From 053af6ac8cda226a62844fc014ed9f133557c111 Mon Sep 17 00:00:00 2001 From: Stéphane Marchesin Date: Wed, 6 Jul 2011 11:44:19 -0700 Subject: i915g: Fix optimization, also make it more generic. --- src/gallium/drivers/i915/i915_fpc_optimize.c | 28 ++++++++++++++++++++++++++-- 1 file changed, 26 insertions(+), 2 deletions(-) (limited to 'src/gallium/drivers/i915/i915_fpc_optimize.c') diff --git a/src/gallium/drivers/i915/i915_fpc_optimize.c b/src/gallium/drivers/i915/i915_fpc_optimize.c index 5c60d952de2..738468d8ff8 100644 --- a/src/gallium/drivers/i915/i915_fpc_optimize.c +++ b/src/gallium/drivers/i915/i915_fpc_optimize.c @@ -54,6 +54,22 @@ static boolean same_src_reg(struct i915_full_src_register* d1, struct i915_full_ d1->Register.Negate == d2->Register.Negate); } +static boolean is_unswizzled(struct i915_full_src_register* r, + int sx, + int sy, + int sz, + int sw) +{ + if (sx && r->Register.SwizzleX != TGSI_SWIZZLE_X) + return FALSE; + if (sy && r->Register.SwizzleY != TGSI_SWIZZLE_Y) + return FALSE; + if (sz && r->Register.SwizzleZ != TGSI_SWIZZLE_Z) + return FALSE; + if (sw && r->Register.SwizzleW != TGSI_SWIZZLE_W) + return FALSE; + return FALSE; +} /* * Optimize away things like: @@ -72,11 +88,15 @@ static void i915_fpc_optimize_mov_after_mul(union i915_full_token* current, unio current->FullInstruction.Dst[0].Register.WriteMask == TGSI_WRITEMASK_XYZ && next->FullInstruction.Dst[0].Register.WriteMask == TGSI_WRITEMASK_W && same_dst_reg(&next->FullInstruction.Dst[0], &next->FullInstruction.Dst[0]) && - same_src_reg(&next->FullInstruction.Src[0], ¤t->FullInstruction.Src[1]) ) + same_src_reg(&next->FullInstruction.Src[0], ¤t->FullInstruction.Src[1]) && + is_unswizzled(¤t->FullInstruction.Src[0], 1, 1, 1, 0) && + is_unswizzled(¤t->FullInstruction.Src[1], 1, 1, 1, 0) && + is_unswizzled(&next->FullInstruction.Src[0], 0, 0, 0, 1) ) { next->FullInstruction.Instruction.Opcode = TGSI_OPCODE_NOP; current->FullInstruction.Dst[0].Register.WriteMask = TGSI_WRITEMASK_XYZW; current->FullInstruction.Src[0].Register.SwizzleW = TGSI_SWIZZLE_ONE; + current->FullInstruction.Src[1].Register.SwizzleW = TGSI_SWIZZLE_W; return; } @@ -87,11 +107,15 @@ static void i915_fpc_optimize_mov_after_mul(union i915_full_token* current, unio current->FullInstruction.Dst[0].Register.WriteMask == TGSI_WRITEMASK_XYZ && next->FullInstruction.Dst[0].Register.WriteMask == TGSI_WRITEMASK_W && same_dst_reg(&next->FullInstruction.Dst[0], &next->FullInstruction.Dst[0]) && - same_src_reg(&next->FullInstruction.Src[0], ¤t->FullInstruction.Src[0]) ) + same_src_reg(&next->FullInstruction.Src[0], ¤t->FullInstruction.Src[0]) && + is_unswizzled(¤t->FullInstruction.Src[0], 1, 1, 1, 0) && + is_unswizzled(¤t->FullInstruction.Src[1], 1, 1, 1, 0) && + is_unswizzled(&next->FullInstruction.Src[0], 0, 0, 0, 1) ) { next->FullInstruction.Instruction.Opcode = TGSI_OPCODE_NOP; current->FullInstruction.Dst[0].Register.WriteMask = TGSI_WRITEMASK_XYZW; current->FullInstruction.Src[1].Register.SwizzleW = TGSI_SWIZZLE_ONE; + current->FullInstruction.Src[0].Register.SwizzleW = TGSI_SWIZZLE_W; return; } } -- cgit v1.2.3