[OpenCL] Clean up and add missing fields for block struct

[yaxunl]

Summary

Currently block is translated to a structure equivalent to

struct Block {
  void *isa;
  int flags;
  int reserved;
  void *invoke;
  void *descriptor;
};

Except invoke, which is the pointer to the block invoke function,
all other fields are useless for OpenCL, which clutter the IR and
also waste memory since the block struct is passed to the block
invoke function as argument.

On the other hand, the size and alignment of the block struct is
not stored in the struct, which causes difficulty to implement
__enqueue_kernel as library function, since the library function
needs to know the size and alignment of the argument which needs
to be passed to the kernel.

This patch removes the useless fields from the block struct and adds
size and align fields. The equivalent block struct will become

struct Block {
  int size;
  int align;
  generic void *invoke;
 /* custom fields */
};

It also changes the pointer to the invoke function to be
a generic pointer since the address space of a function
may not be private on certain targets.

[yaxunl]

Fix bug about calling blocks.

[Anastasia]

Could you please explain a bit more why the alignment have to be put explicitly in the struct? I am just not very convinced this is general enough.

lib/CodeGen/CGBlocks.cpp

314

Why removing this?

test/CodeGenOpenCL/blocks.cl

16–38

We don't need to check other fields too?

[yaxunl]

In D37822#872291, @Anastasia wrote:

Could you please explain a bit more why the alignment have to be put explicitly in the struct? I am just not very convinced this is general enough.

The captured variables are fields of the block literal struct. Due to alignment requirement of these fields, there is alignment requirement of
the block literal struct. The ISA of the block invoke function is generated with the assumption of these alignments. If the block literal is
allocated at a memory address not satisfying the alignment requirement, the kernel behavior is undefined.

Generally, __enqueue_kernel library function needs to prepare the kernel argument before launching the kernel. It usually does this by copying
the block literal to some buffer then pass the address of the buffer to the kernel. Then the address of the buffer has to satisfy the alignment
requirement.

If this block literal struct is not general enough, how about add another field as target reserved size, and leave the remaining space of header for
target specific use. And add a target hook to allow target fill the reserved space, e.g.

struct __opencl_block_literal {
  int total_size;
  int align;
  __generic void *invoke;
  int target_reserved_size; /* round up to 4 bytes */
  int target_reserved[];
  /* captures */
};

lib/CodeGen/CGBlocks.cpp

314

It is not removed. It is moved to line 307.

test/CodeGenOpenCL/blocks.cl

16–38

will add checks.

[Anastasia]

In D37822#872446, @yaxunl wrote:

In D37822#872291, @Anastasia wrote:

Could you please explain a bit more why the alignment have to be put explicitly in the struct? I am just not very convinced this is general enough.

The captured variables are fields of the block literal struct. Due to alignment requirement of these fields, there is alignment requirement of
the block literal struct. The ISA of the block invoke function is generated with the assumption of these alignments. If the block literal is
allocated at a memory address not satisfying the alignment requirement, the kernel behavior is undefined.

Generally, __enqueue_kernel library function needs to prepare the kernel argument before launching the kernel. It usually does this by copying
the block literal to some buffer then pass the address of the buffer to the kernel. Then the address of the buffer has to satisfy the alignment
requirement.

If this block literal struct is not general enough, how about add another field as target reserved size, and leave the remaining space of header for
target specific use. And add a target hook to allow target fill the reserved space, e.g.

struct __opencl_block_literal {
  int total_size;
  int align;
  __generic void *invoke;
  int target_reserved_size; /* round up to 4 bytes */
  int target_reserved[];
  /* captures */
};

I like the idea of the target reserved part actually. But not sure how it could be used without adding any target specific methods?

However, I am still not clear why the alignment of this struct has to be different from any other struct Clang produces. Normally the alignment of objects have to be known during IR generation to put them correctly in the attributes of generated alloca, store and loads. But as a field inside struct I don't know how it can be useful. I would imagine enqueue_kernel would just operate on the block as if it would be an arbitrary buffer of data. Also would size of the struct not account for any padding to make sure the alignment can be deduced based on it correctly?

lib/CodeGen/CGOpenCLRuntime.cpp

108

Should we put an assert of LangOpts.OpenCL?

test/CodeGen/blocks-opencl.cl

1–2

Btw, do you think we need this test any more? And if yes, could this be moved to CodeGenOpenCL?

[yaxunl]

In D37822#873876, @Anastasia wrote:

In D37822#872446, @yaxunl wrote:

In D37822#872291, @Anastasia wrote:

Could you please explain a bit more why the alignment have to be put explicitly in the struct? I am just not very convinced this is general enough.

The captured variables are fields of the block literal struct. Due to alignment requirement of these fields, there is alignment requirement of
the block literal struct. The ISA of the block invoke function is generated with the assumption of these alignments. If the block literal is
allocated at a memory address not satisfying the alignment requirement, the kernel behavior is undefined.

Generally, __enqueue_kernel library function needs to prepare the kernel argument before launching the kernel. It usually does this by copying
the block literal to some buffer then pass the address of the buffer to the kernel. Then the address of the buffer has to satisfy the alignment
requirement.

If this block literal struct is not general enough, how about add another field as target reserved size, and leave the remaining space of header for
target specific use. And add a target hook to allow target fill the reserved space, e.g.

struct __opencl_block_literal {
  int total_size;
  int align;
  __generic void *invoke;
  int target_reserved_size; /* round up to 4 bytes */
  int target_reserved[];
  /* captures */
};

I like the idea of the target reserved part actually. But not sure how it could be used without adding any target specific methods?

If we decide to add target reserved fields, I can add target hooks to fill these fields. However I would suggest to leave this for future since I don't see there is need for other fields for now.

However, I am still not clear why the alignment of this struct has to be different from any other struct Clang produces. Normally the alignment of objects have to be known during IR generation to put them correctly in the attributes of generated alloca, store and loads. But as a field inside struct I don't know how it can be useful. I would imagine enqueue_kernel would just operate on the block as if it would be an arbitrary buffer of data. Also would size of the struct not account for any padding to make sure the alignment can be deduced based on it correctly?

enqueue_kernel needs to pass the block struct to the kernel. Let's assume it does this by copying the block struct to a buffer. If enqueue_kernel does not know the alignment of the struct, it can only put it at an arbitrary address in the buffer. Then the kernel has to copy the struct to an aligned private memory and load the fields. However, if the enqueued_kernel knows the alignment of the struct, it can put it at an address satisfying the alignment. Then the kernel can load the fields directly from the buffer, skips the step of copying to an aligned private memory. Therefore, alignment of the block struct is usually a useful information for enqueue_kernel. I think that's why in the SPIRV spec OpEnqueueKernel requires an alignment operand for the block context.

lib/CodeGen/CGOpenCLRuntime.cpp

108

Will do.

test/CodeGen/blocks-opencl.cl

1–2

I think this one can be removed since what it tests is covered by CodeGenOpenCL/blocks.cl.

[Anastasia]

In D37822#877572, @yaxunl wrote:

In D37822#873876, @Anastasia wrote:

In D37822#872446, @yaxunl wrote:

In D37822#872291, @Anastasia wrote:

Could you please explain a bit more why the alignment have to be put explicitly in the struct? I am just not very convinced this is general enough.

The captured variables are fields of the block literal struct. Due to alignment requirement of these fields, there is alignment requirement of
the block literal struct. The ISA of the block invoke function is generated with the assumption of these alignments. If the block literal is
allocated at a memory address not satisfying the alignment requirement, the kernel behavior is undefined.

Generally, __enqueue_kernel library function needs to prepare the kernel argument before launching the kernel. It usually does this by copying
the block literal to some buffer then pass the address of the buffer to the kernel. Then the address of the buffer has to satisfy the alignment
requirement.

