[X86] Simplify b2b KSHIFTL+KSHIFTR using demanded elts.

llvm-svn: 372155

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -34627,29 +34627,82 @@ bool X86TargetLowering::SimplifyDemandedVectorEltsForTargetNode(
     // TODO convert SrcUndef to KnownUndef.
     break;
   }
-  case X86ISD::KSHIFTL:
+  case X86ISD::KSHIFTL: {
+    SDValue Src = Op.getOperand(0);
+    auto *Amt = cast<ConstantSDNode>(Op.getOperand(1));
+    assert(Amt->getAPIntValue().ult(NumElts) && "Out of range shift amount");
+    unsigned ShiftAmt = Amt->getZExtValue();
+
+    if (ShiftAmt == 0)
+      return TLO.CombineTo(Op, Src);
+
+    // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a
+    // single shift.  We can do this if the bottom bits (which are shifted
+    // out) are never demanded.
+    if (Src.getOpcode() == X86ISD::KSHIFTR) {
+      if (!DemandedElts.intersects(APInt::getLowBitsSet(NumElts, ShiftAmt))) {
+        unsigned C1 = Src.getConstantOperandVal(1);
+        unsigned Opc = X86ISD::KSHIFTL;
+        int Diff = ShiftAmt - C1;
+        if (Diff < 0) {
+          Diff = -Diff;
+          Opc = X86ISD::KSHIFTR;
+        }
+
+        SDLoc dl(Op);
+        SDValue NewSA = TLO.DAG.getConstant(Diff, dl, MVT::i8);
+        return TLO.CombineTo(
+            Op, TLO.DAG.getNode(Opc, dl, VT, Src.getOperand(0), NewSA));
+      }
+    }
+
+    APInt DemandedSrc = DemandedElts.lshr(ShiftAmt);
+    if (SimplifyDemandedVectorElts(Src, DemandedSrc, KnownUndef, KnownZero, TLO,
+                                   Depth + 1))
+      return true;
+
+    KnownUndef = KnownUndef.shl(ShiftAmt);
+    KnownZero = KnownZero.shl(ShiftAmt);
+    KnownZero.setLowBits(ShiftAmt);
+    break;
+  }
   case X86ISD::KSHIFTR: {
     SDValue Src = Op.getOperand(0);
     auto *Amt = cast<ConstantSDNode>(Op.getOperand(1));
     assert(Amt->getAPIntValue().ult(NumElts) && "Out of range shift amount");
     unsigned ShiftAmt = Amt->getZExtValue();
-    bool ShiftLeft = (X86ISD::KSHIFTL == Opc);
 
-    APInt DemandedSrc =
+    if (ShiftAmt == 0)
-        ShiftLeft ? DemandedElts.lshr(ShiftAmt) : DemandedElts.shl(ShiftAmt);
+      return TLO.CombineTo(Op, Src);
+
+    // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a
+    // single shift.  We can do this if the top bits (which are shifted
+    // out) are never demanded.
+    if (Src.getOpcode() == X86ISD::KSHIFTL) {
+      if (!DemandedElts.intersects(APInt::getHighBitsSet(NumElts, ShiftAmt))) {
+        unsigned C1 = Src.getConstantOperandVal(1);
+        unsigned Opc = X86ISD::KSHIFTR;
+        int Diff = ShiftAmt - C1;
+        if (Diff < 0) {
+          Diff = -Diff;
+          Opc = X86ISD::KSHIFTL;
+        }
+
+        SDLoc dl(Op);
+        SDValue NewSA = TLO.DAG.getConstant(Diff, dl, MVT::i8);
+        return TLO.CombineTo(
+            Op, TLO.DAG.getNode(Opc, dl, VT, Src.getOperand(0), NewSA));
+      }
+    }
+
+    APInt DemandedSrc = DemandedElts.shl(ShiftAmt);
     if (SimplifyDemandedVectorElts(Src, DemandedSrc, KnownUndef, KnownZero, TLO,
                                    Depth + 1))
       return true;
 
