[AMDGPU] Mark sin/cos load folding as modifying the function.

[MJDSys]

Summary

When the load value is folded into the sin/cos operation, the
AMDGPU library calls simplifier could still mark the function
as unmodified. Instead ensure if there is an early return,
return whether the load was folded into the sin/cos call.

[arsenm]

Is this an observable problem?

[MJDSys]

I saw this issue when playing with the OpenMM OpenCL test suite with Mesa/Clover. I was using a copy of LLVM with assertions turned on, and it gave me an error about a pass modifying its inputs without marking its inputs as changed.

[rampitec]

LGTM

[MJDSys]

Thanks for the review. Can you please commit this patch? I do not have access.

[rampitec]

In D91401#2395065, @MJDSys wrote:

Thanks for the review. Can you please commit this patch? I do not have access.

Sure, will submit shortly. Thanks for the patch!

llvm/lib/Target/AMDGPU/AMDGPULibCalls.cpp

bool AMDGPULibCalls::fold_sincos(CallInst *CI, IRBuilder<> &B,
   assert(fInfo.getId() == AMDGPULibFunc::EI_SIN ||
          fInfo.getId() == AMDGPULibFunc::EI_COS);
   bool const isSin = fInfo.getId() == AMDGPULibFunc::EI_SIN;
 
   Value *CArgVal = CI->getArgOperand(0);
   BasicBlock * const CBB = CI->getParent();
 
   int const MaxScan = 30;
+  bool Changed = false;
 
   { // fold in load value.
     LoadInst *LI = dyn_cast<LoadInst>(CArgVal);
     if (LI && LI->getParent() == CBB) {
       BasicBlock::iterator BBI = LI->getIterator();
       Value *AvailableVal = FindAvailableLoadedValue(LI, CBB, BBI, MaxScan, AA);
       if (AvailableVal) {
+        Changed = true;
         CArgVal->replaceAllUsesWith(AvailableVal);
         if (CArgVal->getNumUses() == 0)
           LI->eraseFromParent();
         CArgVal = CI->getArgOperand(0);
       }
     }
   }
 
for (int I = MaxScan; I > 0 && BBI != CBB->begin(); --BBI, --I) {
       if (cast<Instruction>(BBI) == XI) {
         UI = XI;
         break;
       }
     }
     if (UI) break;
   }
 
-  if (!UI) return false;
+  if (!UI)
+    return Changed;
 
   // Merge the sin and cos.
 
   // for OpenCL 2.0 we have only generic implementation of sincos
   // function.
   AMDGPULibFunc nf(AMDGPULibFunc::EI_SINCOS, fInfo);
   nf.getLeads()[0].PtrKind = AMDGPULibFunc::getEPtrKindFromAddrSpace(AMDGPUAS::FLAT_ADDRESS);
   FunctionCallee Fsincos = getFunction(M, nf);
-  if (!Fsincos) return false;
+  if (!Fsincos)
+    return Changed;
 
   BasicBlock::iterator ItOld = B.GetInsertPoint();
   AllocaInst *Alloc = insertAlloca(UI, B, "__sincos_");
   B.SetInsertPoint(UI);
 
   Value *P = Alloc;
   Type *PTy = Fsincos.getFunctionType()->getParamType(1);
   // The allocaInst allocates the memory in private address space. This need