fix[venom]: simplify cfg phi rename #4181
Conversation
|
shouldn't we fix up the phi nodes during sccp directly? |
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
HodanPlodky
changed the title
| @@ -305,6 +311,9 @@ def _replace_constants(self, inst: IRInstruction): | |||
| inst.opcode = "jmp" | |||
| inst.operands = [target] | |||
| self.cfg_dirty = True | |||
| for bb in inst.parent.cfg_out: | |||
| if bb.label == target: | |||
| self.cfg_phi_fix_needed.add(bb) | |||
There was a problem hiding this comment.
Why not self._fix_phi_bb_r(bb, OrderedSet()) here and remove all the need for self.cfg_phi_fix_needed and iterating later?
There was a problem hiding this comment.
The reason was not to iterate over some basic block multiple times. Since solutions only detects start from which the occurrence of faulty phi node could be possible but not the end there could be a lot of overlap. In this case since I iterate over possible positions of the faulty phi nodes at the end I visit every possible basic block only once.
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## master #4181 +/- ##
==========================================
- Coverage 91.35% 88.22% -3.14%
==========================================
Files 109 109
Lines 15641 15672 +31
Branches 3443 3454 +11
==========================================
- Hits 14289 13826 -463
- Misses 920 1314 +394
- Partials 432 532 +100 ☔ View full report in Codecov by Sentry. |
| from_src_bb = filter(lambda x: x[0] == src_bb.label, phi_inst.phi_operands) | ||
| operands = list(map(lambda x: x[1], from_src_bb)) |
There was a problem hiding this comment.
please use a list comprehension here
| from_src_bb = filter(lambda x: x[0] == src_bb.label, phi_inst.phi_operands) | |
| operands = list(map(lambda x: x[1], from_src_bb)) | |
| operands = [op for label, op in phi_inst.phi_operands if label == src_bb.label] |
|
|
||
| assert len(operands) == 1 | ||
| assert isinstance(operands[0], IRVariable) | ||
| assert phi_inst.output is not None |
There was a problem hiding this comment.
| assert phi_inst.output is not None | |
| assert phi_inst.output is not None | |
| @@ -305,6 +311,9 @@ def _replace_constants(self, inst: IRInstruction): | |||
| inst.opcode = "jmp" | |||
| inst.operands = [target] | |||
| self.cfg_dirty = True | |||
| for bb in inst.parent.cfg_out: | |||
| if bb.label == target: | |||
| self.cfg_phi_fix_needed.add(bb) | |||
There was a problem hiding this comment.
| self.cfg_phi_fix_needed.add(bb) | |
| self.cfg_phi_fix_needed.add(bb) | |
| break |
| for next_bb in bb.cfg_out: | ||
| self._fix_phi_bb_r(next_bb, visited) | ||
|
|
||
| def _fix_phi_node_inst(self, phi_inst: IRInstruction): |
There was a problem hiding this comment.
| def _fix_phi_node_inst(self, phi_inst: IRInstruction): | |
| def _fix_phi_inst(self, phi_inst: IRInstruction): |
| @@ -329,6 +338,40 @@ def _replace_constants(self, inst: IRInstruction): | |||
| if isinstance(lat, IRLiteral): | |||
| inst.operands[i] = lat | |||
|
|
|||
| def _fix_phi_nodes(self): | |||
| visited: OrderedSet[IRBasicBlock] = OrderedSet() | |||
There was a problem hiding this comment.
why keep track of visited, instead of just iterating over fn.get_basic_blocks()?
| assert len(operands) == 1 | ||
| assert isinstance(operands[0], IRVariable) | ||
| assert phi_inst.output is not None | ||
| phi_inst.output.value = operands[0] |
There was a problem hiding this comment.
why modify the instruction if it gets removed on the next line?
| self._fix_phi_bb_r(next_bb, visited) | ||
|
|
||
| def _fix_phi_node_inst(self, phi_inst: IRInstruction): | ||
| if len(phi_inst.parent.cfg_in) != 1: |
There was a problem hiding this comment.
when can this happen? shouldn't we handle this case?
|
i pushed some cleanup here: HodanPlodky#2. it feels like a kludge though, since it seems like we really should be fixing up these phi nodes inside of |
| assert isinstance(operands[0], IRVariable) | ||
| assert phi_inst.output is not None | ||
| phi_inst.output.value = operands[0] | ||
| phi_inst.parent.remove_instruction(phi_inst) |
There was a problem hiding this comment.
i think it might be a bug to remove the instruction. i think the instruction is dead when the value has been propagated from previous blocks, but it might not be dead if one of the parent blocks has been pruned, but it is not a constant in the other block.
What I did
Fix for issues #4162 and #4070
How I did it
Removed the phi nodes which were not in use after
SCCPpass. This could happen when theSCCPwould replacejnzbyjmpinstruction and by this removing need forphiinstruction (this is the case in #4162). More over the when merging the basic block in theSimplifyCFGPassthere could be problem when renaming the labels inphinodes since it went only into the firstcfg_outbasic block (this is the case in #4070)How to verify it
So far both of the POC from compile correctly
Commit message
fix[venom]: phi label renaming and unused phi nodes removal