Value*> params;
+   params.push_back(val1);
+   params.push_back(val2);
+   CallInst *call = m_builder.CreateCall(m_llvmPow, params.begin(), params.end(),
+                                         name("pow"));
+   call->setCallingConv(CallingConv::C);
+   call->setTailCall(false);
+   return call;
+}
+
+llvm::Value * Instructions::vectorFromVals(llvm::Value *x, llvm::Value *y,
+                                           llvm::Value *z, llvm::Value *w)
+{
+   Constant *const_vec = Constant::getNullValue(m_floatVecType);
+   Value *res = m_builder.CreateInsertElement(const_vec, x,
+                                              m_storage->constantInt(0),
+                                              name("vecx"));
+   res = m_builder.CreateInsertElement(res, y, m_storage->constantInt(1),
+                               name("vecxy"));
+   res = m_builder.CreateInsertElement(res, z, m_storage->constantInt(2),
+                               name("vecxyz"));
+   if (w)
+      res = m_builder.CreateInsertElement(res, w, m_storage->constantInt(3),
+                                          name("vecxyzw"));
+   return res;
+}
+
+llvm::Value * Instructions::constVector(float x, float y, float z, float w)
+{
+   std::vector<Constant*> vec(4);
+   vec[0] = ConstantFP::get(APFloat(x));
+   vec[1] = ConstantFP::get(APFloat(y));
+   vec[2] = ConstantFP::get(APFloat(z));
+   vec[3] = ConstantFP::get(APFloat(w));
+   return ConstantVector::get(m_floatVecType, vec);
+}
+
+llvm::Function * Instructions::declarePrintf()
+{
+   std::vector<const Type*> args;
+   AttrListPtr params;
+   FunctionType* funcTy = FunctionType::get(
+      /*Result=*/IntegerType::get(32),
+      /*Params=*/args,
+      /*isVarArg=*/true);
+   Function* func_printf = Function::Create(
+      /*Type=*/funcTy,
+      /*Linkage=*/GlobalValue::ExternalLinkage,
+      /*Name=*/"printf", m_mod);
+   func_printf->setCallingConv(CallingConv::C);
+   func_printf->setAttributes(params);
+   return func_printf;
 }
 
 llvm::Function * Instructions::declareFunc(int label)
@@ -763,7 +1148,7 @@ llvm::Function * Instructions::declareFunc(int label)
    args.push_back(vecPtr);
    args.push_back(vecPtr);
    args.push_back(vecPtr);
-   PAListPtr params;
+   AttrListPtr params;
    FunctionType *funcType = FunctionType::get(
       /*Result=*/Type::VoidTy,
       /*Params=*/args,
@@ -774,31 +1159,10 @@ llvm::Function * Instructions::declareFunc(int label)
       /*Linkage=*/GlobalValue::ExternalLinkage,
       /*Name=*/name.c_str(), m_mod);
    func->setCallingConv(CallingConv::C);
-   func->setParamAttrs(params);
+   func->setAttributes(params);
    return func;
 }
 
-void Instructions::bgnSub(unsigned label)
-{
-   llvm::Function *func = findFunction(label);
-
-   Function::arg_iterator args = func->arg_begin();
-   Value *ptr_INPUT = args++;
-   ptr_INPUT->setName("INPUT");
-   m_storage->pushArguments(ptr_INPUT);
-
-   llvm::BasicBlock *entry = BasicBlock::Create("entry", func, 0);
-
-   m_func = func;
-   m_builder.SetInsertPoint(entry);
-}
-
-void Instructions::endSub()
-{
-   m_func = 0;
-   m_builder.SetInsertPoint(0);
-}
-
 llvm::Function * Instructions::findFunction(int label)
 {
    llvm::Function *func = m_functions[label];
@@ -809,17 +1173,6 @@ llvm::Function * Instructions::findFunction(int label)
    return func;
 }
 
-llvm::Value * Instructions::constVector(float x, float y, float z, float w)
-{
-   std::vector<Constant*> vec(4);
-   vec[0] = ConstantFP::get(APFloat(x));
-   vec[1] = ConstantFP::get(APFloat(y));
-   vec[2] = ConstantFP::get(APFloat(z));
-   vec[3] = ConstantFP::get(APFloat(w));
-   return ConstantVector::get(m_floatVecType, vec);
-}
-
-
 std::vector<llvm::Value*> Instructions::extractVector(llvm::Value *vec)
 {
    std::vector<llvm::Value*> elems(4);
@@ -834,69 +1187,7 @@ std::vector<llvm::Value*> Instructions::extractVector(llvm::Value *vec)
    return elems;
 }
 
-llvm::Value * Instructions::cmp(llvm::Value *in1, llvm::Value *in2, llvm::Value *in3)
-{
-   llvm::Function *func = m_mod->getFunction("cmp");
-   assert(func);
-
-   std::vector<Value*> params;
-   params.push_back(in1);
-   params.push_back(in2);
-   params.push_back(in3);
-   CallInst *call = m_builder.CreateCall(func, params.begin(), params.end(), name("cmpres"));
-   call->setTailCall(false);
-   return call;
-}
-
-llvm::Value * Instructions::cos(llvm::Value *in)
-{
-#if 0
-   llvm::Function *func = m_mod->getFunction("vcos");
-   assert(func);
 
-   CallInst *call = m_builder.CreateCall(func, in, name("cosres"));
-   call->setTailCall(false);
-   return call;
-#else
-   std::vector<llvm::Value*> elems = extractVector(in);
-   Function *func = m_mod->getFunction("cosf");
-   assert(func);
-   CallInst *cos = m_builder.CreateCall(func, elems[0], name("cosres"));
-   cos->setCallingConv(CallingConv::C);
-   cos->setTailCall(true);
-   return vectorFromVals(cos, cos, cos, cos);
-#endif
-}
-
-llvm::Value * Instructions::scs(llvm::Value *in)
-{
-   llvm::Function *func = m_mod->getFunction("scs");
-   assert(func);
-
-   CallInst *call = m_builder.CreateCall(func, in, name("scsres"));
-   call->setTailCall(false);
-   return call;
-}
-
-llvm::Value * Instructions::kil(llvm::Value *in)
-{
-   llvm::Function *func = m_mod->getFunction("kil");
-   assert(func);
-
-   CallInst *call = m_builder.CreateCall(func, in, name("kilpres"));
-   call->setTailCall(false);
-   return call;
-}
-
-llvm::Value * Instructions::sin(llvm::Value *in)
-{
-   llvm::Function *func = m_mod->getFunction("vsin");
-   assert(func);
-
-   CallInst *call = m_builder.CreateCall(func, in, name("sinres"));
-   call->setTailCall(false);
-   return call;
-}
 #endif //MESA_LLVM
 
