[OpenCL] Always emit alloca in entry block for enqueue_kernel builtin

[scott.linder]

Summary

Ensures the statically sized alloca is not converted to DYNAMIC_STACKALLOC later because it is not in the entry block.

I believe it is valid to insert the alloca in the entry block, but I'm not confident the way I accomplish it is correct.

[yaxunl]

lib/CodeGen/CGBuiltin.cpp

3342

You may try CreateMemTemp. It should handle the insert position and also debug info.

test/CodeGenOpenCL/enqueue-kernel-non-entry-block.cl

3

Can we have a run line for debug info?

[arsenm]

Should this also test for lifetime markers?

[scott.linder]

Address feedback; I hope I understood correctly what debug info to check for.

I don't see where in CreateMemTemp and friends EmitLifetimeStart gets called, and I don't see any lifetime intrinsics in the IR even at -O1.

[svenvh]

You'll probably also need to update test/CodeGenOpenCL/cl20-device-side-enqueue.cl; please verify with make/ninja check-clang.

[scott.linder]

Update test

[svenvh]

LGTM, thanks!

[yaxunl]

In D50104#1184362, @scott.linder wrote:

Address feedback; I hope I understood correctly what debug info to check for.

I don't see where in CreateMemTemp and friends EmitLifetimeStart gets called, and I don't see any lifetime intrinsics in the IR even at -O1.

Emitting lifetime intrinsic is optional. In this case, since the life time of the temp var is just before and after the function call, emitting lifetime intrinsics can help optimizers.

It can be done by code like this:

 if (auto *Size = EmitLifetimeStart(
         CGM.getDataLayout().getTypeAllocSize(Alloca.getElementType()),
         Alloca.getPointer())) {
   pushFullExprCleanup<CallLifetimeEnd>(NormalEHLifetimeMarker, Alloca,
                                          Size);
}

Then the lifetime.start should be emitted before the function call and lifetime.end should be emitted just after the function call.

[scott.linder]

I still don't quite see what you describe; with that change all of the lifetime.end calls pile up just before the enclosing function returns, not after each call to enqueue_kernel. Looking at https://clang.llvm.org/doxygen/EHScopeStack_8h_source.html#l00078 I don't see any option which isn't based on scope. The lifetime.start calls do occur where I would expect, though, so I will update the patch.

[yaxunl]

In D50104#1185920, @scott.linder wrote:

I still don't quite see what you describe; with that change all of the lifetime.end calls pile up just before the enclosing function returns, not after each call to enqueue_kernel. Looking at https://clang.llvm.org/doxygen/EHScopeStack_8h_source.html#l00078 I don't see any option which isn't based on scope. The lifetime.start calls do occur where I would expect, though, so I will update the patch.

Sorry my mistake. In this case, the full expressions seems to be the calling function, so using pushFullExprCleanup to emit lifetime.end does not work well here.

You need to call EmitLifetimeEnd explicitly after emitting the function call.

[scott.linder]

Emit lifetime intrinsics for the sizes temp, and update test

[yaxunl]

LGTM. Thanks!

lib/CodeGen/CGBuiltin.cpp

if (NumArgs == 4) {
       RTCall->setAttributes(ByValAttrSet);
       return RValue::get(RTCall);
     }
     assert(NumArgs >= 5 && "Invalid enqueue_kernel signature");
 
