-/* $Id: x86_cliptest.S,v 1.6 2001/03/28 20:44:44 gareth Exp $ */
/*
* Mesa 3-D graphics library
* CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
*/
+/*
+ * NOTE: Avoid using spaces in between '(' ')' and arguments, especially
+ * with macros like CONST, LLBL that expand to CONCAT(...). Putting spaces
+ * in there will break the build on some platforms.
+ */
+
+#include "assyntax.h"
#include "matypes.h"
#include "clip_args.h"
-#define FP_ONE 1065353216
-#define FP_ZERO 0
-
-#define SRC(i) REGOFF(i * 4, ESI)
-#define DST(i) REGOFF(i * 4, EDI)
-#define MAT(i) REGOFF(i * 4, EDX)
+#define SRC0 REGOFF(0, ESI)
+#define SRC1 REGOFF(4, ESI)
+#define SRC2 REGOFF(8, ESI)
+#define SRC3 REGOFF(12, ESI)
+#define DST0 REGOFF(0, EDI)
+#define DST1 REGOFF(4, EDI)
+#define DST2 REGOFF(8, EDI)
+#define DST3 REGOFF(12, EDI)
+#define MAT0 REGOFF(0, EDX)
+#define MAT1 REGOFF(4, EDX)
+#define MAT2 REGOFF(8, EDX)
+#define MAT3 REGOFF(12, EDX)
/*
* Table for clip test.
*
- * bit6 = SRC(3) < 0
- * bit5 = SRC(2) < 0
+ * bit6 = SRC3 < 0
+ * bit5 = SRC2 < 0
* bit4 = abs(S(2)) > abs(S(3))
- * bit3 = SRC(1) < 0
+ * bit3 = SRC1 < 0
* bit2 = abs(S(1)) > abs(S(3))
- * bit1 = SRC(0) < 0
+ * bit1 = SRC0 < 0
* bit0 = abs(S(0)) > abs(S(3))
*/
SEG_DATA
clip_table:
- D_BYTE 0, 1, 0, 2, 4, 5, 4, 6
- D_BYTE 0, 1, 0, 2, 8, 9, 8, 10
- D_BYTE 32, 33, 32, 34, 36, 37, 36, 38
- D_BYTE 32, 33, 32, 34, 40, 41, 40, 42
- D_BYTE 0, 1, 0, 2, 4, 5, 4, 6
- D_BYTE 0, 1, 0, 2, 8, 9, 8, 10
- D_BYTE 16, 17, 16, 18, 20, 21, 20, 22
- D_BYTE 16, 17, 16, 18, 24, 25, 24, 26
- D_BYTE 63, 61, 63, 62, 55, 53, 55, 54
- D_BYTE 63, 61, 63, 62, 59, 57, 59, 58
- D_BYTE 47, 45, 47, 46, 39, 37, 39, 38
- D_BYTE 47, 45, 47, 46, 43, 41, 43, 42
- D_BYTE 63, 61, 63, 62, 55, 53, 55, 54
- D_BYTE 63, 61, 63, 62, 59, 57, 59, 58
- D_BYTE 31, 29, 31, 30, 23, 21, 23, 22
- D_BYTE 31, 29, 31, 30, 27, 25, 27, 26
+ D_BYTE 0x00, 0x01, 0x00, 0x02, 0x04, 0x05, 0x04, 0x06
+ D_BYTE 0x00, 0x01, 0x00, 0x02, 0x08, 0x09, 0x08, 0x0a
+ D_BYTE 0x20, 0x21, 0x20, 0x22, 0x24, 0x25, 0x24, 0x26
+ D_BYTE 0x20, 0x21, 0x20, 0x22, 0x28, 0x29, 0x28, 0x2a
+ D_BYTE 0x00, 0x01, 0x00, 0x02, 0x04, 0x05, 0x04, 0x06
+ D_BYTE 0x00, 0x01, 0x00, 0x02, 0x08, 0x09, 0x08, 0x0a
+ D_BYTE 0x10, 0x11, 0x10, 0x12, 0x14, 0x15, 0x14, 0x16
+ D_BYTE 0x10, 0x11, 0x10, 0x12, 0x18, 0x19, 0x18, 0x1a
+ D_BYTE 0x3f, 0x3d, 0x3f, 0x3e, 0x37, 0x35, 0x37, 0x36
+ D_BYTE 0x3f, 0x3d, 0x3f, 0x3e, 0x3b, 0x39, 0x3b, 0x3a
+ D_BYTE 0x2f, 0x2d, 0x2f, 0x2e, 0x27, 0x25, 0x27, 0x26