If this block literal struct is not general enough, how about add another field as target reserved size, and leave the remaining space of header for
target specific use. And add a target hook to allow target fill the reserved space, e.g.

struct __opencl_block_literal {
  int total_size;
  int align;
  __generic void *invoke;
  int target_reserved_size; /* round up to 4 bytes */
  int target_reserved[];
  /* captures */
};

I like the idea of the target reserved part actually. But not sure how it could be used without adding any target specific methods?

If we decide to add target reserved fields, I can add target hooks to fill these fields. However I would suggest to leave this for future since I don't see there is need for other fields for now.

I could imagine it can be usefull for some vendor implementations.

However, I am still not clear why the alignment of this struct has to be different from any other struct Clang produces. Normally the alignment of objects have to be known during IR generation to put them correctly in the attributes of generated alloca, store and loads. But as a field inside struct I don't know how it can be useful. I would imagine enqueue_kernel would just operate on the block as if it would be an arbitrary buffer of data. Also would size of the struct not account for any padding to make sure the alignment can be deduced based on it correctly?

enqueue_kernel needs to pass the block struct to the kernel. Let's assume it does this by copying the block struct to a buffer. If enqueue_kernel does not know the alignment of the struct, it can only put it at an arbitrary address in the buffer. Then the kernel has to copy the struct to an aligned private memory and load the fields. However, if the enqueued_kernel knows the alignment of the struct, it can put it at an address satisfying the alignment. Then the kernel can load the fields directly from the buffer, skips the step of copying to an aligned private memory. Therefore, alignment of the block struct is usually a useful information for enqueue_kernel. I think that's why in the SPIRV spec OpEnqueueKernel requires an alignment operand for the block context.

Ok, I just think in C if you use malloc to obtain a pointer to some memory location it doesn't take any alignment information. Then you can use the pointer to copy any data including the struct into the location its pointed to. And the pointer can be used later on correctly. I think the alignment is deduced in this case from the type or the size of an object. Do you know where the alignment information is used for SPIRV call? Also how is the block represented in SPIRV?

[b-sumner]

In D37822#877903, @Anastasia wrote:

In D37822#877572, @yaxunl wrote:

In D37822#873876, @Anastasia wrote:

In D37822#872446, @yaxunl wrote:

In D37822#872291, @Anastasia wrote:

Could you please explain a bit more why the alignment have to be put explicitly in the struct? I am just not very convinced this is general enough.

The captured variables are fields of the block literal struct. Due to alignment requirement of these fields, there is alignment requirement of
the block literal struct. The ISA of the block invoke function is generated with the assumption of these alignments. If the block literal is
allocated at a memory address not satisfying the alignment requirement, the kernel behavior is undefined.

Generally, __enqueue_kernel library function needs to prepare the kernel argument before launching the kernel. It usually does this by copying
the block literal to some buffer then pass the address of the buffer to the kernel. Then the address of the buffer has to satisfy the alignment
requirement.

If this block literal struct is not general enough, how about add another field as target reserved size, and leave the remaining space of header for
target specific use. And add a target hook to allow target fill the reserved space, e.g.

struct __opencl_block_literal {
  int total_size;
  int align;
  __generic void *invoke;
  int target_reserved_size; /* round up to 4 bytes */
  int target_reserved[];
  /* captures */
};

I like the idea of the target reserved part actually. But not sure how it could be used without adding any target specific methods?

If we decide to add target reserved fields, I can add target hooks to fill these fields. However I would suggest to leave this for future since I don't see there is need for other fields for now.

I could imagine it can be usefull for some vendor implementations.

However, I am still not clear why the alignment of this struct has to be different from any other struct Clang produces. Normally the alignment of objects have to be known during IR generation to put them correctly in the attributes of generated alloca, store and loads. But as a field inside struct I don't know how it can be useful. I would imagine enqueue_kernel would just operate on the block as if it would be an arbitrary buffer of data. Also would size of the struct not account for any padding to make sure the alignment can be deduced based on it correctly?

enqueue_kernel needs to pass the block struct to the kernel. Let's assume it does this by copying the block struct to a buffer. If enqueue_kernel does not know the alignment of the struct, it can only put it at an arbitrary address in the buffer. Then the kernel has to copy the struct to an aligned private memory and load the fields. However, if the enqueued_kernel knows the alignment of the struct, it can put it at an address satisfying the alignment. Then the kernel can load the fields directly from the buffer, skips the step of copying to an aligned private memory. Therefore, alignment of the block struct is usually a useful information for enqueue_kernel. I think that's why in the SPIRV spec OpEnqueueKernel requires an alignment operand for the block context.

Ok, I just think in C if you use malloc to obtain a pointer to some memory location it doesn't take any alignment information. Then you can use the pointer to copy any data including the struct into the location its pointed to. And the pointer can be used later on correctly. I think the alignment is deduced in this case from the type or the size of an object. Do you know where the alignment information is used for SPIRV call? Also how is the block represented in SPIRV?

Actually malloc alignment is not sufficient more many uses such as CPU supported vectors, e.g. AVX512 or passed to create buffer with use-host-pointer. In such cases you need posix_memalign or some similar API. Having the alignment means it is available if needed. If an implementation doesn't need it, there is no harm is there?

[yaxunl]

Revise by Anastasia's comments.

[yaxunl]

In D37822#877903, @Anastasia wrote:

In D37822#877572, @yaxunl wrote:

In D37822#873876, @Anastasia wrote:

In D37822#872446, @yaxunl wrote:

In D37822#872291, @Anastasia wrote:

Could you please explain a bit more why the alignment have to be put explicitly in the struct? I am just not very convinced this is general enough.

The captured variables are fields of the block literal struct. Due to alignment requirement of these fields, there is alignment requirement of
the block literal struct. The ISA of the block invoke function is generated with the assumption of these alignments. If the block literal is
allocated at a memory address not satisfying the alignment requirement, the kernel behavior is undefined.

Generally, __enqueue_kernel library function needs to prepare the kernel argument before launching the kernel. It usually does this by copying
the block literal to some buffer then pass the address of the buffer to the kernel. Then the address of the buffer has to satisfy the alignment
requirement.

If this block literal struct is not general enough, how about add another field as target reserved size, and leave the remaining space of header for
target specific use. And add a target hook to allow target fill the reserved space, e.g.

struct __opencl_block_literal {
  int total_size;
  int align;
  __generic void *invoke;
  int target_reserved_size; /* round up to 4 bytes */
  int target_reserved[];
  /* captures */
};

I like the idea of the target reserved part actually. But not sure how it could be used without adding any target specific methods?

If we decide to add target reserved fields, I can add target hooks to fill these fields. However I would suggest to leave this for future since I don't see there is need for other fields for now.

I could imagine it can be usefull for some vendor implementations.

However, I am still not clear why the alignment of this struct has to be different from any other struct Clang produces. Normally the alignment of objects have to be known during IR generation to put them correctly in the attributes of generated alloca, store and loads. But as a field inside struct I don't know how it can be useful. I would imagine enqueue_kernel would just operate on the block as if it would be an arbitrary buffer of data. Also would size of the struct not account for any padding to make sure the alignment can be deduced based on it correctly?

enqueue_kernel needs to pass the block struct to the kernel. Let's assume it does this by copying the block struct to a buffer. If enqueue_kernel does not know the alignment of the struct, it can only put it at an arbitrary address in the buffer. Then the kernel has to copy the struct to an aligned private memory and load the fields. However, if the enqueued_kernel knows the alignment of the struct, it can put it at an address satisfying the alignment. Then the kernel can load the fields directly from the buffer, skips the step of copying to an aligned private memory. Therefore, alignment of the block struct is usually a useful information for enqueue_kernel. I think that's why in the SPIRV spec OpEnqueueKernel requires an alignment operand for the block context.