     // Create a temporary array to hold the sizes of local pointer arguments
     // for the block. \p First is the position of the first size argument.
     auto CreateArrayForSizeVar = [=](unsigned First) {
-      auto *AT = llvm::ArrayType::get(SizeTy, NumArgs - First);
-      auto *Arr = Builder.CreateAlloca(AT);
+      llvm::APInt ArraySize(32, NumArgs - First);
+      QualType SizeArrayTy = getContext().getConstantArrayType(
+          getContext().getSizeType(), ArraySize, ArrayType::Normal,
+          /*IndexTypeQuals=*/0);
+      auto Tmp = CreateMemTemp(SizeArrayTy, "block_sizes");
       llvm::Value *Ptr;
       // Each of the following arguments specifies the size of the corresponding
       // argument passed to the enqueued block.
       auto *Zero = llvm::ConstantInt::get(IntTy, 0);
       for (unsigned I = First; I < NumArgs; ++I) {
         auto *Index = llvm::ConstantInt::get(IntTy, I - First);
-        auto *GEP = Builder.CreateGEP(Arr, {Zero, Index});
+        auto *GEP = Builder.CreateGEP(Tmp.getPointer(), {Zero, Index});
         if (I == First)
           Ptr = GEP;
         auto *V =
             Builder.CreateZExtOrTrunc(EmitScalarExpr(E->getArg(I)), SizeTy);
         Builder.CreateAlignedStore(
             V, GEP, CGM.getDataLayout().getPrefTypeAlignment(SizeTy));
       }
       return Ptr;

test/CodeGenOpenCL/cl20-device-side-enqueue.cl

kernel void device_side_enqueue(global int *a, global int *b, int i) {
   ndrange_t ndrange;
   // COMMON: %clk_event = alloca %opencl.clk_event_t*
   clk_event_t clk_event;
   // COMMON: %event_wait_list = alloca %opencl.clk_event_t*
   clk_event_t event_wait_list;
   // COMMON: %event_wait_list2 = alloca [1 x %opencl.clk_event_t*]
   clk_event_t event_wait_list2[] = {clk_event};
 
-  // Emits block literal on stack and block kernel [[INVLK1]].
   // COMMON: [[NDR:%[a-z0-9]+]] = alloca %struct.ndrange_t, align 4
+
+  // B32: %[[BLOCK_SIZES1:.*]] = alloca [1 x i32]
+  // B64: %[[BLOCK_SIZES1:.*]] = alloca [1 x i64]
+  // B32: %[[BLOCK_SIZES2:.*]] = alloca [1 x i32]
+  // B64: %[[BLOCK_SIZES2:.*]] = alloca [1 x i64]
+  // B32: %[[BLOCK_SIZES3:.*]] = alloca [1 x i32]
+  // B64: %[[BLOCK_SIZES3:.*]] = alloca [1 x i64]
+  // B32: %[[BLOCK_SIZES4:.*]] = alloca [1 x i32]
+  // B64: %[[BLOCK_SIZES4:.*]] = alloca [1 x i64]
+  // B32: %[[BLOCK_SIZES5:.*]] = alloca [1 x i32]
+  // B64: %[[BLOCK_SIZES5:.*]] = alloca [1 x i64]
+  // B32: %[[BLOCK_SIZES6:.*]] = alloca [3 x i32]
+  // B64: %[[BLOCK_SIZES6:.*]] = alloca [3 x i64]
+  // B32: %[[BLOCK_SIZES7:.*]] = alloca [1 x i32]
+  // B64: %[[BLOCK_SIZES7:.*]] = alloca [1 x i64]
+
+  // Emits block literal on stack and block kernel [[INVLK1]].
   // COMMON: [[DEF_Q:%[0-9]+]] = load %opencl.queue_t{{.*}}*, %opencl.queue_t{{.*}}** %default_queue
   // COMMON: [[FLAGS:%[0-9]+]] = load i32, i32* %flags
   // B32: [[BL:%[0-9]+]] = bitcast <{ i32, i32, i32 addrspace(1)*, i32, i32 addrspace(1)* }>* %block to void ()*
   // B64: [[BL:%[0-9]+]] = bitcast <{ i32, i32, i32 addrspace(1)*, i32 addrspace(1)*, i32 }>* %block to void ()*
   // COMMON: [[BL_I8:%[0-9]+]] = addrspacecast void ()* [[BL]] to i8 addrspace(4)*
   // COMMON-LABEL: call i32 @__enqueue_kernel_basic(
   // COMMON-SAME: %opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* byval [[NDR]]{{([0-9]+)?}},
   // COMMON-SAME: i8 addrspace(4)* addrspacecast (i8* bitcast ({{.*}} [[INVLK1:[^ ]+_kernel]] to i8*) to i8 addrspace(4)*),
kernel void device_side_enqueue(global int *a, global int *b, int i) {
   // COMMON: [[WAIT_EVNT:%[0-9]+]] = addrspacecast %opencl.clk_event_t{{.*}}** %event_wait_list to %opencl.clk_event_t{{.*}}* addrspace(4)*
   // COMMON: [[EVNT:%[0-9]+]] = addrspacecast %opencl.clk_event_t{{.*}}** %clk_event to %opencl.clk_event_t{{.*}}* addrspace(4)*
   // COMMON: [[BL:%[0-9]+]] = bitcast <{ i32, i32, i32{{.*}}, i32{{.*}}, i32{{.*}} }>* %block3 to void ()*
   // COMMON: [[BL_I8:%[0-9]+]] = addrspacecast void ()* [[BL]] to i8 addrspace(4)*
   // COMMON-LABEL: call i32 @__enqueue_kernel_basic_events
   // COMMON-SAME: (%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]],  %struct.ndrange_t* {{.*}}, i32 2, %opencl.clk_event_t{{.*}}* addrspace(4)* [[WAIT_EVNT]], %opencl.clk_event_t{{.*}}* addrspace(4)* [[EVNT]],
   // COMMON-SAME: i8 addrspace(4)* addrspacecast (i8* bitcast ({{.*}} [[INVLK2:[^ ]+_kernel]] to i8*) to i8 addrspace(4)*),
   // COMMON-SAME: i8 addrspace(4)* [[BL_I8]])
-
   enqueue_kernel(default_queue, flags, ndrange, 2, &event_wait_list, &clk_event,
                  ^(void) {
                    a[i] = b[i];
                  });
 
