[AMDGPU] Prefer SplitVectorLoad/Store over expandUnalignedLoad/Store

mbrkusanin · mbrkusanin · commit 7c88d13fd1c1 · 2020-10-08T10:17:15.000+02:00
ExpandUnalignedLoad/Store can sometimes produce unnecessary copies to temporary stack slot. We should prefer splitting vectors if possible. Differential Revision: https://reviews.llvm.org/D88882
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -8014,13 +8014,6 @@ SDValue SITargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
   assert(Op.getValueType().getVectorElementType() == MVT::i32 &&
          "Custom lowering for non-i32 vectors hasn't been implemented.");
 
-  if (!allowsMemoryAccessForAlignment(*DAG.getContext(), DAG.getDataLayout(),
-                                      MemVT, *Load->getMemOperand())) {
-    SDValue Ops[2];
-    std::tie(Ops[0], Ops[1]) = expandUnalignedLoad(Load, DAG);
-    return DAG.getMergeValues(Ops, DL);
-  }
-
   unsigned Alignment = Load->getAlignment();
   unsigned AS = Load->getAddressSpace();
   if (Subtarget->hasLDSMisalignedBug() &&
@@ -8132,6 +8125,14 @@ SDValue SITargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
       return SplitVectorLoad(Op, DAG);
     }
   }
+
+  if (!allowsMemoryAccessForAlignment(*DAG.getContext(), DAG.getDataLayout(),
+                                      MemVT, *Load->getMemOperand())) {
+    SDValue Ops[2];
+    std::tie(Ops[0], Ops[1]) = expandUnalignedLoad(Load, DAG);
+    return DAG.getMergeValues(Ops, DL);
+  }
+
   return SDValue();
 }
 
@@ -8537,11 +8538,6 @@ SDValue SITargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
   assert(VT.isVector() &&
          Store->getValue().getValueType().getScalarType() == MVT::i32);
 
-  if (!allowsMemoryAccessForAlignment(*DAG.getContext(), DAG.getDataLayout(),
-                                      VT, *Store->getMemOperand())) {
-    return expandUnalignedStore(Store, DAG);
-  }
-
   unsigned AS = Store->getAddressSpace();
   if (Subtarget->hasLDSMisalignedBug() &&
       AS == AMDGPUAS::FLAT_ADDRESS &&
@@ -8566,6 +8562,11 @@ SDValue SITargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
     // v3 stores not supported on SI.
     if (NumElements == 3 && !Subtarget->hasDwordx3LoadStores())
       return SplitVectorStore(Op, DAG);
+
+    if (!allowsMemoryAccessForAlignment(*DAG.getContext(), DAG.getDataLayout(),
+                                        VT, *Store->getMemOperand()))
+      return expandUnalignedStore(Store, DAG);
+
     return SDValue();