 
diff --git a/src/gallium/auxiliary/gallivm/instructions.h b/src/gallium/auxiliary/gallivm/instructions.h
index d286ce80c78..e18571251ee 100644
--- a/src/gallium/auxiliary/gallivm/instructions.h
+++ b/src/gallium/auxiliary/gallivm/instructions.h
@@ -57,15 +57,24 @@ public:
    llvm::BasicBlock *currentBlock() const;
 
    llvm::Value *abs(llvm::Value *in1);
-   llvm::Value *arl(llvm::Value *in1);
    llvm::Value *add(llvm::Value *in1, llvm::Value *in2);
+   llvm::Value *arl(llvm::Value *in1);
    void         beginLoop();
    void         bgnSub(unsigned);
    void         brk();
    void         cal(int label, llvm::Value *input);
+   llvm::Value *ceil(llvm::Value *in);
+   llvm::Value *clamp(llvm::Value *in);
    llvm::Value *cmp(llvm::Value *in1, llvm::Value *in2, llvm::Value *in3);
+   llvm::Value *cnd(llvm::Value *in1, llvm::Value *in2, llvm::Value *in3);
+   llvm::Value *cnd0(llvm::Value *in1, llvm::Value *in2, llvm::Value *in3);
    llvm::Value *cos(llvm::Value *in);
    llvm::Value *cross(llvm::Value *in1, llvm::Value *in2);
+   llvm::Value *ddx(llvm::Value *in);
+   llvm::Value *ddy(llvm::Value *in);
+   llvm::Value *div(llvm::Value *in1, llvm::Value *in2);
+   llvm::Value *dot2add(llvm::Value *in, llvm::Value *in2, llvm::Value *in3);
+   llvm::Value *dp2(llvm::Value *in1, llvm::Value *in2);
    llvm::Value *dp3(llvm::Value *in1, llvm::Value *in2);
    llvm::Value *dp4(llvm::Value *in1, llvm::Value *in2);
    llvm::Value *dph(llvm::Value *in1, llvm::Value *in2);
@@ -75,6 +84,7 @@ public:
    void         endLoop();
    void         end();
    void         endSub();
+   llvm::Value *exp(llvm::Value *in);
    llvm::Value *ex2(llvm::Value *in);
    llvm::Value *floor(llvm::Value *in);
    llvm::Value *frc(llvm::Value *in);
@@ -82,32 +92,43 @@ public:
    llvm::Value *kil(llvm::Value *in);
    llvm::Value *lerp(llvm::Value *in1, llvm::Value *in2,
                      llvm::Value *in3);
-   llvm::Value *lit(llvm::Value *in);
    llvm::Value *lg2(llvm::Value *in);
+   llvm::Value *lit(llvm::Value *in);
+   llvm::Value *log(llvm::Value *in);
    llvm::Value *madd(llvm::Value *in1, llvm::Value *in2,
                      llvm::Value *in3);
-   llvm::Value *min(llvm::Value *in1, llvm::Value *in2);
    llvm::Value *max(llvm::Value *in1, llvm::Value *in2);
+   llvm::Value *min(llvm::Value *in1, llvm::Value *in2);
    llvm::Value *mul(llvm::Value *in1, llvm::Value *in2);
+   llvm::Value *neg(llvm::Value *in);
+   llvm::Value *nrm(llvm::Value *in);
    llvm::Value *pow(llvm::Value *in1, llvm::Value *in2);
    llvm::Value *rcp(llvm::Value *in);
    llvm::Value *rsq(llvm::Value *in);
    llvm::Value *scs(llvm::Value *in);
+   llvm::Value *seq(llvm::Value *in1, llvm::Value *in2);
+   llvm::Value *sfl(llvm::Value *in1, llvm::Value *in2);
    llvm::Value *sge(llvm::Value *in1, llvm::Value *in2);
    llvm::Value *sgt(llvm::Value *in1, llvm::Value *in2);
    llvm::Value *sin(llvm::Value *in);
+   llvm::Value *sle(llvm::Value *in1, llvm::Value *in2);
    llvm::Value *slt(llvm::Value *in1, llvm::Value *in2);
+   llvm::Value *sne(llvm::Value *in1, llvm::Value *in2);
+   llvm::Value *str(llvm::Value *in1, llvm::Value *in2);
    llvm::Value *sub(llvm::Value *in1, llvm::Value *in2);
    llvm::Value *trunc(llvm::Value *in);
+   llvm::Value *x2d(llvm::Value *in1, llvm::Value *in2, llvm::Value *in3);
 
    void printVector(llvm::Value *val);
 private:
    const char *name(const char *prefix);
 
+   llvm::Value *callCeil(llvm::Value *val);
    llvm::Value *callFAbs(llvm::Value *val);
+   llvm::Value *callFExp(llvm::Value *val);
+   llvm::Value *callFLog(llvm::Value *val);
    llvm::Value *callFloor(llvm::Value *val);
    llvm::Value *callFSqrt(llvm::Value *val);
-   llvm::Value *callFLog(llvm::Value *val);
    llvm::Value *callPow(llvm::Value *val1, llvm::Value *val2);
 
    llvm::Value *vectorFromVals(llvm::Value *x, llvm::Value *y,
@@ -125,16 +146,18 @@ private:
    llvm::Module             *m_mod;
    llvm::Function           *m_func;
    char                      m_name[32];
-   llvm::IRBuilder           m_builder;
+   llvm::IRBuilder<>         m_builder;
    int                       m_idx;
 
    llvm::VectorType *m_floatVecType;
 
+   llvm::Function   *m_llvmCeil;
    llvm::Function   *m_llvmFSqrt;
    llvm::Function   *m_llvmFAbs;
    llvm::Function   *m_llvmPow;
    llvm::Function   *m_llvmFloor;
    llvm::Function   *m_llvmFlog;
+   llvm::Function   *m_llvmFexp;
    llvm::Function   *m_llvmLit;
 
    llvm::Constant   *m_fmtPtr;
diff --git a/src/gallium/auxiliary/gallivm/instructionssoa.cpp b/src/gallium/auxiliary/gallivm/instructionssoa.cpp
index efddc04e818..d5600fd22da 100644
--- a/src/gallium/auxiliary/gallivm/instructionssoa.cpp
+++ b/src/gallium/auxiliary/gallivm/instructionssoa.cpp
@@ -90,68 +90,11 @@ llvm::Value * InstructionsSoa::vectorFromVals(llvm::Value *x, llvm::Value *y,
    return res;
 }
 
