Revert "[LLVM16][StatelessToStateful] Case where BUFFER_OFFSET doesn't seem to be 0"

Revert "[LLVM16][StatelessToStateful] Case where BUFFER_OFFSET doesn't seem to be 0"

(cherry picked from commit e5f2c15)

Author: bokrzesi

IGC/Compiler/Optimizer/OpenCLPasses/StatelessToStateful/StatelessToStateful.cpp

@@ -527,35 +527,15 @@ static alignment_t determinePointerAlignment(Value *Ptr, const DataLayout &DL, A
   // 1) Examine uses: look for loads/stores (which may carry explicit
   //    alignment) or a GEP that reveals an ABI alignment from its element
   //    type.
-
-  // We're saving users into the queue
-  // in order to track bitcasts
-  std::queue<Instruction *> queue;
-
   for (User *U : Ptr->users()) {
-    if (auto *I = dyn_cast<Instruction>(U)) {
-      queue.push(I);
-    }
-  }
-
-  while (queue.size() > 0) {
-    auto U = queue.front();
-    queue.pop();
-
     if (auto *LI = dyn_cast<LoadInst>(U)) {
       // Load has an explicit alignment.
       alignment_t LdAlign = LI->getAlign().value();
       if (LdAlign > BestAlign)
         BestAlign = LdAlign;
-    } else if (auto *cast = dyn_cast<BitCastInst>(U)) {
-      for (User *U : cast->users()) {
-        if (auto *I = dyn_cast<Instruction>(U)) {
-          queue.push(I);
-        }
-      }
     } else if (auto *SI = dyn_cast<StoreInst>(U)) {
       // Store sets alignment only if the pointer we store into is Ptr.
-      if (SI->getPointerOperand()->stripPointerCasts() == Ptr) {
+      if (SI->getPointerOperand() == Ptr) {
         alignment_t StAlign = SI->getAlign().value();
         if (StAlign > BestAlign)
           BestAlign = StAlign;
@@ -1120,12 +1100,6 @@ void StatelessToStateful::finalizeArgInitialValue(Function *F) {
     return;
   }
 
-#if LLVM_VERSION_MAJOR >= 16
-  // Disabling this transformation/optimization because the assumption about BufferOffset being 0 doesn't seem to be true
-  // More details in PR / blame.
-  return;
-#endif
-
   Module *M = F->getParent();
   Type *int32Ty = Type::getInt32Ty(M->getContext());
   Value *ZeroValue = ConstantInt::get(int32Ty, 0);

IGC/Compiler/tests/StatelessToStateful/Bindful/determine_pointer_size_casts.ll

@@ -8,8 +8,7 @@
 ;============================ end_copyright_notice =============================
 ;
 ; REQUIRES: regkeys
-; RUN: igc_opt --typed-pointers --regkey DumpHasNonKernelArgLdSt=1 --regkey EnableOptionalBufferOffset=1 --regkey EnableSupportBufferOffset=1 -enable-debugify -igc-stateless-to-stateful-resolution -igc-serialize-metadata -S < %s 2>&1 | FileCheck %s --check-prefixes=CHECK,%LLVM_DEPENDENT_CHECK_PREFIX%
-; RUN: igc_opt --opaque-pointers --regkey DumpHasNonKernelArgLdSt=1 --regkey EnableOptionalBufferOffset=1 --regkey EnableSupportBufferOffset=1 -enable-debugify -igc-stateless-to-stateful-resolution -igc-serialize-metadata -S < %s 2>&1 | FileCheck %s --check-prefixes=CHECK,%LLVM_DEPENDENT_CHECK_PREFIX%
+; RUN: igc_opt --typed-pointers --regkey DumpHasNonKernelArgLdSt=1 --regkey EnableOptionalBufferOffset=1 --regkey EnableSupportBufferOffset=1 -enable-debugify -igc-stateless-to-stateful-resolution -igc-serialize-metadata -S < %s 2>&1 | FileCheck %s
 ; ------------------------------------------------
 ; StatelessToStateful
 ; ------------------------------------------------
@@ -19,29 +18,18 @@
 ; CHECK-NOT: WARNING
 ; CHECK: CheckModuleDebugify: PASS
 
