From: Tobias Grosser <tobias@grosser.es>
Date: Thu, 4 Aug 2016 13:57:29 +0000 (+0000)
Subject: GPGPU: Support scalars that are mapped to shared memory
X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=f919d8b36071bdacd699e28d7788bb904ba5f23d;p=platform%2Fupstream%2Fllvm.git

GPGPU: Support scalars that are mapped to shared memory

llvm-svn: 277726
---

diff --git a/polly/lib/CodeGen/IslExprBuilder.cpp b/polly/lib/CodeGen/IslExprBuilder.cpp
index 75e3507..30f4b0b 100644
--- a/polly/lib/CodeGen/IslExprBuilder.cpp
+++ b/polly/lib/CodeGen/IslExprBuilder.cpp
@@ -212,7 +212,7 @@ Value *IslExprBuilder::createAccessAddress(isl_ast_expr *Expr) {
          "isl ast expression not of type isl_ast_op");
   assert(isl_ast_expr_get_op_type(Expr) == isl_ast_op_access &&
          "not an access isl ast expression");
-  assert(isl_ast_expr_get_op_n_arg(Expr) >= 2 &&
+  assert(isl_ast_expr_get_op_n_arg(Expr) >= 1 &&
          "We need at least two operands to create a member access.");
 
   Value *Base, *IndexOp, *Access;
@@ -250,6 +250,11 @@ Value *IslExprBuilder::createAccessAddress(isl_ast_expr *Expr) {
         Builder.CreateBitCast(Base, PointerTy, "polly.access.cast." + BaseName);
   }
 
+  if (isl_ast_expr_get_op_n_arg(Expr) == 1) {
+    isl_ast_expr_free(Expr);
+    return Base;
+  }
+
   IndexOp = nullptr;
   for (unsigned u = 1, e = isl_ast_expr_get_op_n_arg(Expr); u < e; u++) {
     Value *NextIndex = create(isl_ast_expr_get_op_arg(Expr, u));
diff --git a/polly/lib/CodeGen/PPCGCodeGeneration.cpp b/polly/lib/CodeGen/PPCGCodeGeneration.cpp
index 8b6679b..481e20b 100644
--- a/polly/lib/CodeGen/PPCGCodeGeneration.cpp
+++ b/polly/lib/CodeGen/PPCGCodeGeneration.cpp
@@ -1252,16 +1252,12 @@ void GPUNodeBuilder::createKernelVariables(ppcg_kernel *Kernel, Function *FN) {
     isl_id *Id = isl_space_get_tuple_id(Var.array->space, isl_dim_set);
     Type *EleTy = ScopArrayInfo::getFromId(Id)->getElementType();
 
+    Type *ArrayTy = EleTy;
     SmallVector<const SCEV *, 4> Sizes;
-    isl_val *V0 = isl_vec_get_element_val(Var.size, 0);
-    long Bound = isl_val_get_num_si(V0);
-    isl_val_free(V0);
-    Sizes.push_back(S.getSE()->getConstant(Builder.getInt64Ty(), Bound));
 
