From 2b957b5a6f228797a871426d48098fee479883ef Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault@amd.com>
Date: Mon, 2 May 2016 20:07:26 +0000
Subject: [PATCH] AMDGPU: Make i64 loads/stores promote to v2i32

Now that unaligned access expansion should not attempt
to produce i64 accesses, we can remove the hack in
PreprocessISelDAG where this is done.

This allows splitting i64 private accesses while
allowing the new add nodes indexing the vector components
can be folded with the base pointer arithmetic.

llvm-svn: 268293
---
 llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp      | 55 ----------------------
 llvm/lib/Target/AMDGPU/SIISelLowering.cpp          | 12 +++++
 .../CodeGen/AMDGPU/bitreverse-inline-immediates.ll |  4 +-
 llvm/test/CodeGen/AMDGPU/fcanonicalize.ll          |  6 +--
 llvm/test/CodeGen/AMDGPU/imm.ll                    |  2 +-
 llvm/test/CodeGen/AMDGPU/shift-and-i128-ubfe.ll    | 20 ++++----
 6 files changed, 26 insertions(+), 73 deletions(-)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
index 015cc8f..8d78bcf 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
@@ -1586,61 +1586,6 @@ void AMDGPUDAGToDAGISel::PreprocessISelDAG() {
       }
     }
   }
-
-  // XXX - Other targets seem to be able to do this without a worklist.
-  SmallVector<LoadSDNode *, 8> LoadsToReplace;
-  SmallVector<StoreSDNode *, 8> StoresToReplace;
-
-  for (SDNode &Node : CurDAG->allnodes()) {
-    if (LoadSDNode *LD = dyn_cast<LoadSDNode>(&Node)) {
-      EVT VT = LD->getValueType(0);
-      if (VT != MVT::i64 || LD->getExtensionType() != ISD::NON_EXTLOAD)
-        continue;
-
-      // To simplify the TableGen patters, we replace all i64 loads with v2i32
-      // loads.  Alternatively, we could promote i64 loads to v2i32 during DAG
-      // legalization, however, so places (ExpandUnalignedLoad) in the DAG
-      // legalizer assume that if i64 is legal, so doing this promotion early
-      // can cause problems.
-      LoadsToReplace.push_back(LD);
-    } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(&Node)) {
-      // Handle i64 stores here for the same reason mentioned above for loads.
-      SDValue Value = ST->getValue();
-      if (Value.getValueType() != MVT::i64 || ST->isTruncatingStore())
-        continue;
-      StoresToReplace.push_back(ST);
-    }
-  }
-
-  for (LoadSDNode *LD : LoadsToReplace) {
-    SDLoc SL(LD);
-
-    SDValue NewLoad = CurDAG->getLoad(MVT::v2i32, SL, LD->getChain(),
-                                      LD->getBasePtr(), LD->getMemOperand());
-    SDValue BitCast = CurDAG->getNode(ISD::BITCAST, SL,
-                                      MVT::i64, NewLoad);
-    CurDAG->ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLoad.getValue(1));
-    CurDAG->ReplaceAllUsesOfValueWith(SDValue(LD, 0), BitCast);
-    Modified = true;
-  }
-
-  for (StoreSDNode *ST : StoresToReplace) {
-    SDValue NewValue = CurDAG->getNode(ISD::BITCAST, SDLoc(ST),
-                                       MVT::v2i32, ST->getValue());
-    const SDValue StoreOps[] = {
-      ST->getChain(),
-      NewValue,
-      ST->getBasePtr(),
-      ST->getOffset()
-    };
-
-    CurDAG->UpdateNodeOperands(ST, StoreOps);
-    Modified = true;
-  }
-
-  // XXX - Is this necessary?
-  if (Modified)
-    CurDAG->RemoveDeadNodes();
 }
 
 void AMDGPUDAGToDAGISel::PostprocessISelDAG() {
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index a79f689..ec0a032 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -95,12 +95,24 @@ SITargetLowering::SITargetLowering(TargetMachine &TM,
   setOperationAction(ISD::LOAD, MVT::v8i32, Custom);
   setOperationAction(ISD::LOAD, MVT::v16i32, Custom);
 