Ok, I just think in C if you use malloc to obtain a pointer to some memory location it doesn't take any alignment information. Then you can use the pointer to copy any data including the struct into the location its pointed to. And the pointer can be used later on correctly. I think the alignment is deduced in this case from the type or the size of an object. Do you know where the alignment information is used for SPIRV call? Also how is the block represented in SPIRV?

If you just use malloc and put your struct in it, there is no guarantee that your struct is aligned at the required alignment, then your kernel cannot load a field directly from that memory. For example, if your first field is an int and the instruction can only load an int from an addr aligned at 4, and your malloc'ed addr is aligned at 1, then you cannot load that int directly. Instead, you need to copy the 4 bytes to an addr aligned at 4, then use that instruction to load it. If you use posix_memalign to get an aligned buffer, then your kernel can generate more efficient code.

OpEnqueueKernel instruction in SPIR-V is for representing OpenCL enqueue_kernel. In SPIR-V block is represented by block invoke function. When enqueue_kernel is translated to OpEnqueueKernel, it is required to provide block invoke function, block context, the size and alignment of the block context.

[yaxunl]

Add custom fields to block and target hooks to fill them.

[Anastasia]

lib/CodeGen/CGBlocks.cpp

310

remove one "that".

311

I think the alignment might not be computed correctly now if there will be custom fields that might have a bigger size than a pointer? Also what happens if we have captures as well?

830

do we need to add numeration to each item name?

1204

If we reorder fields and put this on top we can merge the if statements above and below this point.

[yaxunl]

lib/CodeGen/CGBlocks.cpp

310

will do.

311

Will fix.

The captures will be accounted for by computeBlockInfo and BlockSize and BlockAlign will be updated.

830

yes. will add it.

1204

By convention the size of the whole struct is the first field so that the library function reads the first integer and knows how many bytes to copy.

[yaxunl]

Rebased to ToT and revised by Anastasia's comments.

[Anastasia]

LGTM! Thanks!

test/CodeGenOpenCL/blocks.cl

29

It might be better to give those r0-r7 some names for readability if possible!

[yaxunl]

test/CodeGenOpenCL/blocks.cl

29

Will fix it when committing.

lib/CodeGen/CGBlocks.cpp

 //===--- CGBlocks.cpp - Emit LLVM Code for declarations ---------*- C++ -*-===//
 //
 //                     The LLVM Compiler Infrastructure
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 //
 // This contains code to emit blocks.
 //
 //===----------------------------------------------------------------------===//
 
 #include "CGBlocks.h"
 #include "CGDebugInfo.h"
 #include "CGObjCRuntime.h"
+#include "CGOpenCLRuntime.h"
 #include "CodeGenFunction.h"
 #include "CodeGenModule.h"
 #include "ConstantEmitter.h"
 #include "clang/CodeGen/ConstantInitBuilder.h"
 #include "clang/AST/DeclObjC.h"
 #include "llvm/ADT/SmallSet.h"
 #include "llvm/IR/CallSite.h"
 #include "llvm/IR/DataLayout.h"

 /// Get the low bit of a nonzero character count.  This is the
 /// alignment of the nth byte if the 0th byte is universally aligned.
 static CharUnits getLowBit(CharUnits v) {
   return CharUnits::fromQuantity(v.getQuantity() & (~v.getQuantity() + 1));
 }
 
 static void initializeForBlockHeader(CodeGenModule &CGM, CGBlockInfo &info,
                              SmallVectorImpl<llvm::Type*> &elementTypes) {
+
+  assert(elementTypes.empty());
+  if (CGM.getLangOpts().OpenCL) {
+    // The header is basically 'struct { int; int; generic void *; }'.
+    // Assert that that struct is packed.
+    auto GenPtrAlign = CharUnits::fromQuantity(
+        CGM.getTarget().getPointerAlign(LangAS::opencl_generic) / 8);
+    auto GenPtrSize = CharUnits::fromQuantity(
+        CGM.getTarget().getPointerWidth(LangAS::opencl_generic) / 8);
+    assert(CGM.getIntSize() <= GenPtrSize);
+    assert(CGM.getIntAlign() <= GenPtrAlign);
+    assert((2 * CGM.getIntSize()).isMultipleOf(GenPtrAlign));
+    info.BlockAlign = GenPtrAlign;
+    info.BlockSize = GenPtrSize + 2 * CGM.getIntSize();
+    elementTypes.push_back(CGM.IntTy);
+    elementTypes.push_back(CGM.IntTy);
+    elementTypes.push_back(CGM.getOpenCLRuntime().getGenericVoidPointerType());
+  } else {
     // The header is basically 'struct { void *; int; int; void *; void *; }'.
     // Assert that that struct is packed.
     assert(CGM.getIntSize() <= CGM.getPointerSize());
     assert(CGM.getIntAlign() <= CGM.getPointerAlign());
     assert((2 * CGM.getIntSize()).isMultipleOf(CGM.getPointerAlign()));
-
     info.BlockAlign = CGM.getPointerAlign();
     info.BlockSize = 3 * CGM.getPointerSize() + 2 * CGM.getIntSize();
-
-  assert(elementTypes.empty());
     elementTypes.push_back(CGM.VoidPtrTy);
     elementTypes.push_back(CGM.IntTy);
     elementTypes.push_back(CGM.IntTy);
     elementTypes.push_back(CGM.VoidPtrTy);
     elementTypes.push_back(CGM.getBlockDescriptorType());
   }
+}
 
 static QualType getCaptureFieldType(const CodeGenFunction &CGF,
                                     const BlockDecl::Capture &CI) {
   const VarDecl *VD = CI.getVariable();
 
   // If the variable is captured by an enclosing block or lambda expression,
   // use the type of the capture field.
   if (CGF.BlockInfo && CI.isNested())
llvm::Value *CodeGenFunction::EmitBlockLiteral(const BlockExpr *blockExpr) {
   blockInfo.reset(findAndRemoveBlockInfo(&FirstBlockInfo,
                                          blockExpr->getBlockDecl()));
 
   blockInfo->BlockExpression = blockExpr;
   return EmitBlockLiteral(*blockInfo);
 }
 
 llvm::Value *CodeGenFunction::EmitBlockLiteral(const CGBlockInfo &blockInfo) {
+  bool IsOpenCL = CGM.getContext().getLangOpts().OpenCL;
+  auto GenVoidPtrTy =
+      IsOpenCL ? CGM.getOpenCLRuntime().getGenericVoidPointerType() : VoidPtrTy;
+  unsigned GenVoidPtrAddr = IsOpenCL ? LangAS::opencl_generic : LangAS::Default;
+  auto GenVoidPtrSize = CharUnits::fromQuantity(
+      CGM.getTarget().getPointerWidth(GenVoidPtrAddr) / 8);
   // Using the computed layout, generate the actual block function.
   bool isLambdaConv = blockInfo.getBlockDecl()->isConversionFromLambda();
   llvm::Constant *blockFn
     = CodeGenFunction(CGM, true).GenerateBlockFunction(CurGD, blockInfo,
                                                        LocalDeclMap,
                                                        isLambdaConv);
-  blockFn = llvm::ConstantExpr::getBitCast(blockFn, VoidPtrTy);
+  blockFn = llvm::ConstantExpr::getPointerCast(blockFn, GenVoidPtrTy);
 
   // If there is nothing to capture, we can emit this as a global block.
   if (blockInfo.CanBeGlobal)
     return buildGlobalBlock(CGM, blockInfo, blockFn);
 
   // Otherwise, we have to emit this as a local block.
 
