AMDGPU/GlobalISel: RegBankSelect for llvm.amdgcn.s.buffer.load

llvm-svn: 361023

diff --git a/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
index 6a3e216a4f5b5c1befb03f2ff1946de99b86663d..bee39c957bfe45db2d6845ea41b6b5ebfbb0d376 100644 (file)
--- a/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
@@ -163,6 +163,25 @@ AMDGPURegisterBankInfo::getInstrAlternativeMappingsIntrinsicWSideEffects(
     const std::array<unsigned, 3> RegSrcOpIdx = { 2, 3, 4 };
     return addMappingFromTable<3>(MI, MRI, RegSrcOpIdx, makeArrayRef(Table));
   }
+  case Intrinsic::amdgcn_s_buffer_load: {
+    static const OpRegBankEntry<2> Table[4] = {
+      // Perfectly legal.
+      { { AMDGPU::SGPRRegBankID, AMDGPU::SGPRRegBankID }, 1 },
+
+      // Only need 1 register in loop
+      { { AMDGPU::SGPRRegBankID, AMDGPU::VGPRRegBankID }, 300 },
+
+      // Have to waterfall the resource.
+      { { AMDGPU::VGPRRegBankID, AMDGPU::SGPRRegBankID }, 1000 },
+
+      // Have to waterfall the resource, and the offset.
+      { { AMDGPU::VGPRRegBankID, AMDGPU::VGPRRegBankID }, 1500 }
+    };
+
+    // rsrc, offset
+    const std::array<unsigned, 2> RegSrcOpIdx = { 2, 3 };
+    return addMappingFromTable<2>(MI, MRI, RegSrcOpIdx, makeArrayRef(Table));
+  }
   default:
     return RegisterBankInfo::getInstrAlternativeMappings(MI);
   }
@@ -818,6 +837,10 @@ void AMDGPURegisterBankInfo::applyMappingImpl(
       executeInWaterfallLoop(MI, MRI, { 2 });
       return;
     }
+    case Intrinsic::amdgcn_s_buffer_load: {
+      executeInWaterfallLoop(MI, MRI, { 2, 3 });
+      return;
+    }
     default:
       break;
     }
@@ -1305,6 +1328,27 @@ AMDGPURegisterBankInfo::getInstrMapping(const MachineInstr &MI) const {
       OpdsMapping[6] = nullptr;
       break;
     }
+    case Intrinsic::amdgcn_s_buffer_load: {
+      unsigned RSrc = MI.getOperand(2).getReg();   // SGPR
+      unsigned Offset = MI.getOperand(3).getReg(); // SGPR/imm
+
+      unsigned Size0 = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits();
+      unsigned Size2 = MRI.getType(RSrc).getSizeInBits();
+      unsigned Size3 = MRI.getType(Offset).getSizeInBits();
+
+      unsigned RSrcBank = getRegBankID(RSrc, MRI, *TRI);
+      unsigned OffsetBank = getRegBankID(Offset, MRI, *TRI);
+
+      OpdsMapping[0] = AMDGPU::getValueMapping(AMDGPU::SGPRRegBankID, Size0);
+      OpdsMapping[1] = nullptr; // intrinsic id
+
+      // Lie and claim everything is legal, even though some need to be
+      // SGPRs. applyMapping will have to deal with it as a waterfall loop.
+      OpdsMapping[2] = AMDGPU::getValueMapping(RSrcBank, Size2); // rsrc
+      OpdsMapping[3] = AMDGPU::getValueMapping(OffsetBank, Size3);
+      OpdsMapping[4] = nullptr;
+      break;
+    }
     }
 
     break;

diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-amdgcn-s-buffer-load.mir b/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-amdgcn-s-buffer-load.mir
new file mode 100644 (file)
index 0000000..bf0bd79
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-amdgcn-s-buffer-load.mir
@@ -0,0 +1,151 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -march=amdgcn -mcpu=fiji -run-pass=regbankselect %s -verify-machineinstrs -o - -regbankselect-fast | FileCheck %s
+# RUN: llc -march=amdgcn -mcpu=fiji -run-pass=regbankselect %s -verify-machineinstrs -o - -regbankselect-greedy | FileCheck %s
+
+---
+name: buffer_load_ss
+legalized: true
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4
+
+    ; CHECK-LABEL: name: buffer_load_ss
+    ; CHECK: liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4
+    ; CHECK: [[COPY:%[0-9]+]]:sgpr(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
+    ; CHECK: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr4
+    ; CHECK: [[INT:%[0-9]+]]:sgpr(<4 x s32>) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.s.buffer.load), [[COPY]](<4 x s32>), [[COPY1]](s32), 0
+    %0:_(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
+    %1:_(s32) = COPY $sgpr4
+    %2:_(<4 x s32>) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.s.buffer.load), %0, %1, 0
+
+...