+ D_BYTE 0x2f, 0x2d, 0x2f, 0x2e, 0x2b, 0x29, 0x2b, 0x2a
+ D_BYTE 0x3f, 0x3d, 0x3f, 0x3e, 0x37, 0x35, 0x37, 0x36
+ D_BYTE 0x3f, 0x3d, 0x3f, 0x3e, 0x3b, 0x39, 0x3b, 0x3a
+ D_BYTE 0x1f, 0x1d, 0x1f, 0x1e, 0x17, 0x15, 0x17, 0x16
+ D_BYTE 0x1f, 0x1d, 0x1f, 0x1e, 0x1b, 0x19, 0x1b, 0x1a
SEG_TEXT
* EBP: temp2
*/
-#if defined(__ELF__) && defined(__PIC__) && !defined(ELFPIC)
+#if defined(__ELF__) && defined(__PIC__) && defined(GNU_ASSEMBLER) && !defined(ELFPIC)
#define ELFPIC
#endif
ALIGNTEXT16
GLOBL GLNAME( _mesa_x86_cliptest_points4 )
+HIDDEN(_mesa_x86_cliptest_points4)
GLNAME( _mesa_x86_cliptest_points4 ):
#ifdef ELFPIC
#ifdef ELFPIC
/* store pointer to clip_table on stack */
- CALL( LLBL( ctp4_get_eip ) )
+ CALL( LLBL(ctp4_get_eip) )
ADD_L( CONST(_GLOBAL_OFFSET_TABLE_), EBX )
MOV_L( REGOFF(clip_table@GOT, EBX), EBX )
PUSH_L( EBX )
- JMP( LLBL( ctp4_clip_table_ready ) )
+ JMP( LLBL(ctp4_clip_table_ready) )
-LLBL( ctp4_get_eip ):
+LLBL(ctp4_get_eip):
/* store eip in ebx */
MOV_L( REGIND(ESP), EBX )
RET
-LLBL( ctp4_clip_table_ready ):
+LLBL(ctp4_clip_table_ready):
#endif
MOV_L( ARG_SOURCE, ESI )
MOV_B( REGIND(EBX), AL )
MOV_B( REGIND(EBP), AH )
- JZ( LLBL( ctp4_finish ) )
+ JZ( LLBL(ctp4_finish) )
ALIGNTEXT16
-LLBL( ctp4_top ):
+LLBL(ctp4_top):
- FLD1 /* F3 */
- FDIV_S( SRC(3) )
+ FLD1 /* F3 */
+ FDIV_S( SRC3 ) /* GH: don't care about div-by-zero */
- MOV_L( SRC(3), EBP )
- MOV_L( SRC(2), EBX )
+ MOV_L( SRC3, EBP )
+ MOV_L( SRC2, EBX )
XOR_L( ECX, ECX )
ADD_L( EBP, EBP ) /* ebp = abs(S(3))*2 ; carry = sign of S(3) */
CMP_L( EBX, EBP ) /* carry = abs(S(2))*2 > abs(S(3))*2 */
ADC_L( ECX, ECX )
- MOV_L( SRC(1), EBX )
+ MOV_L( SRC1, EBX )
ADD_L( EBX, EBX ) /* ebx = abs(S(1))*2 ; carry = sign of S(1) */
CMP_L( EBX, EBP ) /* carry = abs(S(1))*2 > abs(S(3))*2 */
ADC_L( ECX, ECX )
- MOV_L( SRC(0), EBX )
+ MOV_L( SRC0, EBX )
ADD_L( EBX, EBX ) /* ebx = abs(S(0))*2 ; carry = sign of S(0) */
TEST_B( CL, CL )
MOV_B( CL, REGIND(EDX) )
- JZ( LLBL( ctp4_proj ) )
+ JZ( LLBL(ctp4_proj) )
+
+LLBL(ctp4_noproj):
FSTP( ST(0) ) /* */
- JMP( LLBL( ctp4_next ) )
-LLBL( ctp4_proj ):
+ MOV_L( CONST(0), DST0 )
+ MOV_L( CONST(0), DST1 )
+ MOV_L( CONST(0), DST2 )
+ MOV_L( CONST(0x3f800000), DST3 )
+
+ JMP( LLBL(ctp4_next) )
- FLD_S( SRC(0) ) /* F0 F3 */
- FMUL2( ST(1), ST(0) )
+LLBL(ctp4_proj):
- FLD_S( SRC(1) ) /* F1 F0 F3 */
- FMUL2( ST(2), ST(0) )
+ FLD_S( SRC0 ) /* F0 F3 */
+ FMUL2( ST(1), ST0 )
- FLD_S( SRC(2) ) /* F2 F1 F0 F3 */
- FMUL2( ST(3), ST(0) )
