[llvm] [AMDGPU] Add support for preloading implicit kernel arguments (PR #83817)

[Matt Arsenault via llvm-commits]

================
@@ -64,6 +64,86 @@ class PreloadKernelArgInfo {
     NumFreeUserSGPRs -= (NumPreloadSGPRs + PaddingSGPRs);
     return true;
   }
+
+  // Try to allocate SGPRs to preload implicit kernel arguments.
+  void tryAllocImplicitArgPreloadSGPRs(unsigned ImplicitArgsBaseOffset,
+                                       IRBuilder<> &Builder) {
+    IntrinsicInst *ImplicitArgPtr = nullptr;
+    for (Function::iterator B = F.begin(), BE = F.end(); B != BE; ++B) {
+      for (BasicBlock::iterator I = B->begin(), IE = B->end(); I != IE; ++I) {
+        if (IntrinsicInst *CI = dyn_cast<IntrinsicInst>(I))
+          if (CI->getIntrinsicID() == Intrinsic::amdgcn_implicitarg_ptr) {
+            ImplicitArgPtr = CI;
+            break;
+          }
+      }
+    }
+
+    if (!ImplicitArgPtr)
+      return;
+
+    const DataLayout &DL = F.getParent()->getDataLayout();
+    // Pair is the load and the load offset.
+    SmallVector<std::pair<LoadInst *, unsigned>, 4> ImplicitArgLoads;
+    for (auto *U : ImplicitArgPtr->users()) {
+      if (!U->hasOneUse())
+        continue;
+
+      int64_t Offset = 0;
+      auto *Load = dyn_cast<LoadInst>(U); // Load from ImplicitArgPtr?
+      if (!Load) {
+        if (GetPointerBaseWithConstantOffset(U, Offset, DL) != ImplicitArgPtr)
+          continue;
+        Load = dyn_cast<LoadInst>(*U->user_begin()); // Load from GEP?
+      }
+
+      if (!Load || !Load->isSimple())
+        continue;
+
+      // FIXME: Expand to handle 64-bit implicit args and large merged loads.
+      if (Load->getType() != Builder.getInt16Ty() &&
+          Load->getType() != Builder.getInt32Ty())
----------------
arsenm wrote:

Better to go on the size. Should be equally easy to handle half, float and v2[if]16 

https://github.com/llvm/llvm-project/pull/83817