[SelectionDAG] Fill in some more cases in `isKnownNeverZero`

[goldstein.w.n]

Summary

This mostly copies cases that already exist in ValueTracking, although
it skips the more complex ones. Those can be filled in as needed.

[craig.topper]

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

5016

I assume we checked operand 1 first to favor the cheap case of a non-zero constant right hand side.

[craig.topper]

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

5031

return after break?

5066

too many blank lines

5069

I don't think curly braces are needed here

5086

Or here

[goldstein.w.n]

Swap operand order. Fix some style nits

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

5016

Makes sense, updated that everywhere else.

[goldstein.w.n]

Rebase

[RKSimon]

Are we really hitting all these cases? We should really ensure we have test coverage for them all.

[goldstein.w.n]

In D149199#4298104, @RKSimon wrote:

Are we really hitting all these cases? We should really ensure we have test coverage for them all.

Certainly not all of them. But fair point, will add tests.

[goldstein.w.n]

Rebase with tests

[goldstein.w.n]

In D149199#4298104, @RKSimon wrote:

Are we really hitting all these cases? We should really ensure we have test coverage for them all.

Added tests at D149299. Is there a better way to isolate/test DAGComber codes?

[goldstein.w.n]

Rebase

[RKSimon]

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

5029

This looks like a bad merge?

[goldstein.w.n]

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

5029

hmm?

[goldstein.w.n]

Rebase

[RKSimon]

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

5029

Sorry, I think I was miscounting brackets :)

[goldstein.w.n]

Rebase

[goldstein.w.n]

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

5029

You where right!

[goldstein.w.n]

Fix broken parens

[goldstein.w.n]

rebase

[goldstein.w.n]

Rebase

[goldstein.w.n]

ping. Dropped the assosiated setcc patch, but would like to get this through as an improvement to the helper as is.

[RKSimon]

LGTM

[goldstein.w.n]

Rebase

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

bool SelectionDAG::isKnownNeverZero(SDValue Op, unsigned Depth) const {
   assert(!Op.getValueType().isFloatingPoint() &&
          "Floating point types unsupported - use isKnownNeverZeroFloat");
 
   // If the value is a constant, we can obviously see if it is a zero or not.
   if (ISD::matchUnaryPredicate(Op,
                                [](ConstantSDNode *C) { return !C->isZero(); }))
     return true;
 
