[llvm-commits] [polly] r155717 - in /polly/trunk: include/polly/CodeGen/BlockGenerators.h lib/CodeGen/BlockGenerators.cpp test/CodeGen/simple_vec_call_2.ll test/CodeGen/simple_vec_impossible.ll
Tobias Grosser
grosser at fim.uni-passau.de
Fri Apr 27 09:36:14 PDT 2012
Author: grosser
Date: Fri Apr 27 11:36:14 2012
New Revision: 155717
URL: http://llvm.org/viewvc/llvm-project?rev=155717&view=rev
Log:
SCEV based code generation
This is an incomplete implementation of the SCEV based code generation.
When finished it will remove the need for -indvars -enable-iv-rewrite.
For the moment it is still disabled. Even though it passes 'make polly-test',
there are still loose ends especially in respect of OpenMP code generation.
Modified:
polly/trunk/include/polly/CodeGen/BlockGenerators.h
polly/trunk/lib/CodeGen/BlockGenerators.cpp
polly/trunk/test/CodeGen/simple_vec_call_2.ll
polly/trunk/test/CodeGen/simple_vec_impossible.ll
Modified: polly/trunk/include/polly/CodeGen/BlockGenerators.h
URL: http://llvm.org/viewvc/llvm-project/polly/trunk/include/polly/CodeGen/BlockGenerators.h?rev=155717&r1=155716&r2=155717&view=diff
==============================================================================
--- polly/trunk/include/polly/CodeGen/BlockGenerators.h (original)
+++ polly/trunk/include/polly/CodeGen/BlockGenerators.h Fri Apr 27 11:36:14 2012
@@ -24,7 +24,8 @@
#include <vector>
namespace llvm {
-class Pass;
+ class Pass;
+ class ScalarEvolution;
}
namespace polly {
@@ -58,9 +59,16 @@
IRBuilder<> &Builder;
ScopStmt &Statement;
Pass *P;
+ ScalarEvolution &SE;
BlockGenerator(IRBuilder<> &B, ScopStmt &Stmt, Pass *P);
+ /// @brief Check if an instruction can be 'SCEV-ignored'
+ ///
+ /// An instruction can be ignored if we can recreate it from its scalar
+ /// evolution expression.
+ bool isSCEVIgnore(const Instruction *Inst);
+
/// @brief Get the new version of a Value.
///
/// @param Old The old Value.
Modified: polly/trunk/lib/CodeGen/BlockGenerators.cpp
URL: http://llvm.org/viewvc/llvm-project/polly/trunk/lib/CodeGen/BlockGenerators.cpp?rev=155717&r1=155716&r2=155717&view=diff
==============================================================================
--- polly/trunk/lib/CodeGen/BlockGenerators.cpp (original)
+++ polly/trunk/lib/CodeGen/BlockGenerators.cpp Fri Apr 27 11:36:14 2012
@@ -17,6 +17,9 @@
#include "polly/CodeGen/BlockGenerators.h"
#include "polly/Support/GICHelper.h"
+#include "llvm/Analysis/LoopInfo.h"
+#include "llvm/Analysis/ScalarEvolution.h"
+#include "llvm/Analysis/ScalarEvolutionExpander.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include "llvm/Support/CommandLine.h"
@@ -37,6 +40,211 @@
cl::desc("Perform grouped unrolling, but don't generate SIMD "
"instuctions"), cl::Hidden, cl::init(false),
cl::ZeroOrMore);
+
+static cl::opt<bool>
+SCEVCodegen("polly-codegen-scev",
+ cl::desc("Use SCEV based code generation."), cl::Hidden,
+ cl::init(false), cl::ZeroOrMore);
+
+/// The SCEVRewriter takes a scalar evolution expression and updates the
+/// following components:
+///
+/// - SCEVUnknown
+///
+/// Values referenced in SCEVUnknown subexpressions are looked up in
+/// two Value to Value maps (GlobalMap and BBMap). If they are found they are
+/// replaced by a reference to the value they map to.
+///
+/// - SCEVAddRecExpr
+///
+/// Based on a Loop -> Value map {Loop_1: %Value}, an expression
+/// {%Base, +, %Step}<Loop_1> is rewritten to %Base + %Value * %Step.
+/// AddRecExpr's with more than two operands can not be translated.
