diff options
| author | Keith Whitwell <keith@tungstengraphics.com> | 2008-05-22 13:48:07 +0100 |
|---|---|---|
| committer | Keith Whitwell <keith@tungstengraphics.com> | 2008-05-23 09:16:58 +0100 |
| commit | 6780a6dede31e7f2eb465e1d7b507b3e64fe6ec9 (patch) | |
| tree | 887e004c9587d403e8e9aecfd56c51eb9bec61a3 /src/gallium/auxiliary/draw/draw_vs_aos.c | |
| parent | 7b25c1a4032960752d8a8e950bdf75740b2de2e8 (diff) | |
draw: shortcircuit shuffle in aos_sse when possible
Diffstat (limited to 'src/gallium/auxiliary/draw/draw_vs_aos.c')
| -rw-r--r-- | src/gallium/auxiliary/draw/draw_vs_aos.c | 18 |
1 files changed, 13 insertions, 5 deletions
diff --git a/src/gallium/auxiliary/draw/draw_vs_aos.c b/src/gallium/auxiliary/draw/draw_vs_aos.c index d60940bb7a..b8fad231ca 100644 --- a/src/gallium/auxiliary/draw/draw_vs_aos.c +++ b/src/gallium/auxiliary/draw/draw_vs_aos.c @@ -809,7 +809,9 @@ static boolean emit_DP3( struct aos_compilation *cp, const struct tgsi_full_inst sse_addss(cp->func, dst, tmp); /* a*x+c*z, b*y, ?, ? */ emit_pshufd(cp, tmp, dst, SHUF(Y,X,W,Z)); sse_addss(cp->func, dst, tmp); - sse_shufps(cp->func, dst, dst, SHUF(X, X, X, X)); + + if (op->FullDstRegisters[0].DstRegister.WriteMask != 0x1) + sse_shufps(cp->func, dst, dst, SHUF(X, X, X, X)); aos_release_xmm_reg(cp, tmp.idx); store_dest(cp, &op->FullDstRegisters[0], dst); @@ -833,7 +835,9 @@ static boolean emit_DP4( struct aos_compilation *cp, const struct tgsi_full_inst sse_addps(cp->func, dst, tmp); /* a*x+c*z, b*y+d*w, a*x+c*z, b*y+d*w */ emit_pshufd(cp, tmp, dst, SHUF(Y,X,W,Z)); sse_addss(cp->func, dst, tmp); - sse_shufps(cp->func, dst, dst, SHUF(X, X, X, X)); + + if (op->FullDstRegisters[0].DstRegister.WriteMask != 0x1) + sse_shufps(cp->func, dst, dst, SHUF(X, X, X, X)); aos_release_xmm_reg(cp, tmp.idx); store_dest(cp, &op->FullDstRegisters[0], dst); @@ -857,7 +861,9 @@ static boolean emit_DPH( struct aos_compilation *cp, const struct tgsi_full_inst sse_addss(cp->func, dst, tmp); emit_pshufd(cp, tmp, arg1, SHUF(W,W,W,W)); sse_addss(cp->func, dst, tmp); - sse_shufps(cp->func, dst, dst, SHUF(X, X, X, X)); + + if (op->FullDstRegisters[0].DstRegister.WriteMask != 0x1) + sse_shufps(cp->func, dst, dst, SHUF(X, X, X, X)); aos_release_xmm_reg(cp, tmp.idx); store_dest(cp, &op->FullDstRegisters[0], dst); @@ -1233,7 +1239,8 @@ static boolean emit_RCP( struct aos_compilation *cp, const struct tgsi_full_inst sse_divss(cp->func, dst, arg0); } - sse_shufps(cp->func, dst, dst, SHUF(X, X, X, X)); + if (op->FullDstRegisters[0].DstRegister.WriteMask != 0x1) + sse_shufps(cp->func, dst, dst, SHUF(X, X, X, X)); store_dest(cp, &op->FullDstRegisters[0], dst); return TRUE; @@ -1249,7 +1256,8 @@ static boolean emit_RSQ( struct aos_compilation *cp, const struct tgsi_full_inst /* Extend precision here... */ - sse_shufps(cp->func, dst, dst, SHUF(X, X, X, X)); + if (op->FullDstRegisters[0].DstRegister.WriteMask != 0x1) + sse_shufps(cp->func, dst, dst, SHUF(X, X, X, X)); store_dest(cp, &op->FullDstRegisters[0], dst); return TRUE; |