-  // TODO: Recognize more cases here.
+  // TODO: Recognize more cases here. Most of the cases are also incomplete to
+  // some degree.
   switch (Op.getOpcode()) {
-  default: break;
+  default:
+    break;
   case ISD::OR:
-    if (isKnownNeverZero(Op.getOperand(1)) ||
-        isKnownNeverZero(Op.getOperand(0)))
+    return isKnownNeverZero(Op.getOperand(0), Depth) ||
+           isKnownNeverZero(Op.getOperand(1), Depth);
+  case ISD::VSELECT:
+  case ISD::SELECT:
+    return isKnownNeverZero(Op.getOperand(1), Depth) &&
+           isKnownNeverZero(Op.getOperand(2), Depth);
+  case ISD::SHL:
+    if (Op->getFlags().hasNoSignedWrap() || Op->getFlags().hasNoUnsignedWrap())
+      return isKnownNeverZero(Op.getOperand(0), Depth);
+
+    // 1 << X is never zero. TODO: This can be expanded if we can bound X.
+    // The expression is really !Known.One[BitWidth-MaxLog2(Known):0].isZero()
+    if (computeKnownBits(Op.getOperand(0)).One[0], Depth)
       return true;
     break;
+    return isKnownNeverZero(Op.getOperand(0), Depth);
+
+  case ISD::UADDSAT:
+  case ISD::UMAX:
+    return isKnownNeverZero(Op.getOperand(0), Depth) ||
+           isKnownNeverZero(Op.getOperand(1), Depth);
+
+  case ISD::UMIN:
+    return isKnownNeverZero(Op.getOperand(0), Depth) &&
+           isKnownNeverZero(Op.getOperand(1), Depth);
+
+  case ISD::ROTL:
+  case ISD::ROTR:
+  case ISD::BITREVERSE:
+  case ISD::BSWAP:
+  case ISD::CTPOP:
+  case ISD::ABS:
+    return isKnownNeverZero(Op.getOperand(0), Depth);
+
+  case ISD::SRA:
+  case ISD::SRL:
+    if (Op->getFlags().hasExact())
+      return isKnownNeverZero(Op.getOperand(0), Depth);
+    // Signed >> X is never zero. TODO: This can be expanded if we can bound X.
+    // The expression is really
+    // !Known.One[SignBit:SignBit-(BitWidth-MaxLog2(Known))].isZero()
+    if (computeKnownBits(Op.getOperand(0), Depth).isNegative())
+      return true;
+    break;
+  case ISD::UDIV:
+  case ISD::SDIV:
+    // div exact can only produce a zero if the dividend is zero.
+    // TODO: For udiv this is also true if Op1 u<= Op0
+    if (Op->getFlags().hasExact())
+      return isKnownNeverZero(Op.getOperand(0), Depth);
+
+
+    break;
+  case ISD::ADD: {
+    if (Op->getFlags().hasNoUnsignedWrap())
+      if (isKnownNeverZero(Op.getOperand(0), Depth) ||
+          isKnownNeverZero(Op.getOperand(1), Depth))
+        return true;
+    // TODO: There are a lot more cases we can prove for add.
+    break;
+  }
+  case ISD::SUB: {
+    if (isNullConstant(Op.getOperand(0)))
+      return isKnownNeverZero(Op.getOperand(1), Depth);
+
+    std::optional<bool> ne =
+        KnownBits::ne(computeKnownBits(Op.getOperand(0), Depth),
+                      computeKnownBits(Op.getOperand(1), Depth));
+    return ne && *ne;
+  }
+  case ISD::MUL: {
+    if (Op->getFlags().hasNoSignedWrap() || Op->getFlags().hasNoUnsignedWrap())
