RAGreedyStats: Ignore identity COPYs; count COPYs from/to physregs

[MatzeB]

Summary

Improve copy statistics:

• Count copies from or to physical registers: They are used to model function parameters and calling conventions and the register allocator optimizes for them.
• Check physical registers assigned to virtual registers and stop counting "identity" COPYs where source and destination is the same physical registers; they will be removed in the virtregmap pass anyway.

[MatzeB]

Thinking about this some more, we should start counting COPYs from PhysReg->VirtReg and VirtReg->PhysReg though... Will adapt this...

[MatzeB]

Updated to also count COPYs from/to physical registers.

llvm/lib/CodeGen/RegAllocGreedy.cpp

2374

Note for reviewers: We don't need the isReg() check as COPYs are only allowed between register operands.

[mtrofin]

llvm/lib/CodeGen/RegAllocGreedy.cpp

2377

nit: how about:

if (SrcReg.isVirtual()) {
  ...
}
if (DestReg.isVirtual()) {
  ...
}
Stats.Copies += (SrcReg != DestReg);

[mtrofin]

llvm/lib/CodeGen/RegAllocGreedy.cpp

2377

Oh, never mind, incorrect suggestion.

llvm/lib/CodeGen/RegAllocGreedy.cpp

RAGreedy::RAGreedyStats RAGreedy::computeStats(MachineBasicBlock &MBB) {
   };
   auto isPatchpointInstr = [](const MachineInstr &MI) {
     return MI.getOpcode() == TargetOpcode::PATCHPOINT ||
            MI.getOpcode() == TargetOpcode::STACKMAP ||
            MI.getOpcode() == TargetOpcode::STATEPOINT;
   };
   for (MachineInstr &MI : MBB) {
     if (MI.isCopy()) {
-      MachineOperand &Dest = MI.getOperand(0);
-      MachineOperand &Src = MI.getOperand(1);
-      if (Dest.isReg() && Src.isReg() && Dest.getReg().isVirtual() &&
-          Src.getReg().isVirtual())
+      const MachineOperand &Dest = MI.getOperand(0);
+      const MachineOperand &Src = MI.getOperand(1);
+      Register SrcReg = Src.getReg();
+      Register DestReg = Dest.getReg();
+      // Only count `COPY`s with a virtual register as source or destination.
+      if (SrcReg.isVirtual() || DestReg.isVirtual()) {
+        if (SrcReg.isVirtual()) {
+          SrcReg = VRM->getPhys(SrcReg);
+          if (Src.getSubReg())
+            SrcReg = TRI->getSubReg(SrcReg, Src.getSubReg());
+        }
+        if (DestReg.isVirtual()) {
+          DestReg = VRM->getPhys(DestReg);
+          if (Dest.getSubReg())
+            DestReg = TRI->getSubReg(DestReg, Dest.getSubReg());
+        }
+        if (SrcReg != DestReg)
           ++Stats.Copies;
+      }
       continue;
     }
 
     SmallVector<const MachineMemOperand *, 2> Accesses;
     if (TII->isLoadFromStackSlot(MI, FI) && MFI.isSpillSlotObjectIndex(FI)) {
       ++Stats.Reloads;
       continue;
     }

llvm/test/CodeGen/AArch64/arm64-spill-remarks.ll

 ; This has two nested loops, each with one value that has to be spilled and
 ; then reloaded.
 
 ; (loop3:)
 ; REMARK: remark: /tmp/kk.c:3:20: 1 spills 1.000000e+02 total spills cost 1 reloads 1.000000e+02 total reloads cost generated in loop{{$}}
 ; (loop2:)
 ; REMARK: remark: /tmp/kk.c:2:20: 1 spills 1.000000e+04 total spills cost 1 reloads 1.000000e+04 total reloads cost generated in loop{{$}}
 ; (loop:)
-; REMARK: remark: /tmp/kk.c:1:20: 2 spills 1.010000e+04 total spills cost 2 reloads 1.010000e+04 total reloads cost generated in loop{{$}}
+; REMARK: remark: /tmp/kk.c:1:20: 2 spills 1.010000e+04 total spills cost 2 reloads 1.010000e+04 total reloads cost 1 virtual registers copies 1.000000e+02 total copies cost generated in loop{{$}}
 ; (func:)
-; REMARK: remark: /tmp/kk.c:1:1: 3 spills 1.020000e+04 total spills cost 3 reloads 1.020000e+04 total reloads cost generated in function
+; REMARK: remark: /tmp/kk.c:1:1: 3 spills 1.020000e+04 total spills cost 3 reloads 1.020000e+04 total reloads cost 3 virtual registers copies 1.020000e+02 total copies cost generated in function{{$}}
 
 ; (loop3:)
 ; HOTNESS: remark: /tmp/kk.c:3:20: 1 spills 1.000000e+02 total spills cost 1 reloads 1.000000e+02 total reloads cost generated in loop (hotness: 300)
 ; (loop2:)
 ; HOTNESS: remark: /tmp/kk.c:2:20: 1 spills 1.000000e+04 total spills cost 1 reloads 1.000000e+04 total reloads cost generated in loop (hotness: 30000)
 ; (loop:)
-; HOTNESS: remark: /tmp/kk.c:1:20: 2 spills 1.010000e+04 total spills cost 2 reloads 1.010000e+04 total reloads cost generated in loop (hotness: 300)
+; HOTNESS: remark: /tmp/kk.c:1:20: 2 spills 1.010000e+04 total spills cost 2 reloads 1.010000e+04 total reloads cost 1 virtual registers copies 1.000000e+02 total copies cost generated in loop (hotness: 300)
 