+
+---
+name: buffer_load_sv
+legalized: true
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $vgpr0
+
+    ; CHECK-LABEL: name: buffer_load_sv
+    ; CHECK: successors: %bb.1(0x80000000)
+    ; CHECK: liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $vgpr0
+    ; CHECK: [[COPY:%[0-9]+]]:sgpr(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
+    ; CHECK: [[COPY1:%[0-9]+]]:vgpr_32(s32) = COPY $vgpr0
+    ; CHECK: [[DEF:%[0-9]+]]:sgpr(<4 x s32>) = G_IMPLICIT_DEF
+    ; CHECK: [[DEF1:%[0-9]+]]:sreg_64_xexec = IMPLICIT_DEF
+    ; CHECK: [[S_MOV_B64_term:%[0-9]+]]:sreg_64_xexec = S_MOV_B64_term $exec
+    ; CHECK: .1:
+    ; CHECK: successors: %bb.2(0x40000000), %bb.1(0x40000000)
+    ; CHECK: [[PHI:%[0-9]+]]:sreg_64 = PHI [[DEF1]], %bb.0, %8, %bb.1
+    ; CHECK: [[PHI1:%[0-9]+]]:sgpr(<4 x s32>) = G_PHI [[DEF]](<4 x s32>), %bb.0, %2(<4 x s32>), %bb.1
+    ; CHECK: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0(s32) = V_READFIRSTLANE_B32 [[COPY1]](s32), implicit $exec
+    ; CHECK: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64 = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_]](s32), [[COPY1]](s32), implicit $exec
+    ; CHECK: [[INT:%[0-9]+]]:sgpr(<4 x s32>) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.s.buffer.load), [[COPY]](<4 x s32>), [[V_READFIRSTLANE_B32_]](s32), 0
+    ; CHECK: [[S_AND_SAVEEXEC_B64_:%[0-9]+]]:sreg_64 = S_AND_SAVEEXEC_B64 killed [[V_CMP_EQ_U32_e64_]], implicit-def $exec, implicit-def $scc, implicit $exec
+    ; CHECK: $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc
+    ; CHECK: S_CBRANCH_EXECNZ %bb.1, implicit $exec
+    ; CHECK: .2:
+    ; CHECK: successors: %bb.3(0x80000000)
+    ; CHECK: $exec = S_MOV_B64_term [[S_MOV_B64_term]]
+    ; CHECK: .3:
+    %0:_(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
+    %1:_(s32) = COPY $vgpr0
+    %2:_(<4 x s32>) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.s.buffer.load), %0, %1, 0
+
+...
