AMDGPU: Push bitcasts through build_vector

This reduces the number of copies and reg_sequences
when using fp constant vectors. This significantly
reduces the code size in local-stack-alloc-bug.ll

llvm-svn: 281822

2 files changed, 96 insertions, 0 deletions

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
index 22871e64a3e..e214164facb 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -2548,6 +2548,33 @@ SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
     break;
   case ISD::BITCAST: {
     EVT DestVT = N->getValueType(0);
+
+    // Push casts through vector builds. This helps avoid emitting a large
+    // number of copies when materializing floating point vector constants.
+    //
+    // vNt1 bitcast (vNt0 (build_vector t0:x, t0:y)) =>
+    //   vnt1 = build_vector (t1 (bitcast t0:x)), (t1 (bitcast t0:y))
+    if (DestVT.isVector()) {
+      SDValue Src = N->getOperand(0);
+      if (Src.getOpcode() == ISD::BUILD_VECTOR) {
+        EVT SrcVT = Src.getValueType();
+        unsigned NElts = DestVT.getVectorNumElements();
+
+        if (SrcVT.getVectorNumElements() == NElts) {
+          EVT DestEltVT = DestVT.getVectorElementType();
+
+          SmallVector<SDValue, 8> CastedElts;
+          SDLoc SL(N);
+          for (unsigned I = 0, E = SrcVT.getVectorNumElements(); I != E; ++I) {
+            SDValue Elt = Src.getOperand(I);
+            CastedElts.push_back(DAG.getNode(ISD::BITCAST, DL, DestEltVT, Elt));
+          }
+
+          return DAG.getBuildVector(DestVT, SL, CastedElts);
+        }
+      }
+    }
+
     if (DestVT.getSizeInBits() != 64 && !DestVT.isVector())
       break;
 
diff --git a/llvm/test/CodeGen/AMDGPU/bitcast-vector-extract.ll b/llvm/test/CodeGen/AMDGPU/bitcast-vector-extract.ll
new file mode 100644
index 00000000000..2482fa761b1
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/bitcast-vector-extract.ll
@@ -0,0 +1,69 @@
+; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
+; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
+
+; The bitcast should be pushed through the bitcasts so the vectors can
+; be broken down and the shared components can be CSEd
+
+; GCN-LABEL: {{^}}store_bitcast_constant_v8i32_to_v8f32:
+; GCN: buffer_store_dwordx4
+; GCN: buffer_store_dwordx4
+; GCN-NOT: v_mov_b32
+; GCN: buffer_store_dwordx4
+; GCN-NOT: v_mov_b32
+; GCN: buffer_store_dwordx4
+define void @store_bitcast_constant_v8i32_to_v8f32(<8 x float> addrspace(1)* %out, <8 x i32> %vec) {
+  %vec0.bc = bitcast <8 x i32> <i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 8> to <8 x float>
+  store volatile <8 x float> %vec0.bc, <8 x float> addrspace(1)* %out
+
+  %vec1.bc = bitcast <8 x i32> <i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 9> to <8 x float>
+  store volatile <8 x float> %vec1.bc, <8 x float> addrspace(1)* %out
+  ret void
+}
+
+; GCN-LABEL: {{^}}store_bitcast_constant_v4i64_to_v8f32:
+; GCN: buffer_store_dwordx4
+; GCN: buffer_store_dwordx4
+; GCN-NOT: v_mov_b32
+; GCN: buffer_store_dwordx4
+; GCN-NOT: v_mov_b32
+; GCN: buffer_store_dwordx4
+define void @store_bitcast_constant_v4i64_to_v8f32(<8 x float> addrspace(1)* %out, <4 x i64> %vec) {
+  %vec0.bc = bitcast <4 x i64> <i64 7, i64 7, i64 7, i64 8> to <8 x float>
+  store volatile <8 x float> %vec0.bc, <8 x float> addrspace(1)* %out
+
+  %vec1.bc = bitcast <4 x i64> <i64 7, i64 7, i64 7, i64 9> to <8 x float>
+  store volatile <8 x float> %vec1.bc, <8 x float> addrspace(1)* %out
+  ret void
+}
+
+; GCN-LABEL: {{^}}store_bitcast_constant_v4i64_to_v4f64:
+; GCN: buffer_store_dwordx4
+; GCN: buffer_store_dwordx4
+; GCN-NOT: v_mov_b32
+; GCN: buffer_store_dwordx4
+; GCN-NOT: v_mov_b32
+; GCN: buffer_store_dwordx4
+define void @store_bitcast_constant_v4i64_to_v4f64(<4 x double> addrspace(1)* %out, <4 x i64> %vec) {
+  %vec0.bc = bitcast <4 x i64> <i64 7, i64 7, i64 7, i64 8> to <4 x double>
+  store volatile <4 x double> %vec0.bc, <4 x double> addrspace(1)* %out
+
+  %vec1.bc = bitcast <4 x i64> <i64 7, i64 7, i64 7, i64 9> to <4 x double>
+  store volatile <4 x double> %vec1.bc, <4 x double> addrspace(1)* %out
+  ret void
+}
+
+; GCN-LABEL: {{^}}store_bitcast_constant_v8i32_to_v16i16:
+; GCN: buffer_store_dwordx4
+; GCN: buffer_store_dwordx4
+; GCN-NOT: v_mov_b32
+; GCN: buffer_store_dwordx4
+; GCN-NOT: v_mov_b32
+; GCN: buffer_store_dwordx4
+define void @store_bitcast_constant_v8i32_to_v16i16(<8 x float> addrspace(1)* %out, <16 x i16> %vec) {
+  %vec0.bc = bitcast <16 x i16> <i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 8> to <8 x float>
+  store volatile <8 x float> %vec0.bc, <8 x float> addrspace(1)* %out
+
+  %vec1.bc = bitcast <16 x i16> <i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 7, i16 9> to <8 x float>
+  store volatile <8 x float> %vec1.bc, <8 x float> addrspace(1)* %out
+  ret void
+}