   // Emits global block literal [[BLG1]] and block kernel [[INVGK1]].
   // COMMON: [[DEF_Q:%[0-9]+]] = load %opencl.queue_t{{.*}}*, %opencl.queue_t{{.*}}** %default_queue
   // COMMON: [[FLAGS:%[0-9]+]] = load i32, i32* %flags
-  // B32: %[[TMP:.*]] = alloca [1 x i32]
-  // B32: %[[TMP1:.*]] = getelementptr [1 x i32], [1 x i32]* %[[TMP]], i32 0, i32 0
-  // B32: store i32 256, i32* %[[TMP1]], align 4
-  // B64: %[[TMP:.*]] = alloca [1 x i64]
-  // B64: %[[TMP1:.*]] = getelementptr [1 x i64], [1 x i64]* %[[TMP]], i32 0, i32 0
-  // B64: store i64 256, i64* %[[TMP1]], align 8
+  // B32: %[[TMP:.*]] = getelementptr [1 x i32], [1 x i32]* %[[BLOCK_SIZES1]], i32 0, i32 0
+  // B32: store i32 256, i32* %[[TMP]], align 4
+  // B64: %[[TMP:.*]] = getelementptr [1 x i64], [1 x i64]* %[[BLOCK_SIZES1]], i32 0, i32 0
+  // B64: store i64 256, i64* %[[TMP]], align 8
   // COMMON-LABEL: call i32 @__enqueue_kernel_varargs(
   // COMMON-SAME: %opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* [[NDR]]{{([0-9]+)?}},
   // COMMON-SAME: i8 addrspace(4)* addrspacecast (i8* bitcast ({{.*}} [[INVGK1:[^ ]+_kernel]] to i8*) to i8 addrspace(4)*),
   // COMMON-SAME: i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i32, i32 } addrspace(1)* [[BLG1]] to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1,
-  // B32-SAME: i32* %[[TMP1]])
-  // B64-SAME: i64* %[[TMP1]])
+  // B32-SAME: i32* %[[TMP]])
+  // B64-SAME: i64* %[[TMP]])
   enqueue_kernel(default_queue, flags, ndrange,
                  ^(local void *p) {
                    return;
                  },
                  256);
+
   char c;
   // Emits global block literal [[BLG2]] and block kernel [[INVGK2]].
   // COMMON: [[DEF_Q:%[0-9]+]] = load %opencl.queue_t{{.*}}*, %opencl.queue_t{{.*}}** %default_queue
   // COMMON: [[FLAGS:%[0-9]+]] = load i32, i32* %flags
-  // B32: %[[TMP:.*]] = alloca [1 x i32]
-  // B32: %[[TMP1:.*]] = getelementptr [1 x i32], [1 x i32]* %[[TMP]], i32 0, i32 0
-  // B32: store i32 %{{.*}}, i32* %[[TMP1]], align 4
-  // B64: %[[TMP:.*]] = alloca [1 x i64]
-  // B64: %[[TMP1:.*]] = getelementptr [1 x i64], [1 x i64]* %[[TMP]], i32 0, i32 0
-  // B64: store i64 %{{.*}}, i64* %[[TMP1]], align 8
+  // B32: %[[TMP:.*]] = getelementptr [1 x i32], [1 x i32]* %[[BLOCK_SIZES2]], i32 0, i32 0
+  // B32: store i32 %{{.*}}, i32* %[[TMP]], align 4
+  // B64: %[[TMP:.*]] = getelementptr [1 x i64], [1 x i64]* %[[BLOCK_SIZES2]], i32 0, i32 0
+  // B64: store i64 %{{.*}}, i64* %[[TMP]], align 8
   // COMMON-LABEL: call i32 @__enqueue_kernel_varargs(
   // COMMON-SAME: %opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* [[NDR]]{{([0-9]+)?}},
   // COMMON-SAME: i8 addrspace(4)* addrspacecast (i8* bitcast ({{.*}} [[INVGK2:[^ ]+_kernel]] to i8*) to i8 addrspace(4)*),
   // COMMON-SAME: i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i32, i32 } addrspace(1)* [[BLG2]] to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1,
-  // B32-SAME: i32* %[[TMP1]])
-  // B64-SAME: i64* %[[TMP1]])
+  // B32-SAME: i32* %[[TMP]])
+  // B64-SAME: i64* %[[TMP]])
   enqueue_kernel(default_queue, flags, ndrange,
                  ^(local void *p) {
                    return;
                  },
                  c);
 
