[PowerPC] Materialize special ConstantFP using instructions instead of load from TOC

[tingwang]

Summary

ConstantFP in the range of [-16, 15] (excluding 0.0, which is done by XXLXOR) that can be exactly converted to integer can be materialized into register using VSPLTISW and XVCVSXWDP. This will reduce TOC usage, get rid of memory reference, and maybe save some TOC pointer calculations.

ConstantFP is expanded into load from TOC during ExpandNode. This patch identifies the opportunity before expansion, and does the manual instruction selection later. LIT test shows this patch exposes some other issues, please see my comments on the test scripts.

Since Power10 use prefixed instructions to materialize ConstantFP, maybe target Power9 for this change.

[tingwang]

llvm/test/CodeGen/PowerPC/handle-f16-storage-type.ll

1255

"beqlr 0" converted into "beq 0, .LBB0_2; .LBB0_2: blr" sequence, introduced dummy jump. This maybe one issue.

llvm/test/CodeGen/PowerPC/scalar_cmp.ll

908

Given the operand is known and isFPImmLegal(<negative>) is checked, fsub is converted into fadd.
Combining: t15: f64 = fsub t2, ConstantFP:f64<1.000000e+00>
Creating fp constant: t17: f64 = ConstantFP<-1.000000e+00>
Creating new node: t18: f64 = fadd t2, ConstantFP:f64<-1.000000e+00>
... into: t18: f64 = fadd t2, ConstantFP:f64<-1.000000e+00>

932

This maybe one issue:
Given below DAG:

  t16: i1 = setcc t2, ConstantFP:f64<1.000000e+00>, setlt:ch                                               
  t18: i1 = setcc t2, ConstantFP:f64<1.000000e+00>, setuo:ch                                                          
t19: i1 = or t16, t18

Some logic optimized the second setcc:
Combining: t18: i1 = setcc t2, ConstantFP:f64<1.000000e+00>, setuo:ch
Creating new node: t23: i1 = setcc t2, t2, setuo:ch

So that resulted two statements which should have been executed in one fcmpu if there is no combine.

  t16: i1 = setcc t2, ConstantFP:f64<1.000000e+00>, setlt:ch
  t23: i1 = setcc t2, t2, setuo:ch
t19: i1 = or t16, t23

[tingwang]

Used clang-format to tidy the code style

[nemanjai]

I am wondering if you have done any performance comparison or even just latency/throughput "back-of-the-envelope" computation for this. It doesn't seem obvious to me that this is better than a CP load. The conversion is a fairly expensive instruction. Another less compelling thing to consider is that in high register pressure situations (for VMX registers), the VSPLTISW may cause a spill.

One thing to note is that this necessarily replaces the cracked lfs instruction which is also expensive, so this transformation may be worthwhile. However, I wonder how this transformation would compare to stopping the conversion from double precision to single precision for constants on Power9 (i.e. return false from ShouldShrinkFPConstant() on Power9). This would of course grow the constant pool in situations where a lot of compact constants are loaded, but presumably this is a somewhat rare situation.

[tingwang]

In D118316#3300414, @nemanjai wrote:

I am wondering if you have done any performance comparison or even just latency/throughput "back-of-the-envelope" computation for this. It doesn't seem obvious to me that this is better than a CP load. The conversion is a fairly expensive instruction. Another less compelling thing to consider is that in high register pressure situations (for VMX registers), the VSPLTISW may cause a spill.

One thing to note is that this necessarily replaces the cracked lfs instruction which is also expensive, so this transformation may be worthwhile. However, I wonder how this transformation would compare to stopping the conversion from double precision to single precision for constants on Power9 (i.e. return false from ShouldShrinkFPConstant() on Power9). This would of course grow the constant pool in situations where a lot of compact constants are loaded, but presumably this is a somewhat rare situation.

Thank you for looking into this. I'm planning to do performance comparison, and will check/compare the performance with stopping the conversion from double to single as you mentioned (LFS vs. LFD). I do agree the 'VSPLTISW' may cause a vector spill is a concern, and that cost is higher than CP load spill fixed point register, although this depends on register usage.

Simple calculation shows that from instruction latency perspective (P9InstrResources.td) VSPLTISW + XVCVSXWDP is 4 + 7 = 11, and CP load ADDI + ADDIS + LFS is 2 + 2 + 7 = 11, which is the same. Although CP load has memory access, its ADDIS maybe saved in consecutive CP loads. Looks like this maybe a draw, so I do need to check the real performance.

[tingwang]

In D118316#3300414, @nemanjai wrote:

I am wondering if you have done any performance comparison or even just latency/throughput "back-of-the-envelope" computation for this. It doesn't seem obvious to me that this is better than a CP load. The conversion is a fairly expensive instruction. Another less compelling thing to consider is that in high register pressure situations (for VMX registers), the VSPLTISW may cause a spill.

One thing to note is that this necessarily replaces the cracked lfs instruction which is also expensive, so this transformation may be worthwhile. However, I wonder how this transformation would compare to stopping the conversion from double precision to single precision for constants on Power9 (i.e. return false from ShouldShrinkFPConstant() on Power9). This would of course grow the constant pool in situations where a lot of compact constants are loaded, but presumably this is a somewhat rare situation.

