[AMDGPU] Do not assume stack size for PAL code object indirect calls

[bsaleil]

Summary

There is no need to set a big default stack size for PAL code object indirect calls. The driver knows the max recursion depth, so it can compute a more accurate value from the minimum scratch size.

[arsenm]

Is

llvm/lib/Target/AMDGPU/AMDGPUResourceUsageAnalysis.cpp

122–124

Dynamic objects should be treated identically

llvm/test/CodeGen/AMDGPU/resource-usage-pal.ll

2

Can you just add a run line to an existing test?

[sebastian-ne]

llvm/lib/Target/AMDGPU/AMDGPUResourceUsageAnalysis.cpp

121–123

Should use explicit brackets around the && clause.

122–124

I think the logic here is that we treat PAL code objects the same as AMDHSL code objects >= 5 (PAL code objects have no amdgpu_code_object_version).
If there is an indirect call, something outside the compiler needs to compute the stack size, in the case of PAL/graphics, that is the driver.

So probably we should use if (AMDGPU::getCodeObjectVersion(M) >= AMDGPU::AMDHSA_COV5 || STI.getTargetTriple().getOS() == Triple::AMDPAL) for both, AssumedStackSizeForDynamicSizeObjects and AssumedStackSizeForExternalCall

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPUResourceUsageAnalysis.cpp

122–124

Is some equivalent to the dynamic bit in the metadata emitted for pal?

[sebastian-ne]

llvm/lib/Target/AMDGPU/AMDGPUResourceUsageAnalysis.cpp

122–124

We emit per-function metadata (https://llvm.org/docs/AMDGPUUsage.html#amdgpu-amdpal-code-object-shader-function-map-table).
An application needs to specify the recursion depth it uses and the driver uses this information to compute the needed scratch allocation.

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPUResourceUsageAnalysis.cpp

122–124

That metadata should be checked in the test

[bsaleil]

Addressed review comments

[arsenm]

llvm/test/CodeGen/AMDGPU/resource-usage-pal.ll

6

Doesn't check a dynamic-is-present metadata field?

[bsaleil]

Add comment to the test case

llvm/test/CodeGen/AMDGPU/resource-usage-pal.ll

6

We don't have such flag in PAL abi.

[arsenm]

llvm/test/CodeGen/AMDGPU/resource-usage-pal.ll

6

Then that is a problem? You would need one to know you need to add some extra?

[sebastian-ne]

llvm/test/CodeGen/AMDGPU/resource-usage-pal.ll

6

The amount of scratch that needs to be allocated is computed outside of LLVM in the graphics driver. The compute equivalent would be the linker/loader that sees all the functions that are linked together and also gets additional data like the maximum recursion depth.
So, the only information needed in PAL metadata is the scratch usage of a function itself, without any callees.

[bsaleil]

This patch is causing buildbot failures because the RUN line of the test case is invalid (missing colon symbol). This is now fixed by 9eea63bc9ccf5bd18a040cc028238ac4b49b77ea

llvm/lib/Target/AMDGPU/AMDGPUResourceUsageAnalysis.cpp

 
 bool AMDGPUResourceUsageAnalysis::runOnModule(Module &M) {
   auto *TPC = getAnalysisIfAvailable<TargetPassConfig>();
   if (!TPC)
     return false;
 
   MachineModuleInfo &MMI = getAnalysis<MachineModuleInfoWrapperPass>().getMMI();
   const TargetMachine &TM = TPC->getTM<TargetMachine>();
+  const MCSubtargetInfo &STI = *TM.getMCSubtargetInfo();
   bool HasIndirectCall = false;
 
   CallGraph CG = CallGraph(M);
   auto End = po_end(&CG);
 
   // By default, for code object v5 and later, track only the minimum scratch
   // size
-  if (AMDGPU::getCodeObjectVersion(M) >= AMDGPU::AMDHSA_COV5) {
+  if (AMDGPU::getCodeObjectVersion(M) >= AMDGPU::AMDHSA_COV5 ||
+      STI.getTargetTriple().getOS() == Triple::AMDPAL) {
     if (!AssumedStackSizeForDynamicSizeObjects.getNumOccurrences())
       AssumedStackSizeForDynamicSizeObjects = 0;
     if (!AssumedStackSizeForExternalCall.getNumOccurrences())
       AssumedStackSizeForExternalCall = 0;
   }
 
   for (auto IT = po_begin(&CG); IT != End; ++IT) {
     Function *F = IT->getFunction();
     if (!F || F->isDeclaration())
       continue;
 
     MachineFunction *MF = MMI.getMachineFunction(*F);
     assert(MF && "function must have been generated already");
 
     auto CI =
         CallGraphResourceInfo.insert(std::pair(F, SIFunctionResourceInfo()));

llvm/test/CodeGen/AMDGPU/resource-usage-pal.ll

+; RUN llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1100 -verify-machineinstrs < %s | FileCheck %s
+
+; Check that we do not assume any default stack size for PAL code object
+; indirect calls. The driver knows the max recursion depth, so it can compute
+; a more accurate value.
+
+; CHECK: ScratchSize: 0
+; CHECK: scratch_memory_size: 0
+define amdgpu_vs void @test() {
+.entry:
+  %0 = call i64 @llvm.amdgcn.s.getpc()
+  %1 = inttoptr i64 %0 to ptr
+  call amdgpu_gfx void %1()
+  ret void
+}
+
+declare i64 @llvm.amdgcn.s.getpc()