diff options
author | Ilia Mirkin <[email protected]> | 2015-09-10 01:54:30 -0400 |
---|---|---|
committer | Ilia Mirkin <[email protected]> | 2015-09-10 03:11:31 -0400 |
commit | a072ef8748a65d286e9b542bb9ea6e020fdcc7f8 (patch) | |
tree | 4347bceab2008d1c12263db27c2e049185589624 /src/gallium | |
parent | 13a974f9aea03a538c2a67417b5bee8bc732cca2 (diff) |
nv50/ir: make edge splitting fix up phi node sources
Unfortunately nv50_ir phi nodes aren't directly connected to the CFG, so
the mapping between source and the actual BB is by inbound edge order.
So when manipulating edges one has to be extremely careful. We were
insufficiently careful when splitting critical edges which resulted in
the phi nodes being confused as to where their sources were coming from.
This primarily manifests itself with the TXL-lowering logic on nv50,
when it is inside of a conditional. I've been unable to trigger the
issue anywhere else so far. This resolves rendering failures
in a number of games like Two Worlds 2, Trine: Enchanted Edition, Trine 2,
XCOM:Enemy Unknown, Stacking. It also improves the situation in
Hearthstone, Sonic Generations, and The Raven: Legacy of a Master Thief.
However more work needs to be done there (splitting a lot more edges
solves it, so it's some other sort of RA-related issue).
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=90887
Signed-off-by: Ilia Mirkin <[email protected]>
Cc: "11.0" <[email protected]>
Diffstat (limited to 'src/gallium')
-rw-r--r-- | src/gallium/drivers/nouveau/codegen/nv50_ir_ra.cpp | 90 |
1 files changed, 77 insertions, 13 deletions
diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_ra.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_ra.cpp index 0cd21cf47f5..400b9f09e51 100644 --- a/src/gallium/drivers/nouveau/codegen/nv50_ir_ra.cpp +++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_ra.cpp @@ -25,6 +25,7 @@ #include <stack> #include <limits> +#include <tr1/unordered_map> namespace nv50_ir { @@ -222,6 +223,7 @@ private: private: virtual bool visit(BasicBlock *); inline bool needNewElseBlock(BasicBlock *b, BasicBlock *p); + inline void splitEdges(BasicBlock *b); }; class ArgumentMovesPass : public Pass { @@ -345,28 +347,55 @@ RegAlloc::PhiMovesPass::needNewElseBlock(BasicBlock *b, BasicBlock *p) return (n == 2); } -// For each operand of each PHI in b, generate a new value by inserting a MOV -// at the end of the block it is coming from and replace the operand with its -// result. This eliminates liveness conflicts and enables us to let values be -// copied to the right register if such a conflict exists nonetheless. +struct PhiMapHash { + size_t operator()(const std::pair<Instruction *, BasicBlock *>& val) const { + return std::tr1::hash<Instruction*>()(val.first) * 31 + + std::tr1::hash<BasicBlock*>()(val.second); + } +}; + +typedef std::tr1::unordered_map< + std::pair<Instruction *, BasicBlock *>, Value *, PhiMapHash> PhiMap; + +// Critical edges need to be split up so that work can be inserted along +// specific edge transitions. Unfortunately manipulating incident edges into a +// BB invalidates all the PHI nodes since their sources are implicitly ordered +// by incident edge order. // -// These MOVs are also crucial in making sure the live intervals of phi srces -// are extended until the end of the loop, since they are not included in the -// live-in sets. -bool -RegAlloc::PhiMovesPass::visit(BasicBlock *bb) +// TODO: Make it so that that is not the case, and PHI nodes store pointers to +// the original BBs. +void +RegAlloc::PhiMovesPass::splitEdges(BasicBlock *bb) { - Instruction *phi, *mov; BasicBlock *pb, *pn; - + Instruction *phi; + Graph::EdgeIterator ei; std::stack<BasicBlock *> stack; + int j = 0; - for (Graph::EdgeIterator ei = bb->cfg.incident(); !ei.end(); ei.next()) { + for (ei = bb->cfg.incident(); !ei.end(); ei.next()) { pb = BasicBlock::get(ei.getNode()); assert(pb); if (needNewElseBlock(bb, pb)) stack.push(pb); } + + // No critical edges were found, no need to perform any work. + if (stack.empty()) + return; + + // We're about to, potentially, reorder the inbound edges. This means that + // we need to hold on to the (phi, bb) -> src mapping, and fix up the phi + // nodes after the graph has been modified. + PhiMap phis; + + j = 0; + for (ei = bb->cfg.incident(); !ei.end(); ei.next(), j++) { + pb = BasicBlock::get(ei.getNode()); + for (phi = bb->getPhi(); phi && phi->op == OP_PHI; phi = phi->next) + phis.insert(std::make_pair(std::make_pair(phi, pb), phi->getSrc(j))); + } + while (!stack.empty()) { pb = stack.top(); pn = new BasicBlock(func); @@ -379,12 +408,47 @@ RegAlloc::PhiMovesPass::visit(BasicBlock *bb) assert(pb->getExit()->op != OP_CALL); if (pb->getExit()->asFlow()->target.bb == bb) pb->getExit()->asFlow()->target.bb = pn; + + for (phi = bb->getPhi(); phi && phi->op == OP_PHI; phi = phi->next) { + PhiMap::iterator it = phis.find(std::make_pair(phi, pb)); + assert(it != phis.end()); + phis.insert(std::make_pair(std::make_pair(phi, pn), it->second)); + phis.erase(it); + } } + // Now go through and fix up all of the phi node sources. + j = 0; + for (ei = bb->cfg.incident(); !ei.end(); ei.next(), j++) { + pb = BasicBlock::get(ei.getNode()); + for (phi = bb->getPhi(); phi && phi->op == OP_PHI; phi = phi->next) { + PhiMap::const_iterator it = phis.find(std::make_pair(phi, pb)); + assert(it != phis.end()); + + phi->setSrc(j, it->second); + } + } +} + +// For each operand of each PHI in b, generate a new value by inserting a MOV +// at the end of the block it is coming from and replace the operand with its +// result. This eliminates liveness conflicts and enables us to let values be +// copied to the right register if such a conflict exists nonetheless. +// +// These MOVs are also crucial in making sure the live intervals of phi srces +// are extended until the end of the loop, since they are not included in the +// live-in sets. +bool +RegAlloc::PhiMovesPass::visit(BasicBlock *bb) +{ + Instruction *phi, *mov; + + splitEdges(bb); + // insert MOVs (phi->src(j) should stem from j-th in-BB) int j = 0; for (Graph::EdgeIterator ei = bb->cfg.incident(); !ei.end(); ei.next()) { - pb = BasicBlock::get(ei.getNode()); + BasicBlock *pb = BasicBlock::get(ei.getNode()); if (!pb->isTerminated()) pb->insertTail(new_FlowInstruction(func, OP_BRA, bb)); |