+  setOperationAction(ISD::LOAD, MVT::f64, Promote);
+  AddPromotedToType(ISD::LOAD, MVT::f64, MVT::v2i32);
+
+  setOperationAction(ISD::LOAD, MVT::i64, Promote);
+  AddPromotedToType(ISD::LOAD, MVT::i64, MVT::v2i32);
+
   setOperationAction(ISD::STORE, MVT::v8i32, Custom);
   setOperationAction(ISD::STORE, MVT::v16i32, Custom);
 
   setOperationAction(ISD::STORE, MVT::i1, Custom);
   setOperationAction(ISD::STORE, MVT::v4i32, Custom);
 
+  setOperationAction(ISD::STORE, MVT::f64, Promote);
+  AddPromotedToType(ISD::STORE, MVT::f64, MVT::v2i32);
+
+  setOperationAction(ISD::STORE, MVT::i64, Promote);
+  AddPromotedToType(ISD::STORE, MVT::i64, MVT::v2i32);
+
   setOperationAction(ISD::SELECT, MVT::i64, Custom);
   setOperationAction(ISD::SELECT, MVT::f64, Promote);
   AddPromotedToType(ISD::SELECT, MVT::f64, MVT::i64);
diff --git a/llvm/test/CodeGen/AMDGPU/bitreverse-inline-immediates.ll b/llvm/test/CodeGen/AMDGPU/bitreverse-inline-immediates.ll
index 0d72adc..150e343 100644
--- a/llvm/test/CodeGen/AMDGPU/bitreverse-inline-immediates.ll
+++ b/llvm/test/CodeGen/AMDGPU/bitreverse-inline-immediates.ll
@@ -14,7 +14,7 @@ define void @materialize_0_i32(i32 addrspace(1)* %out) {
 
 ; GCN-LABEL: {{^}}materialize_0_i64:
 ; GCN: v_mov_b32_e32 v[[LOK:[0-9]+]], 0{{$}}
-; GCN: v_mov_b32_e32 v[[HIK:[0-9]+]], 0{{$}}
+; GCN: v_mov_b32_e32 v[[HIK:[0-9]+]], v[[LOK]]{{$}}
 ; GCN: buffer_store_dwordx2 v{{\[}}[[LOK]]:[[HIK]]{{\]}}
 define void @materialize_0_i64(i64 addrspace(1)* %out) {
   store i64 0, i64 addrspace(1)* %out
@@ -31,7 +31,7 @@ define void @materialize_neg1_i32(i32 addrspace(1)* %out) {
 
 ; GCN-LABEL: {{^}}materialize_neg1_i64:
 ; GCN: v_mov_b32_e32 v[[LOK:[0-9]+]], -1{{$}}
-; GCN: v_mov_b32_e32 v[[HIK:[0-9]+]], -1{{$}}
+; GCN: v_mov_b32_e32 v[[HIK:[0-9]+]], v[[LOK]]{{$}}
 ; GCN: buffer_store_dwordx2 v{{\[}}[[LOK]]:[[HIK]]{{\]}}
 define void @materialize_neg1_i64(i64 addrspace(1)* %out) {
   store i64 -1, i64 addrspace(1)* %out
diff --git a/llvm/test/CodeGen/AMDGPU/fcanonicalize.ll b/llvm/test/CodeGen/AMDGPU/fcanonicalize.ll
index 56b4460..a67e54d 100644
--- a/llvm/test/CodeGen/AMDGPU/fcanonicalize.ll
+++ b/llvm/test/CodeGen/AMDGPU/fcanonicalize.ll
@@ -172,7 +172,7 @@ define void @s_test_canonicalize_var_f64(double addrspace(1)* %out, double %val)
 
 ; GCN-LABEL: {{^}}test_fold_canonicalize_p0_f64:
 ; GCN: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}}
-; GCN: v_mov_b32_e32 v[[HI:[0-9]+]], 0{{$}}
+; GCN: v_mov_b32_e32 v[[HI:[0-9]+]], v[[LO]]{{$}}
 ; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}}
 define void @test_fold_canonicalize_p0_f64(double addrspace(1)* %out) #1 {
   %canonicalized = call double @llvm.canonicalize.f64(double 0.0)
@@ -225,7 +225,7 @@ define void @test_fold_canonicalize_literal_f64(double addrspace(1)* %out) #1 {
 ; DENORM-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0xfffff{{$}}
 