Hi Nemanja, I measured there is no performance change on spec base fprate, so I think should abandon this one. Sorry for the bother. By the way, flip ShouldShrinkFPConstant() improved single-copy base fprate by about 0.2%, though it's single run result.

[tingwang]

No perf gain according to some spec test.

llvm/lib/Target/PowerPC/PPCISelDAGToDAG.cpp

void PPCDAGToDAGISel::Select(SDNode *N) {
 
   case ISD::Constant:
     if (N->getValueType(0) == MVT::i64) {
       ReplaceNode(N, selectI64Imm(CurDAG, N));
       return;
     }
     break;
 
+  case ISD::ConstantFP:
+    if (Subtarget->isISA3_0() && Subtarget->isPPC64() &&
+        !Subtarget->hasPrefixInstrs() &&
+        (N->getValueType(0) == MVT::f64 || N->getValueType(0) == MVT::f32)) {
+      ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(N);
+      bool isExact = false;
+      APSInt Val(5, false);
+      ConstantFP *LLVMC = const_cast<ConstantFP *>(CFP->getConstantFPValue());
+      if (LLVMC->getValueAPF().convertToInteger(Val, APFloat::rmTowardZero,
+                                                &isExact) == APFloat::opOK &&
+          isExact == true && Val.getExtValue() != 0) {
+        SDNode *Op1 = CurDAG->getMachineNode(PPC::VSPLTISW, dl, MVT::v4i32,
+                                             getI32Imm(Val.getExtValue(), dl));
+        SDNode *Op2 = CurDAG->getMachineNode(PPC::XVCVSXWDP, dl, MVT::v2f64,
+                                             SDValue(Op1, 0));
+        CurDAG->SelectNodeTo(N, PPC::COPY_TO_REGCLASS, N->getValueType(0),
+                             SDValue(Op2, 0),
+                             getI32Imm(PPC::VSFRCRegClassID, dl));
+        return;
+      }
+    }
+    break;
+
   case ISD::INTRINSIC_VOID: {
     auto IntrinsicID = N->getConstantOperandVal(1);
     if (IntrinsicID == Intrinsic::ppc_tdw || IntrinsicID == Intrinsic::ppc_tw) {
       unsigned Opcode = IntrinsicID == Intrinsic::ppc_tdw ? PPC::TDI : PPC::TWI;
       SDValue Ops[] = {N->getOperand(4), N->getOperand(2), N->getOperand(3)};
       int16_t SImmOperand2;
       int16_t SImmOperand3;
       int16_t SImmOperand4;

llvm/lib/Target/PowerPC/PPCISelLowering.cpp

default:
     // false. Examples: f16, f80.
     return false;
   case MVT::f32:
   case MVT::f64:
     if (Subtarget.hasPrefixInstrs()) {
       // we can materialize all immediatess via XXSPLTI32DX and XXSPLTIDP.
       return true;
     }
+    if (Subtarget.isISA3_0() && Subtarget.isPPC64() &&
+        !Subtarget.hasPrefixInstrs()) {
+      // Currently target P9 only
+      bool isExact = false;
+      APSInt Value(5, false);
+      if (Imm.convertToInteger(Value, APFloat::rmTowardZero, &isExact) ==
+              APFloat::opOK &&
+          isExact == true && Value.getExtValue() != 0) {
+        // we can materialize FP in range [-16.0, 15.0] excluding zeros that
+        // can be exactly converted to integer via VSPLTISW and XVCVSXWDP.
+        // 0.0 will select XXLXOR, -0.0 is not exact.
+        return true;
+      }
+    }
     LLVM_FALLTHROUGH;
   case MVT::ppcf128:
     return Imm.isPosZero();
   }
 }
 
 // For vector shift operation op, fold
 // (op x, (and y, ((1 << numbits(x)) - 1))) -> (target op x, y)

llvm/test/CodeGen/PowerPC/handle-f16-storage-type.ll

 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    xscvdphp f0, f1
 ; CHECK-NEXT:    xxlxor f1, f1, f1
 ; CHECK-NEXT:    mffprwz r3, f0
 ; CHECK-NEXT:    clrlwi r3, r3, 16
 ; CHECK-NEXT:    mtfprwz f0, r3
 ; CHECK-NEXT:    xscvhpdp f0, f0
 ; CHECK-NEXT:    fcmpu cr0, f0, f1
-; CHECK-NEXT:    beqlr cr0
+; CHECK-NEXT:    beq cr0, .LBB20_2
 ; CHECK-NEXT:  # %bb.1:
-; CHECK-NEXT:    addis r3, r2, .LCPI20_0@toc@ha
-; CHECK-NEXT:    lfs f1, .LCPI20_0@toc@l(r3)
+; CHECK-NEXT:    vspltisw v2, 1
+; CHECK-NEXT:    xvcvsxwdp vs1, vs34
+; CHECK-NEXT:  .LBB20_2:
+; CHECK-NEXT:    # kill: def $f1 killed $f1 killed $vsl1
 ; CHECK-NEXT:    blr
 ;
 ; SOFT-LABEL: PR40273:
 ; SOFT:       # %bb.0:
 ; SOFT-NEXT:    mflr r0
 ; SOFT-NEXT:    std r0, 16(r1)
 ; SOFT-NEXT:    stdu r1, -32(r1)
 ; SOFT-NEXT:    clrldi r3, r3, 48