-    if (ShiftLeft) {
+    KnownUndef = KnownUndef.lshr(ShiftAmt);
-      KnownUndef = KnownUndef.shl(ShiftAmt);
+    KnownZero = KnownZero.lshr(ShiftAmt);
-      KnownZero = KnownZero.shl(ShiftAmt);
+    KnownZero.setHighBits(ShiftAmt);
-      KnownZero.setLowBits(ShiftAmt);
-    } else {
-      KnownUndef = KnownUndef.lshr(ShiftAmt);
-      KnownZero = KnownZero.lshr(ShiftAmt);
-      KnownZero.setHighBits(ShiftAmt);
-    }
     break;
   }
   case X86ISD::CVTSI2P:

llvm/test/CodeGen/X86/avx512-ext.ll

@@ -1974,8 +1974,7 @@ define <64 x i16> @test21(<64 x i16> %x , <64 x i1> %mask) nounwind readnone {
 ; KNL-NEXT:    movb {{[0-9]+}}(%rsp), %al
 ; KNL-NEXT:    kmovw %eax, %k3
 ; KNL-NEXT:    kxorw %k3, %k2, %k2
-; KNL-NEXT:    kshiftlw $15, %k2, %k2
+; KNL-NEXT:    kshiftlw $14, %k2, %k2
-; KNL-NEXT:    kshiftrw $1, %k2, %k2
 ; KNL-NEXT:    kxorw %k2, %k1, %k1
 ; KNL-NEXT:    kshiftlw $1, %k1, %k1
 ; KNL-NEXT:    kshiftrw $1, %k1, %k1
@@ -2074,8 +2073,7 @@ define <64 x i16> @test21(<64 x i16> %x , <64 x i1> %mask) nounwind readnone {
 ; KNL-NEXT:    movb {{[0-9]+}}(%rsp), %al
 ; KNL-NEXT:    kmovw %eax, %k4
 ; KNL-NEXT:    kxorw %k4, %k3, %k3
-; KNL-NEXT:    kshiftlw $15, %k3, %k3
+; KNL-NEXT:    kshiftlw $14, %k3, %k3
-; KNL-NEXT:    kshiftrw $1, %k3, %k3
 ; KNL-NEXT:    kxorw %k3, %k2, %k2
 ; KNL-NEXT:    kshiftlw $1, %k2, %k2
 ; KNL-NEXT:    kshiftrw $1, %k2, %k2
@@ -2174,8 +2172,7 @@ define <64 x i16> @test21(<64 x i16> %x , <64 x i1> %mask) nounwind readnone {
 ; KNL-NEXT:    movb {{[0-9]+}}(%rsp), %al
 ; KNL-NEXT:    kmovw %eax, %k5
 ; KNL-NEXT:    kxorw %k5, %k4, %k4
-; KNL-NEXT:    kshiftlw $15, %k4, %k4
+; KNL-NEXT:    kshiftlw $14, %k4, %k4
-; KNL-NEXT:    kshiftrw $1, %k4, %k4
 ; KNL-NEXT:    kxorw %k4, %k3, %k3
 ; KNL-NEXT:    kshiftlw $1, %k3, %k3
 ; KNL-NEXT:    kshiftrw $1, %k3, %k3
@@ -2274,8 +2271,7 @@ define <64 x i16> @test21(<64 x i16> %x , <64 x i1> %mask) nounwind readnone {
 ; KNL-NEXT:    movb {{[0-9]+}}(%rsp), %al
 ; KNL-NEXT:    kmovw %eax, %k5
 ; KNL-NEXT:    kxorw %k5, %k4, %k4
-; KNL-NEXT:    kshiftlw $15, %k4, %k4
+; KNL-NEXT:    kshiftlw $14, %k4, %k4
-; KNL-NEXT:    kshiftrw $1, %k4, %k4
 ; KNL-NEXT:    kxorw %k4, %k0, %k0
