AMDGPU: Handle "uniform-work-group-size" attribute

[aakanksha555]

Summary

Updated the annotate-kernel-features pass to support the propagation of uniform-work-group attribute. This attribute propagates top down from the callers till the callees. Maintained a list in the increasing order of uses (caller to callee). Propagated the uniform-work-group attribute from called functions for each node to the callees. Any functions which do not have the attribute will be added this attribute with a "false" value after this pass.

[arsenm]

lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp

807 ↗

(On Diff #158825)

We cannot rely on function names, or introduce functions such as this. This is also not necessary for optimizing this

[arsenm]

A more correct way to optimize this would be to have a CallGraphSCC pass that propagates the uniform-work-group-size attribute to callees only reachable from kernels with uniform-work-group-size

[aakanksha555]

Added the tests.

[arsenm]

lib/Target/AMDGPU/AMDGPUAnnotateKernelFeatures.cpp

280

OS check isn't necessary

283–284

I don't see how this prevents propagating the attribute if other callers do not have it

286–288

Should only be looking for == "true"

lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

684

Extra space at end

771–772

Delete

test/CodeGen/AMDGPU/uniform-workgroup-test1.ll

1 ↗

(On Diff #164518)

Tests are missing run lines

25–30 ↗

(On Diff #164518)

This includes attributes added by the pass

[aakanksha555]

lib/Target/AMDGPU/AMDGPUAnnotateKernelFeatures.cpp

283–284

By other callers, do you mean other kernel function?
AttrNames[] does not include "uniform-work-group-size" in the list so it wouldn't get copied to other kernel functions form the callee function.

[aakanksha555]

Added support to ensure the attribute propagates from the caller to the function even within nested function calls.

[aakanksha555]

Updated Test5

[aakanksha555]

Updated the patch to prevent propagating the attribute if other callers do not have it.
Updated Test 5 to show the update.

[aakanksha555]

Found 2 failing tests. Updated code to fix them.

[arsenm]

I think you need to split this into a separate loop before the propagate attributes function instead of adding the recursive call at the same time. This is different from the other attributes because it is inferred top down. You should have a first loop over the CallGraphSCC that adds this. Since the CallGraphSCC should have all of the nodes reachable from each other, this should be some set building / checks from there. You shouldn't need to be looking at the instructions inside the functions and looking for specific call sites

lib/Target/AMDGPU/AMDGPUAnnotateKernelFeatures.cpp

227–228

You don't need to check if it has the attribute and check the value == false. Just check the value == true

238

This shouldn't use recursion. This will fail when there's a recursive call. It also shouldn't be necessary, because callees should be called before callers in an SCC pass

test/CodeGen/AMDGPU/uniform-work-group-test1.ll

1 ↗

(On Diff #166333)

GCN check prefix is usually used for ISA check lines. It doesn't really matter, but would require more changes if for some reason in the future an llc run line were added

4–5 ↗

(On Diff #166333)

Functions and attribute group variables could use better names (as well as the test file). It would also help to add a comment at the top of each test file for what case this is supposed to be

test/CodeGen/AMDGPU/uniform-work-group-test5.ll

1 ↗

(On Diff #166333)

Triple is broken

23 ↗

(On Diff #166333)

This is accepted by the IR parser? I would remove all of these empty attribute groups

[arsenm]

lib/Target/AMDGPU/AMDGPUAnnotateKernelFeatures.cpp

83–85

No globals

217

I don't understand the point of anything this function is doing. You're just copying the same information that's already present in the SCC into a slightly different structure

220

Range loop

223

This leaks and should also be unnecessary

230

std::make_pair

239

Typo propagte

337

Comment should be capitalized, have a space after // and be punctuated

test/CodeGen/AMDGPU/uniform-work-group-nested-function-calls.ll

22

Should also test a recursive loop

[aakanksha555]

Got rid of extra structures and added a recursive test.

[arsenm]

lib/Target/AMDGPU/AMDGPUAnnotateKernelFeatures.cpp

49

llvm:: unnecessary

189

Random whitespace change

220

Space before (

221

You aren't using this as a stack, so this is weird. You could just do a range loop on reverse(SortedNodeList)

338

Comment is misleading since you aren't really sorting it. Also still not clear why the number of uses matter. Is it just to find unused functions?

351

Extra whitespace change

test/CodeGen/AMDGPU/uniform-work-group-resursion-test.ll

1

Missing amdhsa in triple

Testname spelling resursion

7

Better check variable name

test/CodeGen/AMDGPU/uniform-work-group-test.ll

33

A testcase with the attribute on an external function might also be useful

[aakanksha555]

lib/Target/AMDGPU/AMDGPUAnnotateKernelFeatures.cpp

338

I didn't need to explicitly sort it as push_back is doing it for me. The result is a list ranging from the most number of uses to the least.
The uses help to identify Callers and the Callees. For eg. kernel functions which are only calling other functions have just one use.
And that is where we need to start the attribute propagation from. This ensures that attributes get propagated through nested function calls.

[aakanksha555]

lib/Target/AMDGPU/AMDGPUAnnotateKernelFeatures.cpp

221

I used pop_back since I wanted the last element first. I'll change it to a range loop.

[aakanksha555]

-Updated a test case to include an external function.
-Uploaded the correct version of the recursion test.
-Fixed other small errors.

[arsenm]

lib/Target/AMDGPU/AMDGPUAnnotateKernelFeatures.cpp

226–228

You should just deal with this fixme and add a test for it. This should be as easy as checking Callee->mayBeRedefined?

253

It would be cleaner to pull all of this into a function which returns true on changed

[aakanksha555]

lib/Target/AMDGPU/AMDGPUAnnotateKernelFeatures.cpp

226–228

Did you mean Callee->mayBeDerefined() ?

[aakanksha555]

Added a check for external functions and modified the test for it

[arsenm]

lib/Target/AMDGPU/AMDGPUAnnotateKernelFeatures.cpp

49

Remove Sorted from the name now?

226–228

I think so? I'm not sure what the difference is from isInterposable. This also should check F.hasAddressTaken.

234

You don't need this, which was the point of putting this into a function

242–243

You can just return true directly

253

You can just return true

257

You can just return true

260–262

Can you avoid adding the attribute if not originally present?

[aakanksha555]

lib/Target/AMDGPU/AMDGPUAnnotateKernelFeatures.cpp

260–262

If I don't add the attribute if not originally present, it can create a discrepancy in certain scenarios.
For eg. A function is called by two kernels, one without the attribute and the other with uniform-work-group-attribute = true. The function will be set as uniform-work-group-attribute = true, which may not be the correct approach.

[aakanksha555]

-Directly returning "true" from the propagateAttribute function as per the comments.
-Renamed the list to NodeList, dropped "Sorted" from the name

[arsenm]

LGTM with the long line fixed

lib/Target/AMDGPU/AMDGPUAnnotateKernelFeatures.cpp

233

I think this goes over the column limit

[hakzsam]

Hi,

This patch breaks RADV (and probably RadeonSI as well). Here's a backtrace of the problem:

$ gdb --args ./deqp-vk --deqp-case=dEQP-VK.spirv_assembly.instruction.graphics.selection_block_order.out_of_order_frag
GNU gdb (GDB) 8.2
Copyright (C) 2018 Free Software Foundation, Inc.
License GPLv3+: GNU GPL version 3 or later http://gnu.org/licenses/gpl.html
This is free software: you are free to change and redistribute it.
There is NO WARRANTY, to the extent permitted by law.
Type "show copying" and "show warranty" for details.
This GDB was configured as "x86_64-pc-linux-gnu".
Type "show configuration" for configuration details.
For bug reporting instructions, please see:
http://www.gnu.org/software/gdb/bugs/.
Find the GDB manual and other documentation resources online at:

<http://www.gnu.org/software/gdb/documentation/>.

