[AMDGPU] Kernel arg metadata: added support for "__hip_texture" type.

[kpyzhov]

Summary

Added support for HIP texture kernel arguments to the MetadataStreamer.
For pointers to "__hip_texture", return ValueKind::Image.

[kzhuravl]

Does this need to be done for v3 as well? Needs test?

[kzhuravl]

Test for v3?

[kzhuravl]

LGTM, thanks!

[yaxunl]

LGTM

[rampitec]

Both tests fail with compiler crashing. Must always check hasName() before using getName().
I believe this needs to be reverted, this change will not work anyway if name was stripped, which llvm does.

[rampitec]

llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -mattr=-code-object-v3 -amdgpu-dump-hsa-metadata -amdgpu-verify-hsa-metadata -filetype=obj -o - < test/CodeGen/AMDGPU/hsa-metadata-from-llvm-ir-full.ll
llc: /home/stas/work/llvm/llvm/lib/IR/Type.cpp:500: llvm::StringRef llvm::StructType::getName() const: Assertion `!isLiteral() && "Literal structs never have names"' failed.
Stack dump:
0.	Program arguments: /home/stas/work/llvm/bin/bin/llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -mattr=-code-object-v3 -amdgpu-dump-hsa-metadata -amdgpu-verify-hsa-metadata -filetype=obj -o - 
1.	Running pass 'CallGraph Pass Manager' on module '<stdin>'.
2.	Running pass 'AMDGPU Assembly Printer' on function '@test_pointee_align'
 #0 0x0000000005ae8819 llvm::sys::PrintStackTrace(llvm::raw_ostream&) /home/stas/work/llvm/llvm/lib/Support/Unix/Signals.inc:494:11
 #1 0x0000000005ae89c9 PrintStackTraceSignalHandler(void*) /home/stas/work/llvm/llvm/lib/Support/Unix/Signals.inc:555:1
 #2 0x0000000005ae72d6 llvm::sys::RunSignalHandlers() /home/stas/work/llvm/llvm/lib/Support/Signals.cpp:67:5
 #3 0x0000000005ae906b SignalHandler(int) /home/stas/work/llvm/llvm/lib/Support/Unix/Signals.inc:357:1
 #4 0x00007f16e04bdf40 __restore_rt (/lib/x86_64-linux-gnu/libpthread.so.0+0x13f40)
 #5 0x00007f16dff5bed7 raise /build/glibc-KRRWSm/glibc-2.29/signal/../sysdeps/unix/sysv/linux/raise.c:51:1
 #6 0x00007f16dff3d535 abort /build/glibc-KRRWSm/glibc-2.29/stdlib/abort.c:81:7
 #7 0x00007f16dff3d40f _nl_load_domain /build/glibc-KRRWSm/glibc-2.29/intl/loadmsgcat.c:1177:9
 #8 0x00007f16dff4d012 (/lib/x86_64-linux-gnu/libc.so.6+0x35012)
 #9 0x00000000051a5526 llvm::StructType::getName() const /home/stas/work/llvm/llvm/lib/IR/Type.cpp:0:3
#10 0x0000000003308a8d llvm::Type::getStructName() const /home/stas/work/llvm/llvm/include/llvm/IR/DerivedTypes.h:357:34
#11 0x0000000003300804 llvm::AMDGPU::HSAMD::MetadataStreamerV2::getValueKind(llvm::Type*, llvm::StringRef, llvm::StringRef) const /home/stas/work/llvm/llvm/lib/Target/AMDGPU/AMDGPUHSAMetadataStreamer.cpp:108:36
#12 0x0000000003302201 llvm::AMDGPU::HSAMD::MetadataStreamerV2::emitKernelArg(llvm::Argument const&) /home/stas/work/llvm/llvm/lib/Target/AMDGPU/AMDGPUHSAMetadataStreamer.cpp:362:17
#13 0x0000000003301bfe llvm::AMDGPU::HSAMD::MetadataStreamerV2::emitKernelArgs(llvm::Function const&) /home/stas/work/llvm/llvm/lib/Target/AMDGPU/AMDGPUHSAMetadataStreamer.cpp:306:18
#14 0x0000000003303742 llvm::AMDGPU::HSAMD::MetadataStreamerV2::emitKernel(llvm::MachineFunction const&, llvm::SIProgramInfo const&) /home/stas/work/llvm/llvm/lib/Target/AMDGPU/AMDGPUHSAMetadataStreamer.cpp:481:3
#15 0x00000000032f6619 llvm::AMDGPUAsmPrinter::EmitFunctionBodyStart() /home/stas/work/llvm/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp:209:1
#16 0x00000000048ec9c4 llvm::AsmPrinter::EmitFunctionBody() /home/stas/work/llvm/llvm/lib/CodeGen/AsmPrinter/AsmPrinter.cpp:1018:3
#17 0x00000000032f8085 llvm::AMDGPUAsmPrinter::runOnMachineFunction(llvm::MachineFunction&) /home/stas/work/llvm/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp:465:3
#18 0x0000000004c0c957 llvm::MachineFunctionPass::runOnFunction(llvm::Function&) /home/stas/work/llvm/llvm/lib/CodeGen/MachineFunctionPass.cpp:73:8
#19 0x0000000005110109 llvm::FPPassManager::runOnFunction(llvm::Function&) /home/stas/work/llvm/llvm/lib/IR/LegacyPassManager.cpp:1648:23
#20 0x00000000046430ab (anonymous namespace)::CGPassManager::RunPassOnSCC(llvm::Pass*, llvm::CallGraphSCC&, llvm::CallGraph&, bool&, bool&) /home/stas/work/llvm/llvm/lib/Analysis/CallGraphSCCPass.cpp:176:20
#21 0x00000000046429dc (anonymous namespace)::CGPassManager::RunAllPassesOnSCC(llvm::CallGraphSCC&, llvm::CallGraph&, bool&) /home/stas/work/llvm/llvm/lib/Analysis/CallGraphSCCPass.cpp:441:16
#22 0x0000000004642341 (anonymous namespace)::CGPassManager::runOnModule(llvm::Module&) /home/stas/work/llvm/llvm/lib/Analysis/CallGraphSCCPass.cpp:497:18
#23 0x0000000005110cd5 (anonymous namespace)::MPPassManager::runOnModule(llvm::Module&) /home/stas/work/llvm/llvm/lib/IR/LegacyPassManager.cpp:1752:23
#24 0x00000000051107f5 llvm::legacy::PassManagerImpl::run(llvm::Module&) /home/stas/work/llvm/llvm/lib/IR/LegacyPassManager.cpp:1865:16
#25 0x0000000005111251 llvm::legacy::PassManager::run(llvm::Module&) /home/stas/work/llvm/llvm/lib/IR/LegacyPassManager.cpp:1896:3
#26 0x0000000002e3e250 compileModule(char**, llvm::LLVMContext&) /home/stas/work/llvm/llvm/tools/llc/llc.cpp:604:41
#27 0x0000000002e3c8a5 main /home/stas/work/llvm/llvm/tools/llc/llc.cpp:355:13
#28 0x00007f16dff3eb6b __libc_start_main /build/glibc-KRRWSm/glibc-2.29/csu/../csu/libc-start.c:342:3
#29 0x0000000002e3c02a _start (/home/stas/work/llvm/bin/bin/llc+0x2e3c02a)
Aborted (core dumped)