+///
+/// FIXME: The comment above is not yet reality. At the moment we derive
+/// %Value by looking up the canonical IV of the loop and by defining
+/// %Value = GlobalMap[%IV]. This needs to be changed to remove the need for
+/// canonical induction variables.
+///
+///
+/// How can this be used?
+/// ====================
+///
+/// SCEVRewrite based code generation works on virtually independent blocks.
+/// This means we do not run the independent blocks pass to rewrite scalar
+/// instructions, but just ignore instructions that we can analyze with scalar
+/// evolution. Virtually independent blocks are blocks that only reference the
+/// following values:
+///
+/// o Values calculated within a basic block
+/// o Values representable by SCEV
+///
+/// During code generation we can ignore all instructions:
+///
+/// - Ignore all instructions except:
+/// - Load instructions
+/// - Instructions that reference operands already calculated within the
+/// basic block.
+/// - Store instructions
+struct SCEVRewriter : public SCEVVisitor<SCEVRewriter, const SCEV*> {
+public:
+ static const SCEV *rewrite(const SCEV *scev, Scop &S, ScalarEvolution &SE,
+ ValueMapT &GlobalMap, ValueMapT &BBMap) {
+ SCEVRewriter Rewriter(S, SE, GlobalMap, BBMap);
+ return Rewriter.visit(scev);
+ }
+
+ SCEVRewriter(Scop &S, ScalarEvolution &SE, ValueMapT &GlobalMap,
+ ValueMapT &BBMap) : S(S), SE(SE), GlobalMap(GlobalMap),
+ BBMap(BBMap) {}
+
+ const SCEV *visit(const SCEV *Expr) {
+ // FIXME: The parameter handling is incorrect.
+ //
+ // Polly does only detect parameters in Access function and loop iteration
+ // counters, but it does not get parameters that are just used by
+ // instructions within the basic block.
+ //
+ // There are two options to solve this:
+ // o Iterate over all instructions of the SCoP and find the actual
+ // parameters.
+ // o Just check within the SCEVRewriter if Values lay outside of the SCoP
+ // and detect parameters on the fly.
+ //
+ // This is especially important for OpenMP and GPGPU code generation, as
+ // they require us to detect and possibly rewrite the corresponding
+ // parameters.
+ if (isl_id *Id = S.getIdForParam(Expr)) {
+ isl_id_free(Id);
+ return Expr;
+ }
+
+
+ return SCEVVisitor<SCEVRewriter, const SCEV*>::visit(Expr);
+ }
+
+ const SCEV *visitConstant(const SCEVConstant *Constant) {
+ return Constant;
+ }
+
+ const SCEV *visitTruncateExpr(const SCEVTruncateExpr *Expr) {
+ const SCEV *Operand = visit(Expr->getOperand());
+ return SE.getTruncateExpr(Operand, Expr->getType());
+ }
+
+ const SCEV *visitZeroExtendExpr(const SCEVZeroExtendExpr *Expr) {
+ const SCEV *Operand = visit(Expr->getOperand());
+ return SE.getZeroExtendExpr(Operand, Expr->getType());
+ }
+
+ const SCEV *visitSignExtendExpr(const SCEVSignExtendExpr *Expr) {
+ const SCEV *Operand = visit(Expr->getOperand());
+ return SE.getSignExtendExpr(Operand, Expr->getType());
+ }
+
+ const SCEV *visitAddExpr(const SCEVAddExpr *Expr) {
+ SmallVector<const SCEV *, 2> Operands;
+ for (int i = 0, e = Expr->getNumOperands(); i < e; ++i) {
+ const SCEV *Operand = visit(Expr->getOperand(i));
+ Operands.push_back(Operand);
+ }
+
+ return SE.getAddExpr(Operands);
+ }
+
+ const SCEV *visitMulExpr(const SCEVMulExpr *Expr) {
+ SmallVector<const SCEV *, 2> Operands;
+ for (int i = 0, e = Expr->getNumOperands(); i < e; ++i) {
+ const SCEV *Operand = visit(Expr->getOperand(i));
+ Operands.push_back(Operand);
+ }
+
+ return SE.getMulExpr(Operands);
+ }
+
+ const SCEV *visitUDivExpr(const SCEVUDivExpr *Expr) {
+ return SE.getUDivExpr(visit(Expr->getLHS()), visit(Expr->getRHS()));
+ }
+
+ // Return a new induction variable if the loop is within the original SCoP
+ // or NULL otherwise.