For help, type "help".
Type "apropos word" to search for commands related to "word"...
Reading symbols from ./deqp-vk...(no debugging symbols found)...done.
(gdb) r
Starting program: /home/hakzsam/programming/VK-GL-CTS/build/external/vulkancts/modules/vulkan/deqp-vk --deqp-case=dEQP-VK.spirv_assembly.instruction.graphics.selection_block_order.out_of_order_frag
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/usr/lib/libthread_db.so.1".
Writing test log into TestResults.qpa
dEQP Core git-12aa347f43c85df3a0daf930739551d3f53d3d48 (0x12aa347f) starting..

target implementation = 'Default'

[New Thread 0x7ffff1968700 (LWP 4723)]
[Thread 0x7ffff1968700 (LWP 4723) exited]
[New Thread 0x7ffff1968700 (LWP 4724)]

Thread 1 "deqp-vk" received signal SIGSEGV, Segmentation fault.
0x00007ffff447a81a in llvm::ValueHandleBase::AddToExistingUseList (this=this@entry=0x7fffffffbc50, List=0x7ffff5fdab80 <vtable for llvm::AAResults::Model<llvm::ScopedNoAliasAAResult>+16>) at ../lib/IR/Value.cpp:745
745 *List = this;
(gdb) bt
#0 0x00007ffff447a81a in llvm::ValueHandleBase::AddToExistingUseList (this=this@entry=0x7fffffffbc50, List=0x7ffff5fdab80 <vtable for llvm::AAResults::Model<llvm::ScopedNoAliasAAResult>+16>) at ../lib/IR/Value.cpp:745
#1 0x00007ffff572fd87 in llvm::ValueHandleBase::ValueHandleBase (RHS=..., Kind=llvm::ValueHandleBase::WeakTracking, this=0x7fffffffbc50) at ../include/llvm/ADT/PointerIntPair.h:150
#2 llvm::WeakTrackingVH::WeakTrackingVH (RHS=..., this=0x7fffffffbc50) at ../include/llvm/IR/ValueHandle.h:187
#3 std::pair<llvm::WeakTrackingVH, llvm::CallGraphNode*>::pair (this=0x7fffffffbc50) at /usr/include/c++/8.2.1/bits/stl_pair.h:303
#4 (anonymous namespace)::AMDGPUAnnotateKernelFeatures::processUniformWorkGroupAttribute (this=0x555557e87570) at ../lib/Target/AMDGPU/AMDGPUAnnotateKernelFeatures.cpp:224
#5 (anonymous namespace)::AMDGPUAnnotateKernelFeatures::runOnSCC (this=<optimized out>, SCC=...) at ../lib/Target/AMDGPU/AMDGPUAnnotateKernelFeatures.cpp:355
#6 0x00007ffff51d3d47 in (anonymous namespace)::CGPassManager::RunPassOnSCC (DevirtualizedCall=<synthetic pointer>: <optimized out>, CallGraphUpToDate=<synthetic pointer>: <optimized out>, CG=..., CurSCC=..., P=

0x555557e87570, this=0x555557e87650) at ../lib/Analysis/CallGraphSCCPass.cpp:141

#7 (anonymous namespace)::CGPassManager::RunAllPassesOnSCC (DevirtualizedCall=<synthetic pointer>: <optimized out>, CG=..., CurSCC=..., this=0x555557e87650) at ../lib/Analysis/CallGraphSCCPass.cpp:442
#8 (anonymous namespace)::CGPassManager::runOnModule (this=0x555557e87650, M=...) at ../lib/Analysis/CallGraphSCCPass.cpp:498
#9 0x00007ffff4428fba in (anonymous namespace)::MPPassManager::runOnModule (M=..., this=0x555557e65c20) at ../lib/IR/LegacyPassManager.cpp:1744
#10 llvm::legacy::PassManagerImpl::run (this=0x555557e657b0, M=...) at ../lib/IR/LegacyPassManager.cpp:1857
#11 0x00007ffff61fa7a6 in ac_compile_module_to_binary (p=0x555557e65750, module=module@entry=0x555557eb65a0, binary=binary@entry=0x7fffffffc080) at /home/hakzsam/install/llvm/debug/master/include/llvm/IR/Module.h:889
#12 0x00007ffff61b6e2b in radv_llvm_per_thread_info::compile_to_memory_buffer (this=<optimized out>, binary=0x7fffffffc080, module=0x555557eb65a0) at radv_llvm_helper.cpp:97
#13 radv_compile_to_binary (info=info@entry=0x7fffffffc050, module=module@entry=0x7ffff61fa7a6 <ac_compile_module_to_binary(ac_compiler_passes*, LLVMModuleRef, ac_shader_binary*)+22>, binary=binary@entry=0x7fffffffc080)

at radv_llvm_helper.cpp:97

#14 0x00007ffff61b0d81 in ac_llvm_compile (ac_llvm=0x7fffffffc050, binary=0x7fffffffc080, M=0x7ffff61fa7a6 <ac_compile_module_to_binary(ac_compiler_passes*, LLVMModuleRef, ac_shader_binary*)+22>) at radv_nir_to_llvm.c:3660
#15 ac_compile_llvm_module (ac_llvm=ac_llvm@entry=0x7fffffffc050, llvm_module=0x7ffff61fa7a6 <ac_compile_module_to_binary(ac_compiler_passes*, LLVMModuleRef, ac_shader_binary*)+22>, binary=binary@entry=0x7fffffffc080,

config=0x7fffffffc080, config@entry=0x555557f71778, stage=MESA_SHADER_COMPUTE, options=0x7fffffffc140, shader_info=<optimized out>, shader_info=<optimized out>) at radv_nir_to_llvm.c:3684

#16 0x00007ffff61b65d0 in radv_compile_nir_shader (ac_llvm=ac_llvm@entry=0x7fffffffc050, binary=binary@entry=0x7fffffffc080, config=config@entry=0x555557f71778, shader_info=shader_info@entry=0x555557f717a0,

nir=nir@entry=0x7fffffffc388, nir_count=nir_count@entry=1, options=0x7fffffffc140) at radv_nir_to_llvm.c:3808

#17 0x00007ffff61c56db in shader_variant_create (device=device@entry=0x555557e4d920, module=0x7fffffffcd40, shaders=shaders@entry=0x7fffffffc388, shader_count=shader_count@entry=1, stage=MESA_SHADER_COMPUTE,

options=options@entry=0x7fffffffc140, gs_copy_shader=false, code_out=0x7fffffffc3b8, code_size_out=0x7fffffffc2e4) at radv_shader.c:612

#18 0x00007ffff61c5b04 in radv_shader_variant_create (device=device@entry=0x555557e4d920, module=<optimized out>, shaders=shaders@entry=0x7fffffffc388, shader_count=shader_count@entry=1, layout=<optimized out>,

key=key@entry=0x7fffffffc810, code_out=0x7fffffffc3b8, code_size_out=0x7fffffffc2e4) at radv_shader.c:666