 ; KNL-NEXT:    kshiftlw $1, %k0, %k0
 ; KNL-NEXT:    kshiftrw $1, %k0, %k0
@@ -2396,8 +2392,7 @@ define <64 x i16> @test21(<64 x i16> %x , <64 x i1> %mask) nounwind readnone {
 ; AVX512DQNOBW-NEXT:    movb {{[0-9]+}}(%rsp), %al
 ; AVX512DQNOBW-NEXT:    kmovw %eax, %k3
 ; AVX512DQNOBW-NEXT:    kxorw %k3, %k2, %k2
-; AVX512DQNOBW-NEXT:    kshiftlw $15, %k2, %k2
+; AVX512DQNOBW-NEXT:    kshiftlw $14, %k2, %k2
-; AVX512DQNOBW-NEXT:    kshiftrw $1, %k2, %k2
 ; AVX512DQNOBW-NEXT:    kxorw %k2, %k0, %k0
 ; AVX512DQNOBW-NEXT:    kshiftlw $1, %k0, %k0
 ; AVX512DQNOBW-NEXT:    kshiftrw $1, %k0, %k0
@@ -2496,8 +2491,7 @@ define <64 x i16> @test21(<64 x i16> %x , <64 x i1> %mask) nounwind readnone {
 ; AVX512DQNOBW-NEXT:    movb {{[0-9]+}}(%rsp), %al
 ; AVX512DQNOBW-NEXT:    kmovw %eax, %k4
 ; AVX512DQNOBW-NEXT:    kxorw %k4, %k3, %k3
-; AVX512DQNOBW-NEXT:    kshiftlw $15, %k3, %k3
+; AVX512DQNOBW-NEXT:    kshiftlw $14, %k3, %k3
-; AVX512DQNOBW-NEXT:    kshiftrw $1, %k3, %k3
 ; AVX512DQNOBW-NEXT:    kxorw %k3, %k2, %k2
 ; AVX512DQNOBW-NEXT:    kshiftlw $1, %k2, %k2
 ; AVX512DQNOBW-NEXT:    kshiftrw $1, %k2, %k2
@@ -2596,8 +2590,7 @@ define <64 x i16> @test21(<64 x i16> %x , <64 x i1> %mask) nounwind readnone {
 ; AVX512DQNOBW-NEXT:    movb {{[0-9]+}}(%rsp), %al
 ; AVX512DQNOBW-NEXT:    kmovw %eax, %k5
 ; AVX512DQNOBW-NEXT:    kxorw %k5, %k4, %k4
-; AVX512DQNOBW-NEXT:    kshiftlw $15, %k4, %k4
+; AVX512DQNOBW-NEXT:    kshiftlw $14, %k4, %k4
-; AVX512DQNOBW-NEXT:    kshiftrw $1, %k4, %k4
 ; AVX512DQNOBW-NEXT:    kxorw %k4, %k3, %k3
 ; AVX512DQNOBW-NEXT:    kshiftlw $1, %k3, %k3
 ; AVX512DQNOBW-NEXT:    kshiftrw $1, %k3, %k3
@@ -2696,8 +2689,7 @@ define <64 x i16> @test21(<64 x i16> %x , <64 x i1> %mask) nounwind readnone {
 ; AVX512DQNOBW-NEXT:    movb {{[0-9]+}}(%rsp), %al
 ; AVX512DQNOBW-NEXT:    kmovw %eax, %k5
 ; AVX512DQNOBW-NEXT:    kxorw %k5, %k4, %k4
-; AVX512DQNOBW-NEXT:    kshiftlw $15, %k4, %k4
+; AVX512DQNOBW-NEXT:    kshiftlw $14, %k4, %k4
-; AVX512DQNOBW-NEXT:    kshiftrw $1, %k4, %k4
 ; AVX512DQNOBW-NEXT:    kxorw %k4, %k1, %k1
 ; AVX512DQNOBW-NEXT:    kshiftlw $1, %k1, %k1
 ; AVX512DQNOBW-NEXT:    kshiftrw $1, %k1, %k1

