R600/SI: Only use SALU instructions for 64-bit add in a block of CF depth 0

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208886 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/R600/AMDGPUISelDAGToDAG.cpp b/lib/Target/R600/AMDGPUISelDAGToDAG.cpp
index ba705db69dedee3cb704142fb8fad30799c2fec2..f1f0bfa89ac02c43387b478626057c942830bb3d 100644 (file)
--- a/lib/Target/R600/AMDGPUISelDAGToDAG.cpp
+++ b/lib/Target/R600/AMDGPUISelDAGToDAG.cpp
@@ -239,12 +239,13 @@ SDNode *AMDGPUDAGToDAGISel::Select(SDNode *N) {
     AddLoArgs.push_back(SDValue(Lo0, 0));
     AddLoArgs.push_back(SDValue(Lo1, 0));
 
-    SDNode *AddLo = CurDAG->getMachineNode(AMDGPU::S_ADD_I32, DL,
-                                           VTList, AddLoArgs);
+    SDNode *AddLo = CurDAG->getMachineNode(
+        isCFDepth0() ? AMDGPU::S_ADD_I32 : AMDGPU::V_ADD_I32_e32,
+        DL, VTList, AddLoArgs);
     SDValue Carry = SDValue(AddLo, 1);
-    SDNode *AddHi = CurDAG->getMachineNode(AMDGPU::S_ADDC_U32, DL,
-                                           MVT::i32, SDValue(Hi0, 0),
-                                           SDValue(Hi1, 0), Carry);
+    SDNode *AddHi = CurDAG->getMachineNode(
+        isCFDepth0() ? AMDGPU::S_ADDC_U32 : AMDGPU::V_ADDC_U32_e32,
+        DL, MVT::i32, SDValue(Hi0, 0), SDValue(Hi1, 0), Carry);
 
     SDValue Args[5] = {
       CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, MVT::i32),

diff --git a/test/CodeGen/R600/add.ll b/test/CodeGen/R600/add.ll
index e9db52ae4ae1d73150ea20da8640bfbb802c0194..711a2bc417741b06f2a04f17838c4f8b8307bd99 100644 (file)
--- a/test/CodeGen/R600/add.ll
+++ b/test/CodeGen/R600/add.ll
@@ -140,3 +140,28 @@ entry:
   store i64 %1, i64 addrspace(1)* %out
   ret void
 }
+
+; Test i64 add inside a branch.  We don't allow SALU instructions inside of
+; branches.
+; FIXME: We are being conservative here.  We could allow this in some cases.
+; FUNC-LABEL: @add64_in_branch
+; SI-CHECK-NOT: S_ADD_I32
+; SI-CHECK-NOT: S_ADDC_U32
+define void @add64_in_branch(i64 addrspace(1)* %out, i64 addrspace(1)* %in, i64 %a, i64 %b, i64 %c) {
+entry:
+  %0 = icmp eq i64 %a, 0
+  br i1 %0, label %if, label %else
+
+if:
+  %1 = load i64 addrspace(1)* %in
+  br label %endif
+
+else:
+  %2 = add i64 %a, %b
+  br label %endif
+
+endif:
+  %3 = phi i64 [%1, %if], [%2, %else]
+  store i64 %3, i64 addrspace(1)* %out
+  ret void
+}