#19 0x00007ffff61b8aa6 in radv_create_shaders (pipeline=0x555557ef5ca0, device=<optimized out>, cache=0x555557e4d998, key=<optimized out>, pStages=<optimized out>, flags=<optimized out>) at radv_pipeline.c:2151
#20 0x00007ffff61bf7eb in radv_compute_pipeline_create (pPipeline=0x555557e4ef70, pAllocator=<optimized out>, pCreateInfo=0x7fffffffcbd0, _cache=<optimized out>, _device=0x555557e4d920) at radv_pipeline.c:3787
#21 radv_CreateComputePipelines (_device=_device@entry=0x555557e4d920, pipelineCache=pipelineCache@entry=0x555557e4d998, count=count@entry=1, pCreateInfos=pCreateInfos@entry=0x7fffffffcbd0, pAllocator=pAllocator@entry=0x0,

pPipelines=pPipelines@entry=0x555557e4ef70) at radv_pipeline.c:3817

#22 0x00007ffff619653a in radv_device_init_meta_itob_state (device=0x555557e4d920) at radv_private.h:1986
#23 radv_device_init_meta_bufimage_state (device=device@entry=0x555557e4d920) at radv_meta_bufimage.c:1489
#24 0x00007ffff6175a4a in radv_device_init_meta (device=device@entry=0x555557e4d920) at radv_meta.c:365
#25 0x00007ffff61680d0 in radv_CreateDevice (physicalDevice=0x555557d7c0e0, pCreateInfo=0x7fffffffd0d0, pAllocator=<optimized out>, pDevice=0x555557d82ec0) at radv_device.c:1702
#26 0x00007ffff640c574 in ?? () from /usr/lib/libvulkan.so.1
#27 0x00007ffff641599b in ?? () from /usr/lib/libvulkan.so.1
#28 0x00007ffff6419b29 in vkCreateDevice () from /usr/lib/libvulkan.so.1
#29 0x0000555556942f7d in vk::createDevice(vk::PlatformInterface const&, vk::VkInstance_s*, vk::InstanceInterface const&, vk::VkPhysicalDevice_s*, vk::VkDeviceCreateInfo const*, vk::VkAllocationCallbacks const*) ()
#30 0x00005555558a3384 in vkt::DefaultDevice::DefaultDevice(vk::PlatformInterface const&, tcu::CommandLine const&) ()
#31 0x00005555558a40e5 in vkt::Context::Context(tcu::TestContext&, vk::PlatformInterface const&, vk::ProgramCollection<vk::ProgramBinary, vk::BinaryBuildOptions>&) ()
#32 0x000055555588c3e2 in vkt::TestCaseExecutor::TestCaseExecutor(tcu::TestContext&) ()
#33 0x000055555588c552 in vkt::TestPackage::createExecutor() const ()
#34 0x0000555556e04964 in tcu::TestSessionExecutor::iterate() ()
#35 0x0000555556dd89a9 in tcu::App::iterate() ()
#36 0x000055555587e4e8 in main ()

Can you look into it?

Thanks!

[aakanksha555]

In D50200#1329491, @hakzsam wrote:

Hi,

This patch breaks RADV (and probably RadeonSI as well). Here's a backtrace of the problem:

$ gdb --args ./deqp-vk --deqp-case=dEQP-VK.spirv_assembly.instruction.graphics.selection_block_order.out_of_order_frag
GNU gdb (GDB) 8.2
Copyright (C) 2018 Free Software Foundation, Inc.
License GPLv3+: GNU GPL version 3 or later http://gnu.org/licenses/gpl.html
This is free software: you are free to change and redistribute it.
There is NO WARRANTY, to the extent permitted by law.
Type "show copying" and "show warranty" for details.
This GDB was configured as "x86_64-pc-linux-gnu".
Type "show configuration" for configuration details.
For bug reporting instructions, please see:
http://www.gnu.org/software/gdb/bugs/.
Find the GDB manual and other documentation resources online at:

<http://www.gnu.org/software/gdb/documentation/>.

For help, type "help".
Type "apropos word" to search for commands related to "word"...
Reading symbols from ./deqp-vk...(no debugging symbols found)...done.
(gdb) r
Starting program: /home/hakzsam/programming/VK-GL-CTS/build/external/vulkancts/modules/vulkan/deqp-vk --deqp-case=dEQP-VK.spirv_assembly.instruction.graphics.selection_block_order.out_of_order_frag
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/usr/lib/libthread_db.so.1".
Writing test log into TestResults.qpa
dEQP Core git-12aa347f43c85df3a0daf930739551d3f53d3d48 (0x12aa347f) starting..

target implementation = 'Default'

[New Thread 0x7ffff1968700 (LWP 4723)]
[Thread 0x7ffff1968700 (LWP 4723) exited]
[New Thread 0x7ffff1968700 (LWP 4724)]

Thread 1 "deqp-vk" received signal SIGSEGV, Segmentation fault.
0x00007ffff447a81a in llvm::ValueHandleBase::AddToExistingUseList (this=this@entry=0x7fffffffbc50, List=0x7ffff5fdab80 <vtable for llvm::AAResults::Model<llvm::ScopedNoAliasAAResult>+16>) at ../lib/IR/Value.cpp:745
745 *List = this;
(gdb) bt
#0 0x00007ffff447a81a in llvm::ValueHandleBase::AddToExistingUseList (this=this@entry=0x7fffffffbc50, List=0x7ffff5fdab80 <vtable for llvm::AAResults::Model<llvm::ScopedNoAliasAAResult>+16>) at ../lib/IR/Value.cpp:745
#1 0x00007ffff572fd87 in llvm::ValueHandleBase::ValueHandleBase (RHS=..., Kind=llvm::ValueHandleBase::WeakTracking, this=0x7fffffffbc50) at ../include/llvm/ADT/PointerIntPair.h:150
#2 llvm::WeakTrackingVH::WeakTrackingVH (RHS=..., this=0x7fffffffbc50) at ../include/llvm/IR/ValueHandle.h:187
#3 std::pair<llvm::WeakTrackingVH, llvm::CallGraphNode*>::pair (this=0x7fffffffbc50) at /usr/include/c++/8.2.1/bits/stl_pair.h:303
#4 (anonymous namespace)::AMDGPUAnnotateKernelFeatures::processUniformWorkGroupAttribute (this=0x555557e87570) at ../lib/Target/AMDGPU/AMDGPUAnnotateKernelFeatures.cpp:224
#5 (anonymous namespace)::AMDGPUAnnotateKernelFeatures::runOnSCC (this=<optimized out>, SCC=...) at ../lib/Target/AMDGPU/AMDGPUAnnotateKernelFeatures.cpp:355
#6 0x00007ffff51d3d47 in (anonymous namespace)::CGPassManager::RunPassOnSCC (DevirtualizedCall=<synthetic pointer>: <optimized out>, CallGraphUpToDate=<synthetic pointer>: <optimized out>, CG=..., CurSCC=..., P=

0x555557e87570, this=0x555557e87650) at ../lib/Analysis/CallGraphSCCPass.cpp:141

#7 (anonymous namespace)::CGPassManager::RunAllPassesOnSCC (DevirtualizedCall=<synthetic pointer>: <optimized out>, CG=..., CurSCC=..., this=0x555557e87650) at ../lib/Analysis/CallGraphSCCPass.cpp:442
#8 (anonymous namespace)::CGPassManager::runOnModule (this=0x555557e87650, M=...) at ../lib/Analysis/CallGraphSCCPass.cpp:498
#9 0x00007ffff4428fba in (anonymous namespace)::MPPassManager::runOnModule (M=..., this=0x555557e65c20) at ../lib/IR/LegacyPassManager.cpp:1744
#10 llvm::legacy::PassManagerImpl::run (this=0x555557e657b0, M=...) at ../lib/IR/LegacyPassManager.cpp:1857
#11 0x00007ffff61fa7a6 in ac_compile_module_to_binary (p=0x555557e65750, module=module@entry=0x555557eb65a0, binary=binary@entry=0x7fffffffc080) at /home/hakzsam/install/llvm/debug/master/include/llvm/IR/Module.h:889
#12 0x00007ffff61b6e2b in radv_llvm_per_thread_info::compile_to_memory_buffer (this=<optimized out>, binary=0x7fffffffc080, module=0x555557eb65a0) at radv_llvm_helper.cpp:97
#13 radv_compile_to_binary (info=info@entry=0x7fffffffc050, module=module@entry=0x7ffff61fa7a6 <ac_compile_module_to_binary(ac_compiler_passes*, LLVMModuleRef, ac_shader_binary*)+22>, binary=binary@entry=0x7fffffffc080)

at radv_llvm_helper.cpp:97

#14 0x00007ffff61b0d81 in ac_llvm_compile (ac_llvm=0x7fffffffc050, binary=0x7fffffffc080, M=0x7ffff61fa7a6 <ac_compile_module_to_binary(ac_compiler_passes*, LLVMModuleRef, ac_shader_binary*)+22>) at radv_nir_to_llvm.c:3660
#15 ac_compile_llvm_module (ac_llvm=ac_llvm@entry=0x7fffffffc050, llvm_module=0x7ffff61fa7a6 <ac_compile_module_to_binary(ac_compiler_passes*, LLVMModuleRef, ac_shader_binary*)+22>, binary=binary@entry=0x7fffffffc080,

config=0x7fffffffc080, config@entry=0x555557f71778, stage=MESA_SHADER_COMPUTE, options=0x7fffffffc140, shader_info=<optimized out>, shader_info=<optimized out>) at radv_nir_to_llvm.c:3684

