[llvm] AMDGPU/GlobalISel: Uniformity info based regbankselect (PR #73684)

[Pierre van Houtryve via llvm-commits]

================
@@ -695,58 +695,108 @@ static LLT getHalfSizedType(LLT Ty) {
 
 // Build one or more V_READFIRSTLANE_B32 instructions to move the given vector
 // source value into a scalar register.
-Register AMDGPURegisterBankInfo::buildReadFirstLane(MachineIRBuilder &B,
-                                                    MachineRegisterInfo &MRI,
-                                                    Register Src) const {
+Register AMDGPURegisterBankInfo::buildReadFirstLaneSrc(MachineIRBuilder &B,
+                                                       Register Src) const {
+  MachineRegisterInfo &MRI = *B.getMRI();
   LLT Ty = MRI.getType(Src);
   const RegisterBank *Bank = getRegBank(Src, MRI, *TRI);
 
-  if (Bank == &AMDGPU::SGPRRegBank)
-    return Src;
-
-  unsigned Bits = Ty.getSizeInBits();
-  assert(Bits % 32 == 0);
-
   if (Bank != &AMDGPU::VGPRRegBank) {
     // We need to copy from AGPR to VGPR
     Src = B.buildCopy(Ty, Src).getReg(0);
     MRI.setRegBank(Src, AMDGPU::VGPRRegBank);
   }
 
+  Register Dst = MRI.createGenericVirtualRegister(Ty);
+  MRI.setRegBank(Dst, AMDGPU::SGPRRegBank);
+  buildReadFirstLaneForType(B, Dst, Src);
+  return Dst;
+}
+
+// Create new vgpr destination register for MI then move it to current
+// MI's sgpr destination using one or more V_READFIRSTLANE_B32 instructions.
+void AMDGPURegisterBankInfo::buildReadFirstLaneDst(MachineIRBuilder &B,
+                                                   MachineInstr &MI) const {
+  MachineRegisterInfo &MRI = *B.getMRI();
+  Register Dst = MI.getOperand(0).getReg();
+  const RegisterBank *DstBank = getRegBank(Dst, MRI, *TRI);
+  if (DstBank != &AMDGPU::SGPRRegBank)
+    return;
+
+  Register VgprDst = MRI.createGenericVirtualRegister(MRI.getType(Dst));
+  MRI.setRegBank(VgprDst, AMDGPU::VGPRRegBank);
+
+  MI.getOperand(0).setReg(VgprDst);
+  MachineBasicBlock *MBB = MI.getParent();
+  B.setInsertPt(*MBB, std::next(MI.getIterator()));
+  // readFirstLane VgprDst into Dst after MI.
+  return buildReadFirstLaneForType(B, Dst, VgprDst);
+}
+
+void AMDGPURegisterBankInfo::buildReadFirstLaneB32(MachineIRBuilder &B,
+                                                   Register SgprDst,
+                                                   Register VgprSrc) const {
+  MachineRegisterInfo &MRI = *B.getMRI();
+  B.buildInstr(AMDGPU::V_READFIRSTLANE_B32, {SgprDst}, {VgprSrc});
+  MRI.setRegClass(VgprSrc, &AMDGPU::VGPR_32RegClass);
+  MRI.setRegClass(SgprDst, &AMDGPU::SReg_32RegClass);
+}
+
+void AMDGPURegisterBankInfo::buildReadFirstLaneSequenceOfB32(
+    MachineIRBuilder &B, Register SgprDst, Register VgprSrc,
+    unsigned NumElts) const {
+  MachineRegisterInfo &MRI = *B.getMRI();
   LLT S32 = LLT::scalar(32);
-  unsigned NumParts = Bits / 32;
-  SmallVector<Register, 8> SrcParts;
-  SmallVector<Register, 8> DstParts;
+  SmallVector<Register, 8> VgprSrcParts;
+  SmallVector<Register, 8> SgprDstParts;
 
-  if (Bits == 32) {
-    SrcParts.push_back(Src);
-  } else {
-    auto Unmerge = B.buildUnmerge(S32, Src);
-    for (unsigned i = 0; i < NumParts; ++i)
-      SrcParts.push_back(Unmerge.getReg(i));
+  for (unsigned i = 0; i < NumElts; ++i) {
+    VgprSrcParts.push_back(MRI.createGenericVirtualRegister(S32));
+    SgprDstParts.push_back(MRI.createGenericVirtualRegister(S32));
   }
 
-  for (unsigned i = 0; i < NumParts; ++i) {
-    Register SrcPart = SrcParts[i];
-    Register DstPart = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
-    MRI.setType(DstPart, NumParts == 1 ? Ty : S32);
+  B.buildUnmerge(VgprSrcParts, VgprSrc);
+  for (unsigned i = 0; i < NumElts; ++i) {
+    buildReadFirstLaneB32(B, SgprDstParts[i], VgprSrcParts[i]);
+  }
+  B.buildMergeLikeInstr(SgprDst, SgprDstParts);
+}
 
-    const TargetRegisterClass *Constrained =
-        constrainGenericRegister(SrcPart, AMDGPU::VGPR_32RegClass, MRI);
-    (void)Constrained;
-    assert(Constrained && "Failed to constrain readfirstlane src reg");
+void AMDGPURegisterBankInfo::buildReadFirstLaneForType(MachineIRBuilder &B,
+                                                       Register SgprDst,
+                                                       Register VgprSrc) const {
+  MachineRegisterInfo &MRI = *B.getMRI();
+  LLT S16 = LLT::scalar(16);
+  LLT S32 = LLT::scalar(32);
+  LLT S64 = LLT::scalar(64);
+  LLT Ty = MRI.getType(SgprDst);
+
+  if (Ty == S16) {
+    Register VgprSrc32 = MRI.createGenericVirtualRegister(S32);
+    MRI.setRegBank(VgprSrc32, AMDGPU::VGPRRegBank);
+    Register SgprDst32 = MRI.createGenericVirtualRegister(S32);
+    MRI.setRegBank(SgprDst32, AMDGPU::SGPRRegBank);
+
+    B.buildAnyExt(VgprSrc32, VgprSrc);
+    buildReadFirstLaneB32(B, SgprDst32, VgprSrc32);
+    B.buildTrunc(SgprDst, SgprDst32);
+    return;
+  }
 
-    B.buildInstr(AMDGPU::V_READFIRSTLANE_B32, {DstPart}, {SrcPart});
+  if (Ty == S32 || Ty == LLT::pointer(3, 32)) {
+    return buildReadFirstLaneB32(B, SgprDst, VgprSrc);
+  }
 
-    DstParts.push_back(DstPart);
+  if (Ty == S64 || Ty == LLT::pointer(0, 64) || Ty == LLT::pointer(1, 64)) {
+    return buildReadFirstLaneSequenceOfB32(B, SgprDst, VgprSrc, 2);
   }
----------------
Pierre-vh wrote:

`{}` not needed x2

https://github.com/llvm/llvm-project/pull/73684