Loosen scheduling restrictions on the PPC dcbt intrinsic

As with the prefetch intrinsic to which it maps, simply have dcbt
marked as reading from and writing to its arguments instead of having
unmodeled side effects. While this might cause unwanted code motion
(because aliasing checks don't really capture cache-line sharing),
it is more important that prefetches in unrolled loops don't block
the scheduler from rearranging the unrolled loop body.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@171073 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/include/llvm/IntrinsicsPowerPC.td b/include/llvm/IntrinsicsPowerPC.td
index da85bfba8631b8c206cf62317f8e1dd109d910d5..cde39ccd3c523b90f0c2f7aaffcc4ba46b05f78f 100644 (file)
--- a/include/llvm/IntrinsicsPowerPC.td
+++ b/include/llvm/IntrinsicsPowerPC.td
@@ -22,7 +22,8 @@ let TargetPrefix = "ppc" in {  // All intrinsics start with "llvm.ppc.".
   def int_ppc_dcbf  : Intrinsic<[], [llvm_ptr_ty], []>;
   def int_ppc_dcbi  : Intrinsic<[], [llvm_ptr_ty], []>;
   def int_ppc_dcbst : Intrinsic<[], [llvm_ptr_ty], []>;
-  def int_ppc_dcbt  : Intrinsic<[], [llvm_ptr_ty], []>;
+  def int_ppc_dcbt  : Intrinsic<[], [llvm_ptr_ty],
+    [IntrReadWriteArgMem, NoCapture<0>]>;
   def int_ppc_dcbtst: Intrinsic<[], [llvm_ptr_ty], []>;
   def int_ppc_dcbz  : Intrinsic<[], [llvm_ptr_ty], []>;
   def int_ppc_dcbzl : Intrinsic<[], [llvm_ptr_ty], []>;

diff --git a/test/CodeGen/PowerPC/dcbt-sched.ll b/test/CodeGen/PowerPC/dcbt-sched.ll
new file mode 100644 (file)
index 0000000..dfa1b75
--- /dev/null
+++ b/test/CodeGen/PowerPC/dcbt-sched.ll
@@ -0,0 +1,22 @@
+target datalayout = "E-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v128:128:128-n32:64"
+target triple = "powerpc64-unknown-linux-gnu"
+; RUN: llc -mcpu=a2 -enable-misched -enable-aa-sched-mi < %s | FileCheck %s
+
+define i8 @test1(i8* noalias %a, i8* noalias %b, i8* noalias %c) nounwind {
+entry:
+  %q = load i8* %b
+  call void @llvm.prefetch(i8* %a, i32 0, i32 3, i32 1)
+  %r = load i8* %c
+  %s = add i8 %q, %r
+  ret i8 %s
+}
+
+declare void @llvm.prefetch(i8*, i32, i32, i32)
+
+; Test that we've moved the second load to before the dcbt to better
+; hide its latency.
+; CHECK: @test1
+; CHECK: lbz
+; CHECK: lbz
+; CHECK: dcbt
+