[llvm] use multiplication with holes for 32-bit/64-bit clmul fallback (PR #203727)

Folkert de Vries via llvm-commits llvm-commits at lists.llvm.org
Sat Jun 13 13:58:50 PDT 2026


https://github.com/folkertdev created https://github.com/llvm/llvm-project/pull/203727

Well I think I've nerdsniped myself, at least for the simple case that can be based on "holey multiplication" (via [bearssl source](https://github.com/unkaktus/bearssl/blob/6a1dab03beac9a324fdaed36ffdeb5081ac64b9a/src/hash/ghash_ctmul64.c#L27-L56), https://www.bearssl.org/constanttime.html#ghash-for-gcm,  and the [polyval](https://github.com/RustCrypto/universal-hashes/blob/815ddc90ea652b09804565e193e34784c94a3849/polyval/src/field_element/soft.rs#L74-L108) crate).

That still leaves the widening case though, for which this blog post has some good info https://timtaubert.de/blog/2017/06/verified-binary-multiplication-for-ghash/. Also for 8-bit and 16-bit inputs this approach can be generalized but with smaller holes and fewer multiplications. I'll leave that as future work too.

I've tested this locally with a fuzzer against the current LLVM implementation (via the rust standard library) and `pclmulqdq`. So, I think it's correct but I'm not sure how to show that in a convincing way.

I know this needs tests updated for more targets, but I'd like CI to tell me which ones.

CC https://github.com/llvm/llvm-project/issues/203694
CC: @RKSimon, @topperc, @eisenwave, @xarkenz, @artagnon.

>From 039cf96769d244334d441bfd09a809ba0cbb6359 Mon Sep 17 00:00:00 2001
From: Folkert de Vries <folkert at folkertdev.nl>
Date: Sat, 13 Jun 2026 22:04:37 +0200
Subject: [PATCH] use multiplication with holes for clmul fallback

---
 llvm/include/llvm/CodeGen/BasicTTIImpl.h      |   25 +-
 .../CodeGen/SelectionDAG/TargetLowering.cpp   |   51 +
 llvm/test/CodeGen/AArch64/clmul.ll            | 2798 +++----------
 llvm/test/CodeGen/X86/clmul.ll                | 3708 +++++------------
 4 files changed, 1546 insertions(+), 5036 deletions(-)

diff --git a/llvm/include/llvm/CodeGen/BasicTTIImpl.h b/llvm/include/llvm/CodeGen/BasicTTIImpl.h
index e4b6bf51c7a4e..b11057695c2bd 100644
--- a/llvm/include/llvm/CodeGen/BasicTTIImpl.h
+++ b/llvm/include/llvm/CodeGen/BasicTTIImpl.h
@@ -3090,18 +3090,33 @@ class BasicTTIImplBase : public TargetTransformInfoImplCRTPBase<T> {
     case Intrinsic::clmul: {
       // This cost model should match the expansion in
       // TargetLowering::expandCLMUL.
-      InstructionCost PerBitCostMul =
-          thisT()->getArithmeticInstrCost(Instruction::And, RetTy, CostKind) +
-          thisT()->getArithmeticInstrCost(Instruction::Mul, RetTy, CostKind) +
+      unsigned BW = RetTy->getScalarSizeInBits();
+      InstructionCost AndCost =
+          thisT()->getArithmeticInstrCost(Instruction::And, RetTy, CostKind);
+      InstructionCost OrCost =
+          thisT()->getArithmeticInstrCost(Instruction::Or, RetTy, CostKind);
+      InstructionCost XorCost =
           thisT()->getArithmeticInstrCost(Instruction::Xor, RetTy, CostKind);
+      InstructionCost MulCost =
+          thisT()->getArithmeticInstrCost(Instruction::Mul, RetTy, CostKind);
+
+      // When the multiplication with holes approach is used, that emits 16 MULs, 8
+      // + 4 ANDs, 12 XORs and 3 ORs.
+      if (BW >= 32 && BW <= 64 &&
+          TLI->isOperationLegalOrCustom(ISD::MUL,
+                                        TLI->getValueType(DL, RetTy))) {
+        return 16 * MulCost + 12 * AndCost + 12 * XorCost + 3 * OrCost;
+      }
+
+      InstructionCost PerBitCostMul = AndCost + MulCost + XorCost;
       InstructionCost PerBitCostBittest =
-          thisT()->getArithmeticInstrCost(Instruction::And, RetTy, CostKind) +
+          AndCost +
           thisT()->getCmpSelInstrCost(BinaryOperator::Select, RetTy, RetTy,
                                       ICmpInst::BAD_ICMP_PREDICATE, CostKind) +
           thisT()->getCmpSelInstrCost(Instruction::ICmp, RetTy, RetTy,
                                       ICmpInst::ICMP_NE, CostKind);
       InstructionCost PerBitCost = std::min(PerBitCostMul, PerBitCostBittest);
-      return RetTy->getScalarSizeInBits() * PerBitCost;
+      return BW * PerBitCost;
     }
     default:
       break;
diff --git a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
index 95e03a419fb43..175a41aa502dc 100644
--- a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -8899,6 +8899,57 @@ SDValue TargetLowering::expandCLMUL(SDNode *Node, SelectionDAG &DAG) const {
     // calculation in BasicTTIImpl::getTypeBasedIntrinsicInstrCost for
     // Intrinsic::clmul.
 
