diff options
| author | Matt Turner <mattst88@gmail.com> | 2014-09-02 14:43:43 -0700 |
|---|---|---|
| committer | Matt Turner <mattst88@gmail.com> | 2014-09-24 09:42:46 -0700 |
| commit | 72bb3f81c621931e42759148bc8bddc511266dd0 (patch) | |
| tree | 01b4691b95a5fdc368afde0e81944ddf1e8f0ef8 | |
| parent | f0598d413bc8eb7ab02318f1db2dbd446a3c736c (diff) | |
i965/vec4: Don't iterate between blocks with inst->next/prev.
The register coalescing portion of this patch hurts three shaders in
Guacamelee by one instruction each, but examining the diff makes me
believe that what we were generating was (perhaps harmlessly) incorrect.
| -rw-r--r-- | src/mesa/drivers/dri/i965/brw_vec4.cpp | 30 |
1 files changed, 9 insertions, 21 deletions
diff --git a/src/mesa/drivers/dri/i965/brw_vec4.cpp b/src/mesa/drivers/dri/i965/brw_vec4.cpp index ed1200db18e..022ed37a8ef 100644 --- a/src/mesa/drivers/dri/i965/brw_vec4.cpp +++ b/src/mesa/drivers/dri/i965/brw_vec4.cpp @@ -516,12 +516,9 @@ vec4_visitor::dead_code_eliminate() } } - for (exec_node *node = inst->prev, *prev = node->prev; - prev != NULL && dead_channels != 0; - node = prev, prev = prev->prev) { - vec4_instruction *scan_inst = (vec4_instruction *)node; - - if (scan_inst->is_control_flow()) + foreach_inst_in_block_reverse_starting_from(vec4_instruction, scan_inst, + inst, block) { + if (dead_channels == 0) break; if (inst_writes_flag) { @@ -1062,10 +1059,11 @@ vec4_visitor::opt_register_coalesce() * everything writing to the temporary to write into the destination * instead. */ - vec4_instruction *scan_inst; - for (scan_inst = (vec4_instruction *)inst->prev; - scan_inst->prev != NULL; - scan_inst = (vec4_instruction *)scan_inst->prev) { + vec4_instruction *_scan_inst = (vec4_instruction *)inst->prev; + foreach_inst_in_block_reverse_starting_from(vec4_instruction, scan_inst, + inst, block) { + _scan_inst = scan_inst; + if (scan_inst->dst.file == GRF && scan_inst->dst.reg == inst->src[0].reg && scan_inst->dst.reg_offset == inst->src[0].reg_offset) { @@ -1107,16 +1105,6 @@ vec4_visitor::opt_register_coalesce() break; } - /* We don't handle flow control here. Most computation of values - * that could be coalesced happens just before their use. - */ - if (scan_inst->opcode == BRW_OPCODE_DO || - scan_inst->opcode == BRW_OPCODE_WHILE || - scan_inst->opcode == BRW_OPCODE_ELSE || - scan_inst->opcode == BRW_OPCODE_ENDIF) { - break; - } - /* You can't read from an MRF, so if someone else reads our MRF's * source GRF that we wanted to rewrite, that stops us. If it's a * GRF we're trying to coalesce to, we don't actually handle @@ -1169,7 +1157,7 @@ vec4_visitor::opt_register_coalesce() * computing the value. Now go rewrite the instruction stream * between the two. */ - + vec4_instruction *scan_inst = _scan_inst; while (scan_inst != inst) { if (scan_inst->dst.file == GRF && scan_inst->dst.reg == inst->src[0].reg && |