   // Emits global block literal [[BLG3]] and block kernel [[INVGK3]].
   // COMMON: [[DEF_Q:%[0-9]+]] = load %opencl.queue_t{{.*}}*, %opencl.queue_t{{.*}}** %default_queue
   // COMMON: [[FLAGS:%[0-9]+]] = load i32, i32* %flags
   // COMMON: [[AD:%arraydecay[0-9]*]] = getelementptr inbounds [1 x %opencl.clk_event_t*], [1 x %opencl.clk_event_t*]* %event_wait_list2, i32 0, i32 0
   // COMMON: [[WAIT_EVNT:%[0-9]+]] = addrspacecast %opencl.clk_event_t{{.*}}** [[AD]] to %opencl.clk_event_t{{.*}}* addrspace(4)*
   // COMMON: [[EVNT:%[0-9]+]]  = addrspacecast %opencl.clk_event_t{{.*}}** %clk_event to %opencl.clk_event_t{{.*}}* addrspace(4)*
-  // B32: %[[TMP:.*]] = alloca [1 x i32]
-  // B32: %[[TMP1:.*]] = getelementptr [1 x i32], [1 x i32]* %[[TMP]], i32 0, i32 0
-  // B32: store i32 256, i32* %[[TMP1]], align 4
-  // B64: %[[TMP:.*]] = alloca [1 x i64]
-  // B64: %[[TMP1:.*]] = getelementptr [1 x i64], [1 x i64]* %[[TMP]], i32 0, i32 0
-  // B64: store i64 256, i64* %[[TMP1]], align 8
+  // B32: %[[TMP:.*]] = getelementptr [1 x i32], [1 x i32]* %[[BLOCK_SIZES3]], i32 0, i32 0
+  // B32: store i32 256, i32* %[[TMP]], align 4
+  // B64: %[[TMP:.*]] = getelementptr [1 x i64], [1 x i64]* %[[BLOCK_SIZES3]], i32 0, i32 0
+  // B64: store i64 256, i64* %[[TMP]], align 8
   // COMMON-LABEL: call i32 @__enqueue_kernel_events_varargs
   // COMMON-SAME: (%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]],  %struct.ndrange_t* {{.*}}, i32 2, %opencl.clk_event_t{{.*}} [[WAIT_EVNT]], %opencl.clk_event_t{{.*}} [[EVNT]],
   // COMMON-SAME: i8 addrspace(4)* addrspacecast (i8* bitcast ({{.*}} [[INVGK3:[^ ]+_kernel]] to i8*) to i8 addrspace(4)*),
   // COMMON-SAME: i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i32, i32 } addrspace(1)* [[BLG3]] to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1,
-  // B32-SAME: i32* %[[TMP1]])
-  // B64-SAME: i64* %[[TMP1]])
+  // B32-SAME: i32* %[[TMP]])
+  // B64-SAME: i64* %[[TMP]])
   enqueue_kernel(default_queue, flags, ndrange, 2, event_wait_list2, &clk_event,
                  ^(local void *p) {
                    return;
                  },
                  256);
 
