R600/SI: Use DS offsets for constant addresses

Use 0 as the base address for a constant address, so if
we have a constant address we can save moves and form
read2/write2s.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219698 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/R600/AMDGPUISelDAGToDAG.cpp b/lib/Target/R600/AMDGPUISelDAGToDAG.cpp
index 989fc938c08d020f9b3a879311c98d1178ccde69..a1bbfa6553092fab6797a1c2d3ac07cf42869d5f 100644 (file)
--- a/lib/Target/R600/AMDGPUISelDAGToDAG.cpp
+++ b/lib/Target/R600/AMDGPUISelDAGToDAG.cpp
@@ -787,6 +787,18 @@ bool AMDGPUDAGToDAGISel::SelectDS1Addr1Offset(SDValue Addr, SDValue &Base,
     }
   }
 
+  // If we have a constant address, prefer to put the constant into the
+  // offset. This can save moves to load the constant address since multiple
+  // operations can share the zero base address register, and enables merging
+  // into read2 / write2 instructions.
+  if (const ConstantSDNode *CAddr = dyn_cast<ConstantSDNode>(Addr)) {
+    if (isUInt<16>(CAddr->getZExtValue())) {
+      Base = CurDAG->getConstant(0, MVT::i32);
+      Offset = Addr;
+      return true;
+    }
+  }
+
   // default case
   Base = Addr;
   Offset = CurDAG->getTargetConstant(0, MVT::i16);

diff --git a/test/CodeGen/R600/ds_read2.ll b/test/CodeGen/R600/ds_read2.ll
index 4c193987d2ec316e9652d0b0d3ae08006b921820..74d3a598b96c90159a872b2b272070e4a01ed6d9 100644 (file)
--- a/test/CodeGen/R600/ds_read2.ll
+++ b/test/CodeGen/R600/ds_read2.ll
@@ -382,6 +382,30 @@ define void @misaligned_read2_f64(double addrspace(1)* %out, double addrspace(3)
   ret void
 }
 
+@foo = addrspace(3) global [4 x i32] zeroinitializer, align 4
+
+; SI-LABEL: @load_constant_adjacent_offsets
+; SI: V_MOV_B32_e32 [[ZERO:v[0-9]+]], 0{{$}}
+; SI: DS_READ2_B32 v{{\[[0-9]+:[0-9]+\]}}, [[ZERO]] offset0:0 offset1:1
+define void @load_constant_adjacent_offsets(i32 addrspace(1)* %out) {
+  %val0 = load i32 addrspace(3)* getelementptr inbounds ([4 x i32] addrspace(3)* @foo, i32 0, i32 0), align 4
+  %val1 = load i32 addrspace(3)* getelementptr inbounds ([4 x i32] addrspace(3)* @foo, i32 0, i32 1), align 4
+  %sum = add i32 %val0, %val1
+  store i32 %sum, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: @load_constant_disjoint_offsets
+; SI: V_MOV_B32_e32 [[ZERO:v[0-9]+]], 0{{$}}
+; SI: DS_READ2_B32 v{{\[[0-9]+:[0-9]+\]}}, [[ZERO]] offset0:0 offset1:2
+define void @load_constant_disjoint_offsets(i32 addrspace(1)* %out) {
+  %val0 = load i32 addrspace(3)* getelementptr inbounds ([4 x i32] addrspace(3)* @foo, i32 0, i32 0), align 4
+  %val1 = load i32 addrspace(3)* getelementptr inbounds ([4 x i32] addrspace(3)* @foo, i32 0, i32 2), align 4
+  %sum = add i32 %val0, %val1
+  store i32 %sum, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
 @sgemm.lA = internal unnamed_addr addrspace(3) global [264 x float] zeroinitializer, align 4
 @sgemm.lB = internal unnamed_addr addrspace(3) global [776 x float] zeroinitializer, align 4
 

diff --git a/test/CodeGen/R600/ds_write2.ll b/test/CodeGen/R600/ds_write2.ll
index 5dc4b6d7325735a94519eae800c888f9d582e8d5..6e5bcffb62105e1e776382bcf3608dbbba7bfe2b 100644 (file)
--- a/test/CodeGen/R600/ds_write2.ll
+++ b/test/CodeGen/R600/ds_write2.ll
@@ -320,6 +320,27 @@ define void @simple_write2_two_val_f64(double addrspace(1)* %C, double addrspace
   ret void
 }
 
+@foo = addrspace(3) global [4 x i32] zeroinitializer, align 4
+
+; SI-LABEL: @store_constant_adjacent_offsets
+; SI: V_MOV_B32_e32 [[ZERO:v[0-9]+]], 0{{$}}
+; SI: DS_WRITE2_B32 [[ZERO]], v{{[0-9]+}}, v{{[0-9]+}} offset0:0 offset1:1
+define void @store_constant_adjacent_offsets() {
+  store i32 123, i32 addrspace(3)* getelementptr inbounds ([4 x i32] addrspace(3)* @foo, i32 0, i32 0), align 4
+  store i32 123, i32 addrspace(3)* getelementptr inbounds ([4 x i32] addrspace(3)* @foo, i32 0, i32 1), align 4
+  ret void
+}
+
+; SI-LABEL: @store_constant_disjoint_offsets
+; SI-DAG: V_MOV_B32_e32 [[VAL:v[0-9]+]], 0x7b{{$}}
+; SI-DAG: V_MOV_B32_e32 [[ZERO:v[0-9]+]], 0{{$}}
+; SI: DS_WRITE2_B32 [[ZERO]], [[VAL]], [[VAL]] offset0:0 offset1:2
+define void @store_constant_disjoint_offsets() {
+  store i32 123, i32 addrspace(3)* getelementptr inbounds ([4 x i32] addrspace(3)* @foo, i32 0, i32 0), align 4
+  store i32 123, i32 addrspace(3)* getelementptr inbounds ([4 x i32] addrspace(3)* @foo, i32 0, i32 2), align 4
+  ret void
+}
+
 @sgemm.lA = internal unnamed_addr addrspace(3) global [264 x float] zeroinitializer, align 4
 @sgemm.lB = internal unnamed_addr addrspace(3) global [776 x float] zeroinitializer, align 4