-std::vector<llvm::Value*> InstructionsSoa::arl(const std::vector<llvm::Value*> in)
-{
-   std::vector<llvm::Value*> res(4);
-
-   //Extract x's
-   llvm::Value *x1 = m_builder.CreateExtractElement(in[0],
-                                                    m_storage->constantInt(0),
-                                                    name("extractX"));
-   //cast it to an unsigned int
-   x1 = m_builder.CreateFPToUI(x1, IntegerType::get(32), name("x1IntCast"));
-
-   res[0] = x1;//vectorFromVals(x1, x2, x3, x4);
-   //only x is valid. the others shouldn't be necessary
-   /*
-   res[1] = Constant::getNullValue(m_floatVecType);
-   res[2] = Constant::getNullValue(m_floatVecType);
-   res[3] = Constant::getNullValue(m_floatVecType);
-   */
-
-   return res;
-}
-
-
-std::vector<llvm::Value*> InstructionsSoa::add(const std::vector<llvm::Value*> in1,
-                                               const std::vector<llvm::Value*> in2)
-{
-   std::vector<llvm::Value*> res(4);
-
-   res[0] = m_builder.CreateAdd(in1[0], in2[0], name("addx"));
-   res[1] = m_builder.CreateAdd(in1[1], in2[1], name("addy"));
-   res[2] = m_builder.CreateAdd(in1[2], in2[2], name("addz"));
-   res[3] = m_builder.CreateAdd(in1[3], in2[3], name("addw"));
-
-   return res;
-}
-
-std::vector<llvm::Value*> InstructionsSoa::mul(const std::vector<llvm::Value*> in1,
-                                               const std::vector<llvm::Value*> in2)
-{
-   std::vector<llvm::Value*> res(4);
-
-   res[0] = m_builder.CreateMul(in1[0], in2[0], name("mulx"));
-   res[1] = m_builder.CreateMul(in1[1], in2[1], name("muly"));
-   res[2] = m_builder.CreateMul(in1[2], in2[2], name("mulz"));
-   res[3] = m_builder.CreateMul(in1[3], in2[3], name("mulw"));
-
-   return res;
-}
-
 void InstructionsSoa::end()
 {
    m_builder.CreateRetVoid();
 }
 
-std::vector<llvm::Value*> InstructionsSoa::madd(const std::vector<llvm::Value*> in1,
-                                                const std::vector<llvm::Value*> in2,
-                                                const std::vector<llvm::Value*> in3)
-{
-   std::vector<llvm::Value*> res = mul(in1, in2);
-   return add(res, in3);
-}
-
 std::vector<llvm::Value*> InstructionsSoa::extractVector(llvm::Value *vector)
 {
    std::vector<llvm::Value*> res(4);
@@ -171,6 +114,11 @@ std::vector<llvm::Value*> InstructionsSoa::extractVector(llvm::Value *vector)
    return res;
 }
 
+llvm::IRBuilder<>* InstructionsSoa::getIRBuilder()
+{
+   return &m_builder;
+}
+
 void InstructionsSoa::createFunctionMap()
 {
    m_functionsMap[TGSI_OPCODE_ABS]   = "abs";
@@ -181,6 +129,7 @@ void InstructionsSoa::createFunctionMap()
    m_functionsMap[TGSI_OPCODE_POWER] = "pow";
    m_functionsMap[TGSI_OPCODE_LIT]   = "lit";
    m_functionsMap[TGSI_OPCODE_RSQ]   = "rsq";
+   m_functionsMap[TGSI_OPCODE_SLT]   = "slt";
 }
 
 void InstructionsSoa::createDependencies()
@@ -273,6 +222,41 @@ std::vector<llvm::Value*> InstructionsSoa::abs(const std::vector<llvm::Value*> i
    return callBuiltin(func, in1);
 }
 
+std::vector<llvm::Value*> InstructionsSoa::add(const std::vector<llvm::Value*> in1,
+                                               const std::vector<llvm::Value*> in2)
+{
+   std::vector<llvm::Value*> res(4);
+
+   res[0] = m_builder.CreateAdd(in1[0], in2[0], name("addx"));
+   res[1] = m_builder.CreateAdd(in1[1], in2[1], name("addy"));
+   res[2] = m_builder.CreateAdd(in1[2], in2[2], name("addz"));
+   res[3] = m_builder.CreateAdd(in1[3], in2[3], name("addw"));
+
+   return res;
+}
+
+std::vector<llvm::Value*> InstructionsSoa::arl(const std::vector<llvm::Value*> in)
+{
+   std::vector<llvm::Value*> res(4);
+
+   //Extract x's
+   llvm::Value *x1 = m_builder.CreateExtractElement(in[0],
+                                                    m_storage->constantInt(0),
+                                                    name("extractX"));
+   //cast it to an unsigned int
+   x1 = m_builder.CreateFPToUI(x1, IntegerType::get(32), name("x1IntCast"));
+
+   res[0] = x1;//vectorFromVals(x1, x2, x3, x4);
+   //only x is valid. the others shouldn't be necessary
+   /*
+   res[1] = Constant::getNullValue(m_floatVecType);
+   res[2] = Constant::getNullValue(m_floatVecType);
+   res[3] = Constant::getNullValue(m_floatVecType);
+   */
+
+   return res;
+}
+
 std::vector<llvm::Value*> InstructionsSoa::dp3(const std::vector<llvm::Value*> in1,
                                                const std::vector<llvm::Value*> in2)
 {
@@ -280,6 +264,98 @@ std::vector<llvm::Value*> InstructionsSoa::dp3(const std::vector<llvm::Value*> i
    return callBuiltin(func, in1, in2);
 }
 