#16 0x00007ffff61b65d0 in radv_compile_nir_shader (ac_llvm=ac_llvm@entry=0x7fffffffc050, binary=binary@entry=0x7fffffffc080, config=config@entry=0x555557f71778, shader_info=shader_info@entry=0x555557f717a0,

nir=nir@entry=0x7fffffffc388, nir_count=nir_count@entry=1, options=0x7fffffffc140) at radv_nir_to_llvm.c:3808

#17 0x00007ffff61c56db in shader_variant_create (device=device@entry=0x555557e4d920, module=0x7fffffffcd40, shaders=shaders@entry=0x7fffffffc388, shader_count=shader_count@entry=1, stage=MESA_SHADER_COMPUTE,

options=options@entry=0x7fffffffc140, gs_copy_shader=false, code_out=0x7fffffffc3b8, code_size_out=0x7fffffffc2e4) at radv_shader.c:612

#18 0x00007ffff61c5b04 in radv_shader_variant_create (device=device@entry=0x555557e4d920, module=<optimized out>, shaders=shaders@entry=0x7fffffffc388, shader_count=shader_count@entry=1, layout=<optimized out>,

key=key@entry=0x7fffffffc810, code_out=0x7fffffffc3b8, code_size_out=0x7fffffffc2e4) at radv_shader.c:666

#19 0x00007ffff61b8aa6 in radv_create_shaders (pipeline=0x555557ef5ca0, device=<optimized out>, cache=0x555557e4d998, key=<optimized out>, pStages=<optimized out>, flags=<optimized out>) at radv_pipeline.c:2151
#20 0x00007ffff61bf7eb in radv_compute_pipeline_create (pPipeline=0x555557e4ef70, pAllocator=<optimized out>, pCreateInfo=0x7fffffffcbd0, _cache=<optimized out>, _device=0x555557e4d920) at radv_pipeline.c:3787
#21 radv_CreateComputePipelines (_device=_device@entry=0x555557e4d920, pipelineCache=pipelineCache@entry=0x555557e4d998, count=count@entry=1, pCreateInfos=pCreateInfos@entry=0x7fffffffcbd0, pAllocator=pAllocator@entry=0x0,

pPipelines=pPipelines@entry=0x555557e4ef70) at radv_pipeline.c:3817

#22 0x00007ffff619653a in radv_device_init_meta_itob_state (device=0x555557e4d920) at radv_private.h:1986
#23 radv_device_init_meta_bufimage_state (device=device@entry=0x555557e4d920) at radv_meta_bufimage.c:1489
#24 0x00007ffff6175a4a in radv_device_init_meta (device=device@entry=0x555557e4d920) at radv_meta.c:365
#25 0x00007ffff61680d0 in radv_CreateDevice (physicalDevice=0x555557d7c0e0, pCreateInfo=0x7fffffffd0d0, pAllocator=<optimized out>, pDevice=0x555557d82ec0) at radv_device.c:1702
#26 0x00007ffff640c574 in ?? () from /usr/lib/libvulkan.so.1
#27 0x00007ffff641599b in ?? () from /usr/lib/libvulkan.so.1
#28 0x00007ffff6419b29 in vkCreateDevice () from /usr/lib/libvulkan.so.1
#29 0x0000555556942f7d in vk::createDevice(vk::PlatformInterface const&, vk::VkInstance_s*, vk::InstanceInterface const&, vk::VkPhysicalDevice_s*, vk::VkDeviceCreateInfo const*, vk::VkAllocationCallbacks const*) ()
#30 0x00005555558a3384 in vkt::DefaultDevice::DefaultDevice(vk::PlatformInterface const&, tcu::CommandLine const&) ()
#31 0x00005555558a40e5 in vkt::Context::Context(tcu::TestContext&, vk::PlatformInterface const&, vk::ProgramCollection<vk::ProgramBinary, vk::BinaryBuildOptions>&) ()
#32 0x000055555588c3e2 in vkt::TestCaseExecutor::TestCaseExecutor(tcu::TestContext&) ()
#33 0x000055555588c552 in vkt::TestPackage::createExecutor() const ()
#34 0x0000555556e04964 in tcu::TestSessionExecutor::iterate() ()
#35 0x0000555556dd89a9 in tcu::App::iterate() ()
#36 0x000055555587e4e8 in main ()

Can you look into it?

Thanks!

Sorry about that. I have reverted the changes until I fix this.

[aakanksha555]

In D50200#1329491, @hakzsam wrote:

Hi,

This patch breaks RADV (and probably RadeonSI as well). Here's a backtrace of the problem:

$ gdb --args ./deqp-vk --deqp-case=dEQP-VK.spirv_assembly.instruction.graphics.selection_block_order.out_of_order_frag
GNU gdb (GDB) 8.2
Copyright (C) 2018 Free Software Foundation, Inc.
License GPLv3+: GNU GPL version 3 or later http://gnu.org/licenses/gpl.html
This is free software: you are free to change and redistribute it.
There is NO WARRANTY, to the extent permitted by law.
Type "show copying" and "show warranty" for details.
This GDB was configured as "x86_64-pc-linux-gnu".
Type "show configuration" for configuration details.
For bug reporting instructions, please see:
http://www.gnu.org/software/gdb/bugs/.
Find the GDB manual and other documentation resources online at:

<http://www.gnu.org/software/gdb/documentation/>.

For help, type "help".
Type "apropos word" to search for commands related to "word"...
Reading symbols from ./deqp-vk...(no debugging symbols found)...done.
(gdb) r
Starting program: /home/hakzsam/programming/VK-GL-CTS/build/external/vulkancts/modules/vulkan/deqp-vk --deqp-case=dEQP-VK.spirv_assembly.instruction.graphics.selection_block_order.out_of_order_frag
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/usr/lib/libthread_db.so.1".
Writing test log into TestResults.qpa
dEQP Core git-12aa347f43c85df3a0daf930739551d3f53d3d48 (0x12aa347f) starting..

target implementation = 'Default'

[New Thread 0x7ffff1968700 (LWP 4723)]
[Thread 0x7ffff1968700 (LWP 4723) exited]
[New Thread 0x7ffff1968700 (LWP 4724)]