llvm/test/CodeGen/PowerPC/recipest.ll

 ; CHECK-P8-NEXT:    xsmuldp 0, 0, 5
 ; CHECK-P8-NEXT:    xsmuldp 0, 0, 4
 ; CHECK-P8-NEXT:    xsmuldp 1, 1, 0
 ; CHECK-P8-NEXT:    blr
 ;
 ; CHECK-P9-LABEL: foo_fmf:
 ; CHECK-P9:       # %bb.0:
 ; CHECK-P9-NEXT:    xsrsqrtedp 0, 2
+; CHECK-P9-NEXT:    vspltisw 2, -3
 ; CHECK-P9-NEXT:    addis 3, 2, .LCPI0_0@toc@ha
-; CHECK-P9-NEXT:    lfs 4, .LCPI0_0@toc@l(3)
-; CHECK-P9-NEXT:    addis 3, 2, .LCPI0_1@toc@ha
 ; CHECK-P9-NEXT:    xsmuldp 3, 2, 0
+; CHECK-P9-NEXT:    xvcvsxwdp 4, 34
 ; CHECK-P9-NEXT:    fmr 5, 4
 ; CHECK-P9-NEXT:    xsmaddadp 5, 3, 0
-; CHECK-P9-NEXT:    lfs 3, .LCPI0_1@toc@l(3)
+; CHECK-P9-NEXT:    lfs 3, .LCPI0_0@toc@l(3)
 ; CHECK-P9-NEXT:    xsmuldp 0, 0, 3
 ; CHECK-P9-NEXT:    xsmuldp 0, 0, 5
 ; CHECK-P9-NEXT:    xsmuldp 2, 2, 0
 ; CHECK-P9-NEXT:    xsmaddadp 4, 2, 0
 ; CHECK-P9-NEXT:    xsmuldp 0, 0, 3
 ; CHECK-P9-NEXT:    xsmuldp 0, 0, 4
 ; CHECK-P9-NEXT:    xsmuldp 1, 1, 0
 ; CHECK-P9-NEXT:    blr

 ; CHECK-P8-NEXT:    xsmaddasp 4, 2, 0
 ; CHECK-P8-NEXT:    xsmulsp 0, 3, 4
 ; CHECK-P8-NEXT:    xsmuldp 1, 1, 0
 ; CHECK-P8-NEXT:    blr
 ;
 ; CHECK-P9-LABEL: foof_fmf:
 ; CHECK-P9:       # %bb.0:
 ; CHECK-P9-NEXT:    xsrsqrtesp 0, 2
+; CHECK-P9-NEXT:    vspltisw 2, -3
 ; CHECK-P9-NEXT:    addis 3, 2, .LCPI3_0@toc@ha
-; CHECK-P9-NEXT:    lfs 3, .LCPI3_0@toc@l(3)
-; CHECK-P9-NEXT:    addis 3, 2, .LCPI3_1@toc@ha
 ; CHECK-P9-NEXT:    xsmulsp 2, 2, 0
+; CHECK-P9-NEXT:    xvcvsxwdp 3, 34
 ; CHECK-P9-NEXT:    xsmaddasp 3, 2, 0
-; CHECK-P9-NEXT:    lfs 2, .LCPI3_1@toc@l(3)
+; CHECK-P9-NEXT:    lfs 2, .LCPI3_0@toc@l(3)
 ; CHECK-P9-NEXT:    xsmulsp 0, 0, 2
 ; CHECK-P9-NEXT:    xsmulsp 0, 0, 3
 ; CHECK-P9-NEXT:    xsmuldp 1, 1, 0
 ; CHECK-P9-NEXT:    blr
   %x = call contract reassoc arcp float @llvm.sqrt.f32(float %b)
   %y = fpext float %x to double
   %r = fdiv contract reassoc arcp double %a, %y
   ret double %r

 ; CHECK-P8-NEXT:    xsmuldp 0, 0, 4
 ; CHECK-P8-NEXT:    xsrsp 0, 0
 ; CHECK-P8-NEXT:    xsmulsp 1, 1, 0
 ; CHECK-P8-NEXT:    blr
 ;
 ; CHECK-P9-LABEL: food_fmf:
 ; CHECK-P9:       # %bb.0:
 ; CHECK-P9-NEXT:    xsrsqrtedp 0, 2
+; CHECK-P9-NEXT:    vspltisw 2, -3
 ; CHECK-P9-NEXT:    addis 3, 2, .LCPI5_0@toc@ha
-; CHECK-P9-NEXT:    lfs 4, .LCPI5_0@toc@l(3)
-; CHECK-P9-NEXT:    addis 3, 2, .LCPI5_1@toc@ha
 ; CHECK-P9-NEXT:    xsmuldp 3, 2, 0
+; CHECK-P9-NEXT:    xvcvsxwdp 4, 34
 ; CHECK-P9-NEXT:    fmr 5, 4
 ; CHECK-P9-NEXT:    xsmaddadp 5, 3, 0
-; CHECK-P9-NEXT:    lfs 3, .LCPI5_1@toc@l(3)
+; CHECK-P9-NEXT:    lfs 3, .LCPI5_0@toc@l(3)
 ; CHECK-P9-NEXT:    xsmuldp 0, 0, 3
 ; CHECK-P9-NEXT:    xsmuldp 0, 0, 5
 ; CHECK-P9-NEXT:    xsmuldp 2, 2, 0
 ; CHECK-P9-NEXT:    xsmaddadp 4, 2, 0
 ; CHECK-P9-NEXT:    xsmuldp 0, 0, 3
 ; CHECK-P9-NEXT:    xsmuldp 0, 0, 4
 ; CHECK-P9-NEXT:    xsrsp 0, 0
 ; CHECK-P9-NEXT:    xsmulsp 1, 1, 0