+std::vector<llvm::Value*> InstructionsSoa::lit(const std::vector<llvm::Value*> in)
+{
+   llvm::Function *func = function(TGSI_OPCODE_LIT);
+   return callBuiltin(func, in);
+}
+
+std::vector<llvm::Value*> InstructionsSoa::madd(const std::vector<llvm::Value*> in1,
+                                                const std::vector<llvm::Value*> in2,
+                                                const std::vector<llvm::Value*> in3)
+{
+   std::vector<llvm::Value*> res = mul(in1, in2);
+   return add(res, in3);
+}
+
+std::vector<llvm::Value*> InstructionsSoa::max(const std::vector<llvm::Value*> in1,
+                                               const std::vector<llvm::Value*> in2)
+{
+   llvm::Function *func = function(TGSI_OPCODE_MAX);
+   return callBuiltin(func, in1, in2);
+}
+
+std::vector<llvm::Value*> InstructionsSoa::min(const std::vector<llvm::Value*> in1,
+                                               const std::vector<llvm::Value*> in2)
+{
+   llvm::Function *func = function(TGSI_OPCODE_MIN);
+   return callBuiltin(func, in1, in2);
+}
+
+std::vector<llvm::Value*> InstructionsSoa::mul(const std::vector<llvm::Value*> in1,
+                                               const std::vector<llvm::Value*> in2)
+{
+   std::vector<llvm::Value*> res(4);
+
+   res[0] = m_builder.CreateMul(in1[0], in2[0], name("mulx"));
+   res[1] = m_builder.CreateMul(in1[1], in2[1], name("muly"));
+   res[2] = m_builder.CreateMul(in1[2], in2[2], name("mulz"));
+   res[3] = m_builder.CreateMul(in1[3], in2[3], name("mulw"));
+
+   return res;
+}
+
+std::vector<llvm::Value*> InstructionsSoa::pow(const std::vector<llvm::Value*> in1,
+                                               const std::vector<llvm::Value*> in2)
+{
+   llvm::Function *func = function(TGSI_OPCODE_POWER);
+   return callBuiltin(func, in1, in2);
+}
+
+std::vector<llvm::Value*> InstructionsSoa::rsq(const std::vector<llvm::Value*> in)
+{
+   llvm::Function *func = function(TGSI_OPCODE_RSQ);
+   return callBuiltin(func, in);
+}
+
+std::vector<llvm::Value*> InstructionsSoa::slt(const std::vector<llvm::Value*> in1,
+                                               const std::vector<llvm::Value*> in2)
+{
+   llvm::Function *func = function(TGSI_OPCODE_SLT);
+   return callBuiltin(func, in1, in2);
+}
+
+std::vector<llvm::Value*> InstructionsSoa::sub(const std::vector<llvm::Value*> in1,
+                                               const std::vector<llvm::Value*> in2)
+{
+   std::vector<llvm::Value*> res(4);
+
+   res[0] = m_builder.CreateSub(in1[0], in2[0], name("subx"));
+   res[1] = m_builder.CreateSub(in1[1], in2[1], name("suby"));
+   res[2] = m_builder.CreateSub(in1[2], in2[2], name("subz"));
+   res[3] = m_builder.CreateSub(in1[3], in2[3], name("subw"));
+
+   return res;
+}
+
+void checkFunction(Function *func)
+{
+   for (Function::const_iterator BI = func->begin(), BE = func->end();
+        BI != BE; ++BI) {
+      const BasicBlock &BB = *BI;
+      for (BasicBlock::const_iterator II = BB.begin(), IE = BB.end();
+           II != IE; ++II) {
+         const Instruction &I = *II;
+         std::cout<< "Instr = "<<I;
+         for (unsigned op = 0, E = I.getNumOperands(); op != E; ++op) {
+            const Value *Op = I.getOperand(op);
+            std::cout<< "\top = "<<Op<<"("<<op<<")"<<std::endl;
+            //I->setOperand(op, V);
+  }
+      }
+   }
+}
+
 llvm::Value * InstructionsSoa::allocaTemp()
 {
    VectorType *vector   = VectorType::get(Type::FloatTy, 4);
@@ -399,46 +475,6 @@ std::vector<Value*> InstructionsSoa::callBuiltin(llvm::Function *func, const std
    return allocaToResult(allocaPtr);
 }
 
-std::vector<llvm::Value*> InstructionsSoa::pow(const std::vector<llvm::Value*> in1,
-                                               const std::vector<llvm::Value*> in2)
-{
-   llvm::Function *func = function(TGSI_OPCODE_POWER);
-   return callBuiltin(func, in1, in2);
-}
-
-std::vector<llvm::Value*> InstructionsSoa::min(const std::vector<llvm::Value*> in1,
-                                               const std::vector<llvm::Value*> in2)
-{
-   llvm::Function *func = function(TGSI_OPCODE_MIN);
-   return callBuiltin(func, in1, in2);
-}
-
-
-std::vector<llvm::Value*> InstructionsSoa::max(const std::vector<llvm::Value*> in1,
-                                               const std::vector<llvm::Value*> in2)
-{
-   llvm::Function *func = function(TGSI_OPCODE_MAX);
-   return callBuiltin(func, in1, in2);
-}
-
-void checkFunction(Function *func)
-{
-   for (Function::const_iterator BI = func->begin(), BE = func->end();
-        BI != BE; ++BI) {
-      const BasicBlock &BB = *BI;
-      for (BasicBlock::const_iterator II = BB.begin(), IE = BB.end();
-           II != IE; ++II) {
-         const Instruction &I = *II;
-         std::cout<< "Instr = "<<I;
-         for (unsigned op = 0, E = I.getNumOperands(); op != E; ++op) {
-            const Value *Op = I.getOperand(op);
-            std::cout<< "\top = "<<Op<<"("<<op<<")"<<std::endl;
-            //I->setOperand(op, V);
-  }
-      }
-   }
-}
-
 void InstructionsSoa::injectFunction(llvm::Function *originalFunc, int op)
 {
    assert(originalFunc);
@@ -458,8 +494,8 @@ void InstructionsSoa::injectFunction(llvm::Function *originalFunc, int op)
       func = Function::Create(originalFunc->getFunctionType(), GlobalValue::ExternalLinkage,
                               originalFunc->getName(), currentModule());
       func->setCallingConv(CallingConv::C);
-      const PAListPtr pal;
-      func->setParamAttrs(pal);
+      const AttrListPtr pal;
+      func->setAttributes(pal);
       currentModule()->dump();
    } else {
       DenseMap<const Value*, Value *> val;
@@ -483,28 +519,4 @@ void InstructionsSoa::injectFunction(llvm::Function *originalFunc, int op)
    }
 }
 