   // Emits global block literal [[BLG4]] and block kernel [[INVGK4]].
   // COMMON: [[DEF_Q:%[0-9]+]] = load %opencl.queue_t{{.*}}*, %opencl.queue_t{{.*}}** %default_queue
   // COMMON: [[FLAGS:%[0-9]+]] = load i32, i32* %flags
   // COMMON: [[AD:%arraydecay[0-9]*]] = getelementptr inbounds [1 x %opencl.clk_event_t*], [1 x %opencl.clk_event_t*]* %event_wait_list2, i32 0, i32 0
   // COMMON: [[WAIT_EVNT:%[0-9]+]] = addrspacecast %opencl.clk_event_t{{.*}}** [[AD]] to %opencl.clk_event_t{{.*}}* addrspace(4)*
   // COMMON: [[EVNT:%[0-9]+]]  = addrspacecast %opencl.clk_event_t{{.*}}** %clk_event to %opencl.clk_event_t{{.*}}* addrspace(4)*
-  // B32: %[[TMP:.*]] = alloca [1 x i32]
-  // B32: %[[TMP1:.*]] = getelementptr [1 x i32], [1 x i32]* %[[TMP]], i32 0, i32 0
-  // B32: store i32 %{{.*}}, i32* %[[TMP1]], align 4
-  // B64: %[[TMP:.*]] = alloca [1 x i64]
-  // B64: %[[TMP1:.*]] = getelementptr [1 x i64], [1 x i64]* %[[TMP]], i32 0, i32 0
-  // B64: store i64 %{{.*}}, i64* %[[TMP1]], align 8
+  // B32: %[[TMP:.*]] = getelementptr [1 x i32], [1 x i32]* %[[BLOCK_SIZES4]], i32 0, i32 0
+  // B32: store i32 %{{.*}}, i32* %[[TMP]], align 4
+  // B64: %[[TMP:.*]] = getelementptr [1 x i64], [1 x i64]* %[[BLOCK_SIZES4]], i32 0, i32 0
+  // B64: store i64 %{{.*}}, i64* %[[TMP]], align 8
   // COMMON-LABEL: call i32 @__enqueue_kernel_events_varargs
   // COMMON-SAME: (%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]],  %struct.ndrange_t* {{.*}}, i32 2, %opencl.clk_event_t{{.*}}* addrspace(4)* [[WAIT_EVNT]], %opencl.clk_event_t{{.*}}* addrspace(4)* [[EVNT]],
   // COMMON-SAME: i8 addrspace(4)* addrspacecast (i8* bitcast ({{.*}} [[INVGK4:[^ ]+_kernel]] to i8*) to i8 addrspace(4)*),
   // COMMON-SAME: i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i32, i32 } addrspace(1)* [[BLG4]] to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1,
-  // B32-SAME: i32* %[[TMP1]])
-  // B64-SAME: i64* %[[TMP1]])
+  // B32-SAME: i32* %[[TMP]])
+  // B64-SAME: i64* %[[TMP]])
   enqueue_kernel(default_queue, flags, ndrange, 2, event_wait_list2, &clk_event,
                  ^(local void *p) {
                    return;
                  },
                  c);
 
