[PowerPC] Materialize 34 bit constants with pli on Power 10.

[stefanp]

Summary

NOTE: This patch was originally written by Anil Mahmud. His code has been rebased but otherwise left mostly unchanged.

A new instructon on Power 10 allows for the materialization of 34 bit
immediate values. This patch allows the compiler to take advantage of
the new instruction in this situation.

[steven.zhang]

Rebase with patch https://reviews.llvm.org/D92089 ?

[stefanp]

In D92879#2441581, @steven.zhang wrote:

Rebase with patch https://reviews.llvm.org/D92089 ?

Good point. I will rebase on top of that patch since that should go in first.

[stefanp]

Rebased on top of D92089

[amyk]

I just have a minor nit, but overall it looks good to me.

llvm/lib/Target/PowerPC/PPCISelDAGToDAG.cpp

1054

nit: I think the preference is to name this as Subtarget rather than PPCSubTarget (As seen in https://reviews.llvm.org/D81623)

[lei]

llvm/lib/Target/PowerPC/PPCISelDAGToDAG.cpp

1053

doc.

llvm/lib/Target/PowerPC/PPCInstrPrefix.td

2622

tc?

[stefanp]

Added comment for function.
Fixed variable name.
Added test case for 32 bit target.

[stefanp]

Updated some of the test cases for the 32 bit compilation.

[amyk]

Thanks for updating. I just noticed one more minor nit.

llvm/lib/Target/PowerPC/PPCISelDAGToDAG.cpp

1055

s/meterialize/materialize

[stefanp]

Fixed spelling mistake in comment.

[stefanp]

Rebased to top of main branch.

[stefanp]

Adjusted test cast that I forgot to fix on rebase.

llvm/lib/Target/PowerPC/PPCISelDAGToDAG.cpp

static SDNode *selectI64ImmDirect(SelectionDAG *CurDAG, const SDLoc &dl,
   }
 
   return Result;
 }
 
 static SDNode *selectI64Imm(SelectionDAG *CurDAG, const SDLoc &dl,
                             int64_t Imm) {
   unsigned Count = selectI64ImmInstrCountDirect(Imm);
+  const PPCSubtarget &PPCSubTarget =
+      CurDAG->getMachineFunction().getSubtarget<PPCSubtarget>();
+
+  // Use the td file based patterns and use the PLI instruction.
+  if (PPCSubTarget.hasPrefixInstrs() && (Count > 1) && isInt<34>(Imm))
+    return cast<ConstantSDNode>(CurDAG->getConstant(Imm, dl, MVT::i64));
 
   // If the instruction count is 1 or 2, we do not need further analysis
   // since rotate + load constant requires at least 2 instructions.
   if (Count <= 2)
     return selectI64ImmDirect(CurDAG, dl, Imm);
 
   unsigned RMin = 0;
 
case PPC::STBUX8:
         return 0;
       MaxTruncation = std::max(MaxTruncation, 8u);
       continue;
     }
   }
   return MaxTruncation;
 }
 
 // Select a 64-bit constant.
 static SDNode *selectI64Imm(SelectionDAG *CurDAG, SDNode *N) {
   SDLoc dl(N);
 
   // Get 64 bit value.
   int64_t Imm = cast<ConstantSDNode>(N)->getZExtValue();
   if (unsigned MinSize = allUsesTruncate(CurDAG, N)) {
     uint64_t SextImm = SignExtend64(Imm, MinSize);
     SDValue SDImm = CurDAG->getTargetConstant(SextImm, dl, MVT::i64);
     if (isInt<16>(SextImm))
       return CurDAG->getMachineNode(PPC::LI8, dl, MVT::i64, SDImm);
void PPCDAGToDAGISel::Select(SDNode *N) {
   if (tryIntCompareInGPR(N))
     return;
 
   switch (N->getOpcode()) {
   default: break;
 
   case ISD::Constant:
     if (N->getValueType(0) == MVT::i64) {
-      ReplaceNode(N, selectI64Imm(CurDAG, N));
+      SDNode *ResNode = selectI64Imm(CurDAG, N);
+      if (!isa<ConstantSDNode>(ResNode)) {
+        ReplaceNode(N, ResNode);
         return;
       }
+    }
     break;
 
   case ISD::INTRINSIC_WO_CHAIN: {
     if (!Subtarget->isISA3_1())
       break;
     unsigned Opcode = 0;
     switch (N->getConstantOperandVal(0)) {
     default:

llvm/lib/Target/PowerPC/PPCInstrInfo.cpp

bool PPCInstrInfo::isReallyTriviallyReMaterializable(const MachineInstr &MI,
   switch (MI.getOpcode()) {
   default:
     // This function should only be called for opcodes with the ReMaterializable
     // flag set.
     llvm_unreachable("Unknown rematerializable operation!");
     break;
   case PPC::LI:
   case PPC::LI8:
+  case PPC::PLI:
+  case PPC::PLI8:
   case PPC::LIS:
   case PPC::LIS8:
   case PPC::ADDIStocHA:
   case PPC::ADDIStocHA8:
   case PPC::ADDItocL:
   case PPC::LOAD_STACK_GUARD:
   case PPC::XXLXORz:
   case PPC::XXLXORspz:

llvm/lib/Target/PowerPC/PPCInstrInfo.td

 // and get a i32 with the resulting bits.
 def getFPAs32BitInt : SDNodeXForm<fpimm, [{
   APFloat APFloatOfN = N->getValueAPF();
   convertToNonDenormSingle(APFloatOfN);
   return CurDAG->getTargetConstant(APFloatOfN.bitcastToAPInt().getZExtValue(),
                                    SDLoc(N), MVT::i32);
 }]>;
 
+def imm34 : PatLeaf<(imm), [{
+  return isInt<34>(N->getSExtValue());
+}]>;
+
+def getImmAs64BitInt : SDNodeXForm<imm, [{
+  return getI64Imm(N->getSExtValue(), SDLoc(N));
+}]>;
+
 def SHL32 : SDNodeXForm<imm, [{
   // Transformation function: 31 - imm
   return getI32Imm(31 - N->getZExtValue(), SDLoc(N));
 }]>;
 
 def SRL32 : SDNodeXForm<imm, [{
   // Transformation function: 32 - imm
   return N->getZExtValue() ? getI32Imm(32 - N->getZExtValue(), SDLoc(N))

llvm/lib/Target/PowerPC/PPCInstrPrefix.td

 def MMA : Predicate<"Subtarget->hasMMA()">;
 
 def RCCp {
   dag AToVSRC = (COPY_TO_REGCLASS $XA, VSRC);
   dag BToVSRC = (COPY_TO_REGCLASS $XB, VSRC);
 }
 
 let Predicates = [PrefixInstrs] in {
+  let CodeSize = 2 in
   let Interpretation64Bit = 1, isCodeGenOnly = 1 in {
     defm PADDI8 :
       MLS_DForm_R_SI34_RTA5_p<14, (outs g8rc:$RT), (ins g8rc:$RA, s34imm:$SI),
                               (ins immZero:$RA, s34imm_pcrel:$SI),
                               "paddi $RT, $RA, $SI", IIC_LdStLFD>;
     let isReMaterializable = 1, isAsCheapAsAMove = 1, isMoveImm = 1 in {
       def PLI8 : MLS_DForm_SI34_RT5<14, (outs g8rc:$RT),
                                     (ins s34imm:$SI),
def : Pat<(f64 nzFPImmAsi32:$A),
   def : xxevalPattern<(or (vnot_ppc v4i32:$vA), (xor v4i32:$vB, v4i32:$vC)),
                        !sub(255, 9)>;
   // nand(A, nand(B, C))
   def : xxevalPattern<(or (vnot_ppc v4i32:$vA), (and v4i32:$vB, v4i32:$vC)),
                        !sub(255, 14)>;
 }
 
 let Predicates = [PrefixInstrs] in {
+  def : Pat<(i32 imm34:$imm), (PLI (getImmAs64BitInt imm:$imm))>;
+  def : Pat<(i64 imm34:$imm), (PLI8 (getImmAs64BitInt imm:$imm))>;
   def : Pat<(v16i8 (int_ppc_vsx_xxpermx v16i8:$A, v16i8:$B, v16i8:$C, timm:$D)),
             (COPY_TO_REGCLASS (XXPERMX (COPY_TO_REGCLASS $A, VSRC),
                                        (COPY_TO_REGCLASS $B, VSRC),
                                        (COPY_TO_REGCLASS $C, VSRC), $D), VSRC)>;
   def : Pat<(v16i8 (int_ppc_vsx_xxblendvb v16i8:$A, v16i8:$B, v16i8:$C)),
             (COPY_TO_REGCLASS
                    (XXBLENDVB (COPY_TO_REGCLASS $A, VSRC),
                               (COPY_TO_REGCLASS $B, VSRC),