 ; NODENORM: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}}
-; NODENORM: v_mov_b32_e32 v[[HI:[0-9]+]], 0{{$}}
+; NODENORM: v_mov_b32_e32 v[[HI:[0-9]+]], v[[LO]]{{$}}
 ; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}}
 define void @test_fold_canonicalize_denormal0_f64(double addrspace(1)* %out) #1 {
   %canonicalized = call double @llvm.canonicalize.f64(double bitcast (i64 4503599627370495 to double))
@@ -238,7 +238,7 @@ define void @test_fold_canonicalize_denormal0_f64(double addrspace(1)* %out) #1
 ; DENORM-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0x800fffff{{$}}
 
 ; NODENORM: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}}
-; NODENORM: v_mov_b32_e32 v[[HI:[0-9]+]], 0{{$}}
+; NODENORM: v_mov_b32_e32 v[[HI:[0-9]+]], v[[LO]]{{$}}
 ; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}}
 define void @test_fold_canonicalize_denormal1_f64(double addrspace(1)* %out) #1 {
   %canonicalized = call double @llvm.canonicalize.f64(double bitcast (i64 9227875636482146303 to double))
diff --git a/llvm/test/CodeGen/AMDGPU/imm.ll b/llvm/test/CodeGen/AMDGPU/imm.ll
index 298cb41..674ecee 100644
--- a/llvm/test/CodeGen/AMDGPU/imm.ll
+++ b/llvm/test/CodeGen/AMDGPU/imm.ll
@@ -510,7 +510,7 @@ define void @add_inline_imm_64_f64(double addrspace(1)* %out, double %x) {
 
 ; CHECK-LABEL: {{^}}store_inline_imm_0.0_f64:
 ; CHECK: v_mov_b32_e32 v[[LO_VREG:[0-9]+]], 0
-; CHECK: v_mov_b32_e32 v[[HI_VREG:[0-9]+]], 0
+; CHECK: v_mov_b32_e32 v[[HI_VREG:[0-9]+]], v[[LO_VREG]]{{$}}
 ; CHECK: buffer_store_dwordx2 v{{\[}}[[LO_VREG]]:[[HI_VREG]]{{\]}}
 define void @store_inline_imm_0.0_f64(double addrspace(1)* %out) {
   store double 0.0, double addrspace(1)* %out
diff --git a/llvm/test/CodeGen/AMDGPU/shift-and-i128-ubfe.ll b/llvm/test/CodeGen/AMDGPU/shift-and-i128-ubfe.ll
index 60e54b9..ea7357d 100644
--- a/llvm/test/CodeGen/AMDGPU/shift-and-i128-ubfe.ll
+++ b/llvm/test/CodeGen/AMDGPU/shift-and-i128-ubfe.ll
@@ -4,12 +4,11 @@
 ; GCN-LABEL: {{^}}v_uextract_bit_31_i128:
 ; GCN: buffer_load_dword [[VAL:v[0-9]+]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
 
-; GCN: v_mov_b32_e32 v[[ZERO1:[0-9]+]], 0{{$}}
 ; GCN: v_mov_b32_e32 v[[ZERO0:[0-9]+]], 0{{$}}
-; GCN: v_mov_b32_e32 v[[ZERO2:[0-9]+]], 0{{$}}
+; GCN: v_mov_b32_e32 v[[ZERO1:[0-9]+]], v[[ZERO0]]{{$}}
 ; GCN-DAG: v_lshrrev_b32_e32 v[[SHIFT:[0-9]+]], 31, [[VAL]]
 
-; GCN-DAG: buffer_store_dwordx2 v{{\[}}[[ZERO1]]:[[ZERO2]]{{\]}}, v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
+; GCN-DAG: buffer_store_dwordx2 v{{\[}}[[ZERO0]]:[[ZERO1]]{{\]}}, v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
 ; GCN-DAG: buffer_store_dwordx2 v{{\[}}[[SHIFT]]:[[ZERO0]]{{\]}}, v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
 ; GCN: s_endpgm
 define void @v_uextract_bit_31_i128(i128 addrspace(1)* %out, i128 addrspace(1)* %in) #1 {
@@ -27,12 +26,11 @@ define void @v_uextract_bit_31_i128(i128 addrspace(1)* %out, i128 addrspace(1)*
 ; GCN-LABEL: {{^}}v_uextract_bit_63_i128:
 ; GCN: buffer_load_dword [[VAL:v[0-9]+]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4{{$}}
 