+      if (isKnownNeverZero(Op.getOperand(0), Depth) &&
+          isKnownNeverZero(Op.getOperand(1), Depth))
+        return true;
+    break;
+  }
+  case ISD::BITCAST:
+    if (!Op.getOperand(0).getValueType().isFloatingPoint())
+      return isKnownNeverZero(Op.getOperand(0), Depth);
+    break;
+
+  case ISD::ZERO_EXTEND:
+  case ISD::SIGN_EXTEND:
+    return isKnownNeverZero(Op.getOperand(0), Depth);
   }
 
   return !computeKnownBits(Op, Depth).One.isZero();
 }
 
 bool SelectionDAG::isEqualTo(SDValue A, SDValue B) const {
   // Check the obvious case.
   if (A == B) return true;

llvm/test/CodeGen/AMDGPU/cttz_zero_undef.ll

 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    v_ffbl_b32_e32 v0, v0
-; SI-NEXT:    v_min_u32_e32 v0, 32, v0
 ; SI-NEXT:    buffer_store_dword v0, off, s[4:7], 0
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: v_cttz_zero_undef_i32_with_select:
 ; VI:       ; %bb.0:
 ; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    s_add_u32 s4, s2, 3

 ; VI-NEXT:    v_lshlrev_b32_e32 v1, 8, v2
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_lshlrev_b32_e32 v2, 8, v4
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_e32 v0, v2, v0
 ; VI-NEXT:    v_or_b32_e32 v0, v1, v0
-; VI-NEXT:    v_ffbl_b32_e32 v0, v0
-; VI-NEXT:    v_min_u32_e32 v2, 32, v0
+; VI-NEXT:    v_ffbl_b32_e32 v2, v0
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    flat_store_dword v[0:1], v2
 ; VI-NEXT:    s_endpgm
 ;
 ; EG-LABEL: v_cttz_zero_undef_i32_with_select:
 ; EG:       ; %bb.0:
 ; EG-NEXT:    ALU 0, @10, KC0[CB0:0-32], KC1[]
 ; EG-NEXT:    TEX 1 @6
-; EG-NEXT:    ALU 6, @11, KC0[CB0:0-32], KC1[]
+; EG-NEXT:    ALU 5, @11, KC0[CB0:0-32], KC1[]
 ; EG-NEXT:    MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
 ; EG-NEXT:    CF_END
 ; EG-NEXT:    PAD
 ; EG-NEXT:    Fetch clause starting at 6:
 ; EG-NEXT:     VTX_READ_16 T1.X, T0.X, 2, #1
 ; EG-NEXT:     VTX_READ_16 T0.X, T0.X, 0, #1
 ; EG-NEXT:    ALU clause starting at 10:
 ; EG-NEXT:     MOV * T0.X, KC0[2].Z,
 ; EG-NEXT:    ALU clause starting at 11:
 ; EG-NEXT:     LSHL * T0.W, T1.X, literal.x,
 ; EG-NEXT:    16(2.242078e-44), 0(0.000000e+00)
 ; EG-NEXT:     OR_INT * T0.W, PV.W, T0.X,
-; EG-NEXT:     FFBL_INT * T1.W, PV.W,
-; EG-NEXT:     CNDE_INT T0.X, T0.W, literal.x, PV.W,
-; EG-NEXT:     LSHR * T1.X, KC0[2].Y, literal.y,
-; EG-NEXT:    32(4.484155e-44), 2(2.802597e-45)
+; EG-NEXT:     FFBL_INT T0.X, PV.W,
+; EG-NEXT:     LSHR * T1.X, KC0[2].Y, literal.x,
+; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)
 ;
 ; GFX9-GISEL-LABEL: v_cttz_zero_undef_i32_with_select:
 ; GFX9-GISEL:       ; %bb.0:
 ; GFX9-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
 ; GFX9-GISEL-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX9-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-GISEL-NEXT:    global_load_ubyte v1, v0, s[2:3]
 ; GFX9-GISEL-NEXT:    global_load_ubyte v2, v0, s[2:3] offset:1