llvm/test/CodeGen/PowerPC/mma-intrinsics.ll

 }
 
 ; Function Attrs: nounwind
 define void @test_ldst_7(<256 x i1>* %vpp, <256 x i1>* %vp2)  {
 ; FIXME: A prefixed load (plxvp) is expected here as the offset in this
 ; test case is a constant that fits within 34-bits.
 ; CHECK-LABEL: test_ldst_7:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    li r5, 0
-; CHECK-NEXT:    ori r5, r5, 32799
+; CHECK-NEXT:    pli r5, 32799
 ; CHECK-NEXT:    lxvpx vsp0, r3, r5
 ; CHECK-NEXT:    stxvpx vsp0, r4, r5
 ; CHECK-NEXT:    blr
 ;
 ; CHECK-BE-LABEL: test_ldst_7:
 ; CHECK-BE:       # %bb.0: # %entry
-; CHECK-BE-NEXT:    li r5, 0
-; CHECK-BE-NEXT:    ori r5, r5, 32799
+; CHECK-BE-NEXT:    pli r5, 32799
 ; CHECK-BE-NEXT:    lxvpx vsp0, r3, r5
 ; CHECK-BE-NEXT:    stxvpx vsp0, r4, r5
 ; CHECK-BE-NEXT:    blr
 entry:
   %0 = bitcast <256 x i1>* %vpp to i8*
   %1 = getelementptr i8, i8* %0, i64 32799
   %2 = tail call <256 x i1> @llvm.ppc.mma.lxvp(i8* %1)
   %3 = bitcast <256 x i1>* %vp2 to i8*