Thread 1 "deqp-vk" received signal SIGSEGV, Segmentation fault.
0x00007ffff447a81a in llvm::ValueHandleBase::AddToExistingUseList (this=this@entry=0x7fffffffbc50, List=0x7ffff5fdab80 <vtable for llvm::AAResults::Model<llvm::ScopedNoAliasAAResult>+16>) at ../lib/IR/Value.cpp:745
745 *List = this;
(gdb) bt
#0 0x00007ffff447a81a in llvm::ValueHandleBase::AddToExistingUseList (this=this@entry=0x7fffffffbc50, List=0x7ffff5fdab80 <vtable for llvm::AAResults::Model<llvm::ScopedNoAliasAAResult>+16>) at ../lib/IR/Value.cpp:745
#1 0x00007ffff572fd87 in llvm::ValueHandleBase::ValueHandleBase (RHS=..., Kind=llvm::ValueHandleBase::WeakTracking, this=0x7fffffffbc50) at ../include/llvm/ADT/PointerIntPair.h:150
#2 llvm::WeakTrackingVH::WeakTrackingVH (RHS=..., this=0x7fffffffbc50) at ../include/llvm/IR/ValueHandle.h:187
#3 std::pair<llvm::WeakTrackingVH, llvm::CallGraphNode*>::pair (this=0x7fffffffbc50) at /usr/include/c++/8.2.1/bits/stl_pair.h:303
#4 (anonymous namespace)::AMDGPUAnnotateKernelFeatures::processUniformWorkGroupAttribute (this=0x555557e87570) at ../lib/Target/AMDGPU/AMDGPUAnnotateKernelFeatures.cpp:224
#5 (anonymous namespace)::AMDGPUAnnotateKernelFeatures::runOnSCC (this=<optimized out>, SCC=...) at ../lib/Target/AMDGPU/AMDGPUAnnotateKernelFeatures.cpp:355
#6 0x00007ffff51d3d47 in (anonymous namespace)::CGPassManager::RunPassOnSCC (DevirtualizedCall=<synthetic pointer>: <optimized out>, CallGraphUpToDate=<synthetic pointer>: <optimized out>, CG=..., CurSCC=..., P=

0x555557e87570, this=0x555557e87650) at ../lib/Analysis/CallGraphSCCPass.cpp:141

#7 (anonymous namespace)::CGPassManager::RunAllPassesOnSCC (DevirtualizedCall=<synthetic pointer>: <optimized out>, CG=..., CurSCC=..., this=0x555557e87650) at ../lib/Analysis/CallGraphSCCPass.cpp:442
#8 (anonymous namespace)::CGPassManager::runOnModule (this=0x555557e87650, M=...) at ../lib/Analysis/CallGraphSCCPass.cpp:498
#9 0x00007ffff4428fba in (anonymous namespace)::MPPassManager::runOnModule (M=..., this=0x555557e65c20) at ../lib/IR/LegacyPassManager.cpp:1744
#10 llvm::legacy::PassManagerImpl::run (this=0x555557e657b0, M=...) at ../lib/IR/LegacyPassManager.cpp:1857
#11 0x00007ffff61fa7a6 in ac_compile_module_to_binary (p=0x555557e65750, module=module@entry=0x555557eb65a0, binary=binary@entry=0x7fffffffc080) at /home/hakzsam/install/llvm/debug/master/include/llvm/IR/Module.h:889
#12 0x00007ffff61b6e2b in radv_llvm_per_thread_info::compile_to_memory_buffer (this=<optimized out>, binary=0x7fffffffc080, module=0x555557eb65a0) at radv_llvm_helper.cpp:97
#13 radv_compile_to_binary (info=info@entry=0x7fffffffc050, module=module@entry=0x7ffff61fa7a6 <ac_compile_module_to_binary(ac_compiler_passes*, LLVMModuleRef, ac_shader_binary*)+22>, binary=binary@entry=0x7fffffffc080)

at radv_llvm_helper.cpp:97

#14 0x00007ffff61b0d81 in ac_llvm_compile (ac_llvm=0x7fffffffc050, binary=0x7fffffffc080, M=0x7ffff61fa7a6 <ac_compile_module_to_binary(ac_compiler_passes*, LLVMModuleRef, ac_shader_binary*)+22>) at radv_nir_to_llvm.c:3660
#15 ac_compile_llvm_module (ac_llvm=ac_llvm@entry=0x7fffffffc050, llvm_module=0x7ffff61fa7a6 <ac_compile_module_to_binary(ac_compiler_passes*, LLVMModuleRef, ac_shader_binary*)+22>, binary=binary@entry=0x7fffffffc080,

config=0x7fffffffc080, config@entry=0x555557f71778, stage=MESA_SHADER_COMPUTE, options=0x7fffffffc140, shader_info=<optimized out>, shader_info=<optimized out>) at radv_nir_to_llvm.c:3684

#16 0x00007ffff61b65d0 in radv_compile_nir_shader (ac_llvm=ac_llvm@entry=0x7fffffffc050, binary=binary@entry=0x7fffffffc080, config=config@entry=0x555557f71778, shader_info=shader_info@entry=0x555557f717a0,

nir=nir@entry=0x7fffffffc388, nir_count=nir_count@entry=1, options=0x7fffffffc140) at radv_nir_to_llvm.c:3808

#17 0x00007ffff61c56db in shader_variant_create (device=device@entry=0x555557e4d920, module=0x7fffffffcd40, shaders=shaders@entry=0x7fffffffc388, shader_count=shader_count@entry=1, stage=MESA_SHADER_COMPUTE,

options=options@entry=0x7fffffffc140, gs_copy_shader=false, code_out=0x7fffffffc3b8, code_size_out=0x7fffffffc2e4) at radv_shader.c:612

#18 0x00007ffff61c5b04 in radv_shader_variant_create (device=device@entry=0x555557e4d920, module=<optimized out>, shaders=shaders@entry=0x7fffffffc388, shader_count=shader_count@entry=1, layout=<optimized out>,

key=key@entry=0x7fffffffc810, code_out=0x7fffffffc3b8, code_size_out=0x7fffffffc2e4) at radv_shader.c:666

#19 0x00007ffff61b8aa6 in radv_create_shaders (pipeline=0x555557ef5ca0, device=<optimized out>, cache=0x555557e4d998, key=<optimized out>, pStages=<optimized out>, flags=<optimized out>) at radv_pipeline.c:2151
#20 0x00007ffff61bf7eb in radv_compute_pipeline_create (pPipeline=0x555557e4ef70, pAllocator=<optimized out>, pCreateInfo=0x7fffffffcbd0, _cache=<optimized out>, _device=0x555557e4d920) at radv_pipeline.c:3787
#21 radv_CreateComputePipelines (_device=_device@entry=0x555557e4d920, pipelineCache=pipelineCache@entry=0x555557e4d998, count=count@entry=1, pCreateInfos=pCreateInfos@entry=0x7fffffffcbd0, pAllocator=pAllocator@entry=0x0,

pPipelines=pPipelines@entry=0x555557e4ef70) at radv_pipeline.c:3817

#22 0x00007ffff619653a in radv_device_init_meta_itob_state (device=0x555557e4d920) at radv_private.h:1986
#23 radv_device_init_meta_bufimage_state (device=device@entry=0x555557e4d920) at radv_meta_bufimage.c:1489
#24 0x00007ffff6175a4a in radv_device_init_meta (device=device@entry=0x555557e4d920) at radv_meta.c:365
#25 0x00007ffff61680d0 in radv_CreateDevice (physicalDevice=0x555557d7c0e0, pCreateInfo=0x7fffffffd0d0, pAllocator=<optimized out>, pDevice=0x555557d82ec0) at radv_device.c:1702
#26 0x00007ffff640c574 in ?? () from /usr/lib/libvulkan.so.1
#27 0x00007ffff641599b in ?? () from /usr/lib/libvulkan.so.1
#28 0x00007ffff6419b29 in vkCreateDevice () from /usr/lib/libvulkan.so.1
#29 0x0000555556942f7d in vk::createDevice(vk::PlatformInterface const&, vk::VkInstance_s*, vk::InstanceInterface const&, vk::VkPhysicalDevice_s*, vk::VkDeviceCreateInfo const*, vk::VkAllocationCallbacks const*) ()
#30 0x00005555558a3384 in vkt::DefaultDevice::DefaultDevice(vk::PlatformInterface const&, tcu::CommandLine const&) ()
#31 0x00005555558a40e5 in vkt::Context::Context(tcu::TestContext&, vk::PlatformInterface const&, vk::ProgramCollection<vk::ProgramBinary, vk::BinaryBuildOptions>&) ()
#32 0x000055555588c3e2 in vkt::TestCaseExecutor::TestCaseExecutor(tcu::TestContext&) ()
#33 0x000055555588c552 in vkt::TestPackage::createExecutor() const ()
#34 0x0000555556e04964 in tcu::TestSessionExecutor::iterate() ()
#35 0x0000555556dd89a9 in tcu::App::iterate() ()
#36 0x000055555587e4e8 in main ()

Can you look into it?