 ; CHECK-P8-NEXT:    xsmaddasp 4, 2, 0
 ; CHECK-P8-NEXT:    xsmulsp 0, 3, 4
 ; CHECK-P8-NEXT:    xsmulsp 1, 1, 0
 ; CHECK-P8-NEXT:    blr
 ;
 ; CHECK-P9-LABEL: goo_fmf:
 ; CHECK-P9:       # %bb.0:
 ; CHECK-P9-NEXT:    xsrsqrtesp 0, 2
+; CHECK-P9-NEXT:    vspltisw 2, -3
 ; CHECK-P9-NEXT:    addis 3, 2, .LCPI7_0@toc@ha
-; CHECK-P9-NEXT:    lfs 3, .LCPI7_0@toc@l(3)
-; CHECK-P9-NEXT:    addis 3, 2, .LCPI7_1@toc@ha
 ; CHECK-P9-NEXT:    xsmulsp 2, 2, 0
+; CHECK-P9-NEXT:    xvcvsxwdp 3, 34
 ; CHECK-P9-NEXT:    xsmaddasp 3, 2, 0
-; CHECK-P9-NEXT:    lfs 2, .LCPI7_1@toc@l(3)
+; CHECK-P9-NEXT:    lfs 2, .LCPI7_0@toc@l(3)
 ; CHECK-P9-NEXT:    xsmulsp 0, 0, 2
 ; CHECK-P9-NEXT:    xsmulsp 0, 0, 3
 ; CHECK-P9-NEXT:    xsmulsp 1, 1, 0
 ; CHECK-P9-NEXT:    blr
   %x = call contract reassoc arcp float @llvm.sqrt.f32(float %b)
   %r = fdiv contract reassoc arcp float %a, %x
   ret float %r
 }

 ; CHECK-P8-NEXT:    xsnmsubasp 0, 2, 4
 ; CHECK-P8-NEXT:    xsmaddasp 4, 1, 0
 ; CHECK-P8-NEXT:    xsmulsp 1, 3, 4
 ; CHECK-P8-NEXT:    blr
 ;
 ; CHECK-P9-LABEL: rsqrt_fmul_fmf:
 ; CHECK-P9:       # %bb.0:
 ; CHECK-P9-NEXT:    xsrsqrtesp 0, 1
+; CHECK-P9-NEXT:    vspltisw 2, -3
 ; CHECK-P9-NEXT:    addis 3, 2, .LCPI10_0@toc@ha
-; CHECK-P9-NEXT:    lfs 4, .LCPI10_0@toc@l(3)
-; CHECK-P9-NEXT:    addis 3, 2, .LCPI10_1@toc@ha
 ; CHECK-P9-NEXT:    xsmulsp 1, 1, 0
+; CHECK-P9-NEXT:    xvcvsxwdp 4, 34
 ; CHECK-P9-NEXT:    xsmaddasp 4, 1, 0
-; CHECK-P9-NEXT:    lfs 1, .LCPI10_1@toc@l(3)
+; CHECK-P9-NEXT:    lfs 1, .LCPI10_0@toc@l(3)
 ; CHECK-P9-NEXT:    xsmulsp 0, 0, 1
 ; CHECK-P9-NEXT:    xsresp 1, 2
 ; CHECK-P9-NEXT:    xsmulsp 0, 0, 4
 ; CHECK-P9-NEXT:    xsmulsp 4, 0, 1
 ; CHECK-P9-NEXT:    xsnmsubasp 0, 2, 4
 ; CHECK-P9-NEXT:    xsmaddasp 4, 1, 0
 ; CHECK-P9-NEXT:    xsmulsp 1, 3, 4
 ; CHECK-P9-NEXT:    blr

 ; CHECK-P8-NEXT:    xsmuldp 0, 1, 3
 ; CHECK-P8-NEXT:    xsnmsubadp 1, 2, 0
 ; CHECK-P8-NEXT:    xsmaddadp 0, 3, 1
 ; CHECK-P8-NEXT:    fmr 1, 0
 ; CHECK-P8-NEXT:    blr
 ;
 ; CHECK-P9-LABEL: foo2_fmf:
 ; CHECK-P9:       # %bb.0:
-; CHECK-P9-NEXT:    addis 3, 2, .LCPI14_0@toc@ha
+; CHECK-P9-NEXT:    vspltisw 2, -1
 ; CHECK-P9-NEXT:    xsredp 3, 2
-; CHECK-P9-NEXT:    lfs 0, .LCPI14_0@toc@l(3)
+; CHECK-P9-NEXT:    xvcvsxwdp 0, 34
 ; CHECK-P9-NEXT:    xsmaddadp 0, 2, 3
 ; CHECK-P9-NEXT:    xsnmsubadp 3, 3, 0
 ; CHECK-P9-NEXT:    xsmuldp 0, 1, 3
 ; CHECK-P9-NEXT:    xsnmsubadp 1, 2, 0
 ; CHECK-P9-NEXT:    xsmaddadp 0, 3, 1
 ; CHECK-P9-NEXT:    fmr 1, 0
 ; CHECK-P9-NEXT:    blr
   %r = fdiv contract reassoc arcp nsz ninf double %a, %b