   long l;
   // Emits global block literal [[BLG5]] and block kernel [[INVGK5]].
   // COMMON: [[DEF_Q:%[0-9]+]] = load %opencl.queue_t{{.*}}*, %opencl.queue_t{{.*}}** %default_queue
   // COMMON: [[FLAGS:%[0-9]+]] = load i32, i32* %flags
-  // B32: %[[TMP:.*]] = alloca [1 x i32]
-  // B32: %[[TMP1:.*]] = getelementptr [1 x i32], [1 x i32]* %[[TMP]], i32 0, i32 0
-  // B32: store i32 %{{.*}}, i32* %[[TMP1]], align 4
-  // B64: %[[TMP:.*]] = alloca [1 x i64]
-  // B64: %[[TMP1:.*]] = getelementptr [1 x i64], [1 x i64]* %[[TMP]], i32 0, i32 0
-  // B64: store i64 %{{.*}}, i64* %[[TMP1]], align 8
+  // B32: %[[TMP:.*]] = getelementptr [1 x i32], [1 x i32]* %[[BLOCK_SIZES5]], i32 0, i32 0
+  // B32: store i32 %{{.*}}, i32* %[[TMP]], align 4
+  // B64: %[[TMP:.*]] = getelementptr [1 x i64], [1 x i64]* %[[BLOCK_SIZES5]], i32 0, i32 0
+  // B64: store i64 %{{.*}}, i64* %[[TMP]], align 8
   // COMMON-LABEL: call i32 @__enqueue_kernel_varargs
   // COMMON-SAME: (%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* [[NDR]]{{([0-9]+)?}},
   // COMMON-SAME: i8 addrspace(4)* addrspacecast (i8* bitcast ({{.*}} [[INVGK5:[^ ]+_kernel]] to i8*) to i8 addrspace(4)*),
   // COMMON-SAME: i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i32, i32 } addrspace(1)* [[BLG5]] to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1,
-  // B32-SAME: i32* %[[TMP1]])
-  // B64-SAME: i64* %[[TMP1]])
+  // B32-SAME: i32* %[[TMP]])
+  // B64-SAME: i64* %[[TMP]])
   enqueue_kernel(default_queue, flags, ndrange,
                  ^(local void *p) {
                    return;
                  },
                  l);
 