Thanks!

Would you be able to provide LLVM IR? I am trying to reproduce this issue locally but it is taking some time as I haven't worked with RADV yet.

Thanks.

[hakzsam]

Unfortunately, I can't get any LLVM IR because it crashes too early in the process. You can also reproduce the problem with RadeonSI btw.

Thanks for the revert!

[aakanksha555]

Added a fix for the failing deqp-vk test.

lib/Target/AMDGPU/AMDGPUAnnotateKernelFeatures.cpp

 
 using namespace llvm;
 
 namespace {
 
 class AMDGPUAnnotateKernelFeatures : public CallGraphSCCPass {
 private:
   const TargetMachine *TM = nullptr;
+  llvm::SmallVector<CallGraphNode*, 8> SortedNodeList;
 
   bool addFeatureAttributes(Function &F);
+  bool propagateUniformWorkGroupAttribute();
 
 public:
   static char ID;
 
   AMDGPUAnnotateKernelFeatures() : CallGraphSCCPass(ID) {}
 
   bool doInitialization(CallGraph &CG) override;
   bool runOnSCC(CallGraphSCC &SCC) override;

 };
 
 } // end anonymous namespace
 
 char AMDGPUAnnotateKernelFeatures::ID = 0;
 
 char &llvm::AMDGPUAnnotateKernelFeaturesID = AMDGPUAnnotateKernelFeatures::ID;
 
 INITIALIZE_PASS(AMDGPUAnnotateKernelFeatures, DEBUG_TYPE,
                 "Add AMDGPU function attributes", false, false)
 
 
 // The queue ptr is only needed when casting to flat, not from it.
 static bool castRequiresQueuePtr(unsigned SrcAS) {
   return SrcAS == AMDGPUAS::LOCAL_ADDRESS || SrcAS == AMDGPUAS::PRIVATE_ADDRESS;
 }
 
 static bool castRequiresQueuePtr(const AddrSpaceCastInst *ASC) {
   return castRequiresQueuePtr(ASC->getSrcAddressSpace());

 }
 
 static bool handleAttr(Function &Parent, const Function &Callee,
                        StringRef Name) {
   if (Callee.hasFnAttribute(Name)) {
     Parent.addFnAttr(Name);
     return true;
   }
-
   return false;
 }
 
 static void copyFeaturesToFunction(Function &Parent, const Function &Callee,
                                    bool &NeedQueuePtr) {
   // X ids unnecessarily propagated to kernels.
   static const StringRef AttrNames[] = {
     { "amdgpu-work-item-id-x" },
static void copyFeaturesToFunction(Function &Parent, const Function &Callee,
 
   if (handleAttr(Parent, Callee, "amdgpu-queue-ptr"))
     NeedQueuePtr = true;
 
   for (StringRef AttrName : AttrNames)
     handleAttr(Parent, Callee, AttrName);
 }
 
+bool AMDGPUAnnotateKernelFeatures::propagateUniformWorkGroupAttribute() {
+  bool Changed = false;
+
+  while(!SortedNodeList.empty()) {
+    CallGraphNode *Node =  SortedNodeList.pop_back_val();
+    Function *Caller = Node->getFunction();
+
+    for (auto Iter = Node->begin(), End = Node->end(); Iter != End; ++Iter) {
+      Function *Callee = std::get<1>(*Iter)->getFunction();
+
+      // Check if the Caller has the attribute
+      if (Caller->hasFnAttribute("uniform-work-group-size")) {
+        if (Callee) {
+          // Check if the value of the attribute is true
+          if (Caller->getFnAttribute("uniform-work-group-size")
+                  .getValueAsString().equals("true")) {
+            // Propagate the attribute to the Callee, if it does not have it
+            if (!Callee->hasFnAttribute("uniform-work-group-size")) {
+              Callee->addFnAttr("uniform-work-group-size", "true");
+              Changed = true;
+            }
+          } else {
+            Callee->addFnAttr("uniform-work-group-size", "false");
+            Changed = true;
+          }
+        }
+      } else {
+        // If the attribute is absent, set it as false
+        Caller->addFnAttr("uniform-work-group-size", "false");
+        if (Callee)
+          Callee->addFnAttr("uniform-work-group-size", "false");
+        Changed = true;
+      }
+    }
+  }
+  return Changed;
+}
+
 bool AMDGPUAnnotateKernelFeatures::addFeatureAttributes(Function &F) {
   const GCNSubtarget &ST = TM->getSubtarget<GCNSubtarget>(F);
   bool HasFlat = ST.hasFlatAddressSpace();
   bool HasApertureRegs = ST.hasApertureRegs();
   SmallPtrSet<const Constant *, 8> ConstantExprVisited;
 
   bool Changed = false;
   bool NeedQueuePtr = false;
   bool HaveCall = false;
   bool IsFunc = !AMDGPU::isEntryFunctionCC(F.getCallingConv());
 
   for (BasicBlock &BB : F) {
     for (Instruction &I : BB) {
       CallSite CS(&I);
       if (CS) {
         Function *Callee = CS.getCalledFunction();
 
         // TODO: Do something with indirect calls.
         if (!Callee) {
           if (!CS.isInlineAsm())
             HaveCall = true;
           continue;
         }
 
         Intrinsic::ID IID = Callee->getIntrinsicID();
         if (IID == Intrinsic::not_intrinsic) {
           HaveCall = true;
           copyFeaturesToFunction(F, *Callee, NeedQueuePtr);
           Changed = true;
         } else {
           bool NonKernelOnly = false;
           StringRef AttrName = intrinsicToAttrName(IID,
                                                    NonKernelOnly, NeedQueuePtr);
           if (!AttrName.empty() && (IsFunc || !NonKernelOnly)) {
             F.addFnAttr(AttrName);
             Changed = true;
           }
         }
       }
 
       if (NeedQueuePtr || HasApertureRegs)
         continue;
if (HasFlat && !IsFunc && HaveCall) {
     F.addFnAttr("amdgpu-flat-scratch");
     Changed = true;
   }
 
   return Changed;
 }
 
 bool AMDGPUAnnotateKernelFeatures::runOnSCC(CallGraphSCC &SCC) {
-  Module &M = SCC.getCallGraph().getModule();
-  Triple TT(M.getTargetTriple());
-
   bool Changed = false;
+ 
   for (CallGraphNode *I : SCC) {
-    Function *F = I->getFunction();
+    // Build a list of CallGraphNodes that is sorted by the number of uses
+    if (I->getNumReferences())
+      SortedNodeList.push_back(I);
+    else
+      propagateUniformWorkGroupAttribute();
+
+    Function *F = I->getFunction();    
+    // Add feature attributes
     if (!F || F->isDeclaration())
       continue;
-
     Changed |= addFeatureAttributes(*F);
   }
-
   return Changed;
+
 }
 
 bool AMDGPUAnnotateKernelFeatures::doInitialization(CallGraph &CG) {
   auto *TPC = getAnalysisIfAvailable<TargetPassConfig>();
   if (!TPC)
     report_fatal_error("TargetMachine is required");
 