+ Value *getNewIV(const Loop *L) {
+ Value *IV = L->getCanonicalInductionVariable();
+ if (!IV)
+ return NULL;
+
+ ValueMapT::iterator NewIV = GlobalMap.find(IV);
+
+ if (NewIV == GlobalMap.end())
+ return NULL;
+
+ return NewIV->second;
+ }
+
+ const SCEV *visitAddRecExpr(const SCEVAddRecExpr *Expr) {
+ Value *IV;
+
+ IV = getNewIV(Expr->getLoop());
+
+ // The IV is not within the GlobalMaps. So do not rewrite it and also do
+ // not rewrite any descendants.
+ if (!IV)
+ return Expr;
+
+ assert(Expr->getNumOperands() == 2
+ && "An AddRecExpr with more than two operands can not be rewritten.");
+
+ const SCEV *Base, *Step, *IVExpr, *Product;
+
+ Base = visit(Expr->getStart());
+ Step = visit(Expr->getOperand(1));
+ IVExpr = SE.getUnknown(IV);
+ IVExpr = SE.getTruncateOrSignExtend(IVExpr, Step->getType());
+ Product = SE.getMulExpr(Step, IVExpr);
+
+ return SE.getAddExpr(Base, Product);
+ }
+
+ const SCEV *visitSMaxExpr(const SCEVSMaxExpr *Expr) {
+ SmallVector<const SCEV *, 2> Operands;
+ for (int i = 0, e = Expr->getNumOperands(); i < e; ++i) {
+ const SCEV *Operand = visit(Expr->getOperand(i));
+ Operands.push_back(Operand);
+ }
+
+ return SE.getSMaxExpr(Operands);
+ }
+
+ const SCEV *visitUMaxExpr(const SCEVUMaxExpr *Expr) {
+ SmallVector<const SCEV *, 2> Operands;
+ for (int i = 0, e = Expr->getNumOperands(); i < e; ++i) {
+ const SCEV *Operand = visit(Expr->getOperand(i));
+ Operands.push_back(Operand);
+ }
+
+ return SE.getUMaxExpr(Operands);
+ }
+
+ const SCEV *visitUnknown(const SCEVUnknown *Expr) {
+ Value *V = Expr->getValue();
+
+ if (GlobalMap.count(V))
+ return SE.getUnknown(GlobalMap[V]);
+
+ if (BBMap.count(V))
+ return SE.getUnknown(BBMap[V]);
+
+ return Expr;
+ }
+
+private:
+ Scop &S;
+ ScalarEvolution &SE;
+ ValueMapT &GlobalMap;
+ ValueMapT &BBMap;
+};
+
+
// Helper class to generate memory location.
namespace {
class IslGenerator {
@@ -141,7 +349,22 @@
BlockGenerator::BlockGenerator(IRBuilder<> &B, ScopStmt &Stmt, Pass *P):
- Builder(B), Statement(Stmt), P(P) {}
+ Builder(B), Statement(Stmt), P(P), SE(P->getAnalysis<ScalarEvolution>()) {}
+
+bool BlockGenerator::isSCEVIgnore(const Instruction *Inst) {
+ if (SCEVCodegen && SE.isSCEVable(Inst->getType()))
+ if (const SCEV *Scev = SE.getSCEV(const_cast<Instruction*>(Inst)))
+ if (!isa<SCEVCouldNotCompute>(Scev)) {
+ if (const SCEVUnknown *Unknown = dyn_cast<SCEVUnknown>(Scev)) {
+ if (Unknown->getValue() != Inst)
+ return true;
+ } else {
+ return true;
+ }
+ }
+
+ return false;
+}
Value *BlockGenerator::getNewValue(const Value *Old, ValueMapT &BBMap,
ValueMapT &GlobalMap) {
@@ -164,6 +387,20 @@
return BBMap[Old];
}
+ if (SCEVCodegen && SE.isSCEVable(Old->getType()))
+ if (const SCEV *Scev = SE.getSCEV(const_cast<Value*>(Old)))
+ if (!isa<SCEVCouldNotCompute>(Scev)) {
+ const SCEV *NewScev = SCEVRewriter::rewrite(Scev,
+ *Statement.getParent(), SE,
+ GlobalMap, BBMap);
+ SCEVExpander Expander(SE, "polly");
+ Value *Expanded = Expander.expandCodeFor(NewScev, Old->getType(),
+ Builder.GetInsertPoint());
+
+ BBMap[Old] = Expanded;
+ return Expanded;
+ }
+
// 'Old' is within the original SCoP, but was not rewritten.
