[llvm] r260005 - [NVPTX] Mark nvvm synchronizing intrinsics as convergent.

[Justin Lebar via llvm-commits]

Author: jlebar
Date: Sat Feb  6 13:32:44 2016
New Revision: 260005

URL: http://llvm.org/viewvc/llvm-project?rev=260005&view=rev
Log:
[NVPTX] Mark nvvm synchronizing intrinsics as convergent.

Summary:
This is the attribute purpose-made for e.g. __syncthreads.  It appears
that NoDuplicate may not be sufficient to prevent Sink from touching a
call to __syncthreads.

Reviewers: jingyue, hfinkel

Subscribers: llvm-commits, jholewinski, jhen, rnk, tra, majnemer

Differential Revision: http://reviews.llvm.org/D16941

Modified:
    llvm/trunk/include/llvm/IR/IntrinsicsNVVM.td
    llvm/trunk/test/Feature/intrinsic-noduplicate.ll

Modified: llvm/trunk/include/llvm/IR/IntrinsicsNVVM.td
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/include/llvm/IR/IntrinsicsNVVM.td?rev=260005&r1=260004&r2=260005&view=diff
==============================================================================
--- llvm/trunk/include/llvm/IR/IntrinsicsNVVM.td (original)
+++ llvm/trunk/include/llvm/IR/IntrinsicsNVVM.td Sat Feb  6 13:32:44 2016
@@ -729,16 +729,20 @@ def llvm_anyi64ptr_ty     : LLVMAnyPoint
                                       [IntrReadWriteArgMem, NoCapture<0>]>;
 
 // Bar.Sync
+//
+// TODO: Remove NoDuplicate here after fixing up LLVM to handle convergent
+// properly.  See discussion in http://reviews.llvm.org/D16941 and
+// http://reviews.llvm.org/D12246.
   def int_cuda_syncthreads : GCCBuiltin<"__syncthreads">,
-      Intrinsic<[], [], [IntrNoDuplicate]>;
+      Intrinsic<[], [], [IntrNoDuplicate, IntrConvergent]>;
   def int_nvvm_barrier0 : GCCBuiltin<"__nvvm_bar0">,
-      Intrinsic<[], [], [IntrNoDuplicate]>;
+      Intrinsic<[], [], [IntrNoDuplicate, IntrConvergent]>;
   def int_nvvm_barrier0_popc : GCCBuiltin<"__nvvm_bar0_popc">,
-      Intrinsic<[llvm_i32_ty], [llvm_i32_ty], [IntrNoDuplicate]>;
+      Intrinsic<[llvm_i32_ty], [llvm_i32_ty], [IntrNoDuplicate, IntrConvergent]>;
   def int_nvvm_barrier0_and : GCCBuiltin<"__nvvm_bar0_and">,
-      Intrinsic<[llvm_i32_ty], [llvm_i32_ty], [IntrNoDuplicate]>;
+      Intrinsic<[llvm_i32_ty], [llvm_i32_ty], [IntrNoDuplicate, IntrConvergent]>;
   def int_nvvm_barrier0_or : GCCBuiltin<"__nvvm_bar0_or">,
-      Intrinsic<[llvm_i32_ty], [llvm_i32_ty], [IntrNoDuplicate]>;
+      Intrinsic<[llvm_i32_ty], [llvm_i32_ty], [IntrNoDuplicate, IntrConvergent]>;
 
   // Membar
   def int_nvvm_membar_cta : GCCBuiltin<"__nvvm_membar_cta">,

Modified: llvm/trunk/test/Feature/intrinsic-noduplicate.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Feature/intrinsic-noduplicate.ll?rev=260005&r1=260004&r2=260005&view=diff
==============================================================================
--- llvm/trunk/test/Feature/intrinsic-noduplicate.ll (original)
+++ llvm/trunk/test/Feature/intrinsic-noduplicate.ll Sat Feb  6 13:32:44 2016
@@ -1,9 +1,9 @@
 ; RUN: llvm-as < %s | llvm-dis | FileCheck %s
 
-; Make sure LLVM knows about the noduplicate attribute on the
+; Make sure LLVM knows about the convergent and noduplicate attributes on the
 ; llvm.cuda.syncthreads intrinsic.
 
 declare void @llvm.cuda.syncthreads()
 
 ; CHECK: declare void @llvm.cuda.syncthreads() #[[ATTRNUM:[0-9]+]]
-; CHECK: attributes #[[ATTRNUM]] = { noduplicate nounwind }
+; CHECK: attributes #[[ATTRNUM]] = { convergent noduplicate nounwind }