   // Emits global block literal [[BLG6]] and block kernel [[INVGK6]].
   // COMMON: [[DEF_Q:%[0-9]+]] = load %opencl.queue_t{{.*}}*, %opencl.queue_t{{.*}}** %default_queue
   // COMMON: [[FLAGS:%[0-9]+]] = load i32, i32* %flags
-  // B32: %[[TMP:.*]] = alloca [3 x i32]
-  // B32: %[[TMP1:.*]] = getelementptr [3 x i32], [3 x i32]* %[[TMP]], i32 0, i32 0
-  // B32: store i32 1, i32* %[[TMP1]], align 4
-  // B32: %[[TMP2:.*]] = getelementptr [3 x i32], [3 x i32]* %[[TMP]], i32 0, i32 1
-  // B32: store i32 2, i32* %[[TMP2]], align 4
-  // B32: %[[TMP3:.*]] = getelementptr [3 x i32], [3 x i32]* %[[TMP]], i32 0, i32 2
-  // B32: store i32 4, i32* %[[TMP3]], align 4
-  // B64: %[[TMP:.*]] = alloca [3 x i64]
-  // B64: %[[TMP1:.*]] = getelementptr [3 x i64], [3 x i64]* %[[TMP]], i32 0, i32 0
-  // B64: store i64 1, i64* %[[TMP1]], align 8
-  // B64: %[[TMP2:.*]] = getelementptr [3 x i64], [3 x i64]* %[[TMP]], i32 0, i32 1
-  // B64: store i64 2, i64* %[[TMP2]], align 8
-  // B64: %[[TMP3:.*]] = getelementptr [3 x i64], [3 x i64]* %[[TMP]], i32 0, i32 2
-  // B64: store i64 4, i64* %[[TMP3]], align 8
+  // B32: %[[TMP:.*]] = getelementptr [3 x i32], [3 x i32]* %[[BLOCK_SIZES6]], i32 0, i32 0
+  // B32: store i32 1, i32* %[[TMP]], align 4
+  // B32: %[[BLOCK_SIZES62:.*]] = getelementptr [3 x i32], [3 x i32]* %[[BLOCK_SIZES6]], i32 0, i32 1
+  // B32: store i32 2, i32* %[[BLOCK_SIZES62]], align 4
+  // B32: %[[BLOCK_SIZES63:.*]] = getelementptr [3 x i32], [3 x i32]* %[[BLOCK_SIZES6]], i32 0, i32 2
+  // B32: store i32 4, i32* %[[BLOCK_SIZES63]], align 4
+  // B64: %[[TMP:.*]] = getelementptr [3 x i64], [3 x i64]* %[[BLOCK_SIZES6]], i32 0, i32 0
+  // B64: store i64 1, i64* %[[TMP]], align 8
+  // B64: %[[BLOCK_SIZES62:.*]] = getelementptr [3 x i64], [3 x i64]* %[[BLOCK_SIZES6]], i32 0, i32 1
+  // B64: store i64 2, i64* %[[BLOCK_SIZES62]], align 8
+  // B64: %[[BLOCK_SIZES63:.*]] = getelementptr [3 x i64], [3 x i64]* %[[BLOCK_SIZES6]], i32 0, i32 2
+  // B64: store i64 4, i64* %[[BLOCK_SIZES63]], align 8
   // COMMON-LABEL: call i32 @__enqueue_kernel_varargs
   // COMMON-SAME: (%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* [[NDR]]{{([0-9]+)?}},
   // COMMON-SAME: i8 addrspace(4)* addrspacecast (i8* bitcast ({{.*}} [[INVGK6:[^ ]+_kernel]] to i8*) to i8 addrspace(4)*),
   // COMMON-SAME: i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i32, i32 } addrspace(1)* [[BLG6]] to i8 addrspace(1)*) to i8 addrspace(4)*), i32 3,
-  // B32-SAME: i32* %[[TMP1]])
-  // B64-SAME: i64* %[[TMP1]])
+  // B32-SAME: i32* %[[TMP]])
+  // B64-SAME: i64* %[[TMP]])
   enqueue_kernel(default_queue, flags, ndrange,
                  ^(local void *p1, local void *p2, local void *p3) {
                    return;
                  },
                  1, 2, 4);
 
