AMDGPU: Add an option -disable-promote-alloca-to-lds

[yaxunl]

Summary

Add this option for debugging and providing workaround.

By default it is off so no behavior change in backend.

[arsenm]

We already have a flag disable-promote-alloca-to-vector inside the pass itself. It would probably be more useful to add another corresponding one for the LDS promotion part if that's useful

[yaxunl]

Revised by Matt's comments.

[arsenm]

Needs test

[yaxunl]

add a test

[arsenm]

LGTM

lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp

 
 namespace {
 
 static cl::opt<bool> DisablePromoteAllocaToVector(
   "disable-promote-alloca-to-vector",
   cl::desc("Disable promote alloca to vector"),
   cl::init(false));
 
+static cl::opt<bool> DisablePromoteAllocaToLDS(
+  "disable-promote-alloca-to-lds",
+  cl::desc("Disable promote alloca to LDS"),
+  cl::init(false));
+
 // FIXME: This can create globals so should be a module pass.
 class AMDGPUPromoteAlloca : public FunctionPass {
 private:
   const TargetMachine *TM;
   Module *Mod = nullptr;
   const DataLayout *DL = nullptr;
 
   // FIXME: This should be per-kernel.
bool AMDGPUPromoteAlloca::handleAlloca(AllocaInst &I, bool SufficientLDS) {
   // First try to replace the alloca with a vector
   Type *AllocaTy = I.getAllocatedType();
 
   LLVM_DEBUG(dbgs() << "Trying to promote " << I << '\n');
 
   if (tryPromoteAllocaToVector(&I))
     return true; // Promoted to vector.
 
+  if (DisablePromoteAllocaToLDS)
+    return false;
+
   const Function &ContainingFunction = *I.getParent()->getParent();
   CallingConv::ID CC = ContainingFunction.getCallingConv();
 
   // Don't promote the alloca to LDS for shader calling conventions as the work
   // item ID intrinsics are not supported for these calling conventions.
   // Furthermore not all LDS is available for some of the stages.
   switch (CC) {
   case CallingConv::AMDGPU_KERNEL:

test/CodeGen/AMDGPU/promote-alloca-to-lds-icmp.ll

 ; RUN: opt -S -mtriple=amdgcn-unknown-amdhsa -mcpu=kaveri -amdgpu-promote-alloca < %s | FileCheck %s
+; RUN: opt -S -mtriple=amdgcn-unknown-amdhsa -mcpu=kaveri -amdgpu-promote-alloca -disable-promote-alloca-to-lds< %s | FileCheck -check-prefix=NOLDS %s
 
 ; This normally would be fixed by instcombine to be compare to the GEP
 ; indices
 
+; NOLDS-NOT: addrspace(3)
+
 ; CHECK-LABEL: @lds_promoted_alloca_icmp_same_derived_pointer(
 ; CHECK: [[ARRAYGEP:%[0-9]+]] = getelementptr inbounds [256 x [16 x i32]], [256 x [16 x i32]] addrspace(3)* @lds_promoted_alloca_icmp_same_derived_pointer.alloca, i32 0, i32 %{{[0-9]+}}
 ; CHECK: %ptr0 = getelementptr inbounds [16 x i32], [16 x i32] addrspace(3)* [[ARRAYGEP]], i32 0, i32 %a
 ; CHECK: %ptr1 = getelementptr inbounds [16 x i32], [16 x i32] addrspace(3)* [[ARRAYGEP]], i32 0, i32 %b
 ; CHECK: %cmp = icmp eq i32 addrspace(3)* %ptr0, %ptr1
 define amdgpu_kernel void @lds_promoted_alloca_icmp_same_derived_pointer(i32 addrspace(1)* %out, i32 %a, i32 %b) #0 {
   %alloca = alloca [16 x i32], align 4
   %ptr0 = getelementptr inbounds [16 x i32], [16 x i32]* %alloca, i32 0, i32 %a