+    // Strategy 4: multiplication with holes. 
+    //
+    // Uses "holes" (sequences of zeroes) to avoid carry spilling. When carries
+    // do occur, they wind up in a "hole" and are subsequently masked out of the
+    // result.
+    //
+    // A hole of 3 bits is optimal for 32-bit and 64-bit inputs. 128-bit
+    // integers need a larger hole, and for smaller integers the fallback below
+    // is more efficient.
+    //
+    // Based on bmul64 in bearssl and bmul in the rust polyval crate.
+    if (BW >= 32 && BW <= 64 &&
+        isOperationLegalOrCustom(ISD::MUL, getTypeToTransformTo(Ctx, VT))) {
+
+      // Set every fourth bit of each nibble, equivalent to 0b100010001...0001.
+      APInt MaskVal(BW, 0);
+      for (unsigned i = 0; i < BW; i += 4)
+        MaskVal.setBit(i);
+
+      // Create versions of X and Y that keep only the first, second, third, or
+      // fourth bit of every nibble.
+      SDValue M[4], Xp[4], Yp[4];
+      for (unsigned i = 0; i < 4; ++i) {
+        M[i] = DAG.getConstant(MaskVal.shl(i), DL, VT);
+        Xp[i] = DAG.getNode(ISD::AND, DL, VT, X, M[i]);
+        Yp[i] = DAG.getNode(ISD::AND, DL, VT, Y, M[i]);
+      }
+
+      // Codegens these expressions (16 multiplications):
+      //
+      // z0 = (x0 * y0) ^ (x1 * y3) ^ (x2 * y2) ^ (x3 * y1);
+      // z1 = (x0 * y1) ^ (x1 * y0) ^ (x2 * y3) ^ (x3 * y2);
+      // z2 = (x0 * y2) ^ (x1 * y1) ^ (x2 * y0) ^ (x3 * y3);
+      // z3 = (x0 * y3) ^ (x1 * y2) ^ (x2 * y1) ^ (x3 * y0);
+      SDValue Res;
+      for (unsigned i = 0; i < 4; ++i) {
+        SDValue Zi;
+        for (unsigned A = 0; A < 4; ++A) {
+          unsigned B = (i + 4 - A) % 4;
+          SDValue P = DAG.getNode(ISD::MUL, DL, VT, Xp[A], Yp[B]);
+          Zi = Zi ? DAG.getNode(ISD::XOR, DL, VT, Zi, P) : P;
+        }
+
+        // Keep only the bits belonging to this iteration, and bitwise or it all
+        // together.
+        Zi = DAG.getNode(ISD::AND, DL, VT, Zi, M[i]);
+        Res = Res ? DAG.getNode(ISD::OR, DL, VT, Res, Zi) : Zi;
+      }
+      return Res;
+    }
+
     EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(), Ctx, VT);
 
