[Beignet] [PATCH 2/2] GBE: Optmize phi elimination
Ruiling Song
ruiling.song at intel.com
Mon Jun 2 22:53:15 PDT 2014
During phi elimination, we simply insert 3 MOVs for one phi instruction
to avoid lost copy issue. But in fact, only two of them are needed for
most of time. This patch tries to see whether the move from phiCopy
to phi can be avoided.
The patch basically checks whether the phiCopy and phi have live range
interference. If no, then they can be coalesced, thus one instruction
can be optimized.
Signed-off-by: Ruiling Song <ruiling.song at intel.com>
---
backend/src/ir/value.cpp | 11 +++++
backend/src/llvm/llvm_gen_backend.cpp | 83 ++++++++++++++++++++++++++++++++-
2 files changed, 93 insertions(+), 1 deletion(-)
diff --git a/backend/src/ir/value.cpp b/backend/src/ir/value.cpp
index 1dbd4f4..a055bdf 100644
--- a/backend/src/ir/value.cpp
+++ b/backend/src/ir/value.cpp
@@ -523,6 +523,17 @@ namespace ir {
const DefSet &FunctionDAG::getDef(const Instruction *insn, uint32_t srcID) const {
return this->getDef(ValueUse(insn, srcID));
}
+ const UseSet *FunctionDAG::getRegUse(const Register ®) const {
+ auto it = regUse.find(reg);
+ GBE_ASSERT(it != regUse.end());
+ return it->second;
+ }
+ const DefSet *FunctionDAG::getRegDef(const Register ®) const {
+ auto it = regDef.find(reg);
+ GBE_ASSERT(it != regDef.end());
+ return it->second;
+ }
+
const ValueDef *FunctionDAG::getDefAddress(const ValueDef &def) const {
auto it = defName.find(def);
GBE_ASSERT(it != defName.end() && it->second != NULL);
diff --git a/backend/src/llvm/llvm_gen_backend.cpp b/backend/src/llvm/llvm_gen_backend.cpp
index bd111b0..db9e73c 100644
--- a/backend/src/llvm/llvm_gen_backend.cpp
+++ b/backend/src/llvm/llvm_gen_backend.cpp
@@ -148,6 +148,7 @@
#include "ir/context.hpp"
#include "ir/unit.hpp"
#include "ir/liveness.hpp"
+#include "ir/value.hpp"
#include "sys/set.hpp"
#include "sys/cvar.hpp"
@@ -441,6 +442,10 @@ namespace gbe
* compare instructions we need to invert to decrease branch complexity
*/
set<const Value*> conditionSet;
+ /*!
+ * <phi,phiCopy> node information for later optimization
+ */
+ map<const ir::Register, const ir::Register> phiMap;
/*! We visit each function twice. Once to allocate the registers and once to
* emit the Gen IR instructions
*/
@@ -490,6 +495,7 @@ namespace gbe
LI = &getAnalysis<LoopInfo>();
emitFunction(F);
+ phiMap.clear();
return false;
}
@@ -524,6 +530,8 @@ namespace gbe
template <bool isLoad, typename T> void emitLoadOrStore(T &I);
/*! Will try to remove MOVs due to PHI resolution */
void removeMOVs(const ir::Liveness &liveness, ir::Function &fn);
+ /*! Optimize phi move based on liveness information */
+ void optimizePhiCopy(ir::Liveness &liveness, ir::Function &fn);
/*! Will try to remove redundants LOADI in basic blocks */
void removeLOADIs(const ir::Liveness &liveness, ir::Function &fn);
/*! To avoid lost copy, we need two values for PHI. This function create a
@@ -1275,6 +1283,77 @@ namespace gbe
});
}
+ void GenWriter::optimizePhiCopy(ir::Liveness &liveness, ir::Function &fn)
+ {
+ // The overall idea behind is we check whether there is any interference
+ // between phi and phiCopy live range. If there is no point that
+ // phi & phiCopy are both alive, then we can optimize off the move
+ // from phiCopy to phi, and use phiCopy directly instead of phi.
+ using namespace ir;
+ ir::FunctionDAG *dag = new ir::FunctionDAG(liveness);
+
+ for (auto &it : phiMap) {
+ const Register phi = it.first;
+ const Register phiCopy = it.second;
+
+ const ir::DefSet *phiCopyDef = dag->getRegDef(phiCopy);
+ const ir::UseSet *phiUse = dag->getRegUse(phi);
+ const DefSet *phiDef = dag->getRegDef(phi);
+ bool isOpt = true;
+ for (auto &x : *phiCopyDef) {
+ const ir::Instruction * phiCopyDefInsn = x->getInstruction();
+ const ir::BasicBlock *bb = phiCopyDefInsn->getParent();
+ const Liveness::LiveOut &out = liveness.getLiveOut(bb);
+ // phi & phiCopy are both alive at the endpoint of bb,
+ // thus can not be optimized.
+ if (out.contains(phi)) {
+ isOpt = false;
+ break;
+ }
+ // If phi is used in the same BB that define the phiCopy,
+ // we need carefully check the liveness of phi & phiCopy.
+ // Make sure their live ranges do not interfere.
+ bool phiUsedInSameBB = false;
+ for (auto &y : *phiUse) {
+ const ir::Instruction *phiUseInsn = y->getInstruction();
+ const ir::BasicBlock *bb2 = phiUseInsn->getParent();
+ if (bb2 == bb) {
+ phiUsedInSameBB = true;
+ }
+ }
+ // Check phi is not used between phiCopy def point and bb's end point,
+ // which is often referred as 'phi swap issue', just like below:
+ // MOV phiCopy_1, x;
+ // MOV phiCopy_2, phi_1;
+ if (phiUsedInSameBB ) {
+ for (auto it = --bb->end(); it != bb->end() ; --it) {
+ const Instruction &p = *it;
+
+ if (&p == phiCopyDefInsn) break;
+ // we only care MOV here
+ if (p.getSrcNum() == 1 && p.getSrc(0) == phi) {
+ isOpt = false;
+ break;
+ }
+ }
+ }
+ }
+
+ // [MOV phi, phiCopy;] can be removed. So we remove it
+ // and replace phi uses with phiCopy
+ if (isOpt) {
+ for (auto &x : *phiDef) {
+ const_cast<Instruction *>(x->getInstruction())->remove();
+ }
+ for (auto &x : *phiUse) {
+ const Instruction *phiUseInsn = x->getInstruction();
+ replaceSrc(const_cast<Instruction *>(phiUseInsn), phi, phiCopy);
+ }
+ }
+ }
+ delete dag;
+ }
+
void GenWriter::removeMOVs(const ir::Liveness &liveness, ir::Function &fn)
{
// We store the last write and last read for each register
@@ -1579,9 +1658,10 @@ namespace gbe
ctx.endFunction();
// Liveness can be shared when we optimized the immediates and the MOVs
- const ir::Liveness liveness(fn);
+ ir::Liveness liveness(fn);
if (OCL_OPTIMIZE_LOADI) this->removeLOADIs(liveness, fn);
+ if (OCL_OPTIMIZE_PHI_MOVES) this->optimizePhiCopy(liveness, fn);
if (OCL_OPTIMIZE_PHI_MOVES) this->removeMOVs(liveness, fn);
}
@@ -2031,6 +2111,7 @@ namespace gbe
const ir::Register dst = this->getRegister(&I);
const ir::Register src = this->getRegister(copy);
ctx.MOV(type, dst, src);
+ phiMap.insert(std::make_pair(dst, src));
}
void GenWriter::regAllocateBranchInst(BranchInst &I) {}
--
1.7.10.4
More information about the Beignet
mailing list