//
// Such values appear, if they only calculate information already available in
@@ -299,6 +536,9 @@
if (Inst->isTerminator())
return;
+ if (isSCEVIgnore(Inst))
+ return;
+
if (const LoadInst *Load = dyn_cast<LoadInst>(Inst)) {
BBMap[Load] = generateScalarLoad(Load, BBMap, GlobalMap);
return;
@@ -589,6 +829,9 @@
if (Inst->isTerminator())
return;
+ if (isSCEVIgnore(Inst))
+ return;
+
if (const LoadInst *Load = dyn_cast<LoadInst>(Inst)) {
generateLoad(Load, VectorMap, ScalarMaps);
return;
Modified: polly/trunk/test/CodeGen/simple_vec_call_2.ll
URL: http://llvm.org/viewvc/llvm-project/polly/trunk/test/CodeGen/simple_vec_call_2.ll?rev=155717&r1=155716&r2=155717&view=diff
==============================================================================
--- polly/trunk/test/CodeGen/simple_vec_call_2.ll (original)
+++ polly/trunk/test/CodeGen/simple_vec_call_2.ll Fri Apr 27 11:36:14 2012
@@ -1,4 +1,5 @@
-; RUN: opt %loadPolly -basicaa -polly-codegen -enable-polly-vector -dce -S %s | FileCheck %s
+; RUN: opt %loadPolly -basicaa -polly-codegen -enable-polly-vector -polly-codegen-scev=false -dce -S %s | FileCheck %s
+; RUN: opt %loadPolly -basicaa -polly-codegen -enable-polly-vector -polly-codegen-scev=true -dce -S %s | FileCheck %s -check-prefix=CHECK-SCEV
target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
target triple = "x86_64-unknown-linux-gnu"
@@ -35,11 +36,26 @@
; CHECK: %1 = extractelement <4 x float> %value_p_splat, i32 1
; CHECK: %2 = extractelement <4 x float> %value_p_splat, i32 2
; CHECK: %3 = extractelement <4 x float> %value_p_splat, i32 3
-; CHECK: %p_result = tail call float** @foo(float %0) nounwind
-; CHECK: %p_result4 = tail call float** @foo(float %1) nounwind
-; CHECK: %p_result5 = tail call float** @foo(float %2) nounwind
-; CHECK: %p_result6 = tail call float** @foo(float %3) nounwind
-; CHECK: store float** %p_result, float*** %p_scevgep, align 4
-; CHECK: store float** %p_result4, float*** %p_scevgep1, align 4
-; CHECK: store float** %p_result5, float*** %p_scevgep2, align 4
-; CHECK: store float** %p_result6, float*** %p_scevgep3, align 4
+; CHECK: [[RES1:%[a-zA-Z0-9_]+]] = tail call float** @foo(float %0) nounwind
+; CHECK: [[RES2:%[a-zA-Z0-9_]+]] = tail call float** @foo(float %1) nounwind
+; CHECK: [[RES3:%[a-zA-Z0-9_]+]] = tail call float** @foo(float %2) nounwind
+; CHECK: [[RES4:%[a-zA-Z0-9_]+]] = tail call float** @foo(float %3) nounwind
+; CHECK: store float** [[RES1]], float*** %p_scevgep, align 4
+; CHECK: store float** [[RES2]], float*** %p_scevgep1, align 4
+; CHECK: store float** [[RES3]], float*** %p_scevgep2, align 4
+; CHECK: store float** [[RES4]], float*** %p_scevgep3, align 4
+
+; CHECK-SCEV: %value_p_splat_one = load <1 x float>* bitcast ([1024 x float]* @A to <1 x float>*), align 8
+; CHECK-SCEV: %value_p_splat = shufflevector <1 x float> %value_p_splat_one, <1 x float> %value_p_splat_one, <4 x i32> zeroinitializer
+; CHECK-SCEV: %0 = extractelement <4 x float> %value_p_splat, i32 0
+; CHECK-SCEV: %1 = extractelement <4 x float> %value_p_splat, i32 1