-std::vector<llvm::Value*> InstructionsSoa::sub(const std::vector<llvm::Value*> in1,
-                                               const std::vector<llvm::Value*> in2)
-{
-   std::vector<llvm::Value*> res(4);
-
-   res[0] = m_builder.CreateSub(in1[0], in2[0], name("subx"));
-   res[1] = m_builder.CreateSub(in1[1], in2[1], name("suby"));
-   res[2] = m_builder.CreateSub(in1[2], in2[2], name("subz"));
-   res[3] = m_builder.CreateSub(in1[3], in2[3], name("subw"));
-
-   return res;
-}
-
-std::vector<llvm::Value*> InstructionsSoa::lit(const std::vector<llvm::Value*> in)
-{
-   llvm::Function *func = function(TGSI_OPCODE_LIT);
-   return callBuiltin(func, in);
-}
-
-std::vector<llvm::Value*> InstructionsSoa::rsq(const std::vector<llvm::Value*> in)
-{
-   llvm::Function *func = function(TGSI_OPCODE_RSQ);
-   return callBuiltin(func, in);
-}
 
diff --git a/src/gallium/auxiliary/gallivm/instructionssoa.h b/src/gallium/auxiliary/gallivm/instructionssoa.h
index 3e20b652dd3..d6831e0a6b9 100644
--- a/src/gallium/auxiliary/gallivm/instructionssoa.h
+++ b/src/gallium/auxiliary/gallivm/instructionssoa.h
@@ -69,11 +69,14 @@ public:
    std::vector<llvm::Value*> pow(const std::vector<llvm::Value*> in1,
                                  const std::vector<llvm::Value*> in2);
    std::vector<llvm::Value*> rsq(const std::vector<llvm::Value*> in1);
+   std::vector<llvm::Value*> slt(const std::vector<llvm::Value*> in1,
+                                 const std::vector<llvm::Value*> in2);
    std::vector<llvm::Value*> sub(const std::vector<llvm::Value*> in1,
                                  const std::vector<llvm::Value*> in2);
    void         end();
 
    std::vector<llvm::Value*> extractVector(llvm::Value *vector);
+   llvm::IRBuilder<>*  getIRBuilder();
 private:
    const char * name(const char *prefix) const;
    llvm::Value *vectorFromVals(llvm::Value *x, llvm::Value *y,
@@ -96,7 +99,7 @@ private:
                                          const std::vector<llvm::Value*> in3);
    void injectFunction(llvm::Function *originalFunc, int op = TGSI_OPCODE_LAST);
 private:
-   llvm::IRBuilder  m_builder;
+   llvm::IRBuilder<>  m_builder;
    StorageSoa *m_storage;
 
    std::map<int, std::string> m_functionsMap;
diff --git a/src/gallium/auxiliary/gallivm/soabuiltins.c b/src/gallium/auxiliary/gallivm/soabuiltins.c
index 78f84510e29..cb85e1734ec 100644
--- a/src/gallium/auxiliary/gallivm/soabuiltins.c
+++ b/src/gallium/auxiliary/gallivm/soabuiltins.c
@@ -36,6 +36,8 @@ typedef __attribute__(( ext_vector_type(4) )) float float4;
 
 extern float fabsf(float val);
 
+/* helpers */
+
 float4 absvec(float4 vec)
 {
    float4 res;
@@ -47,6 +49,58 @@ float4 absvec(float4 vec)
    return res;
 }
 
+float4 maxvec(float4 a, float4 b)
+{
+   return (float4){(a.x > b.x) ? a.x : b.x,
+         (a.y > b.y) ? a.y : b.y,
+         (a.z > b.z) ? a.z : b.z,
+         (a.w > b.w) ? a.w : b.w};
+}
+
+float4 minvec(float4 a, float4 b)
+{
+   return (float4){(a.x < b.x) ? a.x : b.x,
+         (a.y < b.y) ? a.y : b.y,
+         (a.z < b.z) ? a.z : b.z,
+         (a.w < b.w) ? a.w : b.w};
+}
+
+extern float powf(float num, float p);
+extern float sqrtf(float x);
+
+float4 powvec(float4 vec, float4 q)
+{
+   float4 p;
+   p.x = powf(vec.x, q.x);
+   p.y = powf(vec.y, q.y);
+   p.z = powf(vec.z, q.z);
+   p.w = powf(vec.w, q.w);
+   return p;
+}
+
+float4 sqrtvec(float4 vec)
+{
+   float4 p;
+   p.x = sqrtf(vec.x);
+   p.y = sqrtf(vec.y);
+   p.z = sqrtf(vec.z);
+   p.w = sqrtf(vec.w);
+   return p;
+}
+
+float4 sltvec(float4 v1, float4 v2)
+{
+   float4 p;
+   p.x = (v1.x < v2.x) ? 1.0 : 0.0;
+   p.y = (v1.y < v2.y) ? 1.0 : 0.0;
+   p.z = (v1.z < v2.z) ? 1.0 : 0.0;
+   p.w = (v1.w < v2.w) ? 1.0 : 0.0;
+   return p;
+}
+
+
+/* instructions */
+
 void abs(float4 *res,
          float4 tmp0x, float4 tmp0y, float4 tmp0z, float4 tmp0w)
 {
@@ -69,7 +123,6 @@ void dp3(float4 *res,
    res[3] = dot;
 }
 
-
 void dp4(float4 *res,
          float4 tmp0x, float4 tmp0y, float4 tmp0z, float4 tmp0w,
          float4 tmp1x, float4 tmp1y, float4 tmp1z, float4 tmp1w)
@@ -83,35 +136,25 @@ void dp4(float4 *res,
    res[3] = dot;
 }
 
-extern float powf(float num, float p);
-extern float sqrtf(float x);
-
-float4 powvec(float4 vec, float4 q)
-{
-   float4 p;
-   p.x = powf(vec.x, q.x);
-   p.y = powf(vec.y, q.y);
-   p.z = powf(vec.z, q.z);
-   p.w = powf(vec.w, q.w);
-   return p;
-}
-
-void pow(float4 *res,
-         float4 tmp0x, float4 tmp0y, float4 tmp0z, float4 tmp0w,
-         float4 tmp1x, float4 tmp1y, float4 tmp1z, float4 tmp1w)
+void lit(float4 *res,
+         float4 tmp0x, float4 tmp0y, float4 tmp0z, float4 tmp0w)
 {
-   res[0] = powvec(tmp0x, tmp1x);
-   res[1] = res[0];
-   res[2] = res[0];
-   res[3] = res[0];
-}
+   const float4 zerovec = (float4) {0.0, 0.0, 0.0, 0.0};
+   const float4 min128 = (float4) {-128.f, -128.f, -128.f, -128.f};
+   const float4 plus128 = (float4) {128.f,  128.f,  128.f,  128.f};
 