-  llvm::Constant *isa =
-      (!CGM.getContext().getLangOpts().OpenCL)
-          ? CGM.getNSConcreteStackBlock()
-          : CGM.getNullPointer(VoidPtrPtrTy,
-                               CGM.getContext().getPointerType(
-                                   QualType(CGM.getContext().VoidPtrTy)));
-  isa = llvm::ConstantExpr::getBitCast(isa, VoidPtrTy);
-
-  // Build the block descriptor.
-  llvm::Constant *descriptor = buildBlockDescriptor(CGM, blockInfo);
-
   Address blockAddr = blockInfo.LocalAddress;
   assert(blockAddr.isValid() && "block has no address!");
 
+  llvm::Constant *isa;
+  llvm::Constant *descriptor;
+  BlockFlags flags;
+  if (!IsOpenCL) {
+    isa = llvm::ConstantExpr::getBitCast(CGM.getNSConcreteStackBlock(),
+                                         VoidPtrTy);
+
+    // Build the block descriptor.
+    descriptor = buildBlockDescriptor(CGM, blockInfo);
+
     // Compute the initial on-stack block flags.
-  BlockFlags flags = BLOCK_HAS_SIGNATURE;
-  if (blockInfo.HasCapturedVariableLayout) flags |= BLOCK_HAS_EXTENDED_LAYOUT;
-  if (blockInfo.NeedsCopyDispose) flags |= BLOCK_HAS_COPY_DISPOSE;
-  if (blockInfo.HasCXXObject) flags |= BLOCK_HAS_CXX_OBJ;
-  if (blockInfo.UsesStret) flags |= BLOCK_USE_STRET;
+    flags = BLOCK_HAS_SIGNATURE;
+    if (blockInfo.HasCapturedVariableLayout)
+      flags |= BLOCK_HAS_EXTENDED_LAYOUT;
+    if (blockInfo.NeedsCopyDispose)
+      flags |= BLOCK_HAS_COPY_DISPOSE;
+    if (blockInfo.HasCXXObject)
+      flags |= BLOCK_HAS_CXX_OBJ;
+    if (blockInfo.UsesStret)
+      flags |= BLOCK_USE_STRET;
+  }
 
   auto projectField =
     [&](unsigned index, CharUnits offset, const Twine &name) -> Address {
       return Builder.CreateStructGEP(blockAddr, index, offset, name);
     };
   auto storeField =
     [&](llvm::Value *value, unsigned index, CharUnits offset,
         const Twine &name) {
       Builder.CreateStore(value, projectField(index, offset, name));
     };
 
   // Initialize the block header.
   {
     // We assume all the header fields are densely packed.
     unsigned index = 0;
     CharUnits offset;
     auto addHeaderField =
       [&](llvm::Value *value, CharUnits size, const Twine &name) {
         storeField(value, index, offset, name);
         offset += size;
         index++;
       };
 
+    if (!IsOpenCL) {
       addHeaderField(isa, getPointerSize(), "block.isa");
       addHeaderField(llvm::ConstantInt::get(IntTy, flags.getBitMask()),
                      getIntSize(), "block.flags");
-    addHeaderField(llvm::ConstantInt::get(IntTy, 0),
-                   getIntSize(), "block.reserved");
-    addHeaderField(blockFn, getPointerSize(), "block.invoke");
+      addHeaderField(llvm::ConstantInt::get(IntTy, 0), getIntSize(),
+                     "block.reserved");
+    } else {
+      addHeaderField(
+          llvm::ConstantInt::get(IntTy, blockInfo.BlockSize.getQuantity()),
+          getIntSize(), "block.size");
+      addHeaderField(
+          llvm::ConstantInt::get(IntTy, blockInfo.BlockAlign.getQuantity()),
+          getIntSize(), "block.align");
+    }
+    addHeaderField(blockFn, GenVoidPtrSize, "block.invoke");
+    if (!IsOpenCL)
       addHeaderField(descriptor, getPointerSize(), "block.descriptor");
   }
 
   // Finally, capture all the values into the block.
   const BlockDecl *blockDecl = blockInfo.getBlockDecl();
 
   // First, 'this'.
   if (blockDecl->capturesCXXThis()) {
     Address addr = projectField(blockInfo.CXXThisIndex, blockInfo.CXXThisOffset,
                                 "block.captured-this.addr");
     Builder.CreateStore(LoadCXXThis(), addr);
   }
 
   // Next, captured variables.
   for (const auto &CI : blockDecl->captures()) {
     const VarDecl *variable = CI.getVariable();
llvm::Type *CodeGenModule::getGenericBlockLiteralType() {
   // };
   GenericBlockLiteralType =
       llvm::StructType::create("struct.__block_literal_generic", VoidPtrTy,
                                IntTy, IntTy, VoidPtrTy, BlockDescPtrTy);
 
   return GenericBlockLiteralType;
 }
 
-RValue CodeGenFunction::EmitBlockCallExpr(const CallExpr *E, 
+RValue CodeGenFunction::EmitBlockCallExpr(const CallExpr *E,
                                           ReturnValueSlot ReturnValue) {
   const BlockPointerType *BPT =
     E->getCallee()->getType()->getAs<BlockPointerType>();
 
   llvm::Value *BlockPtr = EmitScalarExpr(E->getCallee());
 
   // Get a pointer to the generic block literal.
   // For OpenCL we generate generic AS void ptr to be able to reuse the same
   // block definition for blocks with captures generated as private AS local
   // variables and without captures generated as global AS program scope
   // variables.
   unsigned AddrSpace = 0;
   if (getLangOpts().OpenCL)
     AddrSpace = getContext().getTargetAddressSpace(LangAS::opencl_generic);
 
   llvm::Type *BlockLiteralTy =
       llvm::PointerType::get(CGM.getGenericBlockLiteralType(), AddrSpace);
 
   // Bitcast the callee to a block literal.
   BlockPtr =
       Builder.CreatePointerCast(BlockPtr, BlockLiteralTy, "block.literal");
 
   // Get the function pointer from the literal.
   llvm::Value *FuncPtr =
-    Builder.CreateStructGEP(CGM.getGenericBlockLiteralType(), BlockPtr, 3);
-
+      Builder.CreateStructGEP(CGM.getGenericBlockLiteralType(), BlockPtr,
+                              CGM.getLangOpts().OpenCL ? 0 : 3);
 
   // Add the block literal.
   CallArgList Args;
 
   QualType VoidPtrQualTy = getContext().VoidPtrTy;
   llvm::Type *GenericVoidPtrTy = VoidPtrTy;
   if (getLangOpts().OpenCL) {
-    GenericVoidPtrTy = Builder.getInt8PtrTy(
-        getContext().getTargetAddressSpace(LangAS::opencl_generic));
+    GenericVoidPtrTy = CGM.getOpenCLRuntime().getGenericVoidPointerType();
     VoidPtrQualTy =
         getContext().getPointerType(getContext().getAddrSpaceQualType(
             getContext().VoidTy, LangAS::opencl_generic));
   }
 
   BlockPtr = Builder.CreatePointerCast(BlockPtr, GenericVoidPtrTy);
   Args.add(RValue::get(BlockPtr), VoidPtrQualTy);
 
   QualType FnType = BPT->getPointeeType();
 
   // And the rest of the arguments.
   EmitCallArgs(Args, FnType->getAs<FunctionProtoType>(), E->arguments());
 
   // Load the function.
   llvm::Value *Func = Builder.CreateAlignedLoad(FuncPtr, getPointerAlign());
 
   const FunctionType *FuncTy = FnType->castAs<FunctionType>();
   const CGFunctionInfo &FnInfo =
     CGM.getTypes().arrangeBlockFunctionCall(Args, FuncTy);
 
   // Cast the function pointer to the right type.
   llvm::Type *BlockFTy = CGM.getTypes().GetFunctionType(FnInfo);
 
   llvm::Type *BlockFTyPtr = llvm::PointerType::getUnqual(BlockFTy);
-  Func = Builder.CreateBitCast(Func, BlockFTyPtr);
+  Func = Builder.CreatePointerCast(Func, BlockFTyPtr);
 