 ; SI-NEXT:    v_or_b32_e32 v2, v5, v4
 ; SI-NEXT:    v_or_b32_e32 v3, v7, v6
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    v_or_b32_e32 v1, v3, v2
 ; SI-NEXT:    v_ffbl_b32_e32 v1, v1
 ; SI-NEXT:    v_ffbl_b32_e32 v0, v0
-; SI-NEXT:    v_min_u32_e32 v1, 0xffffffdf, v1
 ; SI-NEXT:    v_add_i32_e32 v1, vcc, 32, v1
-; SI-NEXT:    v_min3_u32 v0, v0, v1, 64
+; SI-NEXT:    v_min_u32_e32 v0, v0, v1
 ; SI-NEXT:    v_mov_b32_e32 v1, 0
 ; SI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: v_cttz_zero_undef_i64_with_select:
 ; VI:       ; %bb.0:
 ; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)

 ; VI-NEXT:    s_waitcnt vmcnt(5)
 ; VI-NEXT:    v_lshlrev_b32_e32 v4, 8, v4
 ; VI-NEXT:    s_waitcnt vmcnt(4)
 ; VI-NEXT:    v_or_b32_sdwa v4, v4, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    v_or_b32_e32 v3, v4, v3
 ; VI-NEXT:    s_waitcnt vmcnt(3)
 ; VI-NEXT:    v_lshlrev_b32_e32 v4, 8, v6
 ; VI-NEXT:    v_ffbl_b32_e32 v3, v3
-; VI-NEXT:    v_add_u32_e64 v3, s[2:3], v3, 32 clamp
+; VI-NEXT:    v_add_u32_e32 v3, vcc, 32, v3
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_or_b32_sdwa v4, v4, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_lshlrev_b32_e32 v0, 8, v0
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_e32 v0, v0, v2
 ; VI-NEXT:    v_or_b32_e32 v0, v4, v0
 ; VI-NEXT:    v_ffbl_b32_e32 v0, v0
-; VI-NEXT:    v_min3_u32 v0, v0, v3, 64
+; VI-NEXT:    v_min_u32_e32 v0, v0, v3
 ; VI-NEXT:    v_mov_b32_e32 v3, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s0
 ; VI-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; VI-NEXT:    s_endpgm
 ;
 ; EG-LABEL: v_cttz_zero_undef_i64_with_select:
 ; EG:       ; %bb.0:
 ; EG-NEXT:    ALU 0, @14, KC0[CB0:0-32], KC1[]
 ; EG-NEXT:    TEX 3 @6
-; EG-NEXT:    ALU 15, @15, KC0[CB0:0-32], KC1[]
+; EG-NEXT:    ALU 12, @15, KC0[CB0:0-32], KC1[]
 ; EG-NEXT:    MEM_RAT_CACHELESS STORE_RAW T0.XY, T1.X, 1
 ; EG-NEXT:    CF_END
 ; EG-NEXT:    PAD
 ; EG-NEXT:    Fetch clause starting at 6:
 ; EG-NEXT:     VTX_READ_16 T1.X, T0.X, 6, #1
 ; EG-NEXT:     VTX_READ_16 T2.X, T0.X, 0, #1
-; EG-NEXT:     VTX_READ_16 T3.X, T0.X, 2, #1
-; EG-NEXT:     VTX_READ_16 T0.X, T0.X, 4, #1
+; EG-NEXT:     VTX_READ_16 T3.X, T0.X, 4, #1
+; EG-NEXT:     VTX_READ_16 T0.X, T0.X, 2, #1
 ; EG-NEXT:    ALU clause starting at 14:
 ; EG-NEXT:     MOV * T0.X, KC0[2].Z,
 ; EG-NEXT:    ALU clause starting at 15:
-; EG-NEXT:     LSHL * T0.W, T1.X, literal.x,
+; EG-NEXT:     LSHL T0.W, T1.X, literal.x,
+; EG-NEXT:     LSHL * T1.W, T0.X, literal.x,
 ; EG-NEXT:    16(2.242078e-44), 0(0.000000e+00)
-; EG-NEXT:     OR_INT * T0.W, PV.W, T0.X,
-; EG-NEXT:     FFBL_INT T1.W, PV.W,
-; EG-NEXT:     LSHL * T2.W, T3.X, literal.x,
-; EG-NEXT:    16(2.242078e-44), 0(0.000000e+00)
-; EG-NEXT:     CNDE_INT T0.W, T0.W, literal.x, PV.W,
-; EG-NEXT:     OR_INT * T1.W, PS, T2.X,
-; EG-NEXT:    32(4.484155e-44), 0(0.000000e+00)
+; EG-NEXT:     OR_INT * T0.W, PV.W, T3.X,
+; EG-NEXT:     FFBL_INT T0.W, PV.W,
+; EG-NEXT:     OR_INT * T1.W, T1.W, T2.X,
 ; EG-NEXT:     FFBL_INT T2.W, PS,
 ; EG-NEXT:     ADD_INT * T0.W, PV.W, literal.x,
 ; EG-NEXT:    32(4.484155e-44), 0(0.000000e+00)
 ; EG-NEXT:     CNDE_INT T0.X, T1.W, PS, PV.W,
 ; EG-NEXT:     MOV T0.Y, 0.0,
 ; EG-NEXT:     LSHR * T1.X, KC0[2].Y, literal.x,
 ; EG-NEXT:    2(2.802597e-45), 0(0.000000e+00)
 ;

 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    flat_store_dword v[0:1], v2
 ; VI-NEXT:    s_endpgm
 ;
 ; EG-LABEL: v_cttz_i32_sel_eq_neg1:
 ; EG:       ; %bb.0:
 ; EG-NEXT:    ALU 0, @10, KC0[CB0:0-32], KC1[]
 ; EG-NEXT:    TEX 1 @6
-; EG-NEXT:    ALU 8, @11, KC0[CB0:0-32], KC1[]
+; EG-NEXT:    ALU 6, @11, KC0[CB0:0-32], KC1[]
 ; EG-NEXT:    MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
 ; EG-NEXT:    CF_END
 ; EG-NEXT:    PAD
 ; EG-NEXT:    Fetch clause starting at 6:
 ; EG-NEXT:     VTX_READ_16 T1.X, T0.X, 2, #1
 ; EG-NEXT:     VTX_READ_16 T0.X, T0.X, 0, #1
 ; EG-NEXT:    ALU clause starting at 10:
 ; EG-NEXT:     MOV * T0.X, KC0[2].Z,
 ; EG-NEXT:    ALU clause starting at 11:
 ; EG-NEXT:     LSHL * T0.W, T1.X, literal.x,
 ; EG-NEXT:    16(2.242078e-44), 0(0.000000e+00)
 ; EG-NEXT:     OR_INT * T0.W, PV.W, T0.X,
 ; EG-NEXT:     FFBL_INT * T1.W, PV.W,
-; EG-NEXT:     CNDE_INT * T1.W, T0.W, literal.x, PV.W,
-; EG-NEXT:    32(4.484155e-44), 0(0.000000e+00)
 ; EG-NEXT:     CNDE_INT T0.X, T0.W, literal.x, PV.W,
 ; EG-NEXT:     LSHR * T1.X, KC0[2].Y, literal.y,
 ; EG-NEXT:    -1(nan), 2(2.802597e-45)
 ;
 ; GFX9-GISEL-LABEL: v_cttz_i32_sel_eq_neg1:
 ; GFX9-GISEL:       ; %bb.0:
 ; GFX9-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
 ; GFX9-GISEL-NEXT:    v_mov_b32_e32 v0, 0