-    ArrayType *ArrayTy = ArrayType::get(EleTy, Bound);
-    for (unsigned int j = 1; j < Var.array->n_index; ++j) {
+    for (unsigned int j = 0; j < Var.array->n_index; ++j) {
       isl_val *Val = isl_vec_get_element_val(Var.size, j);
-      Bound = isl_val_get_num_si(Val);
+      long Bound = isl_val_get_num_si(Val);
       isl_val_free(Val);
       Sizes.push_back(S.getSE()->getConstant(Builder.getInt64Ty(), Bound));
       ArrayTy = ArrayType::get(ArrayTy, Bound);
@@ -1274,8 +1270,8 @@ void GPUNodeBuilder::createKernelVariables(ppcg_kernel *Kernel, Function *FN) {
           *M, ArrayTy, false, GlobalValue::InternalLinkage, 0, Var.name,
           nullptr, GlobalValue::ThreadLocalMode::NotThreadLocal, 3);
       GlobalVar->setAlignment(EleTy->getPrimitiveSizeInBits() / 8);
-      ConstantAggregateZero *Zero = ConstantAggregateZero::get(ArrayTy);
-      GlobalVar->setInitializer(Zero);
+      GlobalVar->setInitializer(Constant::getNullValue(ArrayTy));
+
       Allocation = GlobalVar;
     } else if (Var.type == ppcg_access_private) {
       Allocation = Builder.CreateAlloca(ArrayTy, 0, "private_array");
diff --git a/polly/test/GPGPU/shared-memory-scalar.ll b/polly/test/GPGPU/shared-memory-scalar.ll
new file mode 100644
index 0000000..d10a6ac
--- /dev/null
+++ b/polly/test/GPGPU/shared-memory-scalar.ll
@@ -0,0 +1,75 @@
+; RUN: opt %loadPolly -polly-codegen-ppcg -polly-acc-dump-code \
+; RUN: -polly-acc-use-shared \
+; RUN: -disable-output < %s | \
+; RUN: FileCheck -check-prefix=CODE %s
+
+; RUN: opt %loadPolly -polly-codegen-ppcg \
+; RUN: -polly-acc-use-shared \
+; RUN: -disable-output -polly-acc-dump-kernel-ir < %s | \
+; RUN: FileCheck -check-prefix=KERNEL %s
+
+; REQUIRES: pollyacc
+
+;    void add(float *A, float alpha) {
+;      for (long i = 0; i < 32; i++)
+;        for (long j = 0; j < 10; j++)
+;          A[i] += alpha;
+;    }
+
+; CODE:  read(t0);
+; CODE-NEXT:  if (t0 == 0)
+; CODE-NEXT:    read();
+; CODE-NEXT:  sync0();
+; CODE-NEXT:  for (int c3 = 0; c3 <= 9; c3 += 1)
+; CODE-NEXT:    Stmt_bb5(t0, c3);
+; CODE-NEXT:  sync1();
+; CODE-NEXT:  write(t0);
+
+
+; KERNEL: @shared_MemRef_alpha = internal addrspace(3) global float 0.000000e+00, align 4
+
+; KERNEL:  %polly.access.cast.MemRef_alpha = bitcast i8* %MemRef_alpha to float*
+; KERNEL-NEXT:  %shared.read1 = load float, float* %polly.access.cast.MemRef_alpha
+; KERNEL-NEXT:  store float %shared.read1, float addrspace(3)* @shared_MemRef_alpha
+
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+
+define void @add(float* %A, float %alpha) {
+bb:
+  br label %bb2
+
+bb2:                                              ; preds = %bb11, %bb
+  %i.0 = phi i64 [ 0, %bb ], [ %tmp12, %bb11 ]
+  %exitcond1 = icmp ne i64 %i.0, 32
+  br i1 %exitcond1, label %bb3, label %bb13
+
+bb3:                                              ; preds = %bb2
+  br label %bb4
+
+bb4:                                              ; preds = %bb8, %bb3
+  %j.0 = phi i64 [ 0, %bb3 ], [ %tmp9, %bb8 ]
+  %exitcond = icmp ne i64 %j.0, 10
+  br i1 %exitcond, label %bb5, label %bb10
+
+bb5:                                              ; preds = %bb4
+  %tmp = getelementptr inbounds float, float* %A, i64 %i.0
+  %tmp6 = load float, float* %tmp, align 4
+  %tmp7 = fadd float %tmp6, %alpha
+  store float %tmp7, float* %tmp, align 4
+  br label %bb8
+
+bb8:                                              ; preds = %bb5
+  %tmp9 = add nuw nsw i64 %j.0, 1
+  br label %bb4
+
+bb10:                                             ; preds = %bb4
+  br label %bb11
+
+bb11:                                             ; preds = %bb10
+  %tmp12 = add nuw nsw i64 %i.0, 1
+  br label %bb2
+
+bb13:                                             ; preds = %bb2
+  ret void
+}