AMDGPU: Avoid using 64-bit shift for i64 (shl x, 32)

[arsenm]

Summary

This can be done only with moves which theoretically
will optimize better later.

Although this transform increases the instruction count,
it should be code size / cycle count neutral in the worst
VALU case. It also seems to slightly improve a couple
of testcases due to other DAG combines this exposes.

This is probably slightly worse for the SALU case, so
it might be better to handle this during moveToVALU,
although then you lose some simplifications like
the load width reducing in the simple testcase.

[arsenm]

Use truncate instead of extract_element x, 0
Add tests for lshr x, 32 case

[arsenm]

ping

[• tstellarAMD]

LGTM

[arsenm]

r242177

lib/Target/AMDGPU/AMDGPUISelLowering.h

   SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
 
   SDValue performStoreCombine(SDNode *N, DAGCombinerInfo &DCI) const;
+  SDValue performShlCombine(SDNode *N, DAGCombinerInfo &DCI) const;
   SDValue performMulCombine(SDNode *N, DAGCombinerInfo &DCI) const;
 
 protected:

lib/Target/AMDGPU/AMDGPUISelLowering.cpp

   setOperationAction(ISD::FNEARBYINT, MVT::f32, Custom);
   setOperationAction(ISD::FNEARBYINT, MVT::f64, Custom);
 
+  setTargetDAGCombine(ISD::SHL);
   setTargetDAGCombine(ISD::MUL);
   setTargetDAGCombine(ISD::SELECT);
   setTargetDAGCombine(ISD::SELECT_CC);

                       SN->getBasePtr(), SN->getMemOperand());
 }
 
+SDValue AMDGPUTargetLowering::performShlCombine(SDNode *N,
+                                                DAGCombinerInfo &DCI) const {
+  if (N->getValueType(0) != MVT::i64)
+    return SDValue();
+
+  // i64 (shl x, 32) -> (build_pair 0, x)
+
+  // Doing this with moves theoretically helps MI optimizations that understand
+  // copies. 2 v_mov_b32_e32 will have the same code size / cycle count as
+  // v_lshl_b64. In the SALU case, I think this is slightly worse since it
+  // doubles the code size and I'm unsure about cycle count.
+  const ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N->getOperand(1));
+  if (!RHS || RHS->getZExtValue() != 32)
+    return SDValue();
+
+  SDValue LHS = N->getOperand(0);
+
+  SDLoc SL(N);
+  SelectionDAG &DAG = DCI.DAG;
+
+  // Extract low 32-bits.
+  SDValue Lo = DAG.getNode(ISD::TRUNCATE, SL, MVT::i32, LHS);
+
+  const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
+  return DAG.getNode(ISD::BUILD_PAIR, SL, MVT::i64, Zero, Lo);
+}
+
 SDValue AMDGPUTargetLowering::performMulCombine(SDNode *N,
                                                 DAGCombinerInfo &DCI) const {
   EVT VT = N->getValueType(0);

   switch(N->getOpcode()) {
   default:
     break;
+  case ISD::SHL: {
+    if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
+      break;
+
+    return performShlCombine(N, DCI);
+  }
   case ISD::MUL:
     return performMulCombine(N, DCI);
   case AMDGPUISD::MUL_I24:

test/CodeGen/AMDGPU/array-ptr-calc-i64.ll

 declare i32 @llvm.SI.tid() readnone
 
 ; SI-LABEL: {{^}}test_array_ptr_calc:
-; SI: v_mul_lo_i32
-; SI: v_mul_hi_i32
+; SI-DAG: v_mul_lo_i32
+; SI-DAG: v_mul_hi_i32
+; SI: s_endpgm
 define void @test_array_ptr_calc(i32 addrspace(1)* noalias %out, [1025 x i32] addrspace(1)* noalias %inA, i32 addrspace(1)* noalias %inB) {
   %tid = call i32 @llvm.SI.tid() readnone
   %a_ptr = getelementptr [1025 x i32], [1025 x i32] addrspace(1)* %inA, i32 %tid, i32 0

test/CodeGen/AMDGPU/mul_uint24.ll

 ; FUNC_LABEL: {{^}}mul24_i64:
 ; EG; MUL_UINT24
 ; EG: MULHI
-; SI: v_mul_u32_u24
 ; FIXME: SI support 24-bit mulhi
-; SI: v_mul_hi_u32
-define void @mul24_i64(i64 addrspace(1)* %out, i64 %a, i64 %b) {
+
+; SI-DAG: v_mul_u32_u24
+; SI-DAG: v_mul_hi_u32
+; SI: s_endpgm
+define void @mul24_i64(i64 addrspace(1)* %out, i64 %a, i64 %b, i64 %c) {
 entry:
-  %0 = shl i64 %a, 40
-  %a_24 = lshr i64 %0, 40
-  %1 = shl i64 %b, 40
-  %b_24 = lshr i64 %1, 40
-  %2 = mul i64 %a_24, %b_24
-  store i64 %2, i64 addrspace(1)* %out
+  %tmp0 = shl i64 %a, 40
+  %a_24 = lshr i64 %tmp0, 40
+  %tmp1 = shl i64 %b, 40
+  %b_24 = lshr i64 %tmp1, 40
+  %tmp2 = mul i64 %a_24, %b_24
+  store i64 %tmp2, i64 addrspace(1)* %out
   ret void
 }

test/CodeGen/AMDGPU/shl.ll

-;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck --check-prefix=EG %s
-;RUN: llc < %s -march=amdgcn -mcpu=verde -verify-machineinstrs | FileCheck --check-prefix=SI %s
-;RUN: llc < %s -march=amdgcn -mcpu=tonga -verify-machineinstrs | FileCheck --check-prefix=VI %s
+; RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck --check-prefix=EG %s
+; RUN: llc < %s -march=amdgcn -mcpu=verde -verify-machineinstrs | FileCheck -check-prefix=GCN -check-prefix=SI %s
+; XUN: llc < %s -march=amdgcn -mcpu=tonga -verify-machineinstrs | FileCheck -check-prefix=GCN -check-prefix=VI %s
+
+declare i32 @llvm.r600.read.tidig.x() #0
+
 