   // Prepare the callee.
   CGCallee Callee(CGCalleeInfo(), Func);
 
   // And call the block.
   return EmitCall(FnInfo, Callee, ReturnValue, Args);
 }
 
CodeGenModule::GetAddrOfGlobalBlock(const BlockExpr *BE,
   llvm::Constant *blockFn;
   {
     CodeGenFunction::DeclMapTy LocalDeclMap;
     blockFn = CodeGenFunction(*this).GenerateBlockFunction(GlobalDecl(),
                                                            blockInfo,
                                                            LocalDeclMap,
                                                            false);
   }
-  blockFn = llvm::ConstantExpr::getBitCast(blockFn, VoidPtrTy);
+  auto GenVoidPtrTy = getContext().getLangOpts().OpenCL
+                          ? getOpenCLRuntime().getGenericVoidPointerType()
+                          : VoidPtrTy;
+
+  blockFn = llvm::ConstantExpr::getPointerCast(blockFn, GenVoidPtrTy);
 
   return buildGlobalBlock(*this, blockInfo, blockFn);
 }
 
 static llvm::Constant *buildGlobalBlock(CodeGenModule &CGM,
                                         const CGBlockInfo &blockInfo,
                                         llvm::Constant *blockFn) {
   assert(blockInfo.CanBeGlobal);
   // Callers should detect this case on their own: calling this function
   // generally requires computing layout information, which is a waste of time
   // if we've already emitted this block.
   assert(!CGM.getAddrOfGlobalBlockIfEmitted(blockInfo.BlockExpression) &&
          "Refusing to re-emit a global block.");
 
   // Generate the constants for the block literal initializer.
   ConstantInitBuilder builder(CGM);
   auto fields = builder.beginStruct();
 
+  bool IsOpenCL = CGM.getLangOpts().OpenCL;
+  if (!IsOpenCL) {
     // isa
-  fields.add((!CGM.getContext().getLangOpts().OpenCL)
-                 ? CGM.getNSConcreteGlobalBlock()
-                 : CGM.getNullPointer(CGM.VoidPtrPtrTy,
-                                      CGM.getContext().getPointerType(QualType(
-                                          CGM.getContext().VoidPtrTy))));
+    fields.add(CGM.getNSConcreteGlobalBlock());
 
     // __flags
     BlockFlags flags = BLOCK_IS_GLOBAL | BLOCK_HAS_SIGNATURE;
-  if (blockInfo.UsesStret) flags |= BLOCK_USE_STRET;
-                                      
+    if (blockInfo.UsesStret)
+      flags |= BLOCK_USE_STRET;
     fields.addInt(CGM.IntTy, flags.getBitMask());
 
     // Reserved
     fields.addInt(CGM.IntTy, 0);
+  } else {
+    fields.addInt(CGM.IntTy, blockInfo.BlockSize.getQuantity());
+    fields.addInt(CGM.IntTy, blockInfo.BlockAlign.getQuantity());
+  }
 
   // Function
   fields.add(blockFn);
 
+  if (!IsOpenCL) {
     // Descriptor
     fields.add(buildBlockDescriptor(CGM, blockInfo));
+  }
 
   unsigned AddrSpace = 0;
   if (CGM.getContext().getLangOpts().OpenCL)
     AddrSpace = CGM.getContext().getTargetAddressSpace(LangAS::opencl_global);
 
   llvm::Constant *literal = fields.finishAndCreateGlobal(
       "__block_literal_global", blockInfo.BlockAlign,
       /*constant*/ true, llvm::GlobalVariable::InternalLinkage, AddrSpace);

lib/CodeGen/CGOpenCLRuntime.h

public:
 
   // \brief Returnes a value which indicates the size in bytes of the pipe
   // element.
   virtual llvm::Value *getPipeElemSize(const Expr *PipeArg);
 
   // \brief Returnes a value which indicates the alignment in bytes of the pipe
   // element.
   virtual llvm::Value *getPipeElemAlign(const Expr *PipeArg);
+
+  /// \return __generic void* type.
+  llvm::PointerType *getGenericVoidPointerType();
 };
 
 }
 }
 
 #endif

lib/CodeGen/CGOpenCLRuntime.cpp

llvm::Value *CGOpenCLRuntime::getPipeElemAlign(const Expr *PipeArg) {
   const PipeType *PipeTy = PipeArg->getType()->getAs<PipeType>();
   // The type of the last (implicit) argument to be passed.
   llvm::Type *Int32Ty = llvm::IntegerType::getInt32Ty(CGM.getLLVMContext());
   unsigned TypeSize = CGM.getContext()
                           .getTypeAlignInChars(PipeTy->getElementType())
                           .getQuantity();
   return llvm::ConstantInt::get(Int32Ty, TypeSize, false);
 }
+
+llvm::PointerType *CGOpenCLRuntime::getGenericVoidPointerType() {
+  return llvm::IntegerType::getInt8PtrTy(
+      CGM.getLLVMContext(),
+      CGM.getContext().getTargetAddressSpace(LangAS::opencl_generic));
+}

test/CodeGen/blocks-opencl.cl

 // RUN: %clang_cc1 -O0 %s -ffake-address-space-map -emit-llvm -o - -fblocks -triple x86_64-unknown-unknown | FileCheck %s
+// RUN: %clang_cc1 -O0 %s -emit-llvm -o - -fblocks -triple amdgcn | FileCheck %s
 // This used to crash due to trying to generate a bitcase from a cstring
 // in the constant address space to i8* in AS0.
 
 void dummy(float (^const op)(float)) {
 }
 
-// CHECK: i8 addrspace(2)* getelementptr inbounds ([9 x i8], [9 x i8] addrspace(2)* @.str, i32 0, i32 0)
+// CHECK-NOT: @.str
 
 kernel void test_block()
 {
   float (^const X)(float) = ^(float x) {
     return x + 42.0f;
   };
   dummy(X);
 }
 

test/CodeGenOpenCL/blocks.cl

-// RUN: %clang_cc1 %s -cl-std=CL2.0 -emit-llvm -o - -O0 -triple spir-unknown-unknown | FileCheck -check-prefix=GENERIC -check-prefix=COMMON %s
-// RUN: %clang_cc1 %s -cl-std=CL2.0 -emit-llvm -o - -O0 -triple amdgcn-amd-amdhsa-opencl | FileCheck -check-prefix=AMD -check-prefix=COMMON %s
+// RUN: %clang_cc1 %s -cl-std=CL2.0 -emit-llvm -o - -O0 -triple spir-unknown-unknown | FileCheck -check-prefixes=COMMON,SPIR %s
+// RUN: %clang_cc1 %s -cl-std=CL2.0 -emit-llvm -o - -O0 -triple amdgcn-amd-amdhsa-opencl | FileCheck -check-prefixes=COMMON,AMD %s
 
-// Checking for null instead of @__NSConcreteGlobalBlock symbol
-// COMMON: @__block_literal_global = internal addrspace(1) constant { i8**, i32, i32, i8*, %struct.__block_descriptor addrspace(2)* } { i8** null
+// SPIR: @__block_literal_global = internal addrspace(1) constant { i32, i32, i8 addrspace(4)* } { i32 12, i32 4, i8 addrspace(4)* addrspacecast (i8* bitcast (void (i8 addrspace(4)*, i8 addrspace(3)*)* @block_A_block_invoke to i8*) to i8 addrspace(4)*) }
+// AMD: @__block_literal_global = internal addrspace(1) constant { i32, i32, i8 addrspace(4)* } { i32 16, i32 8, i8 addrspace(4)* addrspacecast (i8* bitcast (void (i8 addrspace(4)*, i8 addrspace(3)*)* @block_A_block_invoke to i8*) to i8 addrspace(4)*) }
+// COMMON-NOT: .str
+
+// COMMON-LABEL: define internal {{.*}}void @block_A_block_invoke(i8 addrspace(4)* %.block_descriptor, i8 addrspace(3)* %a)
 void (^block_A)(local void *) = ^(local void *a) {
   return;
 };
 