+
 define spir_kernel void @func_b(i32 %n, i32 addrspace(1)* %r, <8 x i32> %r0, <8 x i32> %payloadHeader, i8* %privateBase, i8 addrspace(1)* %s2, i8 addrspace(1)* %s3, i32 %s4, i32 %s5, i32 %bufferOffset) #0 {
 ; CHECK-LABEL: @func_b(
 ; CHECK-NEXT:  entry:
-
-; CHECK-LLVM-16:    %0 = add i32 %bufferOffset, 64
-; CHECK-LLVM-16:    %1 = add i32 %bufferOffset, 64
-
-; CHECK-NEXT:       [[TMP0:%.*]] = getelementptr i32, {{ptr|i32}} addrspace(1){{.*}} [[R:%.*]], i32 16, !dbg [[DBG104:![0-9]+]]
-; CHECK-NEXT:       call void @llvm.dbg.value(metadata {{ptr|i32}} addrspace(1){{.*}} [[TMP0]], metadata [[META100:![0-9]+]], metadata !DIExpression()), !dbg [[DBG104]]
-
-; CHECK-LLVM-14:    [[TMP1:%.*]] = inttoptr i32 64 to {{ptr|i32}} addrspace({{[0-9]+}}){{.*}}, !dbg [[DBG105:![0-9]+]]
-; CHECK-LLVM-15:    [[TMP1:%.*]] = inttoptr i32 64 to {{ptr|i32}} addrspace({{[0-9]+}}){{.*}}, !dbg [[DBG105:![0-9]+]]
-; CHECK-LLVM-16:    [[TMP1:%.*]] = inttoptr i32 %0 to {{ptr|i32}} addrspace({{[0-9]+}}){{.*}}, !dbg [[DBG105:![0-9]+]]
-
-; CHECK-NEXT:       [[TMP2:%.*]] = load i32, {{ptr|i32}} addrspace({{[0-9]+}}){{.*}} [[TMP1]], align 4, !dbg [[DBG105]]
-; CHECK-NEXT:       call void @llvm.dbg.value(metadata i32 [[TMP2]], metadata [[META102:![0-9]+]], metadata !DIExpression()), !dbg [[DBG105]]
-
-; CHECK-LLVM-14:    [[TMP3:%.*]] = inttoptr i32 64 to {{ptr|i32}} addrspace({{[0-9]+}}){{.*}}, !dbg [[DBG106:![0-9]+]]
-; CHECK-LLVM-15:    [[TMP3:%.*]] = inttoptr i32 64 to {{ptr|i32}} addrspace({{[0-9]+}}){{.*}}, !dbg [[DBG106:![0-9]+]]
-; CHECK-LLVM-16:    [[TMP3:%.*]] = inttoptr i32 %1 to {{ptr|i32}} addrspace({{[0-9]+}}){{.*}}, !dbg [[DBG106:![0-9]+]]
-
-; CHECK-NEXT:       store i32 [[N:%.*]], {{ptr|i32}} addrspace({{[0-9]+}}){{.*}} [[TMP3]], align 4, !dbg [[DBG106]]
-; CHECK-NEXT:       ret void, !dbg [[DBG107:![0-9]+]]
+; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr i32, i32 addrspace(1)* [[R:%.*]], i32 16, !dbg [[DBG104:![0-9]+]]
+; CHECK-NEXT:    call void @llvm.dbg.value(metadata i32 addrspace(1)* [[TMP0]], metadata [[META100:![0-9]+]], metadata !DIExpression()), !dbg [[DBG104]]
+; CHECK-NEXT:    [[TMP1:%.*]] = inttoptr i32 64 to i32 addrspace(131072)*, !dbg [[DBG105:![0-9]+]]
+; CHECK-NEXT:    [[TMP2:%.*]] = load i32, i32 addrspace(131072)* [[TMP1]], align 4, !dbg [[DBG105]]
+; CHECK-NEXT:    call void @llvm.dbg.value(metadata i32 [[TMP2]], metadata [[META102:![0-9]+]], metadata !DIExpression()), !dbg [[DBG105]]
+; CHECK-NEXT:    [[TMP3:%.*]] = inttoptr i32 64 to i32 addrspace(131072)*, !dbg [[DBG106:![0-9]+]]
+; CHECK-NEXT:    store i32 [[N:%.*]], i32 addrspace(131072)* [[TMP3]], align 4, !dbg [[DBG106]]
+; CHECK-NEXT:    ret void, !dbg [[DBG107:![0-9]+]]
 ;
 entry:
   %0 = getelementptr i32, i32 addrspace(1)* %r, i32 16