[PowerPC] Custom lower known CR bit spills

[lei]

Summary

For known CRBit spills, CRSET/CRUNSET, it is more efficient to just load and spill the known value instead of extracting the bit.

eg. This sequence is currently used to spill a CRUNSET:

crclr   4*cr5+lt
mfocrf  r3,4
rlwinm  r3,r3,20,0,0
stw     r3,132(r1)

This patch custom lower it to:

li  r3,0
stw r3,132(r1)

[hfinkel]

llvm/lib/Target/PowerPC/PPCRegisterInfo.cpp

725

I think that this makes sense, but I'm a bit concerned that, without a cutoff, this makes the spilling process quadratic. Can you please add a cl::opt search cutoff for this?

[lei]

llvm/lib/Target/PowerPC/PPCRegisterInfo.cpp

725

How about --ppc-max-crbit-spill-dist with an initial value of 20?

[hfinkel]

llvm/lib/Target/PowerPC/PPCRegisterInfo.cpp

725

Sounds good. Make the initial value a bit larger, however. It's easy to have blocks with more than 20 instructions. I'd start with 100. Also, remember to skip the debug instructions when counting, so you don't end up with differences between the debugging-enabled and debugging-disabled cases.

[lei]

llvm/lib/Target/PowerPC/PPCRegisterInfo.cpp

725

I didn't even think about debug instructions... how do I identify if an instruction is a debug instruction?

[hfinkel]

llvm/lib/Target/PowerPC/PPCRegisterInfo.cpp

725

You can, I believe, call Ins->isDebugInstr()

[lei]

Added option to specify cutoff for CR bit definition search with default of 100.

[hfinkel]

llvm/lib/Target/PowerPC/PPCRegisterInfo.cpp

725

Please put spaces around the = here, and the = and == below (to match the general convention here).

[lei]

address spacing issues.

llvm/lib/Target/PowerPC/PPCRegisterInfo.cpp

725

thx!

[hfinkel]

LGTM.

If this test case proves too fragile in the future, we might replace it with an MIR test.

[lei]

In D61754#1500658, @hfinkel wrote:

LGTM.

If this test case proves too fragile in the future, we might replace it with an MIR test.

Hi Hal, I actually tried to create a MIR test first but had a hard time. I was trying to get the MIR prior to my pass via llc -stop-before=regalloc reduced.ll but it gave me LLVM ERROR: "regalloc" pass is not registered.. I just stopped there as I didn't know how to continue after that... let me know if you have suggestions on how to proceed with this kind of issue and I'll try it next time. Thx!

llvm/lib/Target/PowerPC/PPCRegisterInfo.cpp

void PPCRegisterInfo::lowerCRBitSpilling(MachineBasicBlock::iterator II,
                                          unsigned FrameIndex) const {
   // Get the instruction.
   MachineInstr &MI = *II;       // ; SPILL_CRBIT <SrcReg>, <offset>
   // Get the instruction's basic block.
   MachineBasicBlock &MBB = *MI.getParent();
   MachineFunction &MF = *MBB.getParent();
   const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>();
   const TargetInstrInfo &TII = *Subtarget.getInstrInfo();
+  const TargetRegisterInfo* TRI = Subtarget.getRegisterInfo();
   DebugLoc dl = MI.getDebugLoc();
 
   bool LP64 = TM.isPPC64();
   const TargetRegisterClass *G8RC = &PPC::G8RCRegClass;
   const TargetRegisterClass *GPRC = &PPC::GPRCRegClass;
 
   unsigned Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC);
   unsigned SrcReg = MI.getOperand(0).getReg();
 
+  // Search up the BB to find the definition of the CRBit.
+  MachineBasicBlock::reverse_iterator Ins;
+  for (Ins=MI; Ins!=MBB.rend(); Ins++) {
+    if (Ins->modifiesRegister(SrcReg, TRI))
+      break;
+  }
+
+  // Unable to find the definition of the CRBit in the MBB.
+  if (Ins == MBB.rend())
+    Ins=MI;
+
+  // There is no need to extract the CR bit if its value is already known.
+  switch (Ins->getOpcode()) {
+  case PPC::CRUNSET:
+    BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::LI8 : PPC::LI), Reg)
+      .addImm(0);
+    break;
+  case PPC::CRSET:
+    BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::LIS8 : PPC::LIS), Reg)
+      .addImm(-32768);
+    break;
+  default:
     // We need to move the CR field that contains the CR bit we are spilling.
     // The super register may not be explicitly defined (i.e. it can be defined
     // by a CR-logical that only defines the subreg) so we state that the CR
     // field is undef. Also, in order to preserve the kill flag on the CR bit,
     // we add it as an implicit use.
     BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::MFOCRF8 : PPC::MFOCRF), Reg)
       .addReg(getCRFromCRBit(SrcReg), RegState::Undef)
       .addReg(SrcReg,
               RegState::Implicit | getKillRegState(MI.getOperand(0).isKill()));
 
