[SYCL][CUDA] Workaround for the problem with memory reordering (#1334)

Currently there is a bug in LLVM for PTX target. PTX target specific
intrinsics like llvm.nvvm.barrier0 are treated like regular LLVM
intrinsics in Globals AA. As a result, there are situations when Globals
AA produces a result that barrier intrinsic doesn't modify internal
globals. This allows llvm transformations like GVN to perform illegal
memory reordering.

This is a workaround while permanent fix is not implemented in LLVM
project.

Signed-off-by: Artur Gainullin <artur.gainullin@intel.com>

Author: againull

llvm/lib/Analysis/GlobalsModRef.cpp

@@ -534,6 +534,17 @@ void GlobalsAAResult::AnalyzeCallGraph(CallGraph &CG, Module &M) {
           if (!F->isIntrinsic()) {
             KnowNothing = true;
             break;
+          } else if (F->getName().contains("nvvm.barrier") or
+                     F->getName().contains("nvvm.membar")) {
+            // Even if it is an intrinsic, consider that nothing is known for
+            // NVVM barrier itrinsics to prevent illegal optimizations.
+            // This is a workaround for the bug on PTX target: barrier
+            // intrinsics are implemented as llvm intrinsics, as result there
+            // are cases when globals alias analysis can produce a result that
+            // barrier doesn't modify internal global which causes illegal
+            // reordering of memory accesses.
+            KnowNothing = true;
+            break;
           }
         }
         continue;

llvm/test/Analysis/GlobalsModRef/barrier_intrinsic.ll

@@ -0,0 +1,56 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -globals-aa -gvn -S | FileCheck %s
+
+; Check that load from global variable is not moved across barrier.
+
+
+target triple = "nvptx"
+
+@foo.l.0 = internal unnamed_addr addrspace(3) global i32 undef, align 4
+
+define dso_local spir_kernel void @foo(i32 addrspace(1)* nocapture %0) {
+; CHECK-LABEL: @foo(
+; CHECK-NEXT:    [[TMP2:%.*]] = tail call i32 @_Z13get_global_idj(i32 0) #0
+; CHECK-NEXT:    [[TMP3:%.*]] = tail call i32 @_Z12get_local_idj(i32 0) #0
+; CHECK-NEXT:    [[TMP4:%.*]] = icmp eq i32 [[TMP3]], 0
+; CHECK-NEXT:    br i1 [[TMP4]], label [[TMP5:%.*]], label [[TMP7:%.*]]
+; CHECK:       5:
+; CHECK-NEXT:    [[TMP6:%.*]] = add i32 [[TMP2]], 5
+; CHECK-NEXT:    store i32 [[TMP6]], i32 addrspace(3)* @foo.l.0, align 4
+; CHECK-NEXT:    br label [[TMP7]]
+; CHECK:       7:
+; CHECK-NEXT:    tail call void @llvm.nvvm.barrier0() #2
+; CHECK-NEXT:    [[TMP8:%.*]] = load i32, i32 addrspace(3)* @foo.l.0, align 4
+; CHECK-NEXT:    [[TMP9:%.*]] = getelementptr inbounds i32, i32 addrspace(1)* [[TMP0:%.*]], i32 [[TMP2]]
+; CHECK-NEXT:    store i32 [[TMP8]], i32 addrspace(1)* [[TMP9]], align 4
+; CHECK-NEXT:    ret void
+;
+  %2 = tail call i32 @_Z13get_global_idj(i32 0) #0
+  %3 = tail call i32 @_Z12get_local_idj(i32 0) #0
+  %4 = icmp eq i32 %3, 0
+  br i1 %4, label %5, label %7
+
+5:                                                ; preds = %1
+  %6 = add i32 %2, 5
+  store i32 %6, i32 addrspace(3)* @foo.l.0, align 4
+  br label %7
+
+7:                                                ; preds = %5, %1
+  tail call void @llvm.nvvm.barrier0() #1
+  %8 = load i32, i32 addrspace(3)* @foo.l.0, align 4
+  %9 = getelementptr inbounds i32, i32 addrspace(1)* %0, i32 %2
+  store i32 %8, i32 addrspace(1)* %9, align 4
+  ret void
+}
+
+; Function Attrs: convergent nounwind readnone
+declare dso_local i32 @_Z13get_global_idj(i32) local_unnamed_addr #0
+
+; Function Attrs: convergent nounwind readnone
+declare dso_local i32 @_Z12get_local_idj(i32) local_unnamed_addr #0
+
+; Function Attrs: convergent
+declare dso_local void @llvm.nvvm.barrier0() local_unnamed_addr #1
+
+attributes #0 = { convergent nounwind readnone }
+attributes #1 = { convergent }