 ; NO_REMARK-NOT: remark
 
 ; THRESHOLD-NOT: (hotness: 300)
 ; THRESHOLD: remark: /tmp/kk.c:2:20: 1 spills 1.000000e+04 total spills cost 1 reloads 1.000000e+04 total reloads cost generated in loop (hotness: 30000)
 
 ; YAML: --- !Missed
 ; YAML: Pass:            regalloc

 ; YAML:   - NumSpills:       '2'
 ; YAML:   - String:          ' spills '
 ; YAML:   - TotalSpillsCost: '1.010000e+04'
 ; YAML:   - String:          ' total spills cost '
 ; YAML:   - NumReloads:      '2'
 ; YAML:   - String:          ' reloads '
 ; YAML:   - TotalReloadsCost: '1.010000e+04'
 ; YAML:   - String:          ' total reloads cost '
+; YAML:   - NumVRCopies:     '1'
+; YAML:   - String:          ' virtual registers copies '
+; YAML:   - TotalCopiesCost: '1.000000e+02'
+; YAML:   - String:          ' total copies cost '
 ; YAML:   - String:          generated in loop
 ; YAML: ...
 ; YAML: --- !Missed
 ; YAML: Pass:            regalloc
 ; YAML: Name:            SpillReloadCopies
 ; YAML: DebugLoc:        { File: '/tmp/kk.c', Line: 1, Column: 1 }
 ; YAML: Function:        fpr128
 ; YAML: Hotness:         3
 ; YAML: Args:
 ; YAML:   - NumSpills:       '3'
 ; YAML:   - String:          ' spills '
 ; YAML:   - TotalSpillsCost: '1.020000e+04'
 ; YAML:   - String:          ' total spills cost '
 ; YAML:   - NumReloads:      '3'
 ; YAML:   - String:          ' reloads '
 ; YAML:   - TotalReloadsCost: '1.020000e+04'
 ; YAML:   - String:          ' total reloads cost '
+; YAML:   - NumVRCopies:     '3'
+; YAML:   - String:          ' virtual registers copies '
+; YAML:   - TotalCopiesCost: '1.020000e+02'
+; YAML:   - String:          ' total copies cost '
 ; YAML:   - String:          generated in function
 ; YAML: ...
 
 ; THRESHOLD_YAML-NOT: Hotness:         300{{$}}
 ; THRESHOLD_YAML: --- !Missed
 ; THRESHOLD_YAML: Pass:            regalloc
 ; THRESHOLD_YAML: Name:            LoopSpillReloadCopies
 ; THRESHOLD_YAML: DebugLoc:        { File: '/tmp/kk.c', Line: 2, Column: 20 }

llvm/test/CodeGen/X86/statepoint-ra.ll

 ;YAML:   - TotalSpillsCost: '7.000000e+00'
 ;YAML:   - String:          ' total spills cost '
 ;YAML:   - NumReloads:      '7'
 ;YAML:   - String:          ' reloads '
 ;YAML:   - TotalReloadsCost: '3.109004e-15'
 ;YAML:   - String:          ' total reloads cost '
 ;YAML:   - NumZeroCostFoldedReloads: '20'
 ;YAML:   - String:          ' zero cost folded reloads '
-;YAML:   - NumVRCopies:     '2'
-;YAML:   - String:          ' virtual registers copies '
-;YAML:   - TotalCopiesCost: '8.882868e-16'
-;YAML:   - String:          ' total copies cost '
 ;YAML:   - String:          generated in function
 
 define void @barney(ptr addrspace(1) %arg, double %arg1, double %arg2, double %arg3, double %arg4, double %arg5, double %arg6, double %arg7, double %arg8, double %arg9, double %arg10, double %arg11, double %arg12) gc "statepoint-example" personality ptr @widget {
 bb:
   %tmp = call coldcc token (i64, i32, ptr, i32, i32, ...) @llvm.experimental.gc.statepoint.p0(i64 2882400000, i32 0, ptr nonnull elementtype(void ()) @blam, i32 0, i32 0, i32 0, i32 0) [ "deopt"(i32 0, i32 1, i32 0, i32 0, i32 0, i32 26, i32 0, i32 0, ptr addrspace(1) %arg, i32 4, double %arg1, i32 7, ptr null, i32 4, double %arg2, i32 7, ptr null, i32 4, double %arg3, i32 7, ptr null, i32 4, double %arg4, i32 7, ptr null, i32 4, double %arg5, i32 7, ptr null, i32 4, double %arg6, i32 7, ptr null, i32 4, double %arg7, i32 7, ptr null, i32 4, double %arg8, i32 7, ptr null, i32 4, double %arg9, i32 7, ptr null, i32 4, double %arg10, i32 7, ptr null, i32 4, double %arg11, i32 7, ptr null, i32 4, double %arg12, i32 7, ptr null, i32 7, ptr null), "gc-live"(ptr addrspace(1) %arg) ]
   br i1 undef, label %bb13, label %bb15
 
 bb13:                                             ; preds = %bb