llvm/test/CodeGen/X86/avx512-mask-op.ll

@@ -1283,8 +1283,7 @@ define <8 x i1> @test18(i8 %a, i16 %y) {
 ; KNL-NEXT:    kshiftrw $9, %k1, %k1
 ; KNL-NEXT:    kshiftrw $6, %k0, %k3
 ; KNL-NEXT:    kxorw %k1, %k3, %k1
-; KNL-NEXT:    kshiftlw $15, %k1, %k1
+; KNL-NEXT:    kshiftlw $6, %k1, %k1
-; KNL-NEXT:    kshiftrw $9, %k1, %k1
 ; KNL-NEXT:    kxorw %k1, %k0, %k0
 ; KNL-NEXT:    kshiftlw $9, %k0, %k0
 ; KNL-NEXT:    kshiftrw $9, %k0, %k0
@@ -1304,8 +1303,7 @@ define <8 x i1> @test18(i8 %a, i16 %y) {
 ; SKX-NEXT:    kshiftrw $9, %k1, %k1
 ; SKX-NEXT:    kshiftrb $6, %k0, %k3
 ; SKX-NEXT:    kxorb %k1, %k3, %k1
-; SKX-NEXT:    kshiftlb $7, %k1, %k1
+; SKX-NEXT:    kshiftlb $6, %k1, %k1
-; SKX-NEXT:    kshiftrb $1, %k1, %k1
 ; SKX-NEXT:    kxorb %k1, %k0, %k0
 ; SKX-NEXT:    kshiftlb $1, %k0, %k0
 ; SKX-NEXT:    kshiftrb $1, %k0, %k0
@@ -1322,8 +1320,7 @@ define <8 x i1> @test18(i8 %a, i16 %y) {
 ; AVX512BW-NEXT:    kshiftrw $9, %k1, %k1
 ; AVX512BW-NEXT:    kshiftrw $6, %k0, %k3
 ; AVX512BW-NEXT:    kxorw %k1, %k3, %k1
-; AVX512BW-NEXT:    kshiftlw $15, %k1, %k1
+; AVX512BW-NEXT:    kshiftlw $6, %k1, %k1
-; AVX512BW-NEXT:    kshiftrw $9, %k1, %k1
 ; AVX512BW-NEXT:    kxorw %k1, %k0, %k0
 ; AVX512BW-NEXT:    kshiftlw $9, %k0, %k0
 ; AVX512BW-NEXT:    kshiftrw $9, %k0, %k0
@@ -1342,8 +1339,7 @@ define <8 x i1> @test18(i8 %a, i16 %y) {
 ; AVX512DQ-NEXT:    kshiftrw $9, %k1, %k1
 ; AVX512DQ-NEXT:    kshiftrb $6, %k0, %k3
 ; AVX512DQ-NEXT:    kxorb %k1, %k3, %k1
-; AVX512DQ-NEXT:    kshiftlb $7, %k1, %k1
+; AVX512DQ-NEXT:    kshiftlb $6, %k1, %k1
-; AVX512DQ-NEXT:    kshiftrb $1, %k1, %k1
 ; AVX512DQ-NEXT:    kxorb %k1, %k0, %k0
 ; AVX512DQ-NEXT:    kshiftlb $1, %k0, %k0
 ; AVX512DQ-NEXT:    kshiftrb $1, %k0, %k0
@@ -1363,8 +1359,7 @@ define <8 x i1> @test18(i8 %a, i16 %y) {
 ; X86-NEXT:    kshiftrw $9, %k1, %k1
 ; X86-NEXT:    kshiftrb $6, %k0, %k3
 ; X86-NEXT:    kxorb %k1, %k3, %k1
-; X86-NEXT:    kshiftlb $7, %k1, %k1
+; X86-NEXT:    kshiftlb $6, %k1, %k1
-; X86-NEXT:    kshiftrb $1, %k1, %k1
 ; X86-NEXT:    kxorb %k1, %k0, %k0
 ; X86-NEXT:    kshiftlb $1, %k0, %k0
 ; X86-NEXT:    kshiftrb $1, %k0, %k0
@@ -2842,8 +2837,7 @@ define void @store_64i1(<64 x i1>* %a, <64 x i1> %v) {
 ; KNL-NEXT:    movb {{[0-9]+}}(%rsp), %al
 ; KNL-NEXT:    kmovw %eax, %k3
 ; KNL-NEXT:    kxorw %k3, %k2, %k2
-; KNL-NEXT:    kshiftlw $15, %k2, %k2
+; KNL-NEXT:    kshiftlw $14, %k2, %k2
-; KNL-NEXT:    kshiftrw $1, %k2, %k2
 ; KNL-NEXT:    kxorw %k2, %k0, %k0
 ; KNL-NEXT:    kshiftlw $1, %k0, %k0
 ; KNL-NEXT:    kshiftrw $1, %k0, %k0
@@ -2942,8 +2936,7 @@ define void @store_64i1(<64 x i1>* %a, <64 x i1> %v) {
 ; KNL-NEXT:    movb {{[0-9]+}}(%rsp), %al
 ; KNL-NEXT:    kmovw %eax, %k4
 ; KNL-NEXT:    kxorw %k4, %k3, %k3
-; KNL-NEXT:    kshiftlw $15, %k3, %k3
+; KNL-NEXT:    kshiftlw $14, %k3, %k3