-float4 minvec(float4 a, float4 b)
-{
-   return (float4){(a.x < b.x) ? a.x : b.x,
-         (a.y < b.y) ? a.y : b.y,
-         (a.z < b.z) ? a.z : b.z,
-         (a.w < b.w) ? a.w : b.w};
+   res[0] = (float4){1.0, 1.0, 1.0, 1.0};
+   if (tmp0x.x > 0) {
+      float4 tmpy = maxvec(tmp0y, zerovec);
+      float4 tmpw = minvec(tmp0w, plus128);
+      tmpw = maxvec(tmpw, min128);
+      res[1] = tmp0x;
+      res[2] = powvec(tmpy, tmpw);
+   } else {
+      res[1] = zerovec;
+      res[2] = zerovec;
+   }
+   res[3] = (float4){1.0, 1.0, 1.0, 1.0};
 }
 
 void min(float4 *res,
@@ -125,14 +168,6 @@ void min(float4 *res,
 }
 
 
-float4 maxvec(float4 a, float4 b)
-{
-   return (float4){(a.x > b.x) ? a.x : b.x,
-         (a.y > b.y) ? a.y : b.y,
-         (a.z > b.z) ? a.z : b.z,
-         (a.w > b.w) ? a.w : b.w};
-}
-
 void max(float4 *res,
          float4 tmp0x, float4 tmp0y, float4 tmp0z, float4 tmp0w,
          float4 tmp1x, float4 tmp1y, float4 tmp1z, float4 tmp1w)
@@ -143,37 +178,14 @@ void max(float4 *res,
    res[3] = maxvec(tmp0w, tmp1w);
 }
 
-
-void lit(float4 *res,
-         float4 tmp0x, float4 tmp0y, float4 tmp0z, float4 tmp0w)
-{
-   const float4 zerovec = (float4) {0.0, 0.0, 0.0, 0.0};
-   const float4 min128 = (float4) {-128.f, -128.f, -128.f, -128.f};
-   const float4 plus128 = (float4) {128.f,  128.f,  128.f,  128.f};
-
-   res[0] = (float4){1.0, 1.0, 1.0, 1.0};
-   if (tmp0x.x > 0) {
-      float4 tmpy = maxvec(tmp0y, zerovec);
-      float4 tmpw = minvec(tmp0w, plus128);
-      tmpw = maxvec(tmpw, min128);
-      res[1] = tmp0x;
-      res[2] = powvec(tmpy, tmpw);
-   } else {
-      res[1] = zerovec;
-      res[2] = zerovec;
-   }
-   res[3] = (float4){1.0, 1.0, 1.0, 1.0};
-}
-
-
-float4 sqrtvec(float4 vec)
+void pow(float4 *res,
+         float4 tmp0x, float4 tmp0y, float4 tmp0z, float4 tmp0w,
+         float4 tmp1x, float4 tmp1y, float4 tmp1z, float4 tmp1w)
 {
-   float4 p;
-   p.x = sqrtf(vec.x);
-   p.y = sqrtf(vec.y);
-   p.z = sqrtf(vec.z);
-   p.w = sqrtf(vec.w);
-   return p;
+   res[0] = powvec(tmp0x, tmp1x);
+   res[1] = res[0];
+   res[2] = res[0];
+   res[3] = res[0];
 }
 
 void rsq(float4 *res,
@@ -185,3 +197,14 @@ void rsq(float4 *res,
    res[2] = onevec/sqrtvec(absvec(tmp0z));
    res[3] = onevec/sqrtvec(absvec(tmp0w));
 }
+
+void slt(float4 *res,
+         float4 tmp0x, float4 tmp0y, float4 tmp0z, float4 tmp0w,
+         float4 tmp1x, float4 tmp1y, float4 tmp1z, float4 tmp1w)
+{
+   res[0] = sltvec(tmp0x, tmp1x);
+   res[1] = sltvec(tmp0y, tmp1y);
+   res[2] = sltvec(tmp0z, tmp1z);
+   res[3] = sltvec(tmp0w, tmp1w);
+}
+
diff --git a/src/gallium/auxiliary/gallivm/storagesoa.cpp b/src/gallium/auxiliary/gallivm/storagesoa.cpp
index 78d754371f0..4fc075cf6d4 100644
--- a/src/gallium/auxiliary/gallivm/storagesoa.cpp
+++ b/src/gallium/auxiliary/gallivm/storagesoa.cpp
@@ -93,7 +93,7 @@ void StorageSoa::declareImmediates()
       std::vector<float> vals(4);
       std::vector<Constant*> channelArray;
 
-      vals[0] = vec[0]; vals[1] = vec[0]; vals[2] = vec[0]; vals[3] = vec[0];
+      vals[0] = vec[0]; vals[1] = vec[1]; vals[2] = vec[2]; vals[3] = vec[3];
       llvm::Constant *xChannel = createConstGlobalVector(vals);
 
       vals[0] = vec[1]; vals[1] = vec[1]; vals[2] = vec[1]; vals[3] = vec[1];
@@ -144,22 +144,43 @@ std::vector<llvm::Value*> StorageSoa::inputElement(llvm::Value *idx)
    return res;
 }
 
-std::vector<llvm::Value*> StorageSoa::constElement(llvm::Value *idx)
+llvm::Value* StorageSoa::unpackConstElement(llvm::IRBuilder<>* m_builder, llvm::Value* vector, int cc)
 {
-   std::vector<llvm::Value*> res(4);
+   std::vector<llvm::Value*> x(4);
+   x[0] = m_builder->CreateExtractElement(vector,
+                                           constantInt(cc),
+                                           name("x"));
+
+   VectorType  *vectorType = VectorType::get(Type::FloatTy, 4);
+   Constant *constVector = Constant::getNullValue(vectorType);
+   Value *res = m_builder->CreateInsertElement(constVector, x[0],
+                                              constantInt(0),
+                                              name("vecx"));
+   res = m_builder->CreateInsertElement(res, x[0], constantInt(1),
+                               name("vecxx"));
+   res = m_builder->CreateInsertElement(res, x[0], constantInt(2),
+                               name("vecxxx"));
+   res = m_builder->CreateInsertElement(res, x[0], constantInt(3),
+                               name("vecxxxx"));
+   return res;
+}
+
+std::vector<llvm::Value*> StorageSoa::constElement(llvm::IRBuilder<>* m_builder, llvm::Value *idx)
+{
+   llvm::Value* res;
+   std::vector<llvm::Value*> res2(4);
    llvm::Value *xChannel, *yChannel, *zChannel, *wChannel;
 
    xChannel = elementPointer(m_consts, idx, 0);
-   yChannel = elementPointer(m_consts, idx, 1);
-   zChannel = elementPointer(m_consts, idx, 2);
-   wChannel = elementPointer(m_consts, idx, 3);
 