   } else if (AS == AMDGPUAS::PRIVATE_ADDRESS) {
     switch (Subtarget->getMaxPrivateElementSize()) {
@@ -8605,6 +8606,13 @@ SDValue SITargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
       return SplitVectorStore(Op, DAG);
     }
 
+    if (!allowsMemoryAccessForAlignment(*DAG.getContext(), DAG.getDataLayout(),
+                                        VT, *Store->getMemOperand())) {
+      if (VT.isVector())
+        return SplitVectorStore(Op, DAG);
+      return expandUnalignedStore(Store, DAG);
+    }
+
     return SDValue();
   } else {
     llvm_unreachable("unhandled address space");
diff --git a/llvm/test/CodeGen/AMDGPU/load-local-redundant-copies.ll b/llvm/test/CodeGen/AMDGPU/load-local-redundant-copies.ll
@@ -7,35 +7,15 @@
 define amdgpu_vs void @test(<4 x i32> inreg %arg1, <6 x float> addrspace(3)* %arg2) {
 ; CHECK-LABEL: test:
 ; CHECK:       ; %bb.0:
-; CHECK-NEXT:    s_mov_b32 s8, s4
-; CHECK-NEXT:    s_mov_b32 s4, SCRATCH_RSRC_DWORD0
-; CHECK-NEXT:    s_mov_b32 s5, SCRATCH_RSRC_DWORD1
-; CHECK-NEXT:    s_mov_b32 s6, -1
-; CHECK-NEXT:    s_mov_b32 s7, 0xe8f000
-; CHECK-NEXT:    s_add_u32 s4, s4, s8
-; CHECK-NEXT:    s_addc_u32 s5, s5, 0
+; CHECK-NEXT:    v_add_i32_e32 v3, vcc, 12, v0
 ; CHECK-NEXT:    v_add_i32_e32 v1, vcc, 8, v0
-; CHECK-NEXT:    v_add_i32_e32 v2, vcc, 12, v0
+; CHECK-NEXT:    v_add_i32_e32 v4, vcc, 4, v0
 ; CHECK-NEXT:    s_mov_b32 m0, -1
-; CHECK-NEXT:    ds_read_b32 v1, v1
-; CHECK-NEXT:    ds_read_b32 v2, v2
-; CHECK-NEXT:    v_add_i32_e32 v3, vcc, 4, v0
+; CHECK-NEXT:    ds_read_b32 v2, v1
+; CHECK-NEXT:    ds_read_b32 v1, v4
 ; CHECK-NEXT:    ds_read_b32 v3, v3
 ; CHECK-NEXT:    ds_read_b32 v0, v0
-; CHECK-NEXT:    s_waitcnt lgkmcnt(2)
-; CHECK-NEXT:    buffer_store_dword v2, off, s[4:7], 0 offset:28
-; CHECK-NEXT:    buffer_store_dword v1, off, s[4:7], 0 offset:24
-; CHECK-NEXT:    s_waitcnt lgkmcnt(1)
-; CHECK-NEXT:    buffer_store_dword v3, off, s[4:7], 0 offset:20
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
-; CHECK-NEXT:    buffer_store_dword v0, off, s[4:7], 0 offset:16
-; CHECK-NEXT:    s_waitcnt expcnt(1)
-; CHECK-NEXT:    buffer_load_dword v3, off, s[4:7], 0 offset:28
-; CHECK-NEXT:    buffer_load_dword v2, off, s[4:7], 0 offset:24
-; CHECK-NEXT:    buffer_load_dword v1, off, s[4:7], 0 offset:20
-; CHECK-NEXT:    s_waitcnt expcnt(0)
-; CHECK-NEXT:    buffer_load_dword v0, off, s[4:7], 0 offset:16
-; CHECK-NEXT:    s_waitcnt vmcnt(0)
 ; CHECK-NEXT:    exp mrt0 off, off, off, off