   // Emits global block literal [[BLG7]] and block kernel [[INVGK7]].
   // COMMON: [[DEF_Q:%[0-9]+]] = load %opencl.queue_t*, %opencl.queue_t** %default_queue
   // COMMON: [[FLAGS:%[0-9]+]] = load i32, i32* %flags
-  // B32: %[[TMP:.*]] = alloca [1 x i32]
-  // B32: %[[TMP1:.*]] = getelementptr [1 x i32], [1 x i32]* %[[TMP]], i32 0, i32 0
-  // B32: store i32 0, i32* %[[TMP1]], align 4
-  // B64: %[[TMP:.*]] = alloca [1 x i64]
-  // B64: %[[TMP1:.*]] = getelementptr [1 x i64], [1 x i64]* %[[TMP]], i32 0, i32 0
-  // B64: store i64 4294967296, i64* %[[TMP1]], align 8
+  // B32: %[[TMP:.*]] = getelementptr [1 x i32], [1 x i32]* %[[BLOCK_SIZES7]], i32 0, i32 0
+  // B32: store i32 0, i32* %[[TMP]], align 4
+  // B64: %[[TMP:.*]] = getelementptr [1 x i64], [1 x i64]* %[[BLOCK_SIZES7]], i32 0, i32 0
+  // B64: store i64 4294967296, i64* %[[TMP]], align 8
   // COMMON-LABEL: call i32 @__enqueue_kernel_varargs
   // COMMON-SAME: (%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* [[NDR]]{{([0-9]+)?}},
   // COMMON-SAME: i8 addrspace(4)* addrspacecast (i8* bitcast ({{.*}} [[INVGK7:[^ ]+_kernel]] to i8*) to i8 addrspace(4)*),
   // COMMON-SAME: i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i32, i32 } addrspace(1)* [[BLG7]] to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1,
-  // B32-SAME: i32* %[[TMP1]])
-  // B64-SAME: i64* %[[TMP1]])
+  // B32-SAME: i32* %[[TMP]])
+  // B64-SAME: i64* %[[TMP]])
   enqueue_kernel(default_queue, flags, ndrange,
                  ^(local void *p) {
                    return;
                  },
                  4294967296L);
 
   // Emits global block literal [[BLG8]] and invoke function [[INVG8]].
   // The full type of these expressions are long (and repeated elsewhere), so we

test/CodeGenOpenCL/enqueue-kernel-non-entry-block.cl

+// RUN: %clang_cc1 -cl-std=CL2.0 -O0 -emit-llvm -o - -triple amdgcn < %s | FileCheck %s --check-prefixes=COMMON,AMDGPU
+// RUN: %clang_cc1 -cl-std=CL2.0 -O0 -emit-llvm -o - -triple "spir-unknown-unknown" < %s | FileCheck %s --check-prefixes=COMMON,SPIR32
+// RUN: %clang_cc1 -cl-std=CL2.0 -O0 -emit-llvm -o - -triple "spir64-unknown-unknown" < %s | FileCheck %s --check-prefixes=COMMON,SPIR64
+// RUN: %clang_cc1 -cl-std=CL2.0 -O0 -debug-info-kind=limited -emit-llvm -o - -triple amdgcn < %s | FileCheck %s --check-prefixes=CHECK-DEBUG
+
+// Check that the enqueue_kernel array temporary is in the entry block to avoid
+// a dynamic alloca
+
+typedef struct {int a;} ndrange_t;
+
+kernel void test(int i) {
+// COMMON-LABEL: define {{.*}} void @test
+// COMMON-LABEL: entry:
+// AMDGPU: %block_sizes = alloca [1 x i64]
+// SPIR32: %block_sizes = alloca [1 x i32]
+// SPIR64: %block_sizes = alloca [1 x i64]
+// COMMON-LABEL: if.then:
+// COMMON-NOT: alloca
+// CHECK-DEBUG: getelementptr {{.*}} %block_sizes, {{.*}} !dbg !34
+// COMMON-LABEL: if.end
+  queue_t default_queue;
+  unsigned flags = 0;
+  ndrange_t ndrange;
+  if (i)
+    enqueue_kernel(default_queue, flags, ndrange, ^(local void *a) { }, 32);
+}
+
+// Check that the temporary is scoped to the `if`
+
+// CHECK-DEBUG: !32 = distinct !DILexicalBlock(scope: !7, file: !1, line: 24)
+// CHECK-DEBUG: !34 = !DILocation(line: 25, scope: !32)