+; CHECK-SCEV: %2 = extractelement <4 x float> %value_p_splat, i32 2
+; CHECK-SCEV: %3 = extractelement <4 x float> %value_p_splat, i32 3
+; CHECK-SCEV: [[RES1:%[a-zA-Z0-9_]+]] = tail call float** @foo(float %0) nounwind
+; CHECK-SCEV: [[RES2:%[a-zA-Z0-9_]+]] = tail call float** @foo(float %1) nounwind
+; CHECK-SCEV: [[RES3:%[a-zA-Z0-9_]+]] = tail call float** @foo(float %2) nounwind
+; CHECK-SCEV: [[RES4:%[a-zA-Z0-9_]+]] = tail call float** @foo(float %3) nounwind
+; CHECK-SCEV: store float** [[RES1]], float*** getelementptr inbounds ([1024 x float**]* @B, i64 0, i64 0), align 4
+; CHECK-SCEV: store float** [[RES2]], float*** getelementptr inbounds ([1024 x float**]* @B, i64 0, i64 1), align 4
+; CHECK-SCEV: store float** [[RES3]], float*** getelementptr inbounds ([1024 x float**]* @B, i64 0, i64 2), align 4
+; CHECK-SCEV: store float** [[RES4]], float*** getelementptr inbounds ([1024 x float**]* @B, i64 0, i64 3), align 4
Modified: polly/trunk/test/CodeGen/simple_vec_impossible.ll
URL: http://llvm.org/viewvc/llvm-project/polly/trunk/test/CodeGen/simple_vec_impossible.ll?rev=155717&r1=155716&r2=155717&view=diff
==============================================================================
--- polly/trunk/test/CodeGen/simple_vec_impossible.ll (original)
+++ polly/trunk/test/CodeGen/simple_vec_impossible.ll Fri Apr 27 11:36:14 2012
@@ -1,4 +1,5 @@
-; RUN: opt %loadPolly -basicaa -polly-codegen -enable-polly-vector -S %s | FileCheck %s
+; RUN: opt %loadPolly -basicaa -polly-codegen -enable-polly-vector -S -polly-codegen-scev=false %s | FileCheck %s
+; RUN: opt %loadPolly -basicaa -polly-codegen -enable-polly-vector -S -polly-codegen-scev=true %s | FileCheck %s -check-prefix=CHECK-SCEV
target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
target triple = "x86_64-unknown-linux-gnu"
@@ -36,3 +37,12 @@
; CHECK: store float** %value_p_scalar_4, float*** %p_scevgep1, align 4
; CHECK: store float** %value_p_scalar_5, float*** %p_scevgep2, align 4
; CHECK: store float** %value_p_scalar_6, float*** %p_scevgep3, align 4
+
+; CHECK-SCEV: %value_p_scalar_ = load float*** getelementptr inbounds ([1024 x float**]* @A, i64 0, i64 0)
+; CHECK-SCEV: %value_p_scalar_1 = load float*** getelementptr inbounds ([1024 x float**]* @A, i64 0, i64 0)
+; CHECK-SCEV: %value_p_scalar_2 = load float*** getelementptr inbounds ([1024 x float**]* @A, i64 0, i64 0)
+; CHECK-SCEV: %value_p_scalar_3 = load float*** getelementptr inbounds ([1024 x float**]* @A, i64 0, i64 0)
+; CHECK-SCEV: store float** %value_p_scalar_, float*** getelementptr inbounds ([1024 x float**]* @B, i64 0, i64 0), align 4
+; CHECK-SCEV: store float** %value_p_scalar_1, float*** getelementptr inbounds ([1024 x float**]* @B, i64 0, i64 1), align 4
+; CHECK-SCEV: store float** %value_p_scalar_2, float*** getelementptr inbounds ([1024 x float**]* @B, i64 0, i64 2), align 4
+; CHECK-SCEV: store float** %value_p_scalar_3, float*** getelementptr inbounds ([1024 x float**]* @B, i64 0, i64 3), align 4
More information about the llvm-commits
mailing list