     SDValue Res = DAG.getConstant(0, DL, VT);
diff --git a/llvm/test/CodeGen/AArch64/clmul.ll b/llvm/test/CodeGen/AArch64/clmul.ll
index bb42e411c4f99..bfa4b73b4e677 100644
--- a/llvm/test/CodeGen/AArch64/clmul.ll
+++ b/llvm/test/CodeGen/AArch64/clmul.ll
@@ -80,101 +80,49 @@ define i16 @clmul_i16(i16 %x, i16 %y) {
 define i32 @clmul_i32(i32 %x, i32 %y) {
 ; CHECK-NEON-LABEL: clmul_i32:
 ; CHECK-NEON:       // %bb.0:
-; CHECK-NEON-NEXT:    and w8, w1, #0x2
-; CHECK-NEON-NEXT:    and w9, w1, #0x1
-; CHECK-NEON-NEXT:    and w10, w1, #0x4
-; CHECK-NEON-NEXT:    mul w8, w0, w8
-; CHECK-NEON-NEXT:    and w11, w1, #0x8
-; CHECK-NEON-NEXT:    and w12, w1, #0x10
-; CHECK-NEON-NEXT:    mul w9, w0, w9
-; CHECK-NEON-NEXT:    and w13, w1, #0x20
-; CHECK-NEON-NEXT:    and w14, w1, #0x40
-; CHECK-NEON-NEXT:    mul w10, w0, w10
-; CHECK-NEON-NEXT:    and w2, w1, #0x800
-; CHECK-NEON-NEXT:    and w15, w1, #0x80
-; CHECK-NEON-NEXT:    mul w11, w0, w11
-; CHECK-NEON-NEXT:    and w16, w1, #0x100
-; CHECK-NEON-NEXT:    and w17, w1, #0x200
-; CHECK-NEON-NEXT:    mul w12, w0, w12
-; CHECK-NEON-NEXT:    eor w8, w9, w8
-; CHECK-NEON-NEXT:    and w9, w1, #0x1000
-; CHECK-NEON-NEXT:    mul w13, w0, w13
-; CHECK-NEON-NEXT:    and w18, w1, #0x400
-; CHECK-NEON-NEXT:    mul w14, w0, w14
-; CHECK-NEON-NEXT:    eor w10, w10, w11
-; CHECK-NEON-NEXT:    and w11, w1, #0x2000
-; CHECK-NEON-NEXT:    mul w2, w0, w2
-; CHECK-NEON-NEXT:    eor w8, w8, w10
-; CHECK-NEON-NEXT:    and w10, w1, #0x4000
-; CHECK-NEON-NEXT:    mul w9, w0, w9
-; CHECK-NEON-NEXT:    eor w12, w12, w13
-; CHECK-NEON-NEXT:    and w13, w1, #0x8000
-; CHECK-NEON-NEXT:    mul w15, w0, w15
-; CHECK-NEON-NEXT:    eor w12, w12, w14
-; CHECK-NEON-NEXT:    and w14, w1, #0x10000
-; CHECK-NEON-NEXT:    mul w16, w0, w16
-; CHECK-NEON-NEXT:    eor w8, w8, w12
-; CHECK-NEON-NEXT:    and w12, w1, #0x20000
-; CHECK-NEON-NEXT:    mul w11, w0, w11
-; CHECK-NEON-NEXT:    eor w9, w2, w9
-; CHECK-NEON-NEXT:    and w2, w1, #0x400000
-; CHECK-NEON-NEXT:    mul w17, w0, w17
-; CHECK-NEON-NEXT:    mul w10, w0, w10
-; CHECK-NEON-NEXT:    eor w15, w15, w16
-; CHECK-NEON-NEXT:    and w16, w1, #0x40000
-; CHECK-NEON-NEXT:    mul w13, w0, w13
-; CHECK-NEON-NEXT:    eor w9, w9, w11
-; CHECK-NEON-NEXT:    and w11, w1, #0x800000
-; CHECK-NEON-NEXT:    mul w18, w0, w18
-; CHECK-NEON-NEXT:    eor w15, w15, w17
-; CHECK-NEON-NEXT:    and w17, w1, #0x80000
-; CHECK-NEON-NEXT:    mul w14, w0, w14
-; CHECK-NEON-NEXT:    eor w9, w9, w10
-; CHECK-NEON-NEXT:    and w10, w1, #0x1000000
-; CHECK-NEON-NEXT:    mul w12, w0, w12
-; CHECK-NEON-NEXT:    eor w9, w9, w13
-; CHECK-NEON-NEXT:    and w13, w1, #0x2000000
-; CHECK-NEON-NEXT:    mul w16, w0, w16
-; CHECK-NEON-NEXT:    eor w15, w15, w18
