R600/SI: Use S_BFE_I64 for 64-bit sext_inreg

llvm-svn: 222012

5 files changed, 82 insertions, 17 deletions

diff --git a/llvm/lib/Target/R600/SIISelLowering.cpp b/llvm/lib/Target/R600/SIISelLowering.cpp
index 25acaa57187..d8e424936da 100644
--- a/llvm/lib/Target/R600/SIISelLowering.cpp
+++ b/llvm/lib/Target/R600/SIISelLowering.cpp
@@ -136,8 +136,7 @@ SITargetLowering::SITargetLowering(TargetMachine &TM) :
   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i16, Custom);
   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i16, Custom);
 
-  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Custom);
-
+  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::Other, Custom);
 
   setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
diff --git a/llvm/lib/Target/R600/SIInstrInfo.cpp b/llvm/lib/Target/R600/SIInstrInfo.cpp
index 1c15eb8973e..97f2d4b5935 100644
--- a/llvm/lib/Target/R600/SIInstrInfo.cpp
+++ b/llvm/lib/Target/R600/SIInstrInfo.cpp
@@ -1901,8 +1901,13 @@ void SIInstrInfo::moveToVALU(MachineInstr &TopInst) const {
       Inst->eraseFromParent();
       continue;
 
+    case AMDGPU::S_BFE_I64: {
+      splitScalar64BitBFE(Worklist, Inst);
+      Inst->eraseFromParent();
+      continue;
+    }
+
     case AMDGPU::S_BFE_U64:
-    case AMDGPU::S_BFE_I64:
     case AMDGPU::S_BFM_B64:
       llvm_unreachable("Moving this op to VALU not implemented");
     }
@@ -2167,6 +2172,65 @@ void SIInstrInfo::splitScalar64BitBCNT(SmallVectorImpl<MachineInstr *> &Worklist
   Worklist.push_back(Second);
 }
 
+void SIInstrInfo::splitScalar64BitBFE(SmallVectorImpl<MachineInstr *> &Worklist,
+                                      MachineInstr *Inst) const {
+  MachineBasicBlock &MBB = *Inst->getParent();
+  MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
+  MachineBasicBlock::iterator MII = Inst;
+  DebugLoc DL = Inst->getDebugLoc();
+
+  MachineOperand &Dest = Inst->getOperand(0);
+  uint32_t Imm = Inst->getOperand(2).getImm();
+  uint32_t Offset = Imm & 0x3f; // Extract bits [5:0].
+  uint32_t BitWidth = (Imm & 0x7f0000) >> 16; // Extract bits [22:16].
+
+  // Only sext_inreg cases handled.
+  assert(Inst->getOpcode() == AMDGPU::S_BFE_I64 &&
+         BitWidth <= 32 &&
+         Offset == 0 &&
+         "Not implemented");
+
+  if (BitWidth < 32) {
+    unsigned MidRegLo = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
+    unsigned MidRegHi = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
+    unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
+
+    BuildMI(MBB, MII, DL, get(AMDGPU::V_BFE_I32), MidRegLo)
+      .addReg(Inst->getOperand(1).getReg(), 0, AMDGPU::sub0)
+      .addImm(0)
+      .addImm(BitWidth);
+
+    BuildMI(MBB, MII, DL, get(AMDGPU::V_ASHRREV_I32_e32), MidRegHi)
+      .addImm(31)
+      .addReg(MidRegLo);
+
+    BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), ResultReg)
+      .addReg(MidRegLo)
+      .addImm(AMDGPU::sub0)
+      .addReg(MidRegHi)
+      .addImm(AMDGPU::sub1);
+
+    MRI.replaceRegWith(Dest.getReg(), ResultReg);
+    return;
+  }
+
+  MachineOperand &Src = Inst->getOperand(1);
+  unsigned TmpReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
+  unsigned ResultReg = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
+
+  BuildMI(MBB, MII, DL, get(AMDGPU::V_ASHRREV_I32_e64), TmpReg)
+    .addImm(31)
+    .addReg(Src.getReg(), 0, AMDGPU::sub0);
+
+  BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), ResultReg)
+    .addReg(Src.getReg(), 0, AMDGPU::sub0)
+    .addImm(AMDGPU::sub0)
+    .addReg(TmpReg)
+    .addImm(AMDGPU::sub1);
+
+  MRI.replaceRegWith(Dest.getReg(), ResultReg);
+}
+
 void SIInstrInfo::addDescImplicitUseDef(const MCInstrDesc &NewDesc,
                                         MachineInstr *Inst) const {
   // Add the implict and explicit register definitions.
diff --git a/llvm/lib/Target/R600/SIInstrInfo.h b/llvm/lib/Target/R600/SIInstrInfo.h
index 5c5d8476235..ce32fd7fa65 100644
--- a/llvm/lib/Target/R600/SIInstrInfo.h
+++ b/llvm/lib/Target/R600/SIInstrInfo.h
@@ -52,6 +52,8 @@ private:
 