-   res[0] = alignedArrayLoad(xChannel);
-   res[1] = alignedArrayLoad(yChannel);
-   res[2] = alignedArrayLoad(zChannel);
-   res[3] = alignedArrayLoad(wChannel);
+   res = alignedArrayLoad(xChannel);
 
-   return res;
+   res2[0]=unpackConstElement(m_builder, res,0);
+   res2[1]=unpackConstElement(m_builder, res,1);
+   res2[2]=unpackConstElement(m_builder, res,2);
+   res2[3]=unpackConstElement(m_builder, res,3);
+
+   return res2;
 }
 
 std::vector<llvm::Value*> StorageSoa::outputElement(llvm::Value *idx)
@@ -260,6 +281,12 @@ llvm::Module * StorageSoa::currentModule() const
     return m_block->getParent()->getParent();
 }
 
+llvm::Constant * StorageSoa::createConstGlobalFloat(const float val)
+{
+   Constant*c = ConstantFP::get(APFloat(val));
+   return c;
+}
+
 llvm::Constant * StorageSoa::createConstGlobalVector(const std::vector<float> &vec)
 {
    VectorType *vectorType = VectorType::get(Type::FloatTy, 4);
@@ -278,7 +305,7 @@ llvm::Constant * StorageSoa::createConstGlobalVector(const std::vector<float> &v
 }
 
 std::vector<llvm::Value*> StorageSoa::load(enum tgsi_file_type type, int idx, int swizzle,
-                                           llvm::Value *indIdx)
+                                           llvm::IRBuilder<>* m_builder,llvm::Value *indIdx)
 {
    std::vector<llvm::Value*> val(4);
 
@@ -302,7 +329,7 @@ std::vector<llvm::Value*> StorageSoa::load(enum tgsi_file_type type, int idx, in
       val = tempElement(realIndex);
       break;
    case TGSI_FILE_CONSTANT:
-      val = constElement(realIndex);
+      val = constElement(m_builder, realIndex);
       break;
    case TGSI_FILE_IMMEDIATE:
       val = immediateElement(realIndex);
diff --git a/src/gallium/auxiliary/gallivm/storagesoa.h b/src/gallium/auxiliary/gallivm/storagesoa.h
index ae2fc7c6aee..f21ca6ec433 100644
--- a/src/gallium/auxiliary/gallivm/storagesoa.h
+++ b/src/gallium/auxiliary/gallivm/storagesoa.h
@@ -29,6 +29,7 @@
 #define STORAGESOA_H
 
 #include <pipe/p_shader_tokens.h>
+#include <llvm/Support/IRBuilder.h>
 
 #include <vector>
 #include <list>
@@ -56,7 +57,7 @@ public:
 
 
    std::vector<llvm::Value*> load(enum tgsi_file_type type, int idx, int swizzle, 
-                                  llvm::Value *indIdx =0);
+                                  llvm::IRBuilder<>* m_builder, llvm::Value *indIdx =0);
    void store(enum tgsi_file_type type, int idx, const std::vector<llvm::Value*> &val,
               int mask);
 
@@ -76,10 +77,12 @@ private:
    const char *name(const char *prefix) const;
    llvm::Value  *alignedArrayLoad(llvm::Value *val);
    llvm::Module *currentModule() const;
+   llvm::Constant  *createConstGlobalFloat(const float val);
    llvm::Constant  *createConstGlobalVector(const std::vector<float> &vec);
 
    std::vector<llvm::Value*> inputElement(llvm::Value *indIdx);
-   std::vector<llvm::Value*> constElement(llvm::Value *indIdx);
+   llvm::Value* unpackConstElement(llvm::IRBuilder<>* m_builder, llvm::Value *indIdx, int cc);
+   std::vector<llvm::Value*> constElement(llvm::IRBuilder<>* m_builder, llvm::Value *indIdx);
    std::vector<llvm::Value*> outputElement(llvm::Value *indIdx);
    std::vector<llvm::Value*> tempElement(llvm::Value *indIdx);
    std::vector<llvm::Value*> immediateElement(llvm::Value *indIdx);
diff --git a/src/gallium/auxiliary/gallivm/tgsitollvm.cpp b/src/gallium/auxiliary/gallivm/tgsitollvm.cpp
index cc1516a45e0..1191a6cae97 100644
--- a/src/gallium/auxiliary/gallivm/tgsitollvm.cpp
+++ b/src/gallium/auxiliary/gallivm/tgsitollvm.cpp
@@ -52,7 +52,7 @@ static inline FunctionType *vertexShaderFunctionType()
    // pass are castable to the following:
    // [4 x <4 x float>] inputs,
    // [4 x <4 x float>] output,
-   // [4 x [4 x float]] consts,
+   // [4 x [1 x float]] consts,
    // [4 x <4 x float>] temps
 
    std::vector<const Type*> funcArgs;
@@ -61,7 +61,7 @@ static inline FunctionType *vertexShaderFunctionType()
    PointerType *vectorArrayPtr = PointerType::get(vectorArray, 0);
 
    ArrayType   *floatArray     = ArrayType::get(Type::FloatTy, 4);
-   ArrayType   *constsArray    = ArrayType::get(floatArray, 4);
+   ArrayType   *constsArray    = ArrayType::get(floatArray, 1);
    PointerType *constsArrayPtr = PointerType::get(constsArray, 0);
 
