diff options
author | Jason Ekstrand <[email protected]> | 2016-07-20 12:21:41 -0700 |
---|---|---|
committer | Jason Ekstrand <[email protected]> | 2016-08-10 15:19:55 -0700 |
commit | 4c3a6b07e2960266adca634f8607ef38f71b8318 (patch) | |
tree | b2329b975bc774a006b862d9fe54eb52c2c4d678 | |
parent | ac6966360fc2f31b5e862624c6ef0048e11148e7 (diff) |
i965/vec4: Make opt_vector_float reset at the top of each block
The pass isn't really control-flow aware and you can get into case where it
tries to combine instructions from different blocks. This can actually
lead to an assertion failure when removing unneeded instructions if part of
the vector is set in one block and part in another. This prevents
regressions in the next commit.
Signed-off-by: Jason Ekstrand <[email protected]>
Cc: "12.0" <[email protected]>
-rw-r--r-- | src/mesa/drivers/dri/i965/brw_vec4.cpp | 162 |
1 files changed, 82 insertions, 80 deletions
diff --git a/src/mesa/drivers/dri/i965/brw_vec4.cpp b/src/mesa/drivers/dri/i965/brw_vec4.cpp index 162b481573d..2cf0e1759e4 100644 --- a/src/mesa/drivers/dri/i965/brw_vec4.cpp +++ b/src/mesa/drivers/dri/i965/brw_vec4.cpp @@ -354,95 +354,97 @@ vec4_visitor::opt_vector_float() { bool progress = false; - int last_reg = -1, last_reg_offset = -1; - enum brw_reg_file last_reg_file = BAD_FILE; + foreach_block(block, cfg) { + int last_reg = -1, last_reg_offset = -1; + enum brw_reg_file last_reg_file = BAD_FILE; + + uint8_t imm[4] = { 0 }; + int inst_count = 0; + vec4_instruction *imm_inst[4]; + unsigned writemask = 0; + enum brw_reg_type dest_type = BRW_REGISTER_TYPE_F; + + foreach_inst_in_block_safe(vec4_instruction, inst, block) { + int vf = -1; + enum brw_reg_type need_type; + + /* Look for unconditional MOVs from an immediate with a partial + * writemask. Skip type-conversion MOVs other than integer 0, + * where the type doesn't matter. See if the immediate can be + * represented as a VF. + */ + if (inst->opcode == BRW_OPCODE_MOV && + inst->src[0].file == IMM && + inst->predicate == BRW_PREDICATE_NONE && + inst->dst.writemask != WRITEMASK_XYZW && + (inst->src[0].type == inst->dst.type || inst->src[0].d == 0)) { + + vf = brw_float_to_vf(inst->src[0].d); + need_type = BRW_REGISTER_TYPE_D; + + if (vf == -1) { + vf = brw_float_to_vf(inst->src[0].f); + need_type = BRW_REGISTER_TYPE_F; + } + } else { + last_reg = -1; + } - uint8_t imm[4] = { 0 }; - int inst_count = 0; - vec4_instruction *imm_inst[4]; - unsigned writemask = 0; - enum brw_reg_type dest_type = BRW_REGISTER_TYPE_F; + /* If this wasn't a MOV, or the destination register doesn't match, + * or we have to switch destination types, then this breaks our + * sequence. Combine anything we've accumulated so far. + */ + if (last_reg != inst->dst.nr || + last_reg_offset != inst->dst.reg_offset || + last_reg_file != inst->dst.file || + (vf > 0 && dest_type != need_type)) { + + if (inst_count > 1) { + unsigned vf; + memcpy(&vf, imm, sizeof(vf)); + vec4_instruction *mov = MOV(imm_inst[0]->dst, brw_imm_vf(vf)); + mov->dst.type = dest_type; + mov->dst.writemask = writemask; + inst->insert_before(block, mov); + + for (int i = 0; i < inst_count; i++) { + imm_inst[i]->remove(block); + } - foreach_block_and_inst_safe(block, vec4_instruction, inst, cfg) { - int vf = -1; - enum brw_reg_type need_type; + progress = true; + } - /* Look for unconditional MOVs from an immediate with a partial - * writemask. Skip type-conversion MOVs other than integer 0, - * where the type doesn't matter. See if the immediate can be - * represented as a VF. - */ - if (inst->opcode == BRW_OPCODE_MOV && - inst->src[0].file == IMM && - inst->predicate == BRW_PREDICATE_NONE && - inst->dst.writemask != WRITEMASK_XYZW && - (inst->src[0].type == inst->dst.type || inst->src[0].d == 0)) { - - vf = brw_float_to_vf(inst->src[0].d); - need_type = BRW_REGISTER_TYPE_D; - - if (vf == -1) { - vf = brw_float_to_vf(inst->src[0].f); - need_type = BRW_REGISTER_TYPE_F; - } - } else { - last_reg = -1; - } + inst_count = 0; + last_reg = -1; + writemask = 0; + dest_type = BRW_REGISTER_TYPE_F; - /* If this wasn't a MOV, or the destination register doesn't match, - * or we have to switch destination types, then this breaks our - * sequence. Combine anything we've accumulated so far. - */ - if (last_reg != inst->dst.nr || - last_reg_offset != inst->dst.reg_offset || - last_reg_file != inst->dst.file || - (vf > 0 && dest_type != need_type)) { - - if (inst_count > 1) { - unsigned vf; - memcpy(&vf, imm, sizeof(vf)); - vec4_instruction *mov = MOV(imm_inst[0]->dst, brw_imm_vf(vf)); - mov->dst.type = dest_type; - mov->dst.writemask = writemask; - inst->insert_before(block, mov); - - for (int i = 0; i < inst_count; i++) { - imm_inst[i]->remove(block); + for (int i = 0; i < 4; i++) { + imm[i] = 0; } - - progress = true; } - inst_count = 0; - last_reg = -1; - writemask = 0; - dest_type = BRW_REGISTER_TYPE_F; - - for (int i = 0; i < 4; i++) { - imm[i] = 0; + /* Record this instruction's value (if it was representable). */ + if (vf != -1) { + if ((inst->dst.writemask & WRITEMASK_X) != 0) + imm[0] = vf; + if ((inst->dst.writemask & WRITEMASK_Y) != 0) + imm[1] = vf; + if ((inst->dst.writemask & WRITEMASK_Z) != 0) + imm[2] = vf; + if ((inst->dst.writemask & WRITEMASK_W) != 0) + imm[3] = vf; + + writemask |= inst->dst.writemask; + imm_inst[inst_count++] = inst; + + last_reg = inst->dst.nr; + last_reg_offset = inst->dst.reg_offset; + last_reg_file = inst->dst.file; + if (vf > 0) + dest_type = need_type; } } - - /* Record this instruction's value (if it was representable). */ - if (vf != -1) { - if ((inst->dst.writemask & WRITEMASK_X) != 0) - imm[0] = vf; - if ((inst->dst.writemask & WRITEMASK_Y) != 0) - imm[1] = vf; - if ((inst->dst.writemask & WRITEMASK_Z) != 0) - imm[2] = vf; - if ((inst->dst.writemask & WRITEMASK_W) != 0) - imm[3] = vf; - - writemask |= inst->dst.writemask; - imm_inst[inst_count++] = inst; - - last_reg = inst->dst.nr; - last_reg_offset = inst->dst.reg_offset; - last_reg_file = inst->dst.file; - if (vf > 0) - dest_type = need_type; - } } if (progress) |