   void splitScalar64BitBCNT(SmallVectorImpl<MachineInstr *> &Worklist,
                             MachineInstr *Inst) const;
+  void splitScalar64BitBFE(SmallVectorImpl<MachineInstr *> &Worklist,
+                           MachineInstr *Inst) const;
 
   void addDescImplicitUseDef(const MCInstrDesc &Desc, MachineInstr *MI) const;
 
diff --git a/llvm/lib/Target/R600/SIInstrInfo.td b/llvm/lib/Target/R600/SIInstrInfo.td
index b84a2b12149..2c9ffaffca8 100644
--- a/llvm/lib/Target/R600/SIInstrInfo.td
+++ b/llvm/lib/Target/R600/SIInstrInfo.td
@@ -350,6 +350,11 @@ class SOP2_64 <bits<7> op, string opName, list<dag> pattern> : SOP2 <
   opName#" $dst, $src0, $src1", pattern
 >;
 
+class SOP2_64_32 <bits<7> op, string opName, list<dag> pattern> : SOP2 <
+  op, (outs SReg_64:$dst), (ins SSrc_64:$src0, SSrc_32:$src1),
+  opName#" $dst, $src0, $src1", pattern
+>;
+
 class SOP2_SHIFT_64 <bits<7> op, string opName, list<dag> pattern> : SOP2 <
   op, (outs SReg_64:$dst), (ins SSrc_64:$src0, SSrc_32:$src1),
   opName#" $dst, $src0, $src1", pattern
diff --git a/llvm/lib/Target/R600/SIInstructions.td b/llvm/lib/Target/R600/SIInstructions.td
index 193a083085a..6dcb6de0797 100644
--- a/llvm/lib/Target/R600/SIInstructions.td
+++ b/llvm/lib/Target/R600/SIInstructions.td
@@ -297,7 +297,7 @@ def S_MUL_I32 : SOP2_32 <0x00000026, "s_mul_i32",
 def S_BFE_U32 : SOP2_32 <0x00000027, "s_bfe_u32", []>;
 def S_BFE_I32 : SOP2_32 <0x00000028, "s_bfe_i32", []>;
 def S_BFE_U64 : SOP2_64 <0x00000029, "s_bfe_u64", []>;
-def S_BFE_I64 : SOP2_64 <0x0000002a, "s_bfe_i64", []>;
+def S_BFE_I64 : SOP2_64_32 <0x0000002a, "s_bfe_i64", []>;
 //def S_CBRANCH_G_FORK : SOP2_ <0x0000002b, "s_cbranch_g_fork", []>;
 def S_ABSDIFF_I32 : SOP2_32 <0x0000002c, "s_absdiff_i32", []>;
 
@@ -2972,30 +2972,25 @@ defm : SI_INDIRECT_Pattern <v16i32, i32, SI_INDIRECT_DST_V16>;
 def : Pat<(i32 (sext_inreg i32:$src, i1)),
   (S_BFE_I32 i32:$src, 65536)>; // 0 | 1 << 16
 
-// TODO: Match 64-bit BFE. SI has a 64-bit BFE, but it's scalar only so it
-// might not be worth the effort, and will need to expand to shifts when
-// fixing SGPR copies.
-
 // Handle sext_inreg in i64
 def : Pat <
   (i64 (sext_inreg i64:$src, i1)),
-  (REG_SEQUENCE SReg_64,
-    (S_BFE_I32 (EXTRACT_SUBREG i64:$src, sub0), 65536), sub0, // 0 | 1 << 16
-    (S_MOV_B32 -1), sub1)
+  (S_BFE_I64 i64:$src, 0x10000) // 0 | 1 << 16
 >;
 
 def : Pat <
   (i64 (sext_inreg i64:$src, i8)),
-  (REG_SEQUENCE SReg_64,
-    (S_SEXT_I32_I8 (EXTRACT_SUBREG i64:$src, sub0)), sub0,
-    (S_MOV_B32 -1), sub1)
+  (S_BFE_I64 i64:$src, 0x80000) // 0 | 8 << 16
 >;
 
 def : Pat <
   (i64 (sext_inreg i64:$src, i16)),
-  (REG_SEQUENCE SReg_64,
-    (S_SEXT_I32_I16 (EXTRACT_SUBREG i64:$src, sub0)), sub0,
-    (S_MOV_B32 -1), sub1)
+  (S_BFE_I64 i64:$src, 0x100000) // 0 | 16 << 16
+>;
+
+def : Pat <
+  (i64 (sext_inreg i64:$src, i32)),
+  (S_BFE_I64 i64:$src, 0x200000) // 0 | 32 << 16
 >;
 
 class ZExt_i64_i32_Pat <SDNode ext> : Pat <