    funcArgs.push_back(vectorArrayPtr);//inputs
@@ -246,6 +246,7 @@ translate_instruction(llvm::Module *module,
          val = storage->constElement(src->SrcRegister.Index, indIdx);
       } else if (src->SrcRegister.File == TGSI_FILE_INPUT) {
          val = storage->inputElement(src->SrcRegister.Index, indIdx);
+      // FIXME we should not be generating elements for temporaries, this creates useless memory writes
       } else if (src->SrcRegister.File == TGSI_FILE_TEMPORARY) {
          val = storage->tempElement(src->SrcRegister.Index);
       } else if (src->SrcRegister.File == TGSI_FILE_OUTPUT) {
@@ -286,9 +287,13 @@ translate_instruction(llvm::Module *module,
       out = instr->rsq(inputs[0]);
    }
       break;
-   case TGSI_OPCODE_EXP:
+   case TGSI_OPCODE_EXP: {
+      out = instr->exp(inputs[0]);
+   }
       break;
-   case TGSI_OPCODE_LOG:
+   case TGSI_OPCODE_LOG: {
+      out = instr->log(inputs[0]);
+   }
       break;
    case TGSI_OPCODE_MUL: {
       out = instr->mul(inputs[0], inputs[1]);
@@ -338,21 +343,31 @@ translate_instruction(llvm::Module *module,
       out = instr->lerp(inputs[0], inputs[1], inputs[2]);
    }
       break;
-   case TGSI_OPCODE_CND:
+   case TGSI_OPCODE_CND: {
+      out = instr->cnd(inputs[0], inputs[1], inputs[2]);
+   }
       break;
-   case TGSI_OPCODE_CND0:
+   case TGSI_OPCODE_CND0: {
+      out = instr->cnd0(inputs[0], inputs[1], inputs[2]);
+   }
       break;
-   case TGSI_OPCODE_DOT2ADD:
+   case TGSI_OPCODE_DOT2ADD: {
+      out = instr->dot2add(inputs[0], inputs[1], inputs[2]);
+   }
       break;
    case TGSI_OPCODE_INDEX:
       break;
-   case TGSI_OPCODE_NEGATE:
+   case TGSI_OPCODE_NEGATE: {
+      out = instr->neg(inputs[0]);
+   }
       break;
    case TGSI_OPCODE_FRAC: {
       out = instr->frc(inputs[0]);
    }
       break;
-   case TGSI_OPCODE_CLAMP:
+   case TGSI_OPCODE_CLAMP: {
+      out = instr->clamp(inputs[0]);
+   }
       break;
    case TGSI_OPCODE_FLOOR: {
       out = instr->floor(inputs[0]);
@@ -392,9 +407,13 @@ translate_instruction(llvm::Module *module,
       out = instr->cos(inputs[0]);
    }
       break;
-   case TGSI_OPCODE_DDX:
+   case TGSI_OPCODE_DDX: {
+      out = instr->ddx(inputs[0]);
+   }
       break;
-   case TGSI_OPCODE_DDY:
+   case TGSI_OPCODE_DDY: {
+      out = instr->ddy(inputs[0]);
+   }
       break;
    case TGSI_OPCODE_KILP:
       break;
@@ -408,9 +427,13 @@ translate_instruction(llvm::Module *module,
       break;
    case TGSI_OPCODE_RFL:
       break;
-   case TGSI_OPCODE_SEQ:
+   case TGSI_OPCODE_SEQ: {
+      out = instr->seq(inputs[0], inputs[1]);
+   }
       break;
-   case TGSI_OPCODE_SFL:
+   case TGSI_OPCODE_SFL: {
+      out = instr->sfl(inputs[0], inputs[1]);
+   }
       break;
    case TGSI_OPCODE_SGT: {
       out = instr->sgt(inputs[0], inputs[1]);
@@ -420,11 +443,17 @@ translate_instruction(llvm::Module *module,
       out = instr->sin(inputs[0]);
    }
       break;
-   case TGSI_OPCODE_SLE:
+   case TGSI_OPCODE_SLE: {
+      out = instr->sle(inputs[0], inputs[1]);
+   }
       break;
-   case TGSI_OPCODE_SNE:
+   case TGSI_OPCODE_SNE: {
+      out = instr->sne(inputs[0], inputs[1]);
+   }
       break;
-   case TGSI_OPCODE_STR:
+   case TGSI_OPCODE_STR: {
+      out = instr->str(inputs[0], inputs[1]);
+   }
       break;
    case TGSI_OPCODE_TEX:
       break;
@@ -438,7 +467,9 @@ translate_instruction(llvm::Module *module,
       break;
    case TGSI_OPCODE_UP4UB:
       break;
-   case TGSI_OPCODE_X2D:
+   case TGSI_OPCODE_X2D: {
+      out = instr->x2d(inputs[0], inputs[1], inputs[2]);
+   }
       break;
    case TGSI_OPCODE_ARA:
       break;
@@ -468,11 +499,18 @@ translate_instruction(llvm::Module *module,
       break;
    case TGSI_OPCODE_TXB:
       break;
-   case TGSI_OPCODE_NRM:
+   case TGSI_OPCODE_NRM4:
+   case TGSI_OPCODE_NRM: {
+      out = instr->nrm(inputs[0]);
+   }
       break;
-   case TGSI_OPCODE_DIV:
+   case TGSI_OPCODE_DIV: {
+      out = instr->div(inputs[0], inputs[1]);
+   }
       break;
-   case TGSI_OPCODE_DP2:
+   case TGSI_OPCODE_DP2: {
+      out = instr->dp2(inputs[0], inputs[1]);
+   }
       break;
    case TGSI_OPCODE_TXL:
       break;
@@ -590,8 +628,6 @@ translate_instruction(llvm::Module *module,
       break;
    case TGSI_OPCODE_M3X2:
       break;
-   case TGSI_OPCODE_NRM4:
-      break;
    case TGSI_OPCODE_CALLNZ:
       break;
    case TGSI_OPCODE_IFC:
@@ -641,6 +677,7 @@ translate_instruction(llvm::Module *module,
 
       if (dst->DstRegister.File == TGSI_FILE_OUTPUT) {
          storage->setOutputElement(dst->DstRegister.Index, out, dst->DstRegister.WriteMask);
+      // FIXME we should not be generating elements for temporaries, this creates useless memory writes
       } else if (dst->DstRegister.File == TGSI_FILE_TEMPORARY) {
          storage->setTempElement(dst->DstRegister.Index, out, dst->DstRegister.WriteMask);
       } else if (dst->DstRegister.File == TGSI_FILE_ADDRESS) {
@@ -672,9 +709,8 @@ translate_instructionir(llvm::Module *module,
       if (src->SrcRegister.Indirect) {
          indIdx = storage->addrElement(src->SrcRegisterInd.Index);
       }
-
       val = storage->load((enum tgsi_file_type)src->SrcRegister.File,
-                          src->SrcRegister.Index, swizzle, indIdx);
+                          src->SrcRegister.Index, swizzle, instr->getIRBuilder(), indIdx);
 
       inputs[i] = val;
    }
@@ -732,6 +768,7 @@ translate_instructionir(llvm::Module *module,
    }
       break;
    case TGSI_OPCODE_SLT: {
+      out = instr->slt(inputs[0], inputs[1]);
    }
       break;
    case TGSI_OPCODE_SGE: {
@@ -989,7 +1026,6 @@ translate_instructionir(llvm::Module *module,
    /* store results  */
    for (int i = 0; i < inst->Instruction.NumDstRegs; ++i) {
       struct tgsi_full_dst_register *dst = &inst->FullDstRegisters[i];
-
       storage->store((enum tgsi_file_type)dst->DstRegister.File,
                      dst->DstRegister.Index, out, dst->DstRegister.WriteMask);
    }