+// COMMON-LABEL: define {{.*}}void @foo()
 void foo(){
   int i;
-// Checking for null instead of @_NSConcreteStackBlock symbol
-// COMMON: store i8* null, i8** %block.isa
+  // COMMON-NOT: store i8* null, i8** %block.isa
+  // COMMON: %[[block_size:.*]] = getelementptr inbounds <{ i32, i32, i8 addrspace(4)*, i32 }>, <{ i32, i32, i8 addrspace(4)*, i32 }>* %block, i32 0, i32 0
+  // SPIR: store i32 16, i32* %[[block_size]]
+  // AMD: store i32 20, i32* %[[block_size]]
+  // COMMON: %[[block_align:.*]] = getelementptr inbounds <{ i32, i32, i8 addrspace(4)*, i32 }>, <{ i32, i32, i8 addrspace(4)*, i32 }>* %block, i32 0, i32 1
+  // SPIR: store i32 4, i32* %[[block_align]]
+  // AMD: store i32 8, i32* %[[block_align]]
+  // COMMON: %[[block_invoke:.*]] = getelementptr inbounds <{ i32, i32, i8 addrspace(4)*, i32 }>, <{ i32, i32, i8 addrspace(4)*, i32 }>* %[[block:.*]], i32 0, i32 2
+  // COMMON: store i8 addrspace(4)* addrspacecast (i8* bitcast (i32 (i8 addrspace(4)*)* @__foo_block_invoke to i8*) to i8 addrspace(4)*), i8 addrspace(4)** %[[block_invoke]]
+  // COMMON: %[[block_captured:.*]] = getelementptr inbounds <{ i32, i32, i8 addrspace(4)*, i32 }>, <{ i32, i32, i8 addrspace(4)*, i32 }>* %[[block]], i32 0, i32 3
+  // COMMON: %[[r0:.*]] = load i32, i32* %i
+  // COMMON: store i32 %[[r0]], i32* %[[block_captured]],
+  // COMMON: %[[r1:.*]] = bitcast <{ i32, i32, i8 addrspace(4)*, i32 }>* %[[block]] to i32 ()*
+  // COMMON: %[[r2:.*]] = addrspacecast i32 ()* %[[r1]] to i32 () addrspace(4)*
+  // COMMON: store i32 () addrspace(4)* %[[r2]], i32 () addrspace(4)** %[[block_B:.*]],
+  // COMMON: %[[r3:.*]] = load i32 () addrspace(4)*, i32 () addrspace(4)** %[[block_B]]
+  // COMMON: %[[block_literal:.*]] = bitcast i32 () addrspace(4)* %3 to %struct.__block_literal_generic addrspace(4)*
+  // COMMON: %[[r4:.*]] = getelementptr inbounds %struct.__block_literal_generic, %struct.__block_literal_generic addrspace(4)* %[[block_literal]], i32 0, i32 0
+  // COMMON: %[[r5:.*]] = bitcast %struct.__block_literal_generic addrspace(4)* %[[block_literal]] to i8 addrspace(4)*
+  // COMMON: %[[r6:.*]] = load i8*, i8* addrspace(4)* %[[r4]]
+  // COMMON: %[[r7:.*]] = bitcast i8* %[[r6]] to i32 (i8 addrspace(4)*)*
+  // COMMON: call {{.*}}i32 %[[r7]](i8 addrspace(4)* %[[r5]])
+
   int (^ block_B)(void) = ^{
     return i;
   };
+  block_B();
 }
+
+// COMMON-LABEL: define internal {{.*}}i32 @__foo_block_invoke(i8 addrspace(4)* %.block_descriptor)
+// COMMON:  %[[block:.*]] = bitcast i8 addrspace(4)* %.block_descriptor to <{ i32, i32, i8 addrspace(4)*, i32 }> addrspace(4)*
+// COMMON:  %[[block_capture_addr:.*]] = getelementptr inbounds <{ i32, i32, i8 addrspace(4)*, i32 }>, <{ i32, i32, i8 addrspace(4)*, i32 }> addrspace(4)* %[[block]], i32 0, i32 3
+// COMMON:  %[[r1:.*]] = load i32, i32 addrspace(4)* %[[block_capture_addr]]

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

 // RUN: %clang_cc1 %s -cl-std=CL2.0 -ffake-address-space-map -O0 -emit-llvm -o - -triple "spir-unknown-unknown" | FileCheck %s --check-prefix=COMMON --check-prefix=B32
 // RUN: %clang_cc1 %s -cl-std=CL2.0 -ffake-address-space-map -O0 -emit-llvm -o - -triple "spir64-unknown-unknown" | FileCheck %s --check-prefix=COMMON --check-prefix=B64
 
 #pragma OPENCL EXTENSION cl_khr_subgroups : enable
 
 typedef void (^bl_t)(local void *);
 typedef struct {int a;} ndrange_t;
 
 // N.B. The check here only exists to set BL_GLOBAL
-// COMMON: @block_G =  addrspace(1) constant void (i8 addrspace(3)*) addrspace(4)* addrspacecast (void (i8 addrspace(3)*) addrspace(1)* bitcast ({ i8**, i32, i32, i8*, %struct.__block_descriptor addrspace(2)* } addrspace(1)* [[BL_GLOBAL:@__block_literal_global(\.[0-9]+)?]] to void (i8 addrspace(3)*) addrspace(1)*) to void (i8 addrspace(3)*) addrspace(4)*)
+// COMMON: @block_G =  addrspace(1) constant void (i8 addrspace(3)*) addrspace(4)* addrspacecast (void (i8 addrspace(3)*) addrspace(1)* bitcast ({ i32, i32, i8 addrspace(4)* } addrspace(1)* [[BL_GLOBAL:@__block_literal_global(\.[0-9]+)?]] to void (i8 addrspace(3)*) addrspace(1)*) to void (i8 addrspace(3)*) addrspace(4)*)
 const bl_t block_G = (bl_t) ^ (local void *a) {};
 
 kernel void device_side_enqueue(global int *a, global int *b, int i) {
   // COMMON: %default_queue = alloca %opencl.queue_t*
   queue_t default_queue;
   // COMMON: %flags = alloca i32
   unsigned flags = 0;
   // COMMON: %ndrange = alloca %struct.ndrange_t
   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};
 
   // COMMON: [[NDR:%[a-z0-9]+]] = alloca %struct.ndrange_t, align 4
   // COMMON: [[DEF_Q:%[0-9]+]] = load %opencl.queue_t{{.*}}*, %opencl.queue_t{{.*}}** %default_queue
   // COMMON: [[FLAGS:%[0-9]+]] = load i32, i32* %flags
-  // COMMON: [[BL:%[0-9]+]] = bitcast <{ i8*, i32, i32, i8*, %struct.__block_descriptor addrspace(2)*, i32{{.*}}, i32{{.*}}, i32{{.*}} }>* %block to void ()*
+  // B32: [[BL:%[0-9]+]] = bitcast <{ i32, i32, i8 addrspace(4)*, i32 addrspace(1)*, i32, i32 addrspace(1)* }>* %block to void ()*
+  // B64: [[BL:%[0-9]+]] = bitcast <{ i32, i32, i8 addrspace(4)*, i32 addrspace(1)*, i32 addrspace(1)*, i32 }>* %block to void ()*
   // COMMON: [[BL_I8:%[0-9]+]] = addrspacecast void ()* [[BL]] to i8 addrspace(4)*
-  // COMMON: call i32 @__enqueue_kernel_basic(%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* byval [[NDR]]{{(.[0-9]+)?}}, i8 addrspace(4)* [[BL_I8]])
+  // COMMON: call i32 @__enqueue_kernel_basic(%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* byval [[NDR]]{{([0-9]+)?}}, i8 addrspace(4)* [[BL_I8]])
   enqueue_kernel(default_queue, flags, ndrange,
                  ^(void) {
                    a[i] = b[i];
                  });
 