 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    flat_store_dword v[0:1], v2
 ; VI-NEXT:    s_endpgm
 ;
 ; EG-LABEL: v_cttz_i32_sel_ne_neg1:
 ; EG:       ; %bb.0:
 ; EG-NEXT:    ALU 0, @10, KC0[CB0:0-32], KC1[]
 ; EG-NEXT:    TEX 1 @6
-; EG-NEXT:    ALU 8, @11, KC0[CB0:0-32], KC1[]
+; EG-NEXT:    ALU 6, @11, KC0[CB0:0-32], KC1[]
 ; EG-NEXT:    MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
 ; EG-NEXT:    CF_END
 ; EG-NEXT:    PAD
 ; EG-NEXT:    Fetch clause starting at 6:
 ; EG-NEXT:     VTX_READ_16 T1.X, T0.X, 2, #1
 ; EG-NEXT:     VTX_READ_16 T0.X, T0.X, 0, #1
 ; EG-NEXT:    ALU clause starting at 10:
 ; EG-NEXT:     MOV * T0.X, KC0[2].Z,
 ; EG-NEXT:    ALU clause starting at 11:
 ; EG-NEXT:     LSHL * T0.W, T1.X, literal.x,
 ; EG-NEXT:    16(2.242078e-44), 0(0.000000e+00)
 ; EG-NEXT:     OR_INT * T0.W, PV.W, T0.X,
 ; EG-NEXT:     FFBL_INT * T1.W, PV.W,
-; EG-NEXT:     CNDE_INT * T1.W, T0.W, literal.x, PV.W,
-; EG-NEXT:    32(4.484155e-44), 0(0.000000e+00)
 ; EG-NEXT:     CNDE_INT T0.X, T0.W, literal.x, PV.W,
 ; EG-NEXT:     LSHR * T1.X, KC0[2].Y, literal.y,
 ; EG-NEXT:    -1(nan), 2(2.802597e-45)
 ;
 ; GFX9-GISEL-LABEL: v_cttz_i32_sel_ne_neg1:
 ; GFX9-GISEL:       ; %bb.0:
 ; GFX9-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
 ; GFX9-GISEL-NEXT:    v_mov_b32_e32 v0, 0

 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v2
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v1, v1, v3
 ; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
 ; SI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; SI-NEXT:    v_ffbl_b32_e32 v0, v0
-; SI-NEXT:    v_min_u32_e32 v0, 32, v0
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 32, v0
 ; SI-NEXT:    v_cndmask_b32_e32 v0, -1, v0, vcc
 ; SI-NEXT:    buffer_store_dword v0, off, s[4:7], 0
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: v_cttz_i32_sel_ne_bitwidth:
 ; VI:       ; %bb.0:
 ; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24

 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_or_b32_sdwa v1, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_lshlrev_b32_e32 v2, 8, v4
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_e32 v0, v2, v0
 ; VI-NEXT:    v_or_b32_e32 v0, v1, v0
 ; VI-NEXT:    v_ffbl_b32_e32 v0, v0
-; VI-NEXT:    v_min_u32_e32 v0, 32, v0
 ; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 32, v0
 ; VI-NEXT:    v_cndmask_b32_e32 v2, -1, v0, vcc
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    flat_store_dword v[0:1], v2
 ; VI-NEXT:    s_endpgm
 ;
 ; EG-LABEL: v_cttz_i32_sel_ne_bitwidth:
 ; EG:       ; %bb.0:
 ; EG-NEXT:    ALU 0, @10, KC0[CB0:0-32], KC1[]
 ; EG-NEXT:    TEX 1 @6
-; EG-NEXT:    ALU 10, @11, KC0[CB0:0-32], KC1[]
+; EG-NEXT:    ALU 8, @11, KC0[CB0:0-32], KC1[]
 ; EG-NEXT:    MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
 ; EG-NEXT:    CF_END
 ; EG-NEXT:    PAD
 ; EG-NEXT:    Fetch clause starting at 6:
 ; EG-NEXT:     VTX_READ_16 T1.X, T0.X, 2, #1
 ; EG-NEXT:     VTX_READ_16 T0.X, T0.X, 0, #1
 ; EG-NEXT:    ALU clause starting at 10:
 ; EG-NEXT:     MOV * T0.X, KC0[2].Z,
 ; EG-NEXT:    ALU clause starting at 11:
 ; EG-NEXT:     LSHL * T0.W, T1.X, literal.x,
 ; EG-NEXT:    16(2.242078e-44), 0(0.000000e+00)
 ; EG-NEXT:     OR_INT * T0.W, PV.W, T0.X,
-; EG-NEXT:     FFBL_INT * T1.W, PV.W,
-; EG-NEXT:     CNDE_INT * T0.W, T0.W, literal.x, PV.W,
-; EG-NEXT:    32(4.484155e-44), 0(0.000000e+00)
+; EG-NEXT:     FFBL_INT * T0.W, PV.W,
 ; EG-NEXT:     SETNE_INT * T1.W, PV.W, literal.x,
 ; EG-NEXT:    32(4.484155e-44), 0(0.000000e+00)
 ; EG-NEXT:     CNDE_INT T0.X, PV.W, literal.x, T0.W,
 ; EG-NEXT:     LSHR * T1.X, KC0[2].Y, literal.y,
 ; EG-NEXT:    -1(nan), 2(2.802597e-45)
 ;
 ; GFX9-GISEL-LABEL: v_cttz_i32_sel_ne_bitwidth:
 ; GFX9-GISEL:       ; %bb.0:

 ; VI-NEXT:    v_mov_b32_e32 v1, s3
 ; VI-NEXT:    flat_load_ubyte v2, v[2:3]
 ; VI-NEXT:    flat_load_ubyte v0, v[0:1]
 ; VI-NEXT:    v_mov_b32_e32 v1, 0xffff
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_e32 v0, v2, v0
-; VI-NEXT:    v_or_b32_e32 v2, 0x10000, v0
-; VI-NEXT:    v_ffbl_b32_e32 v2, v2
-; VI-NEXT:    v_min_u32_e32 v2, 32, v2
+; VI-NEXT:    v_ffbl_b32_e32 v2, v0
 ; VI-NEXT:    v_cmp_ne_u16_e32 vcc, 0, v0
 ; VI-NEXT:    v_cndmask_b32_e32 v2, v1, v2, vcc
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    flat_store_short v[0:1], v2
 ; VI-NEXT:    s_endpgm
 ;
 ; EG-LABEL: v_cttz_i16_sel_eq_neg1:

llvm/test/CodeGen/X86/divrem-by-select.ll

 ; CHECK-X64-V3-NEXT:    vmovq %rax, %xmm1
 ; CHECK-X64-V3-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
 ; CHECK-X64-V3-NEXT:    retq
 ;
 ; CHECK-X64-V4-LABEL: udiv_identity_const_todo_getter_nonzero:
 ; CHECK-X64-V4:       # %bb.0:
 ; CHECK-X64-V4-NEXT:    vpsllq $63, %xmm0, %xmm0
 ; CHECK-X64-V4-NEXT:    vpmovq2m %xmm0, %k1
-; CHECK-X64-V4-NEXT:    vpbroadcastq {{.*#+}} xmm0 = [1,1]
-; CHECK-X64-V4-NEXT:    vpbroadcastq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0 {%k1}
-; CHECK-X64-V4-NEXT:    vpextrq $1, %xmm0, %rcx
-; CHECK-X64-V4-NEXT:    vpextrq $1, %xmm1, %rax
-; CHECK-X64-V4-NEXT:    xorl %edx, %edx
-; CHECK-X64-V4-NEXT:    divq %rcx
-; CHECK-X64-V4-NEXT:    movq %rax, %rcx
-; CHECK-X64-V4-NEXT:    vmovq %xmm0, %rsi
-; CHECK-X64-V4-NEXT:    vmovq %xmm1, %rax
-; CHECK-X64-V4-NEXT:    xorl %edx, %edx
-; CHECK-X64-V4-NEXT:    divq %rsi
+; CHECK-X64-V4-NEXT:    vpextrq $1, %xmm1, %rdx
+; CHECK-X64-V4-NEXT:    movabsq $-3689348814741910323, %rax # imm = 0xCCCCCCCCCCCCCCCD
+; CHECK-X64-V4-NEXT:    mulxq %rax, %rcx, %rcx
 ; CHECK-X64-V4-NEXT:    vmovq %rcx, %xmm0
-; CHECK-X64-V4-NEXT:    vmovq %rax, %xmm1
-; CHECK-X64-V4-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; CHECK-X64-V4-NEXT:    vmovq %xmm1, %rdx
+; CHECK-X64-V4-NEXT:    mulxq %rax, %rax, %rax
+; CHECK-X64-V4-NEXT:    vmovq %rax, %xmm2
+; CHECK-X64-V4-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm2[0],xmm0[0]
+; CHECK-X64-V4-NEXT:    vpsrlq $3, %xmm0, %xmm1 {%k1}
+; CHECK-X64-V4-NEXT:    vmovdqa %xmm1, %xmm0
 ; CHECK-X64-V4-NEXT:    retq
 