-; KNL-NEXT:    kshiftrw $1, %k3, %k3
 ; KNL-NEXT:    kxorw %k3, %k2, %k2
 ; KNL-NEXT:    kshiftlw $1, %k2, %k2
 ; KNL-NEXT:    kshiftrw $1, %k2, %k2
@@ -3042,8 +3035,7 @@ define void @store_64i1(<64 x i1>* %a, <64 x i1> %v) {
 ; KNL-NEXT:    movb {{[0-9]+}}(%rsp), %al
 ; KNL-NEXT:    kmovw %eax, %k5
 ; KNL-NEXT:    kxorw %k5, %k4, %k4
-; KNL-NEXT:    kshiftlw $15, %k4, %k4
+; KNL-NEXT:    kshiftlw $14, %k4, %k4
-; KNL-NEXT:    kshiftrw $1, %k4, %k4
 ; KNL-NEXT:    kxorw %k4, %k3, %k3
 ; KNL-NEXT:    kshiftlw $1, %k3, %k3
 ; KNL-NEXT:    kshiftrw $1, %k3, %k3
@@ -3142,8 +3134,7 @@ define void @store_64i1(<64 x i1>* %a, <64 x i1> %v) {
 ; KNL-NEXT:    movb {{[0-9]+}}(%rsp), %al
 ; KNL-NEXT:    kmovw %eax, %k5
 ; KNL-NEXT:    kxorw %k5, %k4, %k4
-; KNL-NEXT:    kshiftlw $15, %k4, %k4
+; KNL-NEXT:    kshiftlw $14, %k4, %k4
-; KNL-NEXT:    kshiftrw $1, %k4, %k4
 ; KNL-NEXT:    kxorw %k4, %k1, %k1
 ; KNL-NEXT:    kshiftlw $1, %k1, %k1
 ; KNL-NEXT:    kshiftrw $1, %k1, %k1
@@ -3264,8 +3255,7 @@ define void @store_64i1(<64 x i1>* %a, <64 x i1> %v) {
 ; AVX512DQ-NEXT:    movb {{[0-9]+}}(%rsp), %al
 ; AVX512DQ-NEXT:    kmovw %eax, %k3
 ; AVX512DQ-NEXT:    kxorw %k3, %k2, %k2
-; AVX512DQ-NEXT:    kshiftlw $15, %k2, %k2
+; AVX512DQ-NEXT:    kshiftlw $14, %k2, %k2
-; AVX512DQ-NEXT:    kshiftrw $1, %k2, %k2
 ; AVX512DQ-NEXT:    kxorw %k2, %k0, %k0
 ; AVX512DQ-NEXT:    kshiftlw $1, %k0, %k0
 ; AVX512DQ-NEXT:    kshiftrw $1, %k0, %k0
@@ -3364,8 +3354,7 @@ define void @store_64i1(<64 x i1>* %a, <64 x i1> %v) {
 ; AVX512DQ-NEXT:    movb {{[0-9]+}}(%rsp), %al
 ; AVX512DQ-NEXT:    kmovw %eax, %k4
 ; AVX512DQ-NEXT:    kxorw %k4, %k3, %k3
-; AVX512DQ-NEXT:    kshiftlw $15, %k3, %k3
+; AVX512DQ-NEXT:    kshiftlw $14, %k3, %k3
-; AVX512DQ-NEXT:    kshiftrw $1, %k3, %k3
 ; AVX512DQ-NEXT:    kxorw %k3, %k2, %k2
 ; AVX512DQ-NEXT:    kshiftlw $1, %k2, %k2
 ; AVX512DQ-NEXT:    kshiftrw $1, %k2, %k2
@@ -3464,8 +3453,7 @@ define void @store_64i1(<64 x i1>* %a, <64 x i1> %v) {
 ; AVX512DQ-NEXT:    movb {{[0-9]+}}(%rsp), %al
 ; AVX512DQ-NEXT:    kmovw %eax, %k5
 ; AVX512DQ-NEXT:    kxorw %k5, %k4, %k4
-; AVX512DQ-NEXT:    kshiftlw $15, %k4, %k4
+; AVX512DQ-NEXT:    kshiftlw $14, %k4, %k4
-; AVX512DQ-NEXT:    kshiftrw $1, %k4, %k4
 ; AVX512DQ-NEXT:    kxorw %k4, %k3, %k3
 ; AVX512DQ-NEXT:    kshiftlw $1, %k3, %k3
 ; AVX512DQ-NEXT:    kshiftrw $1, %k3, %k3
@@ -3564,8 +3552,7 @@ define void @store_64i1(<64 x i1>* %a, <64 x i1> %v) {
 ; AVX512DQ-NEXT:    movb {{[0-9]+}}(%rsp), %al
 ; AVX512DQ-NEXT:    kmovw %eax, %k5
 ; AVX512DQ-NEXT:    kxorw %k5, %k4, %k4
-; AVX512DQ-NEXT:    kshiftlw $15, %k4, %k4
+; AVX512DQ-NEXT:    kshiftlw $14, %k4, %k4
-; AVX512DQ-NEXT:    kshiftrw $1, %k4, %k4
 ; AVX512DQ-NEXT:    kxorw %k4, %k1, %k1
 ; AVX512DQ-NEXT:    kshiftlw $1, %k1, %k1
 ; AVX512DQ-NEXT:    kshiftrw $1, %k1, %k1