 ; CHECK-NEXT:    v_mov_b32_e32 v4, 0
 ; CHECK-NEXT:    tbuffer_store_format_xyzw v[0:3], v4, s[0:3], 0 format:[BUF_DATA_FORMAT_32_32_32_32,BUF_NUM_FORMAT_FLOAT] idxen
@@ -50,42 +30,25 @@ define amdgpu_vs void @test(<4 x i32> inreg %arg1, <6 x float> addrspace(3)* %ar
 define amdgpu_vs void @test_2(<4 x i32> inreg %arg1, i32 %arg2, i32 inreg %arg3, <8 x float> addrspace(3)* %arg4) {
 ; CHECK-LABEL: test_2:
 ; CHECK:       ; %bb.0:
-; CHECK-NEXT:    s_mov_b32 s8, SCRATCH_RSRC_DWORD0
-; CHECK-NEXT:    s_mov_b32 s9, SCRATCH_RSRC_DWORD1
-; CHECK-NEXT:    s_mov_b32 s10, -1
-; CHECK-NEXT:    s_mov_b32 s11, 0xe8f000
-; CHECK-NEXT:    s_add_u32 s8, s8, s5
-; CHECK-NEXT:    s_addc_u32 s9, s9, 0
+; CHECK-NEXT:    v_add_i32_e32 v5, vcc, 28, v1
 ; CHECK-NEXT:    v_add_i32_e32 v2, vcc, 24, v1
-; CHECK-NEXT:    v_add_i32_e32 v3, vcc, 28, v1
+; CHECK-NEXT:    v_add_i32_e32 v3, vcc, 20, v1
 ; CHECK-NEXT:    v_add_i32_e32 v6, vcc, 16, v1
-; CHECK-NEXT:    v_add_i32_e32 v7, vcc, 20, v1
+; CHECK-NEXT:    v_add_i32_e32 v7, vcc, 12, v1
 ; CHECK-NEXT:    v_add_i32_e32 v8, vcc, 8, v1
-; CHECK-NEXT:    v_add_i32_e32 v9, vcc, 12, v1
 ; CHECK-NEXT:    v_add_i32_e32 v10, vcc, 4, v1
 ; CHECK-NEXT:    s_mov_b32 m0, -1
 ; CHECK-NEXT:    ds_read_b32 v4, v2
-; CHECK-NEXT:    ds_read_b32 v5, v3
+; CHECK-NEXT:    ds_read_b32 v3, v3
 ; CHECK-NEXT:    ds_read_b32 v2, v6
-; CHECK-NEXT:    ds_read_b32 v3, v7
+; CHECK-NEXT:    ds_read_b32 v9, v7
 ; CHECK-NEXT:    ds_read_b32 v8, v8
-; CHECK-NEXT:    ds_read_b32 v9, v9
 ; CHECK-NEXT:    ds_read_b32 v7, v10
 ; CHECK-NEXT:    ds_read_b32 v6, v1
-; CHECK-NEXT:    s_waitcnt lgkmcnt(6)
-; CHECK-NEXT:    buffer_store_dword v5, off, s[8:11], 0 offset:28
-; CHECK-NEXT:    buffer_store_dword v4, off, s[8:11], 0 offset:24
-; CHECK-NEXT:    s_waitcnt lgkmcnt(4)
-; CHECK-NEXT:    buffer_store_dword v3, off, s[8:11], 0 offset:20
-; CHECK-NEXT:    buffer_store_dword v2, off, s[8:11], 0 offset:16
-; CHECK-NEXT:    s_waitcnt lgkmcnt(2)
-; CHECK-NEXT:    buffer_store_dword v9, off, s[8:11], 0 offset:44
-; CHECK-NEXT:    buffer_store_dword v8, off, s[8:11], 0 offset:40
+; CHECK-NEXT:    ds_read_b32 v5, v5
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(1)
-; CHECK-NEXT:    buffer_store_dword v7, off, s[8:11], 0 offset:36
-; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
-; CHECK-NEXT:    buffer_store_dword v6, off, s[8:11], 0 offset:32
 ; CHECK-NEXT:    tbuffer_store_format_xyzw v[6:9], v0, s[0:3], s4 format:[BUF_DATA_FORMAT_32_32_32,BUF_NUM_FORMAT_UINT] idxen glc slc
+; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
 ; CHECK-NEXT:    tbuffer_store_format_xyzw v[2:5], v0, s[0:3], s4 format:[BUF_DATA_FORMAT_32_32_32,BUF_NUM_FORMAT_UINT] idxen offset:16 glc slc
 ; CHECK-NEXT:    s_endpgm
   %load = load <8 x float>, <8 x float> addrspace(3)* %arg4, align 4
@@ -99,65 +62,42 @@ define amdgpu_vs void @test_2(<4 x i32> inreg %arg1, i32 %arg2, i32 inreg %arg3,
 define amdgpu_vs void @test_3(i32 inreg %arg1, i32 inreg %arg2, <4 x i32> inreg %arg3, i32 %arg4, <6 x float> addrspace(3)* %arg5, <6 x float> addrspace(3)* %arg6) {
 ; CHECK-LABEL: test_3:
 ; CHECK:       ; %bb.0:
-; CHECK-NEXT:    s_mov_b32 s8, SCRATCH_RSRC_DWORD0
-; CHECK-NEXT:    s_mov_b32 s9, SCRATCH_RSRC_DWORD1
-; CHECK-NEXT:    s_mov_b32 s10, -1
-; CHECK-NEXT:    s_mov_b32 s11, 0xe8f000
-; CHECK-NEXT:    s_add_u32 s8, s8, s6
-; CHECK-NEXT:    s_addc_u32 s9, s9, 0
 ; CHECK-NEXT:    s_mov_b32 s7, s5
 ; CHECK-NEXT:    s_mov_b32 s6, s4
 ; CHECK-NEXT:    s_mov_b32 s5, s3
 ; CHECK-NEXT:    s_mov_b32 s4, s2
-; CHECK-NEXT:    v_add_i32_e32 v0, vcc, 8, v1
-; CHECK-NEXT:    v_add_i32_e32 v3, vcc, 12, v1
-; CHECK-NEXT:    v_add_i32_e32 v4, vcc, 4, v1
-; CHECK-NEXT:    v_add_i32_e32 v7, vcc, 16, v1
+; CHECK-NEXT:    v_add_i32_e32 v0, vcc, 16, v1
+; CHECK-NEXT:    v_add_i32_e32 v6, vcc, 12, v1
+; CHECK-NEXT:    v_add_i32_e32 v4, vcc, 8, v1
+; CHECK-NEXT:    v_add_i32_e32 v7, vcc, 4, v1
 ; CHECK-NEXT:    v_add_i32_e32 v8, vcc, 20, v1
 ; CHECK-NEXT:    v_mov_b32_e32 v9, s0
-; CHECK-NEXT:    v_add_i32_e32 v10, vcc, 8, v2
+; CHECK-NEXT:    v_add_i32_e32 v10, vcc, 16, v2
 ; CHECK-NEXT:    v_add_i32_e32 v11, vcc, 12, v2
-; CHECK-NEXT:    v_add_i32_e32 v12, vcc, 4, v2
-; CHECK-NEXT:    v_add_i32_e32 v13, vcc, 16, v2
-; CHECK-NEXT:    v_add_i32_e32 v14, vcc, 20, v2
+; CHECK-NEXT:    v_add_i32_e32 v12, vcc, 8, v2
 ; CHECK-NEXT:    s_mov_b32 m0, -1
-; CHECK-NEXT:    ds_read_b32 v5, v0
-; CHECK-NEXT:    ds_read_b32 v6, v3
-; CHECK-NEXT:    ds_read_b32 v4, v4
-; CHECK-NEXT:    ds_read_b32 v8, v8
-; CHECK-NEXT:    ds_read_b32 v7, v7
 ; CHECK-NEXT:    ds_read_b32 v3, v1
-; CHECK-NEXT:    s_waitcnt lgkmcnt(4)
-; CHECK-NEXT:    buffer_store_dword v6, off, s[8:11], 0 offset:44
-; CHECK-NEXT:    buffer_store_dword v5, off, s[8:11], 0 offset:40
-; CHECK-NEXT:    s_waitcnt lgkmcnt(3)
-; CHECK-NEXT:    buffer_store_dword v4, off, s[8:11], 0 offset:36
-; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
-; CHECK-NEXT:    buffer_store_dword v3, off, s[8:11], 0 offset:32