   TM = &TPC->getTM<TargetMachine>();
   return false;
 }
 
 Pass *llvm::createAMDGPUAnnotateKernelFeaturesPass() {
   return new AMDGPUAnnotateKernelFeatures();
 }

lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

void AMDGPUPassConfig::addIRPasses() {
   //   %1 = shl %a, 2
   //
   // but EarlyCSE can do neither of them.
   if (getOptLevel() != CodeGenOpt::None)
     addEarlyCSEOrGVNPass();
 }
 
 void AMDGPUPassConfig::addCodeGenPrepare() {
+  if (TM->getTargetTriple().getArch() == Triple::amdgcn)
+    addPass(createAMDGPUAnnotateKernelFeaturesPass());
   if (TM->getTargetTriple().getArch() == Triple::amdgcn &&
       EnableLowerKernelArguments)
     addPass(createAMDGPULowerKernelArgumentsPass());
 
   TargetPassConfig::addCodeGenPrepare();
 
   if (EnableLoadStoreVectorizer)
     addPass(createLoadStoreVectorizerPass());

 bool GCNPassConfig::addPreISel() {
   AMDGPUPassConfig::addPreISel();
 
   if (EnableAtomicOptimizations) {
     addPass(createAMDGPUAtomicOptimizerPass());
   }
 
   // FIXME: We need to run a pass to propagate the attributes when calls are
   // supported.
-  addPass(createAMDGPUAnnotateKernelFeaturesPass());
 
   // Merge divergent exit nodes. StructurizeCFG won't recognize the multi-exit
   // regions formed by them.
   addPass(&AMDGPUUnifyDivergentExitNodesID);
   if (!LateCFGStructurize) {
     addPass(createStructurizeCFGPass(true)); // true -> SkipUniformRegions
   }
   addPass(createSinkingPass());
   addPass(createAMDGPUAnnotateUniformValues());

test/CodeGen/AMDGPU/annotate-kernel-features-hsa-call.ll

 
 ; HSA: define amdgpu_kernel void @kern_use_implicitarg_ptr() #15 {
 define amdgpu_kernel void @kern_use_implicitarg_ptr() #1 {
   %implicitarg.ptr = call i8 addrspace(4)* @llvm.amdgcn.implicitarg.ptr()
   store volatile i8 addrspace(4)* %implicitarg.ptr, i8 addrspace(4)* addrspace(1)* undef
   ret void
 }
 
-; HSA: define void @use_implicitarg_ptr() #15 {
+; HSA: define void @use_implicitarg_ptr() #16 {
 define void @use_implicitarg_ptr() #1 {
   %implicitarg.ptr = call i8 addrspace(4)* @llvm.amdgcn.implicitarg.ptr()
   store volatile i8 addrspace(4)* %implicitarg.ptr, i8 addrspace(4)* addrspace(1)* undef
   ret void
 }
 
-; HSA: define void @func_indirect_use_implicitarg_ptr() #15 {
+; HSA: define void @func_indirect_use_implicitarg_ptr() #16 {
 define void @func_indirect_use_implicitarg_ptr() #1 {
   call void @use_implicitarg_ptr()
   ret void
 }
 
-; HSA: declare void @external.func() #16
+; HSA: declare void @external.func() #17
 declare void @external.func() #3
 
-; HSA: define internal void @defined.func() #16 {
+; HSA: define internal void @defined.func() #17 {
 define internal void @defined.func() #3 {
   ret void
 }
 
-; HSA: define void @func_call_external() #16 {
+; HSA: define void @func_call_external() #17 {
 define void @func_call_external() #3 {
   call void @external.func()
   ret void
 }
 
-; HSA: define void @func_call_defined() #16 {
+; HSA: define void @func_call_defined() #17 {
 define void @func_call_defined() #3 {
   call void @defined.func()
   ret void
 }
 
-; HSA: define void @func_call_asm() #16 {
+; HSA: define void @func_call_asm() #17 {
 define void @func_call_asm() #3 {
   call void asm sideeffect "", ""() #3
   ret void
 }
 
-; HSA: define amdgpu_kernel void @kern_call_external() #17 {
+; HSA: define amdgpu_kernel void @kern_call_external() #18 {
 define amdgpu_kernel void @kern_call_external() #3 {
   call void @external.func()
   ret void
 }
 
-; HSA: define amdgpu_kernel void @func_kern_defined() #17 {
+; HSA: define amdgpu_kernel void @func_kern_defined() #18 {
 define amdgpu_kernel void @func_kern_defined() #3 {
   call void @defined.func()
   ret void
 }
 
 attributes #0 = { nounwind readnone speculatable }
 attributes #1 = { nounwind "target-cpu"="fiji" }
 attributes #2 = { nounwind "target-cpu"="gfx900" }
 attributes #3 = { nounwind }
 
 ; HSA: attributes #0 = { nounwind readnone speculatable }
-; HSA: attributes #1 = { nounwind "amdgpu-work-item-id-x" "target-cpu"="fiji" }
-; HSA: attributes #2 = { nounwind "amdgpu-work-item-id-y" "target-cpu"="fiji" }
-; HSA: attributes #3 = { nounwind "amdgpu-work-item-id-z" "target-cpu"="fiji" }
-; HSA: attributes #4 = { nounwind "amdgpu-work-group-id-x" "target-cpu"="fiji" }
-; HSA: attributes #5 = { nounwind "amdgpu-work-group-id-y" "target-cpu"="fiji" }
-; HSA: attributes #6 = { nounwind "amdgpu-work-group-id-z" "target-cpu"="fiji" }
-; HSA: attributes #7 = { nounwind "amdgpu-dispatch-ptr" "target-cpu"="fiji" }
-; HSA: attributes #8 = { nounwind "amdgpu-queue-ptr" "target-cpu"="fiji" }
-; HSA: attributes #9 = { nounwind "amdgpu-dispatch-id" "target-cpu"="fiji" }
+; HSA: attributes #1 = { nounwind "amdgpu-work-item-id-x" "target-cpu"="fiji" "uniform-work-group-size"="false" }
+; HSA: attributes #2 = { nounwind "amdgpu-work-item-id-y" "target-cpu"="fiji" "uniform-work-group-size"="false" }
+; HSA: attributes #3 = { nounwind "amdgpu-work-item-id-z" "target-cpu"="fiji" "uniform-work-group-size"="false" }
+; HSA: attributes #4 = { nounwind "amdgpu-work-group-id-x" "target-cpu"="fiji" "uniform-work-group-size"="false" }
+; HSA: attributes #5 = { nounwind "amdgpu-work-group-id-y" "target-cpu"="fiji" "uniform-work-group-size"="false" }
+; HSA: attributes #6 = { nounwind "amdgpu-work-group-id-z" "target-cpu"="fiji" "uniform-work-group-size"="false" }
+; HSA: attributes #7 = { nounwind "amdgpu-dispatch-ptr" "target-cpu"="fiji" "uniform-work-group-size"="false" }
+; HSA: attributes #8 = { nounwind "amdgpu-queue-ptr" "target-cpu"="fiji" "uniform-work-group-size"="false" }
+; HSA: attributes #9 = { nounwind "amdgpu-dispatch-id" "target-cpu"="fiji" "uniform-work-group-size"="false" }
 ; HSA: attributes #10 = { nounwind "amdgpu-work-group-id-y" "amdgpu-work-group-id-z" "target-cpu"="fiji" }
-; HSA: attributes #11 = { nounwind "target-cpu"="fiji" }
-; HSA: attributes #12 = { nounwind "target-cpu"="gfx900" }
-; HSA: attributes #13 = { nounwind "amdgpu-queue-ptr" "target-cpu"="gfx900" }
-; HSA: attributes #14 = { nounwind "amdgpu-kernarg-segment-ptr" "target-cpu"="fiji" }
+; HSA: attributes #11 = { nounwind "target-cpu"="fiji" "uniform-work-group-size"="false" }
+; HSA: attributes #12 = { nounwind "target-cpu"="gfx900" "uniform-work-group-size"="false" }
+; HSA: attributes #13 = { nounwind "amdgpu-queue-ptr" "target-cpu"="gfx900" "uniform-work-group-size"="false" }
+; HSA: attributes #14 = { nounwind "amdgpu-kernarg-segment-ptr" "target-cpu"="fiji" "uniform-work-group-size"="false" }
 ; HSA: attributes #15 = { nounwind "amdgpu-implicitarg-ptr" "target-cpu"="fiji" }
-; HSA: attributes #16 = { nounwind }
-; HSA: attributes #17 = { nounwind "amdgpu-flat-scratch" }
+; HSA: attributes #16 = { nounwind "amdgpu-implicitarg-ptr" "target-cpu"="fiji" "uniform-work-group-size"="false" }
+; HSA: attributes #17 = { nounwind "uniform-work-group-size"="false" }