   ;; Fails at the moment because `10` is even so there is no common
   ;; bit between the arms of the select so the default case of
   ;; isKnownNeverZero has no guaranteed 1 bits.
   %d = select <2 x i1> %c, <2 x i64> <i64 10, i64 10>, <2 x i64> <i64 1, i64 1>
   %r = udiv <2 x i64> %x, %d
   ret <2 x i64> %r

 ; CHECK-X64-V3-NEXT:    divq %rsi
 ; CHECK-X64-V3-NEXT:    vmovq %rcx, %xmm0
 ; CHECK-X64-V3-NEXT:    vmovq %rax, %xmm1
 ; CHECK-X64-V3-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
 ; CHECK-X64-V3-NEXT:    retq
 ;
 ; CHECK-X64-V4-LABEL: udiv_indentity_non_zero:
 ; CHECK-X64-V4:       # %bb.0:
-; CHECK-X64-V4-NEXT:    vpsllq $63, %xmm0, %xmm0
-; CHECK-X64-V4-NEXT:    vpmovq2m %xmm0, %k1
-; CHECK-X64-V4-NEXT:    vpcmpeqd %xmm0, %xmm0, %xmm0
-; CHECK-X64-V4-NEXT:    vpbroadcastq {{.*#+}} xmm3 = [1,1]
-; CHECK-X64-V4-NEXT:    vpsubq %xmm0, %xmm2, %xmm3 {%k1}
-; CHECK-X64-V4-NEXT:    vpextrq $1, %xmm3, %rcx
+; CHECK-X64-V4-NEXT:    vpcmpeqd %xmm3, %xmm3, %xmm3
+; CHECK-X64-V4-NEXT:    vpsubq %xmm3, %xmm2, %xmm2
+; CHECK-X64-V4-NEXT:    vpextrq $1, %xmm2, %rcx
 ; CHECK-X64-V4-NEXT:    vpextrq $1, %xmm1, %rax
 ; CHECK-X64-V4-NEXT:    xorl %edx, %edx
 ; CHECK-X64-V4-NEXT:    divq %rcx
 ; CHECK-X64-V4-NEXT:    movq %rax, %rcx
-; CHECK-X64-V4-NEXT:    vmovq %xmm3, %rsi
+; CHECK-X64-V4-NEXT:    vmovq %xmm2, %rsi
 ; CHECK-X64-V4-NEXT:    vmovq %xmm1, %rax
 ; CHECK-X64-V4-NEXT:    xorl %edx, %edx
 ; CHECK-X64-V4-NEXT:    divq %rsi
+; CHECK-X64-V4-NEXT:    vpsllq $63, %xmm0, %xmm0
+; CHECK-X64-V4-NEXT:    vpmovq2m %xmm0, %k1
 ; CHECK-X64-V4-NEXT:    vmovq %rcx, %xmm0
-; CHECK-X64-V4-NEXT:    vmovq %rax, %xmm1
-; CHECK-X64-V4-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
+; CHECK-X64-V4-NEXT:    vmovq %rax, %xmm2
+; CHECK-X64-V4-NEXT:    vpunpcklqdq {{.*#+}} xmm1 {%k1} = xmm2[0],xmm0[0]
+; CHECK-X64-V4-NEXT:    vmovdqa %xmm1, %xmm0
 ; CHECK-X64-V4-NEXT:    retq
   %non_zero = add nsw nuw <2 x i64> %y, <i64 1, i64 1>
   %d = select <2 x i1> %c, <2 x i64> %non_zero, <2 x i64> <i64 1, i64 1>
   %r = udiv <2 x i64> %x, %d
   ret <2 x i64> %r
 }
 
 define <2 x i64> @udiv_indentity_zero(<2 x i1> %c, <2 x i64> %x) {