 ;EG: {{^}}shl_v2i32:
 ;EG: LSHL {{\*? *}}T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}

   store <4 x i64> %result, <4 x i64> addrspace(1)* %out
   ret void
 }
+
+; Make sure load width gets reduced to i32 load.
+; GCN-LABEL: {{^}}s_shl_32_i64:
+; GCN-DAG: s_load_dword [[LO_A:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xb{{$}}
+; GCN-DAG: s_mov_b32 s[[SLO:[0-9]+]], 0{{$}}
+; GCN-DAG: v_mov_b32_e32 v[[VLO:[0-9]+]], s[[SLO]]
+; GCN-DAG: v_mov_b32_e32 v[[VHI:[0-9]+]], [[LO_A]]
+; GCN: buffer_store_dwordx2 v{{\[}}[[VLO]]:[[VHI]]{{\]}}
+define void @s_shl_32_i64(i64 addrspace(1)* %out, i64 %a) {
+  %result = shl i64 %a, 32
+  store i64 %result, i64 addrspace(1)* %out
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_shl_32_i64:
+; GCN-DAG: buffer_load_dword v[[LO_A:[0-9]+]],
+; GCN-DAG: v_mov_b32_e32 v[[VLO:[0-9]+]], 0{{$}}
+; GCN: buffer_store_dwordx2 v{{\[}}[[VLO]]:[[LO_A]]{{\]}}
+define void @v_shl_32_i64(i64 addrspace(1)* %out, i64 addrspace(1)* %in) {
+  %tid = call i32 @llvm.r600.read.tidig.x() #0
+  %gep.in = getelementptr i64, i64 addrspace(1)* %in, i32 %tid
+  %gep.out = getelementptr i64, i64 addrspace(1)* %out, i32 %tid
+  %a = load i64, i64 addrspace(1)* %gep.in
+  %result = shl i64 %a, 32
+  store i64 %result, i64 addrspace(1)* %gep.out
+  ret void
+}
+
+attributes #0 = { nounwind readnone }

test/CodeGen/AMDGPU/srl.ll

-; RUN: llc -march=amdgcn -mcpu=verde -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
-; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=VI -check-prefix=FUNC %s
+; RUN: llc -march=amdgcn -mcpu=verde -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=GCN -check-prefix=FUNC %s
+; XUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=VI -check-prefix=GCN -check-prefix=FUNC %s
 ; RUN: llc -march=r600 -mcpu=redwood < %s | FileCheck -check-prefix=EG -check-prefix=FUNC %s
 
+declare i32 @llvm.r600.read.tidig.x() #0
+
 ; FUNC-LABEL: {{^}}lshr_i32:
 ; SI: v_lshrrev_b32_e32 v{{[0-9]+, v[0-9]+, v[0-9]+}}
 ; VI: v_lshrrev_b32_e32 v{{[0-9]+, v[0-9]+, v[0-9]+}}

   store <4 x i64> %result, <4 x i64> addrspace(1)* %out
   ret void
 }
+
+; Make sure load width gets reduced to i32 load.
+; GCN-LABEL: {{^}}s_lshr_32_i64:
+; GCN-DAG: s_load_dword [[HI_A:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xc{{$}}
+; GCN-DAG: s_mov_b32 s[[SHI:[0-9]+]], 0{{$}}
+; GCN-DAG: v_mov_b32_e32 v[[VHI:[0-9]+]], s[[SHI]]
+; GCN-DAG: v_mov_b32_e32 v[[VLO:[0-9]+]], [[HI_A]]
+; GCN: buffer_store_dwordx2 v{{\[}}[[VLO]]:[[VHI]]{{\]}}
+define void @s_lshr_32_i64(i64 addrspace(1)* %out, i64 %a) {
+  %result = lshr i64 %a, 32
+  store i64 %result, i64 addrspace(1)* %out
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_lshr_32_i64:
+; GCN-DAG: buffer_load_dword v[[HI_A:[0-9]+]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
+; GCN-DAG: v_mov_b32_e32 v[[VHI:[0-9]+]], 0{{$}}
+; GCN: buffer_store_dwordx2 v{{\[}}[[HI_A]]:[[VHI]]{{\]}}
+define void @v_lshr_32_i64(i64 addrspace(1)* %out, i64 addrspace(1)* %in) {
+  %tid = call i32 @llvm.r600.read.tidig.x() #0
+  %gep.in = getelementptr i64, i64 addrspace(1)* %in, i32 %tid
+  %gep.out = getelementptr i64, i64 addrspace(1)* %out, i32 %tid
+  %a = load i64, i64 addrspace(1)* %gep.in
+  %result = lshr i64 %a, 32
+  store i64 %result, i64 addrspace(1)* %gep.out
+  ret void
+}
+
+attributes #0 = { nounwind readnone }