 ; CHECK-P8-NEXT:    blr
 ;
 ; CHECK-P9-LABEL: foo3_fmf:
 ; CHECK-P9:       # %bb.0:
 ; CHECK-P9-NEXT:    xstsqrtdp 0, 1
 ; CHECK-P9-NEXT:    bc 12, 2, .LBB20_2
 ; CHECK-P9-NEXT:  # %bb.1:
 ; CHECK-P9-NEXT:    xsrsqrtedp 0, 1
+; CHECK-P9-NEXT:    vspltisw 2, -3
 ; CHECK-P9-NEXT:    addis 3, 2, .LCPI20_0@toc@ha
-; CHECK-P9-NEXT:    lfs 3, .LCPI20_0@toc@l(3)
-; CHECK-P9-NEXT:    addis 3, 2, .LCPI20_1@toc@ha
 ; CHECK-P9-NEXT:    xsmuldp 2, 1, 0
+; CHECK-P9-NEXT:    xvcvsxwdp 3, 34
 ; CHECK-P9-NEXT:    fmr 4, 3
 ; CHECK-P9-NEXT:    xsmaddadp 4, 2, 0
-; CHECK-P9-NEXT:    lfs 2, .LCPI20_1@toc@l(3)
+; CHECK-P9-NEXT:    lfs 2, .LCPI20_0@toc@l(3)
 ; CHECK-P9-NEXT:    xsmuldp 0, 0, 2
 ; CHECK-P9-NEXT:    xsmuldp 0, 0, 4
 ; CHECK-P9-NEXT:    xsmuldp 1, 1, 0
 ; CHECK-P9-NEXT:    xsmaddadp 3, 1, 0
 ; CHECK-P9-NEXT:    xsmuldp 0, 1, 2
 ; CHECK-P9-NEXT:    xsmuldp 1, 0, 3
 ; CHECK-P9-NEXT:    blr
 ; CHECK-P9-NEXT:  .LBB20_2:

 ; CHECK-P8-NEXT:    xsmuldp 1, 0, 3
 ; CHECK-P8-NEXT:    blr
 ; CHECK-P8-NEXT:  .LBB21_2:
 ; CHECK-P8-NEXT:    xssqrtdp 1, 1
 ; CHECK-P8-NEXT:    blr
 ;
 ; CHECK-P9-LABEL: foo3_fmf_crbits_off:
 ; CHECK-P9:       # %bb.0:
-; CHECK-P9-NEXT:    addis 3, 2, .LCPI21_2@toc@ha
+; CHECK-P9-NEXT:    addis 3, 2, .LCPI21_1@toc@ha
 ; CHECK-P9-NEXT:    xsabsdp 0, 1
-; CHECK-P9-NEXT:    lfd 2, .LCPI21_2@toc@l(3)
+; CHECK-P9-NEXT:    lfd 2, .LCPI21_1@toc@l(3)
 ; CHECK-P9-NEXT:    xscmpudp 0, 0, 2
 ; CHECK-P9-NEXT:    blt 0, .LBB21_2
 ; CHECK-P9-NEXT:  # %bb.1:
 ; CHECK-P9-NEXT:    xsrsqrtedp 0, 1
+; CHECK-P9-NEXT:    vspltisw 2, -3
 ; CHECK-P9-NEXT:    addis 3, 2, .LCPI21_0@toc@ha
-; CHECK-P9-NEXT:    lfs 3, .LCPI21_0@toc@l(3)
-; CHECK-P9-NEXT:    addis 3, 2, .LCPI21_1@toc@ha
 ; CHECK-P9-NEXT:    xsmuldp 2, 1, 0
+; CHECK-P9-NEXT:    xvcvsxwdp 3, 34
 ; CHECK-P9-NEXT:    fmr 4, 3
 ; CHECK-P9-NEXT:    xsmaddadp 4, 2, 0
-; CHECK-P9-NEXT:    lfs 2, .LCPI21_1@toc@l(3)
+; CHECK-P9-NEXT:    lfs 2, .LCPI21_0@toc@l(3)
 ; CHECK-P9-NEXT:    xsmuldp 0, 0, 2
 ; CHECK-P9-NEXT:    xsmuldp 0, 0, 4
 ; CHECK-P9-NEXT:    xsmuldp 1, 1, 0
 ; CHECK-P9-NEXT:    xsmaddadp 3, 1, 0
 ; CHECK-P9-NEXT:    xsmuldp 0, 1, 2
 ; CHECK-P9-NEXT:    xsmuldp 1, 0, 3
 ; CHECK-P9-NEXT:    blr
 ; CHECK-P9-NEXT:  .LBB21_2:

 ; CHECK-P8-NEXT:    xsmulsp 0, 1, 3
 ; CHECK-P8-NEXT:    xsmulsp 0, 0, 2
 ; CHECK-P8-NEXT:  .LBB23_2:
 ; CHECK-P8-NEXT:    fmr 1, 0
 ; CHECK-P8-NEXT:    blr
 ;
 ; CHECK-P9-LABEL: goo3_fmf:
 ; CHECK-P9:       # %bb.0:
-; CHECK-P9-NEXT:    addis 3, 2, .LCPI23_2@toc@ha
+; CHECK-P9-NEXT:    addis 3, 2, .LCPI23_1@toc@ha
 ; CHECK-P9-NEXT:    xsabsdp 0, 1
-; CHECK-P9-NEXT:    lfs 2, .LCPI23_2@toc@l(3)
+; CHECK-P9-NEXT:    lfs 2, .LCPI23_1@toc@l(3)
 ; CHECK-P9-NEXT:    fcmpu 0, 0, 2
 ; CHECK-P9-NEXT:    xxlxor 0, 0, 0
 ; CHECK-P9-NEXT:    blt 0, .LBB23_2
 ; CHECK-P9-NEXT:  # %bb.1:
 ; CHECK-P9-NEXT:    xsrsqrtesp 0, 1
+; CHECK-P9-NEXT:    vspltisw 2, -3
 ; CHECK-P9-NEXT:    addis 3, 2, .LCPI23_0@toc@ha
-; CHECK-P9-NEXT:    lfs 2, .LCPI23_0@toc@l(3)
-; CHECK-P9-NEXT:    addis 3, 2, .LCPI23_1@toc@ha
 ; CHECK-P9-NEXT:    xsmulsp 1, 1, 0
+; CHECK-P9-NEXT:    xvcvsxwdp 2, 34
 ; CHECK-P9-NEXT:    xsmaddasp 2, 1, 0
-; CHECK-P9-NEXT:    lfs 0, .LCPI23_1@toc@l(3)
+; CHECK-P9-NEXT:    lfs 0, .LCPI23_0@toc@l(3)
 ; CHECK-P9-NEXT:    xsmulsp 0, 1, 0
 ; CHECK-P9-NEXT:    xsmulsp 0, 0, 2
 ; CHECK-P9-NEXT:  .LBB23_2:
 ; CHECK-P9-NEXT:    fmr 1, 0
 ; CHECK-P9-NEXT:    blr
   %r = call contract reassoc ninf afn float @llvm.sqrt.f32(float %a)
   ret float %r
 }

llvm/test/CodeGen/PowerPC/scalar_cmp.ll

 ; FAST-P8-NEXT:    addis r3, r2, .LCPI24_0@toc@ha
 ; FAST-P8-NEXT:    lfs f0, .LCPI24_0@toc@l(r3)
 ; FAST-P8-NEXT:    xssubdp f0, f1, f0
 ; FAST-P8-NEXT:    fsel f1, f0, f2, f3
 ; FAST-P8-NEXT:    blr
 ;
 ; FAST-P9-LABEL: onecmp1:
 ; FAST-P9:       # %bb.0: # %entry
-; FAST-P9-NEXT:    addis r3, r2, .LCPI24_0@toc@ha
-; FAST-P9-NEXT:    lfs f0, .LCPI24_0@toc@l(r3)
-; FAST-P9-NEXT:    xssubdp f0, f1, f0
+; FAST-P9-NEXT:    vspltisw v2, -1
+; FAST-P9-NEXT:    xvcvsxwdp vs0, vs34
+; FAST-P9-NEXT:    xsadddp f0, f1, f0
 ; FAST-P9-NEXT:    fsel f1, f0, f2, f3
 ; FAST-P9-NEXT:    blr
 ;
 ; NO-FAST-P8-LABEL: onecmp1:
 ; NO-FAST-P8:       # %bb.0: # %entry
 ; NO-FAST-P8-NEXT:    addis r3, r2, .LCPI24_0@toc@ha
 ; NO-FAST-P8-NEXT:    lfs f0, .LCPI24_0@toc@l(r3)
 ; NO-FAST-P8-NEXT:    fcmpu cr0, f1, f0
 ; NO-FAST-P8-NEXT:    cror 4*cr5+lt, lt, un
 ; NO-FAST-P8-NEXT:    bc 12, 4*cr5+lt, .LBB24_2
 ; NO-FAST-P8-NEXT:  # %bb.1: # %entry
 ; NO-FAST-P8-NEXT:    fmr f3, f2
 ; NO-FAST-P8-NEXT:  .LBB24_2: # %entry
 ; NO-FAST-P8-NEXT:    fmr f1, f3
 ; NO-FAST-P8-NEXT:    blr
 ;
 ; NO-FAST-P9-LABEL: onecmp1:
 ; NO-FAST-P9:       # %bb.0: # %entry
-; NO-FAST-P9-NEXT:    addis r3, r2, .LCPI24_0@toc@ha
-; NO-FAST-P9-NEXT:    lfs f0, .LCPI24_0@toc@l(r3)
+; NO-FAST-P9-NEXT:    vspltisw v2, 1
+; NO-FAST-P9-NEXT:    xvcvsxwdp vs0, vs34
 ; NO-FAST-P9-NEXT:    fcmpu cr0, f1, f0
-; NO-FAST-P9-NEXT:    cror 4*cr5+lt, lt, un
-; NO-FAST-P9-NEXT:    bc 12, 4*cr5+lt, .LBB24_2
+; NO-FAST-P9-NEXT:    bc 12, lt, .LBB24_3
 ; NO-FAST-P9-NEXT:  # %bb.1: # %entry
+; NO-FAST-P9-NEXT:    fcmpu cr0, f1, f1
+; NO-FAST-P9-NEXT:    bc 12, un, .LBB24_3
+; NO-FAST-P9-NEXT:  # %bb.2: # %entry
 ; NO-FAST-P9-NEXT:    fmr f3, f2
-; NO-FAST-P9-NEXT:  .LBB24_2: # %entry
+; NO-FAST-P9-NEXT:  .LBB24_3: # %entry
 ; NO-FAST-P9-NEXT:    fmr f1, f3
 ; NO-FAST-P9-NEXT:    blr
 entry:
   %cmp = fcmp ult double %a, 1.000000e+00
   %z.y = select i1 %cmp, double %z, double %y
   ret double %z.y
 }
 