llvm/test/CodeGen/PowerPC/p10-constants.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -verify-machineinstrs -mtriple=powerpc64le-unknown-linux-gnu \
+; RUN:   -mcpu=pwr10 -ppc-asm-full-reg-names -ppc-vsr-nums-as-vr < %s | \
+; RUN:   FileCheck %s
+; RUN: llc -verify-machineinstrs -mtriple=powerpc64-unknown-linux-gnu \
+; RUN:   -mcpu=pwr10 -ppc-asm-full-reg-names -ppc-vsr-nums-as-vr < %s | \
+; RUN:   FileCheck %s
+
+; These test cases aim to test constant materialization using the pli instruction on Power10.
+
+define  signext i32 @t_16BitsMinRequiring34Bits() {
+; CHECK-LABEL: t_16BitsMinRequiring34Bits:
+; CHECK:  pli r3, 32768
+; CHECK-NEXT: blr
+
+entry:
+  ret i32 32768
+}
+
+define  signext i32 @t_16Bits() {
+; CHECK-LABEL: t_16Bits:
+; CHECK:  pli r3, 62004
+; CHECK-NEXT: blr
+
+entry:
+  ret i32 62004
+}
+
+define  signext i32 @t_lt32gt16BitsNonShiftable() {
+; CHECK-LABEL: t_lt32gt16BitsNonShiftable:
+; CHECK:  pli r3, 1193046
+; CHECK-NEXT: blr
+
+entry:
+  ret i32 1193046
+}
+
+define  signext i32 @t_32Bits() {
+; CHECK-LABEL: t_32Bits:
+; CHECK:  pli r3, -231451016
+; CHECK-NEXT: blr
+
+entry:
+  ret i32 -231451016
+}
+
+define  i64 @t_34BitsLargestPositive() {
+; CHECK-LABEL: t_34BitsLargestPositive:
+; CHECK:  pli r3, 8589934591
+; CHECK-NEXT: blr
+
+entry:
+  ret i64 8589934591
+}
+
+define  i64 @t_neg34Bits() {
+; CHECK-LABEL: t_neg34Bits:
+; CHECK:  pli r3, -8284514696
+; CHECK-NEXT: blr
+
+entry:
+  ret i64 -8284514696
+}
+
+define  signext i32 @t_16BitsMinRequiring34BitsMinusOne() {
+; CHECK-LABEL: t_16BitsMinRequiring34BitsMinusOne:
+; CHECK:  li r3, 32767
+; CHECK-NEXT: blr
+
+entry:
+  ret i32 32767
+}
+
+define  signext i32 @t_lt16Bits() {
+; CHECK-LABEL: t_lt16Bits:
+; CHECK:  li r3, 291
+; CHECK-NEXT: blr
+
+entry:
+  ret i32 291
+}
+
+define  signext i32 @t_neglt16Bits() {
+; CHECK-LABEL: t_neglt16Bits:
+; CHECK:  li r3, -3805
+; CHECK-NEXT: blr
+
+entry:
+  ret i32 -3805
+}
+
+define  signext i32 @t_neg16Bits() {
+; CHECK-LABEL: t_neg16Bits:
+; CHECK:  li r3, -32204
+; CHECK-NEXT: blr
+
+entry:
+  ret i32 -32204
+}
+
+define  signext i32 @t_lt32gt16BitsShiftable() {
+; CHECK-LABEL: t_lt32gt16BitsShiftable:
+; CHECK:  lis r3, 18
+; CHECK-NEXT: blr
+
+entry:
+  ret i32 1179648
+}
+
+define  signext i32 @t_32gt16BitsShiftable() {
+; CHECK-LABEL: t_32gt16BitsShiftable:
+; CHECK:  lis r3, -3532
+; CHECK-NEXT: blr
+
+entry:
+  ret i32 -231473152
+}
+
+define  signext i32 @t_32BitsZero() {
+; CHECK-LABEL: t_32BitsZero:
+; CHECK:  li r3, 0
+; CHECK-NEXT: blr
+
+entry:
+  ret i32 0
+}
+
+define  signext i32 @t_32BitsAllOnes() {
+; CHECK-LABEL: t_32BitsAllOnes:
+; CHECK:  li r3, -1
+; CHECK-NEXT: blr
+
+entry:
+  ret i32 -1
+}
+
+define  i64 @t_34BitsLargestPositivePlus() {
+; CHECK-LABEL: t_34BitsLargestPositivePlus:
+; CHECK:  li r3, 1
+; CHECK-NEXT: sldi r3, r3, 33
+; CHECK-NEXT: blr
+
+entry:
+  ret i64 8589934592
+}
+
+define  i64 @t_34Bits() {
+; CHECK-LABEL: t_34Bits:
+; CHECK:  lis r3, 25158
+; CHECK-NEXT: ori r3, r3, 35535
+; CHECK-NEXT: sldi r3, r3, 3
+; CHECK-NEXT: blr
+
+entry:
+  ret i64 13190321784
+}
+
+define  i64 @t_35Bits() {
+; CHECK-LABEL: t_35Bits:
+; CHECK:  li r3, 7
+; CHECK-NEXT: sldi r3, r3, 32
+; CHECK-NEXT: oris r3, r3, 62004
+; CHECK-NEXT: ori r3, r3, 22136
+; CHECK-NEXT: blr
+
+entry:
+  ret i64 34128287352
+}
+
+; The load immediates resulting from phi-nodes are needed to test whether
+; li/lis is preferred to pli by the instruction selector.
+define dso_local void @t_phiNode() {
+; CHECK-LABEL: t_phiNode:
+; CHECK:  lis r6, 18
+; CHECK-NEXT: li r5, 291
+; CHECK-NEXT: li r4, 0
+; CHECK-NEXT:   cmpwi r3, 1
+; CHECK-NEXT: li r3, -1
+; CHECK:  pli r6, 2147483647
+; CHECK-NEXT: pli r5, 1193046
+; CHECK-NEXT: pli r4, 32768
+; CHECK-NEXT: pli r3, -231451016
+
+entry:
+  br label %while.body
+
+while.body:                                       ; preds = %if.else.i, %entry
+  br label %while.body.i
+
+while.body.i:                                     ; preds = %sw.epilog.i, %while.body
+  %a.1.i = phi i32 [ %a.2.i, %sw.epilog.i ], [ -1, %while.body ]
+  %b.1.i = phi i32 [ %b.2.i, %sw.epilog.i ], [ 0, %while.body ]
+  %c.1.i = phi i32 [ %c.2.i, %sw.epilog.i ], [ 291, %while.body ]
+  %d.1.i = phi i32 [ %d.2.i, %sw.epilog.i ], [ 1179648, %while.body ]
+  %0 = load i8, i8* null, align 1
+  %cmp1.i = icmp eq i8 %0, 1
+  br i1 %cmp1.i, label %if.then.i, label %if.else.i
+
+if.then.i:                                        ; preds = %while.body.i
+  switch i8 undef, label %sw.default.i [
+    i8 3, label %sw.epilog.i
+    i8 2, label %sw.bb1.i
+  ]
+
+sw.bb1.i:                                        ; preds = %if.then.i
+  br label %sw.epilog.i
+
+sw.default.i:                                     ; preds = %if.then.i
+  unreachable
+
+sw.epilog.i:                                      ; preds = %sw.bb2.i, %sw.bb1.i, %if.then.i
+  %a.2.i = phi i32 [ -231451016, %sw.bb1.i ], [ %a.1.i, %if.then.i ]
+  %b.2.i = phi i32 [ 32768, %sw.bb1.i ], [ %b.1.i, %if.then.i ]
+  %c.2.i = phi i32 [ 1193046, %sw.bb1.i ], [ %c.1.i, %if.then.i ]
+  %d.2.i = phi i32 [ 2147483647, %sw.bb1.i ], [ %d.1.i, %if.then.i ]
+  br label %while.body.i
+
+if.else.i:                                     ; preds = %while.body.i
+  call void @func2(i32 signext %a.1.i, i32 signext %b.1.i, i32 signext %c.1.i, i32 signext %d.1.i)
+  br label %while.body
+}
+
+declare void @func2(i32, i32, i32, i32)