-  // If the saved register wasn't CR0LT, shift the bits left so that the bit to
-  // store is the first one. Mask all but that bit.
+    // If the saved register wasn't CR0LT, shift the bits left so that the bit
+    // to store is the first one. Mask all but that bit.
     unsigned Reg1 = Reg;
     Reg = MF.getRegInfo().createVirtualRegister(LP64 ? G8RC : GPRC);
 
     // rlwinm rA, rA, ShiftBits, 0, 0.
     BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::RLWINM8 : PPC::RLWINM), Reg)
       .addReg(Reg1, RegState::Kill)
       .addImm(getEncodingValue(SrcReg))
       .addImm(0).addImm(0);
-
+  }
   addFrameReference(BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::STW8 : PPC::STW))
                     .addReg(Reg, RegState::Kill),
                     FrameIndex);
 
   // Discard the pseudo instruction.
   MBB.erase(II);
 }
 

llvm/test/CodeGen/PowerPC/knowCRBitSpill.ll

+; RUN: llc -verify-machineinstrs -mtriple=powerpc64le-unknown-linux-gnu \
+; RUN:     -ppc-asm-full-reg-names -ppc-vsr-nums-as-vr < %s | FileCheck %s
+; RUN: llc -verify-machineinstrs -mtriple=powerpc64-unknown-linux-gnu \
+; RUN:     -ppc-asm-full-reg-names -ppc-vsr-nums-as-vr < %s | FileCheck %s
+
+
+; For known CRBit spills, CRSET/CRUNSET, it is more efficient to just load and
+; spill the known value.  These tests verify that for CRSET and CRUNSET spills
+; we do not extract the bit for spilling.
+
+%struct.anon = type { i32 }
+
+@b = common dso_local global %struct.anon* null, align 8
+@a = common dso_local global i64 0, align 8
+
+; Function Attrs: nounwind
+define dso_local signext i32 @spillCRSET(i32 signext %p1, i32 signext %p2) {
+; CHECK-LABEL: spillCRSET:
+; CHECK:       # %bb.0: # %entry
+; CHECK:        lis [[REG1:.*]], -32768
+; CHECK-DAG:    creqv [[CREG:.*]]*cr5+lt, [[CREG]]*cr5+lt, [[CREG]]*cr5+lt
+; CHECK-NOT:    mfocrf [[REG2:.*]], [[CREG]]
+; CHECK-NOT:    rlwinm [[REG2]], [[REG2]]
+; CHECK:        stw [[REG1]]
+; CHECK:  .LBB0_1: # %redo_first_pass
+entry:
+  %tobool = icmp eq i32 %p2, 0
+  %tobool2 = icmp eq i32 %p1, 0
+  br label %redo_first_pass
+
+redo_first_pass:                                  ; preds = %for.end, %entry
+  br i1 %tobool, label %if.end, label %if.then
+
+if.then:                                          ; preds = %redo_first_pass
+  %call = tail call signext i32 bitcast (i32 (...)* @fn2 to i32 ()*)() #2
+  %tobool1 = icmp ne i32 %call, 0
+  br label %if.end
+
+if.end:                                           ; preds = %redo_first_pass, %if.then
+  %c.1.off0 = phi i1 [ %tobool1, %if.then ], [ true, %redo_first_pass ]
+  br i1 %tobool2, label %if.end4, label %if.then3
+
+if.then3:                                         ; preds = %if.end
+  %0 = load %struct.anon*, %struct.anon** @b, align 8
+  %contains_i = getelementptr inbounds %struct.anon, %struct.anon* %0, i64 0, i32 0
+  store i32 1, i32* %contains_i, align 4
+  br label %if.end4
+
+if.end4:                                          ; preds = %if.end, %if.then3
+  tail call void asm sideeffect "#DO_NOTHING", "~{cr0},~{cr1},~{cr2},~{cr3},~{cr4},~{cr5},~{cr6},~{cr7}"()
+  br i1 %c.1.off0, label %if.then6, label %if.end13
+
+if.then6:                                         ; preds = %if.end4