-; GCN: v_mov_b32_e32 v[[ZERO1:[0-9]+]], 0{{$}}
-; GCN: v_mov_b32_e32 v[[ZERO2:[0-9]+]], 0{{$}}
 ; GCN: v_mov_b32_e32 v[[ZERO0:[0-9]+]], 0{{$}}
+; GCN: v_mov_b32_e32 v[[ZERO1:[0-9]+]], v[[ZERO0]]{{$}}
 ; GCN-DAG: v_lshrrev_b32_e32 v[[SHIFT:[0-9]+]], 31, [[VAL]]
 
-; GCN-DAG: buffer_store_dwordx2 v{{\[}}[[ZERO1]]:[[ZERO2]]{{\]}}, v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
+; GCN-DAG: buffer_store_dwordx2 v{{\[}}[[ZERO0]]:[[ZERO1]]{{\]}}, v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
 ; GCN-DAG: buffer_store_dwordx2 v{{\[}}[[SHIFT]]:[[ZERO0]]{{\]}}, v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
 ; GCN: s_endpgm
 define void @v_uextract_bit_63_i128(i128 addrspace(1)* %out, i128 addrspace(1)* %in) #1 {
@@ -50,12 +48,11 @@ define void @v_uextract_bit_63_i128(i128 addrspace(1)* %out, i128 addrspace(1)*
 ; GCN-LABEL: {{^}}v_uextract_bit_95_i128:
 ; GCN: buffer_load_dword [[VAL:v[0-9]+]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
 
-; GCN: v_mov_b32_e32 v[[ZERO1:[0-9]+]], 0{{$}}
-; GCN: v_mov_b32_e32 v[[ZERO2:[0-9]+]], 0{{$}}
 ; GCN: v_mov_b32_e32 v[[ZERO0:[0-9]+]], 0{{$}}
+; GCN: v_mov_b32_e32 v[[ZERO1:[0-9]+]], v[[ZERO0]]{{$}}
 ; GCN-DAG: v_lshrrev_b32_e32 v[[SHIFT:[0-9]+]], 31, [[VAL]]
 
-; GCN-DAG: buffer_store_dwordx2 v{{\[}}[[ZERO1]]:[[ZERO2]]{{\]}}, v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
+; GCN-DAG: buffer_store_dwordx2 v{{\[}}[[ZERO0]]:[[ZERO1]]{{\]}}, v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
 ; GCN-DAG: buffer_store_dwordx2 v{{\[}}[[SHIFT]]:[[ZERO0]]{{\]}}, v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
 ; GCN: s_endpgm
 define void @v_uextract_bit_95_i128(i128 addrspace(1)* %out, i128 addrspace(1)* %in) #1 {
@@ -73,12 +70,11 @@ define void @v_uextract_bit_95_i128(i128 addrspace(1)* %out, i128 addrspace(1)*
 ; GCN-LABEL: {{^}}v_uextract_bit_127_i128:
 ; GCN: buffer_load_dword [[VAL:v[0-9]+]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:12{{$}}
 
-; GCN: v_mov_b32_e32 v[[ZERO1:[0-9]+]], 0{{$}}
-; GCN: v_mov_b32_e32 v[[ZERO2:[0-9]+]], 0{{$}}
 ; GCN: v_mov_b32_e32 v[[ZERO0:[0-9]+]], 0{{$}}
+; GCN: v_mov_b32_e32 v[[ZERO1:[0-9]+]], v[[ZERO0]]{{$}}
 ; GCN-DAG: v_lshrrev_b32_e32 v[[SHIFT:[0-9]+]], 31, [[VAL]]
 
-; GCN-DAG: buffer_store_dwordx2 v{{\[}}[[ZERO1]]:[[ZERO2]]{{\]}}, v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
+; GCN-DAG: buffer_store_dwordx2 v{{\[}}[[ZERO0]]:[[ZERO1]]{{\]}}, v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8{{$}}
 ; GCN-DAG: buffer_store_dwordx2 v{{\[}}[[SHIFT]]:[[ZERO0]]{{\]}}, v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
 ; GCN: s_endpgm
 define void @v_uextract_bit_127_i128(i128 addrspace(1)* %out, i128 addrspace(1)* %in) #1 {
-- 
2.7.4