+ FLD_S( SRC1 ) /* F1 F0 F3 */
+ FMUL2( ST(2), ST0 )
+
+ FLD_S( SRC2 ) /* F2 F1 F0 F3 */
+ FMUL2( ST(3), ST0 )
FXCH( ST(2) ) /* F0 F1 F2 F3 */
- FSTP_S( DST(0) ) /* F1 F2 F3 */
- FSTP_S( DST(1) ) /* F2 F3 */
- FSTP_S( DST(2) ) /* F3 */
- FSTP_S( DST(3) ) /* */
+ FSTP_S( DST0 ) /* F1 F2 F3 */
+ FSTP_S( DST1 ) /* F2 F3 */
+ FSTP_S( DST2 ) /* F3 */
+ FSTP_S( DST3 ) /* */
-LLBL( ctp4_next ):
+LLBL(ctp4_next):
INC_L( EDX )
ADD_L( CONST(16), EDI )
ADD_L( ARG_SOURCE, ESI )
CMP_L( EDX, ARG_CLIP )
- JNZ( LLBL( ctp4_top ) )
+ JNZ( LLBL(ctp4_top) )
MOV_L( ARG_OR, ECX )
MOV_L( ARG_AND, EDX )
MOV_B( AL, REGIND(ECX) )
MOV_B( AH, REGIND(EDX) )
-LLBL( ctp4_finish ):
+LLBL(ctp4_finish):
MOV_L( ARG_DEST, EAX )
#ifdef ELFPIC
ALIGNTEXT16
GLOBL GLNAME( _mesa_x86_cliptest_points4_np )
+HIDDEN(_mesa_x86_cliptest_points4_np)
GLNAME( _mesa_x86_cliptest_points4_np ):
#ifdef ELFPIC
#ifdef ELFPIC
/* store pointer to clip_table on stack */
- CALL( LLBL( ctp4_np_get_eip ) )
+ CALL( LLBL(ctp4_np_get_eip) )
ADD_L( CONST(_GLOBAL_OFFSET_TABLE_), EBX )
MOV_L( REGOFF(clip_table@GOT, EBX), EBX )
PUSH_L( EBX )
- JMP( LLBL( ctp4_np_clip_table_ready ) )
+ JMP( LLBL(ctp4_np_clip_table_ready) )
-LLBL( ctp4_np_get_eip ):
+LLBL(ctp4_np_get_eip):
/* store eip in ebx */
MOV_L( REGIND(ESP), EBX )
RET
-LLBL( ctp4_np_clip_table_ready ):
+LLBL(ctp4_np_clip_table_ready):
#endif
MOV_L( ARG_SOURCE, ESI )
MOV_B( REGIND(EBX), AL )
MOV_B( REGIND(EBP), AH )
- JZ( LLBL( ctp4_np_finish ) )
+ JZ( LLBL(ctp4_np_finish) )
ALIGNTEXT16
-LLBL( ctp4_np_top ):
+LLBL(ctp4_np_top):
- MOV_L( SRC(3), EBP )
- MOV_L( SRC(2), EBX )
+ MOV_L( SRC3, EBP )
+ MOV_L( SRC2, EBX )
XOR_L( ECX, ECX )
ADD_L( EBP, EBP ) /* ebp = abs(S(3))*2 ; carry = sign of S(3) */
CMP_L( EBX, EBP ) /* carry = abs(S(2))*2 > abs(S(3))*2 */
ADC_L( ECX, ECX )
- MOV_L( SRC(1), EBX )
+ MOV_L( SRC1, EBX )
ADD_L( EBX, EBX ) /* ebx = abs(S(1))*2 ; carry = sign of S(1) */
CMP_L( EBX, EBP ) /* carry = abs(S(1))*2 > abs(S(3))*2 */
ADC_L( ECX, ECX )
- MOV_L( SRC(0), EBX )
+ MOV_L( SRC0, EBX )
ADD_L( EBX, EBX ) /* ebx = abs(S(0))*2 ; carry = sign of S(0) */
ADD_L( ARG_DEST, ESI )
CMP_L( EDX, EDI )
- JNZ( LLBL( ctp4_np_top ) )
+ JNZ( LLBL(ctp4_np_top) )
MOV_L( ARG_OR, ECX )
MOV_L( ARG_AND, EDX )
MOV_B( AL, REGIND(ECX) )
MOV_B( AH, REGIND(EDX) )
-LLBL( ctp4_np_finish ):
+LLBL(ctp4_np_finish):
MOV_L( ARG_SOURCE, EAX )
#ifdef ELFPIC
POP_L( ESI )
RET
+
+#if defined (__ELF__) && defined (__linux__)
+ .section .note.GNU-stack,"",%progbits
+#endif