+
+---
+name: buffer_load_vs
+legalized: true
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1_vgpr2_vgpr3, $sgpr0
+
+    ; CHECK-LABEL: name: buffer_load_vs
+    ; CHECK: successors: %bb.1(0x80000000)
+    ; CHECK: liveins: $vgpr0_vgpr1_vgpr2_vgpr3, $sgpr0
+    ; CHECK: [[COPY:%[0-9]+]]:vgpr(<4 x s32>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
+    ; CHECK: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
+    ; CHECK: [[DEF:%[0-9]+]]:sgpr(<4 x s32>) = G_IMPLICIT_DEF
+    ; CHECK: [[DEF1:%[0-9]+]]:sreg_64_xexec = IMPLICIT_DEF
+    ; CHECK: [[UV:%[0-9]+]]:vreg_64(s64), [[UV1:%[0-9]+]]:vreg_64(s64) = G_UNMERGE_VALUES [[COPY]](<4 x s32>)
+    ; CHECK: [[S_MOV_B64_term:%[0-9]+]]:sreg_64_xexec = S_MOV_B64_term $exec
+    ; CHECK: .1:
+    ; CHECK: successors: %bb.2(0x40000000), %bb.1(0x40000000)
+    ; CHECK: [[PHI:%[0-9]+]]:sreg_64 = PHI [[DEF1]], %bb.0, %8, %bb.1
+    ; CHECK: [[PHI1:%[0-9]+]]:sgpr(<4 x s32>) = G_PHI [[DEF]](<4 x s32>), %bb.0, %2(<4 x s32>), %bb.1
+    ; CHECK: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0(s32) = V_READFIRSTLANE_B32 [[UV]].sub0(s64), implicit $exec
+    ; CHECK: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0(s32) = V_READFIRSTLANE_B32 [[UV]].sub1(s64), implicit $exec
+    ; CHECK: [[MV:%[0-9]+]]:sreg_64_xexec(s64) = G_MERGE_VALUES [[V_READFIRSTLANE_B32_]](s32), [[V_READFIRSTLANE_B32_1]](s32)
+    ; CHECK: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[MV]](s64), [[UV]](s64), implicit $exec
+    ; CHECK: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0(s32) = V_READFIRSTLANE_B32 [[UV1]].sub0(s64), implicit $exec
+    ; CHECK: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0(s32) = V_READFIRSTLANE_B32 [[UV1]].sub1(s64), implicit $exec
+    ; CHECK: [[MV1:%[0-9]+]]:sreg_64_xexec(s64) = G_MERGE_VALUES [[V_READFIRSTLANE_B32_2]](s32), [[V_READFIRSTLANE_B32_3]](s32)
+    ; CHECK: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[MV1]](s64), [[UV1]](s64), implicit $exec
+    ; CHECK: [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_1]], [[V_CMP_EQ_U64_e64_]], implicit-def $scc
+    ; CHECK: [[BUILD_VECTOR:%[0-9]+]]:sgpr(<4 x s32>) = G_BUILD_VECTOR [[V_READFIRSTLANE_B32_]](s32), [[V_READFIRSTLANE_B32_1]](s32), [[V_READFIRSTLANE_B32_2]](s32), [[V_READFIRSTLANE_B32_3]](s32)
+    ; CHECK: [[INT:%[0-9]+]]:sgpr(<4 x s32>) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.s.buffer.load), [[BUILD_VECTOR]](<4 x s32>), [[COPY1]](s32), 0
+    ; CHECK: [[S_AND_SAVEEXEC_B64_:%[0-9]+]]:sreg_64 = S_AND_SAVEEXEC_B64 killed [[S_AND_B64_]], implicit-def $exec, implicit-def $scc, implicit $exec
+    ; CHECK: $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc
+    ; CHECK: S_CBRANCH_EXECNZ %bb.1, implicit $exec
+    ; CHECK: .2:
+    ; CHECK: successors: %bb.3(0x80000000)
+    ; CHECK: $exec = S_MOV_B64_term [[S_MOV_B64_term]]
+    ; CHECK: .3:
+    %0:_(<4 x s32>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
+    %1:_(s32) = COPY $sgpr0
+    %2:_(<4 x s32>) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.s.buffer.load), %0, %1, 0
+
+...