   // COMMON: [[DEF_Q:%[0-9]+]] = load %opencl.queue_t{{.*}}*, %opencl.queue_t{{.*}}** %default_queue
   // COMMON: [[FLAGS:%[0-9]+]] = load i32, i32* %flags
   // 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 <{ i8*, i32, i32, i8*, %struct.__block_descriptor addrspace(2)*, i32{{.*}}, i32{{.*}}, i32{{.*}} }>* %block3 to void ()*
+  // COMMON: [[BL:%[0-9]+]] = bitcast <{ i32, i32, i8 addrspace(4)*, i32{{.*}}, i32{{.*}}, i32{{.*}} }>* %block3 to void ()*
   // COMMON: [[BL_I8:%[0-9]+]] = addrspacecast void ()* [[BL]] to i8 addrspace(4)*
   // COMMON: call i32 @__enqueue_kernel_basic_events(%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]], i8 addrspace(4)* [[BL_I8]])
   enqueue_kernel(default_queue, flags, ndrange, 2, &event_wait_list, &clk_event,
                  ^(void) {
                    a[i] = b[i];
                  });
 
   // 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
-  // B32: call i32 @__enqueue_kernel_vaargs(%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* [[NDR]]{{(.[0-9]+)?}}, i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i8**, i32, i32, i8*, %struct.__block_descriptor addrspace(2)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1, i32* %[[TMP1]])
+  // B32: call i32 @__enqueue_kernel_vaargs(%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* [[NDR]]{{([0-9]+)?}}, i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i32, i32, i8 addrspace(4)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1, i32* %[[TMP1]])
   // 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
-  // B64: call i32 @__enqueue_kernel_vaargs(%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* [[NDR]]{{(.[0-9]+)?}}, i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i8**, i32, i32, i8*, %struct.__block_descriptor addrspace(2)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1, i64* %[[TMP1]])
+  // B64: call i32 @__enqueue_kernel_vaargs(%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* [[NDR]]{{([0-9]+)?}}, i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i32, i32, i8 addrspace(4)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1, i64* %[[TMP1]])
   enqueue_kernel(default_queue, flags, ndrange,
                  ^(local void *p) {
                    return;
                  },
                  256);
   char c;
   // 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
-  // B32: call i32 @__enqueue_kernel_vaargs(%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* [[NDR]]{{(.[0-9]+)?}}, i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i8**, i32, i32, i8*, %struct.__block_descriptor addrspace(2)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1, i32* %[[TMP1]])
+  // B32: call i32 @__enqueue_kernel_vaargs(%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* [[NDR]]{{([0-9]+)?}}, i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i32, i32, i8 addrspace(4)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1, i32* %[[TMP1]])
   // 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
-  // B64: call i32 @__enqueue_kernel_vaargs(%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* [[NDR]]{{(.[0-9]+)?}}, i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i8**, i32, i32, i8*, %struct.__block_descriptor addrspace(2)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1, i64* %[[TMP1]])
+  // B64: call i32 @__enqueue_kernel_vaargs(%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* [[NDR]]{{([0-9]+)?}}, i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i32, i32, i8 addrspace(4)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1, i64* %[[TMP1]])
   enqueue_kernel(default_queue, flags, ndrange,
                  ^(local void *p) {
                    return;
                  },
                  c);
 
   // 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
-  // B32: call i32 @__enqueue_kernel_events_vaargs(%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]],  %struct.ndrange_t* {{.*}}, i32 2, %opencl.clk_event_t{{.*}} [[WAIT_EVNT]], %opencl.clk_event_t{{.*}} [[EVNT]], i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i8**, i32, i32, i8*, %struct.__block_descriptor addrspace(2)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1, i32* %[[TMP1]])
+  // B32: call i32 @__enqueue_kernel_events_vaargs(%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]],  %struct.ndrange_t* {{.*}}, i32 2, %opencl.clk_event_t{{.*}} [[WAIT_EVNT]], %opencl.clk_event_t{{.*}} [[EVNT]], i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i32, i32, i8 addrspace(4)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1, i32* %[[TMP1]])
   // 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
-  // B64: call i32 @__enqueue_kernel_events_vaargs(%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]],  %struct.ndrange_t* {{.*}}, i32 2, %opencl.clk_event_t{{.*}} [[WAIT_EVNT]], %opencl.clk_event_t{{.*}} [[EVNT]], i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i8**, i32, i32, i8*, %struct.__block_descriptor addrspace(2)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1, i64* %[[TMP1]])
+  // B64: call i32 @__enqueue_kernel_events_vaargs(%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]],  %struct.ndrange_t* {{.*}}, i32 2, %opencl.clk_event_t{{.*}} [[WAIT_EVNT]], %opencl.clk_event_t{{.*}} [[EVNT]], i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i32, i32, i8 addrspace(4)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1, i64* %[[TMP1]])
   enqueue_kernel(default_queue, flags, ndrange, 2, event_wait_list2, &clk_event,
                  ^(local void *p) {
                    return;
                  },
                  256);
 
   // 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
-  // B32: call i32 @__enqueue_kernel_events_vaargs(%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]], i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i8**, i32, i32, i8*, %struct.__block_descriptor addrspace(2)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1, i32* %[[TMP1]])
+  // B32: call i32 @__enqueue_kernel_events_vaargs(%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]], i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i32, i32, i8 addrspace(4)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1, i32* %[[TMP1]])
   // 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
-  // B64: call i32 @__enqueue_kernel_events_vaargs(%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]], i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i8**, i32, i32, i8*, %struct.__block_descriptor addrspace(2)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1, i64* %[[TMP1]])
+  // B64: call i32 @__enqueue_kernel_events_vaargs(%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]], i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i32, i32, i8 addrspace(4)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1, i64* %[[TMP1]])
   enqueue_kernel(default_queue, flags, ndrange, 2, event_wait_list2, &clk_event,
                  ^(local void *p) {
                    return;
                  },
                  c);
 
   long l;
   // 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
-  // B32: call i32 @__enqueue_kernel_vaargs(%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* [[NDR]]{{(.[0-9]+)?}}, i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i8**, i32, i32, i8*, %struct.__block_descriptor addrspace(2)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1, i32* %[[TMP1]])
+  // B32: call i32 @__enqueue_kernel_vaargs(%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* [[NDR]]{{([0-9]+)?}}, i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i32, i32, i8 addrspace(4)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1, i32* %[[TMP1]])
   // 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
-  // B64: call i32 @__enqueue_kernel_vaargs(%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* [[NDR]]{{(.[0-9]+)?}}, i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i8**, i32, i32, i8*, %struct.__block_descriptor addrspace(2)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1, i64* %[[TMP1]])
+  // B64: call i32 @__enqueue_kernel_vaargs(%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* [[NDR]]{{([0-9]+)?}}, i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i32, i32, i8 addrspace(4)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1, i64* %[[TMP1]])
   enqueue_kernel(default_queue, flags, ndrange,
                  ^(local void *p) {
                    return;
                  },
                  l);
 