llvm/test/CodeGen/X86/avx512vl-vec-masked-cmp.ll

@@ -23526,8 +23526,7 @@ define i8 @mask_zero_lower(<4 x i32> %a) {
 ; NoVLX:       # %bb.0:
 ; NoVLX-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; NoVLX-NEXT:    vptestmd %zmm0, %zmm0, %k0
-; NoVLX-NEXT:    kshiftlw $12, %k0, %k0
+; NoVLX-NEXT:    kshiftlw $4, %k0, %k0
-; NoVLX-NEXT:    kshiftrw $8, %k0, %k0
 ; NoVLX-NEXT:    kmovw %k0, %eax
 ; NoVLX-NEXT:    # kill: def $al killed $al killed $eax
 ; NoVLX-NEXT:    vzeroupper

llvm/test/CodeGen/X86/vec_smulo.ll

@@ -2252,8 +2252,7 @@ define <4 x i32> @smulo_v4i1(<4 x i1> %a0, <4 x i1> %a1, <4 x i1>* %p2) nounwind
 ; AVX512-NEXT:    kxorw %k0, %k2, %k2
 ; AVX512-NEXT:    kshiftrw $2, %k2, %k3
 ; AVX512-NEXT:    kxorw %k1, %k3, %k1
-; AVX512-NEXT:    kshiftlw $15, %k1, %k1
+; AVX512-NEXT:    kshiftlw $2, %k1, %k1
-; AVX512-NEXT:    kshiftrw $13, %k1, %k1
 ; AVX512-NEXT:    kxorw %k1, %k2, %k1
 ; AVX512-NEXT:    kshiftlw $13, %k1, %k1
 ; AVX512-NEXT:    kshiftrw $13, %k1, %k1

llvm/test/CodeGen/X86/vec_umulo.ll

@@ -1988,8 +1988,7 @@ define <4 x i32> @umulo_v4i1(<4 x i1> %a0, <4 x i1> %a1, <4 x i1>* %p2) nounwind
 ; AVX512-NEXT:    kxorw %k0, %k2, %k2
 ; AVX512-NEXT:    kshiftrw $2, %k2, %k3
 ; AVX512-NEXT:    kxorw %k1, %k3, %k1
-; AVX512-NEXT:    kshiftlw $15, %k1, %k1
+; AVX512-NEXT:    kshiftlw $2, %k1, %k1
-; AVX512-NEXT:    kshiftrw $13, %k1, %k1
 ; AVX512-NEXT:    kxorw %k1, %k2, %k1
 ; AVX512-NEXT:    kshiftlw $13, %k1, %k1
 ; AVX512-NEXT:    kshiftrw $13, %k1, %k1