[Polly] Put the parallel context into the function entry block

[jdoerfert]

Summary

We use lifetime markers to limit the actual life range.

[grosser]

No test case changes?

Otherwise LGTM

lib/CodeGen/LoopGenerators.cpp

296 ↗

(On Diff #14333)

I don't think this XXX is even needed. AFAIK clang will always allocate all stack space in the entry block.

[dpeixott]

Needs test case. Otherwise LGTM.

lib/CodeGen/LoopGenerators.cpp

177 ↗

(On Diff #14333)

Use cast<> instead of dyn_cast<> when you don't check for null return result.

302 ↗

(On Diff #14333)

Use cast<> instead of dyn_cast<>

[jdoerfert]

Added test and use DataLayout for best results

[jdoerfert]

Closed by commit rL219005 (authored by @jdoerfert).

polly/trunk/lib/CodeGen/LoopGenerators.cpp

Value *ParallelLoopGenerator::createParallelLoop(
   // whereas the codegenForSequential function creates a <= comparison.
   UB = Builder.CreateAdd(UB, ConstantInt::get(LongType, 1));
 
   // Tell the runtime we start a parallel loop
   createCallSpawnThreads(SubFn, SubFnParam, LB, UB, Stride);
   Builder.CreateCall(SubFn, SubFnParam);
   createCallJoinThreads();
 
+  // Mark the end of the lifetime for the parameter struct.
+  Type *Ty = Struct->getType();
+  ConstantInt *SizeOf = Builder.getInt64(DL.getTypeAllocSize(Ty));
+  Builder.CreateLifetimeEnd(Struct, SizeOf);
+
   return IV;
 }
 
 void ParallelLoopGenerator::createCallSpawnThreads(Value *SubFn,
                                                    Value *SubFnParam, Value *LB,
                                                    Value *UB, Value *Stride) {
   const std::string Name = "GOMP_parallel_loop_runtime_start";
 

 
 Value *
 ParallelLoopGenerator::storeValuesIntoStruct(SetVector<Value *> &Values) {
   SmallVector<Type *, 8> Members;
 
   for (Value *V : Values)
     Members.push_back(V->getType());
 
+  // We do not want to allocate the alloca inside any loop, thus we allocate it
+  // in the entry block of the function and use annotations to denote the actual
+  // live span (similar to clang).
+  BasicBlock &EntryBB = Builder.GetInsertBlock()->getParent()->getEntryBlock();
+  Instruction *IP = EntryBB.getFirstInsertionPt();
   StructType *Ty = StructType::get(Builder.getContext(), Members);
-  Value *Struct =
-      new AllocaInst(Ty, 0, "polly.par.userContext", Builder.GetInsertPoint());
+  Value *Struct = new AllocaInst(Ty, 0, "polly.par.userContext", IP);
+
+  // Mark the start of the lifetime for the parameter struct.
+  ConstantInt *SizeOf = Builder.getInt64(DL.getTypeAllocSize(Ty));
+  Builder.CreateLifetimeStart(Struct, SizeOf);
 
   for (unsigned i = 0; i < Values.size(); i++) {
     Value *Address = Builder.CreateStructGEP(Struct, i);
     Builder.CreateStore(Values[i], Address);
   }
 
   return Struct;
 }

polly/trunk/test/Cloog/CodeGen/OpenMP/lifetime_markers.ll

+; RUN: opt %loadPolly -S -polly-codegen -enable-polly-openmp < %s | FileCheck %s
+;
+; Check that we allocate the parallel context in the entry block and use
+; lifetime markers to mark the live range.
+;
+; CHECK: entry:
+; CHECK:   %polly.par.userContext = alloca { i32* }
+; CHECK:   br label %while.cond
+;
+; CHECK:       polly.start:
+; CHECK-NEXT:    %[[BC1:[._0-9a-zA-Z]*]] = bitcast { i32* }* %polly.par.userContext to i8*
+; CHECK-NEXT:    call void @llvm.lifetime.start(i64 8, i8* %[[BC1]])
+; CHECK-NEXT:    %[[GEP:[._0-9a-zA-Z]*]] = getelementptr inbounds { i32* }* %polly.par.userContext, i32 0, i32 0
+; CHECK-NEXT:    store i32* %A, i32** %[[GEP]]
+; CHECK-NEXT:    %polly.par.userContext{{[0-9]*}} = bitcast { i32* }* %polly.par.userContext to i8*
+; CHECK-NEXT:    call void @GOMP_parallel_loop_runtime_start(void (i8*)* @jd.polly.subfn, i8* %polly.par.userContext{{[0-9]*}}, i64 0, i64 0, i64 1024, i64 1)
+; CHECK-NEXT:    call void @jd.polly.subfn(i8* %polly.par.userContext{{[0-9]*}})
+; CHECK-NEXT:    call void @GOMP_parallel_end()
+; CHECK-NEXT:    %[[BC2:[._0-9a-zA-Z]*]] = bitcast { i32* }* %polly.par.userContext to i8*
+; CHECK-NEXT:    call void @llvm.lifetime.end(i64 8, i8* %[[BC2]])
+; CHECK-NEXT:    br label %polly.merge_new_and_old
+
+;    int cond();
+;    void jd(int *A) {
+;      while (cond())
+;        for (int j = 0; j < 1024; j++)
+;          A[j] = 1;
+;    }
+;
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+
+define void @jd(i32* %A) {
+entry:
+  br label %while.cond
+
+while.cond:                                       ; preds = %for.end, %entry
+  %call = call i32 (...)* @cond() #2
+  %tobool = icmp eq i32 %call, 0
+  br i1 %tobool, label %while.end, label %while.body
+
+while.body:                                       ; preds = %while.cond
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.inc, %while.body
+  %indvars.iv = phi i64 [ %indvars.iv.next, %for.inc ], [ 0, %while.body ]
+  %exitcond = icmp ne i64 %indvars.iv, 1024
+  br i1 %exitcond, label %for.body, label %for.end
+
+for.body:                                         ; preds = %for.cond
+  %arrayidx = getelementptr inbounds i32* %A, i64 %indvars.iv
+  store i32 1, i32* %arrayidx, align 4
+  br label %for.inc
+
+for.inc:                                          ; preds = %for.body
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+  br label %for.cond
+
+for.end:                                          ; preds = %for.cond
+  br label %while.cond
+
+while.end:                                        ; preds = %while.cond
+  ret void
+}
+
+declare i32 @cond(...) #1