   // 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
-  // B32: call i32 @__enqueue_kernel_vaargs(%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* [[NDR]]{{(.[0-9]+)?}}, i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i8**, i32, i32, i8*, %struct.__block_descriptor addrspace(2)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 3, i32* %[[TMP1]])
+  // B32: call i32 @__enqueue_kernel_vaargs(%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* [[NDR]]{{([0-9]+)?}}, i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i32, i32, i8 addrspace(4)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 3, i32* %[[TMP1]])
   // 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
-  // B64: call i32 @__enqueue_kernel_vaargs(%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* [[NDR]]{{(.[0-9]+)?}}, i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i8**, i32, i32, i8*, %struct.__block_descriptor addrspace(2)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 3, i64* %[[TMP1]])
+  // B64: call i32 @__enqueue_kernel_vaargs(%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* [[NDR]]{{([0-9]+)?}}, i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i32, i32, i8 addrspace(4)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 3, i64* %[[TMP1]])
   enqueue_kernel(default_queue, flags, ndrange,
                  ^(local void *p1, local void *p2, local void *p3) {
                    return;
                  },
                  1, 2, 4);
 
   // 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
-  // B32: call i32 @__enqueue_kernel_vaargs(%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* [[NDR]]{{(.[0-9]+)?}}, i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i8**, i32, i32, i8*, %struct.__block_descriptor addrspace(2)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1, i32* %[[TMP1]])
+  // B32: call i32 @__enqueue_kernel_vaargs(%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* [[NDR]]{{([0-9]+)?}}, i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i32, i32, i8 addrspace(4)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1, i32* %[[TMP1]])
   // 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
-  // B64: call i32 @__enqueue_kernel_vaargs(%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* [[NDR]]{{(.[0-9]+)?}}, i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i8**, i32, i32, i8*, %struct.__block_descriptor addrspace(2)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1, i64* %[[TMP1]])
+  // B64: call i32 @__enqueue_kernel_vaargs(%opencl.queue_t{{.*}}* [[DEF_Q]], i32 [[FLAGS]], %struct.ndrange_t* [[NDR]]{{([0-9]+)?}}, i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i32, i32, i8 addrspace(4)* } addrspace(1)* @__block_literal_global{{(.[0-9]+)?}} to i8 addrspace(1)*) to i8 addrspace(4)*), i32 1, i64* %[[TMP1]])
   enqueue_kernel(default_queue, flags, ndrange,
                  ^(local void *p) {
                    return;
                  },
                  4294967296L);
 
   // The full type of these expressions are long (and repeated elsewhere), so we
   // capture it as part of the regex for convenience and clarity.
-  // COMMON: store void () addrspace(4)* addrspacecast (void () addrspace(1)* bitcast ({ i8**, i32, i32, i8*, %struct.__block_descriptor addrspace(2)* } addrspace(1)* [[BL_A:@__block_literal_global(\.[0-9]+)?]] to void () addrspace(1)*) to void () addrspace(4)*), void () addrspace(4)** %block_A
+  // COMMON: store void () addrspace(4)* addrspacecast (void () addrspace(1)* bitcast ({ i32, i32, i8 addrspace(4)* } addrspace(1)* [[BL_A:@__block_literal_global(\.[0-9]+)?]] to void () addrspace(1)*) to void () addrspace(4)*), void () addrspace(4)** %block_A
   void (^const block_A)(void) = ^{
     return;
   };
 
-  // COMMON: store void (i8 addrspace(3)*) addrspace(4)* addrspacecast (void (i8 addrspace(3)*) addrspace(1)* bitcast ({ i8**, i32, i32, i8*, %struct.__block_descriptor addrspace(2)* } addrspace(1)* [[BL_B:@__block_literal_global(\.[0-9]+)?]] to void (i8 addrspace(3)*) addrspace(1)*) to void (i8 addrspace(3)*) addrspace(4)*), void (i8 addrspace(3)*) addrspace(4)** %block_B
+  // COMMON: store void (i8 addrspace(3)*) addrspace(4)* addrspacecast (void (i8 addrspace(3)*) addrspace(1)* bitcast ({ i32, i32, i8 addrspace(4)* } addrspace(1)* [[BL_B:@__block_literal_global(\.[0-9]+)?]] to void (i8 addrspace(3)*) addrspace(1)*) to void (i8 addrspace(3)*) addrspace(4)*), void (i8 addrspace(3)*) addrspace(4)** %block_B
   void (^const block_B)(local void *) = ^(local void *a) {
     return;
   };
 
-  // COMMON: call i32 @__get_kernel_work_group_size_impl(i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i8**, i32, i32, i8*, %struct.__block_descriptor addrspace(2)* } addrspace(1)* [[BL_A]] to i8 addrspace(1)*) to i8 addrspace(4)*))
+  // COMMON: call i32 @__get_kernel_work_group_size_impl(i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i32, i32, i8 addrspace(4)* } addrspace(1)* [[BL_A]] to i8 addrspace(1)*) to i8 addrspace(4)*))
   unsigned size = get_kernel_work_group_size(block_A);
-  // COMMON: call i32 @__get_kernel_work_group_size_impl(i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i8**, i32, i32, i8*, %struct.__block_descriptor addrspace(2)* } addrspace(1)* [[BL_B]] to i8 addrspace(1)*) to i8 addrspace(4)*))
+  // COMMON: call i32 @__get_kernel_work_group_size_impl(i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i32, i32, i8 addrspace(4)* } addrspace(1)* [[BL_B]] to i8 addrspace(1)*) to i8 addrspace(4)*))
   size = get_kernel_work_group_size(block_B);
-  // COMMON: call i32 @__get_kernel_preferred_work_group_multiple_impl(i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i8**, i32, i32, i8*, %struct.__block_descriptor addrspace(2)* } addrspace(1)* [[BL_A]] to i8 addrspace(1)*) to i8 addrspace(4)*))
+  // COMMON: call i32 @__get_kernel_preferred_work_group_multiple_impl(i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i32, i32, i8 addrspace(4)* } addrspace(1)* [[BL_A]] to i8 addrspace(1)*) to i8 addrspace(4)*))
   size = get_kernel_preferred_work_group_size_multiple(block_A);
-  // COMMON: call i32 @__get_kernel_preferred_work_group_multiple_impl(i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i8**, i32, i32, i8*, %struct.__block_descriptor addrspace(2)* } addrspace(1)* [[BL_GLOBAL]] to i8 addrspace(1)*) to i8 addrspace(4)*))
+  // COMMON: call i32 @__get_kernel_preferred_work_group_multiple_impl(i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i32, i32, i8 addrspace(4)* } addrspace(1)* [[BL_GLOBAL]] to i8 addrspace(1)*) to i8 addrspace(4)*))
   size = get_kernel_preferred_work_group_size_multiple(block_G);
 
-  // COMMON: call i32 @__get_kernel_max_sub_group_size_for_ndrange_impl(%struct.ndrange_t* {{.*}}, i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i8**, i32, i32, i8*, %struct.__block_descriptor addrspace(2)* } addrspace(1)* {{.*}} to i8 addrspace(1)*) to i8 addrspace(4)*))
+  // COMMON: call i32 @__get_kernel_max_sub_group_size_for_ndrange_impl(%struct.ndrange_t* {{.*}}, i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i32, i32, i8 addrspace(4)* } addrspace(1)* {{.*}} to i8 addrspace(1)*) to i8 addrspace(4)*))
   size = get_kernel_max_sub_group_size_for_ndrange(ndrange, ^(){});
-  // COMMON: call i32 @__get_kernel_sub_group_count_for_ndrange_impl(%struct.ndrange_t* {{.*}}, i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i8**, i32, i32, i8*, %struct.__block_descriptor addrspace(2)* } addrspace(1)* {{.*}} to i8 addrspace(1)*) to i8 addrspace(4)*))
+  // COMMON: call i32 @__get_kernel_sub_group_count_for_ndrange_impl(%struct.ndrange_t* {{.*}}, i8 addrspace(4)* addrspacecast (i8 addrspace(1)* bitcast ({ i32, i32, i8 addrspace(4)* } addrspace(1)* {{.*}} to i8 addrspace(1)*) to i8 addrspace(4)*))
   size = get_kernel_sub_group_count_for_ndrange(ndrange, ^(){});
 }