+  %1 = load i64, i64* @a, align 8
+  %cmp21 = icmp eq i64 %1, 0
+  br i1 %cmp21, label %if.end13, label %for.body
+
+for.body:                                         ; preds = %if.then6, %for.body
+  %s.122 = phi i64 [ %inc, %for.body ], [ 0, %if.then6 ]
+  %call7 = tail call signext i32 bitcast (i32 (...)* @fn3 to i32 ()*)()
+  %inc = add nuw i64 %s.122, 1
+  %exitcond = icmp eq i64 %inc, %1
+  br i1 %exitcond, label %for.end, label %for.body
+
+for.end:                                          ; preds = %for.body
+  br i1 %cmp21, label %if.end13, label %redo_first_pass
+
+if.end13:                                         ; preds = %if.then6, %for.end, %if.end4
+  ret i32 0
+}
+
+%struct.p5rx = type { i32 }
+
+; Function Attrs: nounwind
+define dso_local signext i32 @spillCRUNSET(%struct.p5rx* readonly %p1, i32 signext %p2, i32 signext %p3) {
+; CHECK-LABEL: spillCRUNSET:
+; CHECK:       # %bb.0: # %entry
+; CHECK-DAG:    crxor [[CREG:.*]]*cr5+lt, [[CREG]]*cr5+lt, [[CREG]]*cr5+lt
+; CHECK-DAG:    li [[REG1:.*]], 0
+; CHECK-NOT:    mfocrf [[REG2:.*]], [[CREG]]
+; CHECK-NOT:    rlwinm [[REG2]], [[REG2]]
+; CHECK:        stw [[REG1]]
+; CHECK:        .LBB1_1: # %redo_first_pass
+entry:
+  %and = and i32 %p3, 128
+  %tobool = icmp eq i32 %and, 0
+  %tobool2 = icmp eq %struct.p5rx* %p1, null
+  %sv_any = getelementptr inbounds %struct.p5rx, %struct.p5rx* %p1, i64 0, i32 0
+  %tobool12 = icmp eq i32 %p2, 0
+  br label %redo_first_pass
+
+redo_first_pass:                                  ; preds = %if.end11, %entry
+  %a.0.off0 = phi i1 [ false, %entry ], [ %a.1.off0, %if.end11 ]
+  br i1 %tobool, label %if.end, label %if.then
+
+if.then:                                          ; preds = %redo_first_pass
+  %call = tail call signext i32 bitcast (i32 (...)* @fn2 to i32 ()*)()
+  %tobool1 = icmp ne i32 %call, 0
+  br label %if.end
+
+if.end:                                           ; preds = %redo_first_pass, %if.then
+  %a.1.off0 = phi i1 [ %tobool1, %if.then ], [ %a.0.off0, %redo_first_pass ]
+  tail call void asm sideeffect "#DO_NOTHING", "~{cr0},~{cr1},~{cr2},~{cr3},~{cr4},~{cr5},~{cr6},~{cr7}"()
+  br i1 %tobool2, label %if.end11, label %land.lhs.true
+
+land.lhs.true:                                    ; preds = %if.end
+  %call3 = tail call signext i32 bitcast (i32 (...)* @fn3 to i32 ()*)()
+  %tobool4 = icmp eq i32 %call3, 0
+  br i1 %tobool4, label %if.end11, label %land.lhs.true5
+
+land.lhs.true5:                                   ; preds = %land.lhs.true
+  %0 = load i32, i32* %sv_any, align 4
+  %tobool6 = icmp eq i32 %0, 0
+  %a.1.off0.not = xor i1 %a.1.off0, true
+  %brmerge = or i1 %tobool6, %a.1.off0.not
+  br i1 %brmerge, label %if.end11, label %if.then9
+
+if.then9:                                         ; preds = %land.lhs.true5
+  %call10 = tail call signext i32 bitcast (i32 (...)* @fn4 to i32 ()*)()
+  br label %if.end11
+
+if.end11:                                         ; preds = %land.lhs.true5, %land.lhs.true, %if.end, %if.then9
+  br i1 %tobool12, label %if.end14, label %redo_first_pass
+
+if.end14:                                         ; preds = %if.end11
+  ret i32 0
+}
+
+declare signext i32 @fn2(...)
+declare signext i32 @fn3(...)
+declare signext i32 @fn4(...)