 define double @onecmp2(double %a, double %y, double %z) {
 ; FAST-P8-LABEL: onecmp2:
 ; FAST-P8:       # %bb.0: # %entry
 ; FAST-P8-NEXT:    addis r3, r2, .LCPI25_0@toc@ha
 ; FAST-P8-NEXT:    lfs f0, .LCPI25_0@toc@l(r3)
 ; FAST-P8-NEXT:    xssubdp f0, f0, f1
 ; FAST-P8-NEXT:    fsel f1, f0, f3, f2
 ; FAST-P8-NEXT:    blr
 ;
 ; FAST-P9-LABEL: onecmp2:
 ; FAST-P9:       # %bb.0: # %entry
-; FAST-P9-NEXT:    addis r3, r2, .LCPI25_0@toc@ha
-; FAST-P9-NEXT:    lfs f0, .LCPI25_0@toc@l(r3)
+; FAST-P9-NEXT:    vspltisw v2, 1
+; FAST-P9-NEXT:    xvcvsxwdp vs0, vs34
 ; FAST-P9-NEXT:    xssubdp f0, f0, f1
 ; FAST-P9-NEXT:    fsel f1, f0, f3, f2
 ; FAST-P9-NEXT:    blr
 ;
 ; NO-FAST-P8-LABEL: onecmp2:
 ; NO-FAST-P8:       # %bb.0: # %entry
 ; NO-FAST-P8-NEXT:    addis r3, r2, .LCPI25_0@toc@ha
 ; NO-FAST-P8-NEXT:    lfs f0, .LCPI25_0@toc@l(r3)
 ; NO-FAST-P8-NEXT:    xscmpudp cr0, f1, f0
 ; NO-FAST-P8-NEXT:    fmr f1, f2
 ; NO-FAST-P8-NEXT:    bgtlr cr0
 ; NO-FAST-P8-NEXT:  # %bb.1: # %entry
 ; NO-FAST-P8-NEXT:    fmr f1, f3
 ; NO-FAST-P8-NEXT:    blr
 ;
 ; NO-FAST-P9-LABEL: onecmp2:
 ; NO-FAST-P9:       # %bb.0: # %entry
-; NO-FAST-P9-NEXT:    addis r3, r2, .LCPI25_0@toc@ha
-; NO-FAST-P9-NEXT:    lfs f0, .LCPI25_0@toc@l(r3)
+; NO-FAST-P9-NEXT:    vspltisw v2, 1
+; NO-FAST-P9-NEXT:    xvcvsxwdp vs0, vs34
 ; NO-FAST-P9-NEXT:    xscmpudp cr0, f1, f0
 ; NO-FAST-P9-NEXT:    bgt cr0, .LBB25_2
 ; NO-FAST-P9-NEXT:  # %bb.1: # %entry
 ; NO-FAST-P9-NEXT:    fmr f2, f3
 ; NO-FAST-P9-NEXT:  .LBB25_2: # %entry
 ; NO-FAST-P9-NEXT:    fmr f1, f2
 ; NO-FAST-P9-NEXT:    blr
 entry:

 ; FAST-P8-NEXT:    xssubdp f0, f1, f0
 ; FAST-P8-NEXT:    xsnegdp f1, f0
 ; FAST-P8-NEXT:    fsel f0, f0, f2, f3
 ; FAST-P8-NEXT:    fsel f1, f1, f0, f3
 ; FAST-P8-NEXT:    blr
 ;
 ; FAST-P9-LABEL: onecmp3:
 ; FAST-P9:       # %bb.0: # %entry
-; FAST-P9-NEXT:    addis r3, r2, .LCPI26_0@toc@ha
-; FAST-P9-NEXT:    lfs f0, .LCPI26_0@toc@l(r3)
-; FAST-P9-NEXT:    xssubdp f0, f1, f0
+; FAST-P9-NEXT:    vspltisw v2, -1
+; FAST-P9-NEXT:    xvcvsxwdp vs0, vs34
+; FAST-P9-NEXT:    xsadddp f0, f1, f0
 ; FAST-P9-NEXT:    fsel f1, f0, f2, f3
 ; FAST-P9-NEXT:    xsnegdp f0, f0
 ; FAST-P9-NEXT:    fsel f1, f0, f1, f3
 ; FAST-P9-NEXT:    blr
 ;
 ; NO-FAST-P8-LABEL: onecmp3:
 ; NO-FAST-P8:       # %bb.0: # %entry
 ; NO-FAST-P8-NEXT:    addis r3, r2, .LCPI26_0@toc@ha
 ; NO-FAST-P8-NEXT:    lfs f0, .LCPI26_0@toc@l(r3)
 ; NO-FAST-P8-NEXT:    xscmpudp cr0, f1, f0
 ; NO-FAST-P8-NEXT:    fmr f1, f2
 ; NO-FAST-P8-NEXT:    beqlr cr0
 ; NO-FAST-P8-NEXT:  # %bb.1: # %entry
 ; NO-FAST-P8-NEXT:    fmr f1, f3
 ; NO-FAST-P8-NEXT:    blr
 ;
 ; NO-FAST-P9-LABEL: onecmp3:
 ; NO-FAST-P9:       # %bb.0: # %entry
-; NO-FAST-P9-NEXT:    addis r3, r2, .LCPI26_0@toc@ha
-; NO-FAST-P9-NEXT:    lfs f0, .LCPI26_0@toc@l(r3)
+; NO-FAST-P9-NEXT:    vspltisw v2, 1
+; NO-FAST-P9-NEXT:    xvcvsxwdp vs0, vs34
 ; NO-FAST-P9-NEXT:    xscmpudp cr0, f1, f0
 ; NO-FAST-P9-NEXT:    beq cr0, .LBB26_2
 ; NO-FAST-P9-NEXT:  # %bb.1: # %entry
 ; NO-FAST-P9-NEXT:    fmr f2, f3
 ; NO-FAST-P9-NEXT:  .LBB26_2: # %entry
 ; NO-FAST-P9-NEXT:    fmr f1, f2
 ; NO-FAST-P9-NEXT:    blr
 entry:
   %cmp = fcmp oeq double %a, 1.000000e+00
   %y.z = select i1 %cmp, double %y, double %z
   ret double %y.z
 }

