[llvm] [AMDGPU] Work around gfx1250 VOPD3 FP64 OPY hazard (PR #224741)

[Wooseok Lee via llvm-commits]

================
@@ -105,6 +105,46 @@ static bool canMapVOP3PToVOPD(const MachineInstr &MI) {
          getNamedOp(MI, AMDGPU::OpName::src2).getReg();
 }
 
+// In a VOPD3 whose OPX is a 64-bit operation an OPY VGPR source operand reads
+// back the wrong value if it names the last VGPR the wave owns.
+//
+// The number of VGPRs a wave owns is always a whole number of allocation
+// granules, so the last VGPR it owns is always the one just below a granule
+// boundary. A source which is not there cannot be it, whatever the wave's VGPR
+// count turns out to be. That count is deliberately not consulted: it is only
+// known once the whole call graph has been seen, at emission time.
+static bool isVOPD3F64OPYSrcHazard(const SIInstrInfo &TII,
+                                   const MachineInstr &MIX,
+                                   const MachineInstr &MIY) {
+  const MachineFunction &MF = *MIX.getMF();
+  const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
+  if (!ST.hasVOPD3F64OPYSrcHazard())
+    return false;
+
+  // Every 64-bit VOPD3 OPX opcode has a 64-bit vdst, and no 32-bit one has.
+  int VDstIdx =
+      AMDGPU::getNamedOperandIdx(MIX.getOpcode(), AMDGPU::OpName::vdst);
+  assert(VDstIdx != -1 && "VOPD3 OPX component must have a vdst");
+  if (TII.getOpSize(MIX, VDstIdx) != 8)
+    return false;
+
+  // Targets with this hazard do not support dynamic VGPR allocation.
+  unsigned Granule =
+      AMDGPU::IsaInfo::getVGPRAllocGranule(ST, /*DynamicVGPRBlockSize=*/0);
+  const SIRegisterInfo *TRI = ST.getRegisterInfo();
+  for (AMDGPU::OpName Name :
+       {AMDGPU::OpName::src0, AMDGPU::OpName::src1, AMDGPU::OpName::src2}) {
+    const MachineOperand *Src = TII.getNamedOperand(MIY, Name);
+    // Every OPY source which can be a VGPR is 32 bits wide.
+    if (!Src || !Src->isReg() ||
+        !AMDGPU::VGPR_32RegClass.contains(Src->getReg()))
+      continue;
+    if ((TRI->getHWRegIndex(Src->getReg()) + 1) % Granule == 0)
----------------
wooseoklee wrote:

I believe rejecting all VOPD3 pair if any OPY src register index that sits at a granule boundary because VGPR count is not known is too restrictive.

There are three cases. 
(a) VGPR count increase
(b) VGPR count unchanged
(c) VGPR count decrease

For case (a) and (b), checking whether a register is actually the wave's last VGPR and rejecting the VOPD3 pairing would be fine. 
Considering this is post RA pass, AFAIK, case (c) is not going to happen, right? 

https://github.com/llvm/llvm-project/pull/224741