+
+---
+name: buffer_load_vv
+legalized: true
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1_vgpr2_vgpr3, $vgpr4
+
+    ; CHECK-LABEL: name: buffer_load_vv
+    ; CHECK: successors: %bb.1(0x80000000)
+    ; CHECK: liveins: $vgpr0_vgpr1_vgpr2_vgpr3, $vgpr4
+    ; CHECK: [[COPY:%[0-9]+]]:vgpr(<4 x s32>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
+    ; CHECK: [[COPY1:%[0-9]+]]:vgpr_32(s32) = COPY $vgpr4
+    ; CHECK: [[DEF:%[0-9]+]]:sgpr(<4 x s32>) = G_IMPLICIT_DEF
+    ; CHECK: [[DEF1:%[0-9]+]]:sreg_64_xexec = IMPLICIT_DEF
+    ; CHECK: [[UV:%[0-9]+]]:vreg_64(s64), [[UV1:%[0-9]+]]:vreg_64(s64) = G_UNMERGE_VALUES [[COPY]](<4 x s32>)
+    ; CHECK: [[S_MOV_B64_term:%[0-9]+]]:sreg_64_xexec = S_MOV_B64_term $exec
+    ; CHECK: .1:
+    ; CHECK: successors: %bb.2(0x40000000), %bb.1(0x40000000)
+    ; CHECK: [[PHI:%[0-9]+]]:sreg_64 = PHI [[DEF1]], %bb.0, %8, %bb.1
+    ; CHECK: [[PHI1:%[0-9]+]]:sgpr(<4 x s32>) = G_PHI [[DEF]](<4 x s32>), %bb.0, %2(<4 x s32>), %bb.1
+    ; CHECK: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0(s32) = V_READFIRSTLANE_B32 [[UV]].sub0(s64), implicit $exec
+    ; CHECK: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0(s32) = V_READFIRSTLANE_B32 [[UV]].sub1(s64), implicit $exec
+    ; CHECK: [[MV:%[0-9]+]]:sreg_64_xexec(s64) = G_MERGE_VALUES [[V_READFIRSTLANE_B32_]](s32), [[V_READFIRSTLANE_B32_1]](s32)
+    ; CHECK: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[MV]](s64), [[UV]](s64), implicit $exec
+    ; CHECK: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0(s32) = V_READFIRSTLANE_B32 [[UV1]].sub0(s64), implicit $exec
+    ; CHECK: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0(s32) = V_READFIRSTLANE_B32 [[UV1]].sub1(s64), implicit $exec
+    ; CHECK: [[MV1:%[0-9]+]]:sreg_64_xexec(s64) = G_MERGE_VALUES [[V_READFIRSTLANE_B32_2]](s32), [[V_READFIRSTLANE_B32_3]](s32)
+    ; CHECK: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[MV1]](s64), [[UV1]](s64), implicit $exec
+    ; CHECK: [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_1]], [[V_CMP_EQ_U64_e64_]], implicit-def $scc
+    ; CHECK: [[BUILD_VECTOR:%[0-9]+]]:sgpr(<4 x s32>) = G_BUILD_VECTOR [[V_READFIRSTLANE_B32_]](s32), [[V_READFIRSTLANE_B32_1]](s32), [[V_READFIRSTLANE_B32_2]](s32), [[V_READFIRSTLANE_B32_3]](s32)
+    ; CHECK: [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sreg_32_xm0(s32) = V_READFIRSTLANE_B32 [[COPY1]](s32), implicit $exec
+    ; CHECK: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64 = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]](s32), [[COPY1]](s32), implicit $exec
+    ; CHECK: [[S_AND_B64_1:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U32_e64_]], [[S_AND_B64_]], implicit-def $scc
+    ; CHECK: [[INT:%[0-9]+]]:sgpr(<4 x s32>) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.s.buffer.load), [[BUILD_VECTOR]](<4 x s32>), [[V_READFIRSTLANE_B32_4]](s32), 0
+    ; CHECK: [[S_AND_SAVEEXEC_B64_:%[0-9]+]]:sreg_64 = S_AND_SAVEEXEC_B64 killed [[S_AND_B64_1]], implicit-def $exec, implicit-def $scc, implicit $exec
+    ; CHECK: $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc
+    ; CHECK: S_CBRANCH_EXECNZ %bb.1, implicit $exec
+    ; CHECK: .2:
+    ; CHECK: successors: %bb.3(0x80000000)
+    ; CHECK: $exec = S_MOV_B64_term [[S_MOV_B64_term]]
+    ; CHECK: .3:
+    %0:_(<4 x s32>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
+    %1:_(s32) = COPY $vgpr4
+    %2:_(<4 x s32>) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.s.buffer.load), %0, %1, 0
+
+...