From 45d21ca58755098e8fd47e1083c68e244c4f903c Mon Sep 17 00:00:00 2001 From: Brian Paul Date: Mon, 22 Oct 2001 01:21:16 +0000 Subject: applied Andrew Lewycky's patches --- src/mesa/x86/3dnow_xform4.S | 101 +++++++++++++++----------------------------- 1 file changed, 35 insertions(+), 66 deletions(-) (limited to 'src/mesa/x86/3dnow_xform4.S') diff --git a/src/mesa/x86/3dnow_xform4.S b/src/mesa/x86/3dnow_xform4.S index 06d7c6950a..0336b708bc 100644 --- a/src/mesa/x86/3dnow_xform4.S +++ b/src/mesa/x86/3dnow_xform4.S @@ -1,4 +1,4 @@ -/* $Id: 3dnow_xform4.S,v 1.1 2001/03/29 06:46:16 gareth Exp $ */ +/* $Id: 3dnow_xform4.S,v 1.2 2001/10/22 01:21:16 brianp Exp $ */ /* * Mesa 3-D graphics library @@ -59,17 +59,17 @@ GLNAME( _mesa_3dnow_transform_points4_general ): PREFETCHW ( REGIND(EDX) ) - MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */ - MOVQ ( REGOFF(8, EAX), MM4 ) /* x3 | x2 */ - - ADD_L ( EDI, EAX ) /* next vertex */ - ALIGNTEXT16 LLBL( G3TPGR_1 ): PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ - PREFETCH ( REGIND(EAX) ) + MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */ + MOVQ ( REGOFF(8, EAX), MM4 ) /* x3 | x2 */ + + ADD_L ( EDI, EAX ) /* next vertex */ + PREFETCH ( REGIND(EAX) ) + MOVQ ( MM0, MM2 ) /* x1 | x0 */ MOVQ ( MM4, MM6 ) /* x3 | x2 */ @@ -110,10 +110,6 @@ LLBL( G3TPGR_1 ): MOVQ ( MM6, REGOFF(-16, EDX) ) MOVQ ( MM7, REGOFF(-8, EDX) ) - MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */ - - MOVQ ( REGOFF(8, EAX), MM4 ) /* x3 | x2 */ - ADD_L ( EDI, EAX ) /* next vertex */ DEC_L ( ESI ) /* decrement vertex counter */ JNZ ( LLBL( G3TPGR_1 ) ) /* cnt > 0 ? -> process next vertex */ @@ -165,16 +161,16 @@ GLNAME( _mesa_3dnow_transform_points4_perspective ): MOVQ ( REGOFF(32, ECX), MM2 ) /* m21 | m20 */ PXOR ( MM7, MM7 ) /* 0 | 0 */ +ALIGNTEXT16 +LLBL( G3TPPR_1 ): + + PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ + MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */ MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */ MOVD ( REGOFF(8, EAX), MM3 ) /* | x2 */ ADD_L ( EDI, EAX ) /* next vertex */ - -ALIGNTEXT16 -LLBL( G3TPPR_1 ): - - PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ PREFETCH ( REGOFF(32, EAX) ) /* hopefully stride is zero */ MOVQ ( MM5, MM6 ) /* x3 | x2 */ @@ -193,12 +189,6 @@ LLBL( G3TPPR_1 ): MOVQ ( MM5, REGOFF(-16, EDX) ) /* write r0, r1 */ MOVQ ( MM6, REGOFF(-8, EDX) ) /* write r2, r3 */ - MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */ - - MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */ - MOVD ( REGOFF(8, EAX), MM3 ) /* | x2 */ - - ADD_L ( EDI, EAX ) /* next vertex */ DEC_L ( ESI ) /* decrement vertex counter */ JNZ ( LLBL( G3TPPR_1 ) ) /* cnt > 0 ? -> process next vertex */ @@ -244,15 +234,15 @@ GLNAME( _mesa_3dnow_transform_points4_3d ): MOVD ( REGOFF(40, ECX), MM7 ) /* | m10 */ PUNPCKLDQ ( REGOFF(56, ECX), MM7 ) /* m14 | m10 */ - MOVQ ( REGIND(EAX), MM2 ) /* x1 | x0 */ - MOVQ ( REGOFF(8, EAX), MM3 ) /* x3 | x2 */ - ALIGNTEXT16 LLBL( G3TP3R_1 ): PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ PREFETCH ( REGOFF(32, EAX) ) /* hopefully array is tightly packed */ + MOVQ ( REGIND(EAX), MM2 ) /* x1 | x0 */ + MOVQ ( REGOFF(8, EAX), MM3 ) /* x3 | x2 */ + MOVQ ( MM2, MM0 ) /* x1 | x0 */ MOVQ ( MM3, MM4 ) /* x3 | x2 */ @@ -289,9 +279,6 @@ LLBL( G3TP3R_1 ): MOVQ ( MM4, REGOFF(-16, EDX) ) /* write r0, r1 */ MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */ - MOVQ ( REGIND(EAX), MM2 ) /* x1 | x0 */ - MOVQ ( REGOFF(8, EAX), MM3 ) /* x3 | x2 */ - DEC_L ( ESI ) /* decrement vertex counter */ JNZ ( LLBL( G3TP3R_1 ) ) /* cnt > 0 ? -> process next vertex */ @@ -337,16 +324,16 @@ GLNAME( _mesa_3dnow_transform_points4_3d_no_rot ): MOVQ ( REGOFF(48, ECX), MM1 ) /* m31 | m30 */ +ALIGNTEXT16 +LLBL( G3TP3NRR_1 ): + + PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ + MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */ MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */ MOVD ( REGOFF(12, EAX), MM7 ) /* | x3 */ ADD_L ( EDI, EAX ) /* next vertex */ - -ALIGNTEXT16 -LLBL( G3TP3NRR_1 ): - - PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ PREFETCH ( REGOFF(32, EAX) ) /* hopefully stride is zero */ MOVQ ( MM5, MM6 ) /* x3 | x2 */ @@ -364,12 +351,6 @@ LLBL( G3TP3NRR_1 ): MOVQ ( MM4, REGOFF(-16, EDX) ) /* write r0, r1 */ MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */ - MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */ - MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */ - - MOVD ( REGOFF(12, EAX), MM7 ) /* | x3 */ - ADD_L ( EDI, EAX ) /* next vertex */ - DEC_L ( ESI ) /* decrement vertex counter */ JNZ ( LLBL( G3TP3NRR_1 ) ) /* cnt > 0 ? -> process next vertex */ @@ -416,15 +397,15 @@ GLNAME( _mesa_3dnow_transform_points4_2d ): MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */ - MOVQ ( REGIND(EAX), MM3 ) /* x1 | x0 */ - MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */ - - ADD_L ( EDI, EAX ) /* next vertex */ - ALIGNTEXT16 LLBL( G3TP2R_1 ): PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ + + MOVQ ( REGIND(EAX), MM3 ) /* x1 | x0 */ + MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */ + + ADD_L ( EDI, EAX ) /* next vertex */ PREFETCH ( REGIND(EAX) ) MOVQ ( MM3, MM4 ) /* x1 | x0 */ @@ -443,10 +424,6 @@ LLBL( G3TP2R_1 ): MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */ MOVQ ( MM3, REGOFF(-16, EDX) ) /* write r0, r1 */ - MOVQ ( REGIND(EAX), MM3 ) /* x1 | x0 */ - - MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */ - ADD_L ( EDI, EAX ) /* next vertex */ DEC_L ( ESI ) /* decrement vertex counter */ JNZ ( LLBL( G3TP2R_1 ) ) /* cnt > 0 ? -> process next vertex */ @@ -491,15 +468,15 @@ GLNAME( _mesa_3dnow_transform_points4_2d_no_rot ): MOVQ ( REGOFF(48, ECX), MM1 ) /* m31 | m30 */ - MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */ - MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */ - - ADD_L ( EDI, EAX ) /* next vertex */ - ALIGNTEXT16 LLBL( G3TP2NRR_2 ): PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ + + MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */ + MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */ + + ADD_L ( EDI, EAX ) /* next vertex */ PREFETCH ( REGIND(EAX) ) PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */ @@ -514,10 +491,6 @@ LLBL( G3TP2NRR_2 ): MOVQ ( MM6, REGOFF(-16, EDX) ) /* write r0, r1 */ MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */ - MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */ - MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */ - - ADD_L ( EDI, EAX ) /* next vertex */ DEC_L ( ESI ) /* decrement vertex counter */ JNZ ( LLBL( G3TP2NRR_2 ) ) /* cnt > 0 ? -> process next vertex */ @@ -557,25 +530,21 @@ GLNAME( _mesa_3dnow_transform_points4_identity ): TEST_L ( ESI, ESI ) JZ ( LLBL( G3TPIR_2 ) ) - MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */ - MOVQ ( REGOFF(8, EAX), MM1 ) /* x3 | x2 */ - - ADD_L ( EDI, EAX ) /* next vertex */ - ALIGNTEXT16 LLBL( G3TPIR_1 ): PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */ + + MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */ + MOVQ ( REGOFF(8, EAX), MM1 ) /* x3 | x2 */ + + ADD_L ( EDI, EAX ) /* next vertex */ PREFETCH ( REGIND(EAX) ) ADD_L ( CONST(16), EDX ) /* next r */ MOVQ ( MM0, REGOFF(-16, EDX) ) /* r1 | r0 */ MOVQ ( MM1, REGOFF(-8, EDX) ) /* r3 | r2 */ - MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */ - - MOVQ ( REGOFF(8, EAX), MM1 ) /* x3 | x2 */ - ADD_L ( EDI, EAX ) /* next vertex */ DEC_L ( ESI ) /* decrement vertex counter */ JNZ ( LLBL( G3TPIR_1 ) ) /* cnt > 0 ? -> process next vertex */ -- cgit v1.2.3