AMDGPU: Account for implicit argument alignment for kernarg segment

If a kernel had no formal arguments but did have the implicit
arguments, we were reporting a required kernarg alignment of 4. For
some reason we require an 8-byte alignment for this, even though
there's no real advantage and I don't see where this is documented in
the ABI.

The code object header code also claims the minimum alignment is 16,
which is what I thought you always got at runtime anyway so I don't
know why this matters.

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp b/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp
index ecb6008..bb2e723 100644 (file)
--- a/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp
@@ -1092,6 +1092,7 @@ void AMDGPUAsmPrinter::getAmdKernelCode(amd_kernel_code_t &Out,
 
   // kernarg_segment_alignment is specified as log of the alignment.
   // The minimum alignment is 16.
+  // FIXME: The metadata treats the minimum as 4?
   Out.kernarg_segment_alignment = Log2(std::max(Align(16), MaxKernArgAlign));
 }
 

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUHSAMetadataStreamer.cpp b/llvm/lib/Target/AMDGPU/AMDGPUHSAMetadataStreamer.cpp
index 3b4f92f..b9c59f4 100644 (file)
--- a/llvm/lib/Target/AMDGPU/AMDGPUHSAMetadataStreamer.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUHSAMetadataStreamer.cpp
@@ -201,10 +201,11 @@ MetadataStreamerV2::getHSACodeProps(const MachineFunction &MF,
   Align MaxKernArgAlign;
   HSACodeProps.mKernargSegmentSize = STM.getKernArgSegmentSize(F,
                                                                MaxKernArgAlign);
+  HSACodeProps.mKernargSegmentAlign =
+    std::max(MaxKernArgAlign, Align(4)).value();
+
   HSACodeProps.mGroupSegmentFixedSize = ProgramInfo.LDSSize;
   HSACodeProps.mPrivateSegmentFixedSize = ProgramInfo.ScratchSize;
-  HSACodeProps.mKernargSegmentAlign =
-      std::max(MaxKernArgAlign, Align(4)).value();
   HSACodeProps.mWavefrontSize = STM.getWavefrontSize();
   HSACodeProps.mNumSGPRs = ProgramInfo.NumSGPR;
   HSACodeProps.mNumVGPRs = ProgramInfo.NumVGPR;
@@ -867,6 +868,8 @@ MetadataStreamerV3::getHSAKernelProps(const MachineFunction &MF,
       Kern.getDocument()->getNode(ProgramInfo.LDSSize);
   Kern[".private_segment_fixed_size"] =
       Kern.getDocument()->getNode(ProgramInfo.ScratchSize);
+
+  // FIXME: The metadata treats the minimum as 16?
   Kern[".kernarg_segment_align"] =
       Kern.getDocument()->getNode(std::max(Align(4), MaxKernArgAlign).value());
   Kern[".wavefront_size"] =

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
index d19431a..0655b43 100644 (file)
--- a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
@@ -688,6 +688,7 @@ unsigned AMDGPUSubtarget::getKernArgSegmentSize(const Function &F,
   if (ImplicitBytes != 0) {
     const Align Alignment = getAlignmentForImplicitArgPtr();
     TotalSize = alignTo(ExplicitArgBytes, Alignment) + ImplicitBytes;
+    MaxAlign = std::max(MaxAlign, Alignment);
   }
 
   // Being able to dereference past the end is useful for emitting scalar loads.

diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.implicitarg.ptr.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.implicitarg.ptr.ll
index 7431832..e3ad549 100644 (file)
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.implicitarg.ptr.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.implicitarg.ptr.ll
@@ -281,7 +281,7 @@ define amdgpu_kernel void @kernel_implicitarg_no_struct_align_padding(<16 x i32>
 
 ; HSA-LABEL: - Name:            opencl_kernel_implicitarg_ptr_empty
 ; HSA: KernargSegmentSize: 48
-; HSA: KernargSegmentAlign: 4
+; HSA: KernargSegmentAlign: 8
 
 ; HSA-LABEL: - Name:            kernel_implicitarg_ptr
 ; HSA: KernargSegmentSize: 112
@@ -289,7 +289,7 @@ define amdgpu_kernel void @kernel_implicitarg_no_struct_align_padding(<16 x i32>
 
 ; HSA-LABEL: - Name:            opencl_kernel_implicitarg_ptr
 ; HSA: KernargSegmentSize: 160
-; HSA: KernargSegmentAlign: 4
+; HSA: KernargSegmentAlign: 8
 
 ; HSA-LABEL: - Name:            kernel_call_implicitarg_ptr_func_empty
 ; HSA: KernargSegmentSize: 0
@@ -301,7 +301,7 @@ define amdgpu_kernel void @kernel_implicitarg_no_struct_align_padding(<16 x i32>
 
 ; HSA-LABEL:  - Name:            opencl_kernel_call_implicitarg_ptr_func_empty
 ; HSA: KernargSegmentSize: 48
-; HSA: KernargSegmentAlign: 4
+; HSA: KernargSegmentAlign: 8
 
 ; HSA-LABEL:  - Name:            kernel_call_implicitarg_ptr_func
 ; HSA: KernargSegmentSize: 112
@@ -309,7 +309,7 @@ define amdgpu_kernel void @kernel_implicitarg_no_struct_align_padding(<16 x i32>
 
 ; HSA-LABEL:  - Name:            opencl_kernel_call_implicitarg_ptr_func
 ; HSA: KernargSegmentSize: 160
-; HSA: KernargSegmentAlign: 4
+; HSA: KernargSegmentAlign: 8
 
 ; HSA-LABEL: - Name:            kernel_call_kernarg_implicitarg_ptr_func
 ; HSA: KernargSegmentSize: 112