[arsenm]

llvm/trunk/lib/Target/AMDGPU/AMDGPUHSAMetadataStreamer.cpp

107–109 ↗

(On Diff #207856)

You cannot rely on the pointer element type, or the name of IR types

[jdoerfert]

In D63850#1569264, @rampitec wrote:

Both tests fail with compiler crashing. Must always check hasName() before using getName().
I believe this needs to be reverted, this change will not work anyway if name was stripped, which llvm does.

Especially Release builds strip all unnecessary names.

llvm/lib/Target/AMDGPU/AMDGPUHSAMetadataStreamer.cpp

MetadataStreamerV2::getAddressSpaceQualifier(
   }
 }
 
 ValueKind MetadataStreamerV2::getValueKind(Type *Ty, StringRef TypeQual,
                                            StringRef BaseTypeName) const {
   if (TypeQual.find("pipe") != StringRef::npos)
     return ValueKind::Pipe;
 
+  if (isa<PointerType>(Ty) && Ty->getPointerElementType()->isStructTy() &&
+      Ty->getPointerElementType()->getStructName() == "struct.__hip_texture") {
+    return ValueKind::Image;
+  }
+
   return StringSwitch<ValueKind>(BaseTypeName)
              .Case("image1d_t", ValueKind::Image)
              .Case("image1d_array_t", ValueKind::Image)
              .Case("image1d_buffer_t", ValueKind::Image)
              .Case("image2d_t", ValueKind::Image)
              .Case("image2d_array_t", ValueKind::Image)
              .Case("image2d_array_depth_t", ValueKind::Image)
              .Case("image2d_array_msaa_t", ValueKind::Image)
MetadataStreamerV3::getAddressSpaceQualifier(unsigned AddressSpace) const {
   }
 }
 