llvm/test/CodeGen/PowerPC/toc-float.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -relocation-model=pic -verify-machineinstrs -mtriple=powerpc64le-unknown-linux-gnu -mcpu=pwr9 < %s | FileCheck -check-prefix=CHECK-P9 %s
 ; RUN: llc -relocation-model=pic -verify-machineinstrs -mtriple=powerpc64le-unknown-linux-gnu -mcpu=pwr8 < %s | FileCheck -check-prefix=CHECK-P8 %s
 
 ; As the constant could be represented as float, a float is
 ; loaded from constant pool.
 define double @doubleConstant1() {
 ; CHECK-P9-LABEL: doubleConstant1:
 ; CHECK-P9:       # %bb.0:
-; CHECK-P9-NEXT:    addis 3, 2, .LCPI0_0@toc@ha
-; CHECK-P9-NEXT:    lfs 1, .LCPI0_0@toc@l(3)
+; CHECK-P9-NEXT:    vspltisw 2, 14
+; CHECK-P9-NEXT:    xvcvsxwdp 1, 34
+; CHECK-P9-NEXT:    # kill: def $f1 killed $f1 killed $vsl1
 ; CHECK-P9-NEXT:    blr
 ;
 ; CHECK-P8-LABEL: doubleConstant1:
 ; CHECK-P8:       # %bb.0:
 ; CHECK-P8-NEXT:    addis 3, 2, .LCPI0_0@toc@ha
 ; CHECK-P8-NEXT:    lfs 1, .LCPI0_0@toc@l(3)
 ; CHECK-P8-NEXT:    blr
   ret double 1.400000e+01

llvm/test/CodeGen/PowerPC/vec_extract_p9.ll

 }
 
 define double @test10(<4 x i32> %a, <4 x i32> %b) {
 ; CHECK-LE-LABEL: test10:
 ; CHECK-LE:       # %bb.0: # %entry
 ; CHECK-LE-NEXT:    addis 3, 2, .LCPI9_0@toc@ha
 ; CHECK-LE-NEXT:    addi 3, 3, .LCPI9_0@toc@l
 ; CHECK-LE-NEXT:    lxv 36, 0(3)
-; CHECK-LE-NEXT:    addis 3, 2, .LCPI9_1@toc@ha
-; CHECK-LE-NEXT:    lfs 0, .LCPI9_1@toc@l(3)
 ; CHECK-LE-NEXT:    vperm 2, 3, 2, 4
+; CHECK-LE-NEXT:    vspltisw 3, 1
+; CHECK-LE-NEXT:    xvcvsxwdp 0, 35
 ; CHECK-LE-NEXT:    xsadddp 1, 34, 0
 ; CHECK-LE-NEXT:    blr
 ;
 ; CHECK-BE-LABEL: test10:
 ; CHECK-BE:       # %bb.0: # %entry
-; CHECK-BE-NEXT:    addis 3, 2, .LCPI9_0@toc@ha
 ; CHECK-BE-NEXT:    vmrghw 3, 3, 3
-; CHECK-BE-NEXT:    lfs 0, .LCPI9_0@toc@l(3)
 ; CHECK-BE-NEXT:    vmrglw 2, 3, 2
+; CHECK-BE-NEXT:    vspltisw 3, 1
+; CHECK-BE-NEXT:    xvcvsxwdp 0, 35
 ; CHECK-BE-NEXT:    xsadddp 1, 34, 0
 ; CHECK-BE-NEXT:    blr
 entry:
   %shuffle = shufflevector <4 x i32> %a, <4 x i32> %b, <4 x i32> <i32 5, i32 2, i32 3, i32 7>
   %cast = bitcast <4 x i32> %shuffle to <2 x double>
   %extract = extractelement <2 x double> %cast, i32 0
   %add = fadd double %extract, 1.0000
   ret double %add
 }