+; CHECK-NEXT:    ds_read_b32 v5, v4
+; CHECK-NEXT:    ds_read_b32 v4, v7
+; CHECK-NEXT:    ds_read_b32 v1, v8
+; CHECK-NEXT:    ds_read_b32 v6, v6
+; CHECK-NEXT:    ds_read_b32 v0, v0
+; CHECK-NEXT:    v_add_i32_e32 v7, vcc, 4, v2
+; CHECK-NEXT:    v_add_i32_e32 v8, vcc, 20, v2
+; CHECK-NEXT:    s_waitcnt lgkmcnt(1)
 ; CHECK-NEXT:    tbuffer_store_format_xyzw v[3:6], v9, s[4:7], s1 format:[BUF_DATA_FORMAT_32_32_32,BUF_NUM_FORMAT_UINT] idxen offset:264 glc slc
-; CHECK-NEXT:    tbuffer_store_format_xy v[7:8], v9, s[4:7], s1 format:[BUF_DATA_FORMAT_INVALID,BUF_NUM_FORMAT_UINT] idxen offset:280 glc slc
-; CHECK-NEXT:    ds_read_b32 v0, v10
-; CHECK-NEXT:    ds_read_b32 v1, v11
-; CHECK-NEXT:    s_waitcnt expcnt(1)
-; CHECK-NEXT:    ds_read_b32 v3, v12
-; CHECK-NEXT:    ds_read_b32 v4, v13
-; CHECK-NEXT:    ds_read_b32 v2, v2
-; CHECK-NEXT:    s_waitcnt lgkmcnt(3)
-; CHECK-NEXT:    buffer_store_dword v1, off, s[8:11], 0 offset:28
-; CHECK-NEXT:    buffer_store_dword v0, off, s[8:11], 0 offset:24
-; CHECK-NEXT:    s_waitcnt lgkmcnt(2)
-; CHECK-NEXT:    buffer_store_dword v3, off, s[8:11], 0 offset:20
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
-; CHECK-NEXT:    buffer_store_dword v2, off, s[8:11], 0 offset:16
-; CHECK-NEXT:    s_waitcnt expcnt(1)
-; CHECK-NEXT:    buffer_load_dword v3, off, s[8:11], 0 offset:28
+; CHECK-NEXT:    tbuffer_store_format_xy v[0:1], v9, s[4:7], s1 format:[BUF_DATA_FORMAT_INVALID,BUF_NUM_FORMAT_UINT] idxen offset:280 glc slc
 ; CHECK-NEXT:    s_waitcnt expcnt(0)
-; CHECK-NEXT:    buffer_load_dword v2, off, s[8:11], 0 offset:24
-; CHECK-NEXT:    buffer_load_dword v1, off, s[8:11], 0 offset:20
-; CHECK-NEXT:    buffer_load_dword v0, off, s[8:11], 0 offset:16
-; CHECK-NEXT:    ds_read_b32 v5, v14
-; CHECK-NEXT:    s_waitcnt vmcnt(0)
+; CHECK-NEXT:    ds_read_b32 v0, v2
+; CHECK-NEXT:    ds_read_b32 v2, v12
+; CHECK-NEXT:    ds_read_b32 v1, v7
+; CHECK-NEXT:    ds_read_b32 v5, v8
+; CHECK-NEXT:    ds_read_b32 v3, v11
+; CHECK-NEXT:    ds_read_b32 v4, v10
+; CHECK-NEXT:    s_waitcnt lgkmcnt(5)
 ; CHECK-NEXT:    exp mrt0 off, off, off, off
+; CHECK-NEXT:    s_waitcnt lgkmcnt(1)
 ; CHECK-NEXT:    tbuffer_store_format_xyzw v[0:3], v9, s[4:7], s1 format:[BUF_DATA_FORMAT_32_32_32,BUF_NUM_FORMAT_UINT] idxen offset:240 glc slc
 ; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
 ; CHECK-NEXT:    tbuffer_store_format_xy v[4:5], v9, s[4:7], s1 format:[BUF_DATA_FORMAT_INVALID,BUF_NUM_FORMAT_UINT] idxen offset:256 glc slc