 StringRef MetadataStreamerV3::getValueKind(Type *Ty, StringRef TypeQual,
                                            StringRef BaseTypeName) const {
   if (TypeQual.find("pipe") != StringRef::npos)
     return "pipe";
 
+  if (isa<PointerType>(Ty) && Ty->getPointerElementType()->isStructTy() &&
+      Ty->getPointerElementType()->getStructName() == "struct.__hip_texture") {
+    return "image";
+  }
+
   return StringSwitch<StringRef>(BaseTypeName)
       .Case("image1d_t", "image")
       .Case("image1d_array_t", "image")
       .Case("image1d_buffer_t", "image")
       .Case("image2d_t", "image")
       .Case("image2d_array_t", "image")
       .Case("image2d_array_depth_t", "image")
       .Case("image2d_array_msaa_t", "image")

llvm/test/CodeGen/AMDGPU/hsa-metadata-from-llvm-ir-full.ll

 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 -mattr=-code-object-v3 -filetype=obj -o - < %s | llvm-readelf --notes | FileCheck --check-prefix=CHECK --check-prefix=GFX700 --check-prefix=NOTES %s
 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx802 -mattr=-code-object-v3 -filetype=obj -o - < %s | llvm-readelf --notes | FileCheck --check-prefix=CHECK --check-prefix=GFX802 --check-prefix=NOTES %s
 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -mattr=-code-object-v3 -filetype=obj -o - < %s | llvm-readelf --notes | FileCheck --check-prefix=CHECK --check-prefix=GFX900 --check-prefix=NOTES %s
 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx700 -mattr=-code-object-v3 -amdgpu-dump-hsa-metadata -amdgpu-verify-hsa-metadata -filetype=obj -o - < %s 2>&1 | FileCheck --check-prefix=PARSER %s
 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx802 -mattr=-code-object-v3 -amdgpu-dump-hsa-metadata -amdgpu-verify-hsa-metadata -filetype=obj -o - < %s 2>&1 | FileCheck --check-prefix=PARSER %s
 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -mattr=-code-object-v3 -amdgpu-dump-hsa-metadata -amdgpu-verify-hsa-metadata -filetype=obj -o - < %s 2>&1 | FileCheck --check-prefix=PARSER %s
 
 %struct.A = type { i8, float }
 %opencl.image1d_t = type opaque
 %opencl.image2d_t = type opaque
 %opencl.image3d_t = type opaque
 %opencl.queue_t = type opaque
 %opencl.pipe_t = type opaque
 %struct.B = type { i32 addrspace(1)*}
 %opencl.clk_event_t = type opaque
+%struct.__hip_texture = type opaque
 
 @__test_block_invoke_kernel_runtime_handle = external addrspace(1) externally_initialized constant i8 addrspace(1)*
 
 ; CHECK: ---
 ; CHECK:  Version: [ 1, 0 ]
 ; CHECK:  Printf:
 ; CHECK:    - '1:1:4:%d\n'
 ; CHECK:    - '2:1:8:%g\n'

 ; CHECK-NEXT: Size:            8
 ; CHECK-NEXT: Align:           8
 ; CHECK-NEXT: ValueKind:       GlobalBuffer
 ; CHECK-NEXT: ValueType:       I32
 define amdgpu_kernel void @unknown_addrspace_kernarg(i32 addrspace(12345)* %ptr) #0 {
   ret void
 }
 
+; CHECK: - Name:            _Z5EmptyP13__hip_texture
+; CHECK: Args:
+; CHECK-NEXT:      - Size:            8
+; CHECK-NEXT:        Align:           8
+; CHECK-NEXT:        ValueKind:       Image
+; CHECK-NEXT:        ValueType:       Struct
+define protected amdgpu_kernel void @_Z5EmptyP13__hip_texture(%struct.__hip_texture* nocapture) local_unnamed_addr #1 {
+  ret void
+}
+
 attributes #0 = { "amdgpu-implicitarg-num-bytes"="48" }
 attributes #1 = { "amdgpu-implicitarg-num-bytes"="48" "runtime-handle"="__test_block_invoke_kernel_runtime_handle" }
 attributes #2 = { "amdgpu-implicitarg-num-bytes"="48" "calls-enqueue-kernel" }
 
 !llvm.printf.fmts = !{!100, !101}
 
 !1 = !{i32 0}
 !2 = !{!"none"}