test/CodeGen/AMDGPU/uniform-work-group-attribute-missing.ll

+; RUN: opt -S -mtriple=amdgcn-amd- -amdgpu-annotate-kernel-features %s | FileCheck %s 
+
+; Test 1
+; "If the kernel does not have the uniform-work-group-attribute, set both callee and caller as false"
+; CHECK: define void @foo() #[[FOO:[0-9]+]] {
+define void @foo() #0 {
+  ret void
+}
+
+; CHECK: define amdgpu_kernel void @kernel1() #[[FOO]] {
+define amdgpu_kernel void @kernel1() #1 {
+  call void @foo()
+  ret void
+}
+
+attributes #0 = { "uniform-work-group-size"="true" }
+
+; CHECK: attributes #[[FOO]] = { "uniform-work-group-size"="false" }

test/CodeGen/AMDGPU/uniform-work-group-nested-function-calls.ll

+; RUN: opt -S -mtriple=amdgcn-amd- -amdgpu-annotate-kernel-features %s | FileCheck %s 
+
+; Test 3
+; "Test to verify if the attribute gets propagated across nested function calls"
+; CHECK: define void @func1() #[[FUNC:[0-9]+]] {
+define void @func1() #0 {
+  ret void
+}
+
+; CHECK: define void @func2() #[[FUNC]] {
+define void @func2() #1 {
+  call void @func1()
+  ret void
+}
+
+; CHECK: define amdgpu_kernel void @kernel3() #[[FUNC:[0-9]+]] {
+define amdgpu_kernel void @kernel3() #2 {
+  call void @func2()
+  ret void
+}
+
+attributes #2 = { "uniform-work-group-size"="true" }
+
+; CHECK: attributes #[[FUNC]] = { "uniform-work-group-size"="true" }

test/CodeGen/AMDGPU/uniform-work-group-prevent-attribute-propagation.ll

+; RUN: opt -S -mtriple=amdgcn-amd- -amdgpu-annotate-kernel-features %s | FileCheck %s 
+
+; Test 5
+; "Two kernels with different values of the uniform-work-group-attribute call the same function"
+; CHECK: define void @func() #[[FUNC:[0-9]+]] {
+define void @func() #0 {
+  ret void
+}
+
+; CHECK: define amdgpu_kernel void @kernel1() #[[KERNEL1:[0-9]+]] {
+define amdgpu_kernel void @kernel1() #1 {
+  call void @func()
+  ret void
+}
+
+; CHECK: define amdgpu_kernel void @kernel2() #[[FUNC]] {
+define amdgpu_kernel void @kernel2() #2 {
+  call void @func()
+  ret void
+}
+
+attributes #1 = { "uniform-work-group-size"="true" }
+
+; CHECK: attributes #[[FUNC]] = { "uniform-work-group-size"="false" }
+; CHECK: attributes #[[KERNEL1]] = { "uniform-work-group-size"="true" }

test/CodeGen/AMDGPU/uniform-work-group-propagate-attribute.ll

+; RUN: opt -S -mtriple=amdgcn-amd- -amdgpu-annotate-kernel-features %s | FileCheck %s 
+
+; Test 2
+; "Propagate the uniform-work-group-attribute from the kernel to callee if it doesn't have it"
+; CHECK: define void @func() #[[FUNC:[0-9]+]] {
+define void @func() #0 {
+  ret void
+}
+
+; CHECK: define amdgpu_kernel void @kernel() #[[KERNEL:[0-9]+]] {
+define amdgpu_kernel void @kernel() #1 {
+  call void @func()
+  ret void
+}
+
+attributes #0 = { nounwind }
+attributes #1 = { "uniform-work-group-size"="false" }
+
+; CHECK: attributes #[[FUNC]] = { nounwind "uniform-work-group-size"="false" }
+; CHECK: attributes #[[KERNEL]] = { "uniform-work-group-size"="false" }

test/CodeGen/AMDGPU/uniform-work-group-resursion-test.ll

+; RUN: opt -S -mtriple=amdgcn-amd- -amdgpu-annotate-kernel-features %s | FileCheck %s 
+
+; Test 5
+; Test to ensure recursive functions exhibit proper behaviour
+; Add all numbers below n
+
+; CHECK: define i32 @fib(i32 %n) #[[FOO:[0-9]+]] {
+define i32 @fib(i32 %n) #0 {
+  %cmp1 = icmp eq i32 %n, 0
+  br i1 %cmp1, label %exit, label %cont1
+
+cont1:
+  %cmp2 = icmp eq i32 %n, 1
+  br i1 %cmp2, label %exit, label %cont2
+
+cont2:
+  %nm1 = sub i32 %n, 1
+  %fibm1 = call i32 @fib(i32 %nm1)
+  %nm2 = sub i32 %n, 2
+  %fibm2 = call i32 @fib(i32 %nm2)
+  %retval = add i32 %fibm1, %fibm2
+
+  ret i32 %retval
+
+exit:
+  ret i32 1
+
+}
+
+; CHECK: define amdgpu_kernel void @kernel1(i32 addrspace(1)* %m) #[[FOO]] {
+define amdgpu_kernel void @kernel1(i32 addrspace(1)* %m) #1 {
+  %r = call i32 @fib(i32 5)
+  store i32 %r, i32 addrspace(1)* %m
+  ret void
+}
+
+attributes #1 = { "uniform-work-group-size"="true" }
+
+; CHECK: attributes #[[FOO]] = { "uniform-work-group-size"="true" }

test/CodeGen/AMDGPU/uniform-work-group-test.ll

+; RUN: opt -S -mtriple=amdgcn-amd- -amdgpu-annotate-kernel-features %s | FileCheck %s 
+
+; Test 3
+; "Test to verify if the attribute gets propagated across nested function calls"
+; CHECK: define void @func1() #[[FUNC:[0-9]+]] {
+define void @func1() #0 {
+  ret void
+}
+; CHECK: define void @func4() #[[FUNC]] {
+define void @func4() #1 {
+  ret void
+}
+
+; CHECK: define void @func2() #[[FUNC]] {
+define void @func2() #1 {
+  call void @func4()
+  call void @func1()
+  ret void
+}
+
+; CHECK: define void @func3() #[[FUNC]] {
+define void @func3() #1 {
+  call void @func1()
+  ret void
+}
+
+; CHECK: define amdgpu_kernel void @kernel3() #[[FUNC:[0-9]+]] {
+define amdgpu_kernel void @kernel3() #2 {
+  call void @func2()
+  call void @func3()
+  ret void
+}
+
+attributes #2 = { "uniform-work-group-size"="true" }
+
+; CHECK: attributes #[[FUNC]] = { "uniform-work-group-size"="true" }