-; CHECK-NEON-NEXT:    and w18, w1, #0x100000
-; CHECK-NEON-NEXT:    mul w2, w0, w2
-; CHECK-NEON-NEXT:    eor w8, w8, w15
-; CHECK-NEON-NEXT:    and w15, w1, #0x200000
-; CHECK-NEON-NEXT:    mul w11, w0, w11
+; CHECK-NEON-NEXT:    and w8, w1, #0x11111111
+; CHECK-NEON-NEXT:    and w9, w0, #0x22222222
+; CHECK-NEON-NEXT:    and w10, w1, #0x22222222
+; CHECK-NEON-NEXT:    and w11, w0, #0x11111111
+; CHECK-NEON-NEXT:    and w13, w1, #0x88888888
+; CHECK-NEON-NEXT:    and w15, w0, #0x44444444
+; CHECK-NEON-NEXT:    and w17, w1, #0x44444444
+; CHECK-NEON-NEXT:    and w18, w0, #0x88888888
+; CHECK-NEON-NEXT:    mul w12, w9, w8
+; CHECK-NEON-NEXT:    mul w14, w11, w10
+; CHECK-NEON-NEXT:    mul w16, w15, w13
+; CHECK-NEON-NEXT:    mul w0, w18, w17
+; CHECK-NEON-NEXT:    mul w1, w9, w13
 ; CHECK-NEON-NEXT:    eor w12, w14, w12
-; CHECK-NEON-NEXT:    and w14, w1, #0x4000000
-; CHECK-NEON-NEXT:    mul w17, w0, w17
-; CHECK-NEON-NEXT:    eor w12, w12, w16
-; CHECK-NEON-NEXT:    and w16, w1, #0x8000000
-; CHECK-NEON-NEXT:    mul w10, w0, w10
-; CHECK-NEON-NEXT:    eor w8, w8, w9
-; CHECK-NEON-NEXT:    mul w13, w0, w13
-; CHECK-NEON-NEXT:    eor w11, w2, w11
-; CHECK-NEON-NEXT:    and w2, w1, #0x20000000
-; CHECK-NEON-NEXT:    mul w18, w0, w18
-; CHECK-NEON-NEXT:    eor w12, w12, w17
-; CHECK-NEON-NEXT:    and w17, w1, #0x10000000
-; CHECK-NEON-NEXT:    mul w14, w0, w14
-; CHECK-NEON-NEXT:    eor w10, w11, w10
-; CHECK-NEON-NEXT:    and w11, w1, #0x40000000
-; CHECK-NEON-NEXT:    mul w15, w0, w15
-; CHECK-NEON-NEXT:    eor w10, w10, w13
-; CHECK-NEON-NEXT:    and w13, w1, #0x80000000
-; CHECK-NEON-NEXT:    mul w16, w0, w16
-; CHECK-NEON-NEXT:    eor w12, w12, w18
-; CHECK-NEON-NEXT:    mul w17, w0, w17
-; CHECK-NEON-NEXT:    eor w10, w10, w14
-; CHECK-NEON-NEXT:    mul w2, w0, w2
-; CHECK-NEON-NEXT:    eor w9, w12, w15
-; CHECK-NEON-NEXT:    mul w11, w0, w11
-; CHECK-NEON-NEXT:    eor w10, w10, w16
-; CHECK-NEON-NEXT:    eor w8, w8, w9
-; CHECK-NEON-NEXT:    mul w13, w0, w13
-; CHECK-NEON-NEXT:    eor w9, w10, w17
-; CHECK-NEON-NEXT:    eor w8, w8, w9
-; CHECK-NEON-NEXT:    eor w10, w2, w11
-; CHECK-NEON-NEXT:    eor w9, w10, w13
-; CHECK-NEON-NEXT:    eor w0, w8, w9
+; CHECK-NEON-NEXT:    mul w2, w11, w8
+; CHECK-NEON-NEXT:    mul w3, w15, w17
+; CHECK-NEON-NEXT:    eor w14, w16, w0
+; CHECK-NEON-NEXT:    mul w4, w18, w10
+; CHECK-NEON-NEXT:    eor w12, w12, w14
+; CHECK-NEON-NEXT:    mul w5, w9, w10
+; CHECK-NEON-NEXT:    eor w14, w2, w1
+; CHECK-NEON-NEXT:    and w12, w12, #0x22222222
+; CHECK-NEON-NEXT:    mul w9, w9, w17
+; CHECK-NEON-NEXT:    mul w17, w11, w17
+; CHECK-NEON-NEXT:    eor w16, w3, w4
+; CHECK-NEON-NEXT:    mul w10, w15, w10
+; CHECK-NEON-NEXT:    mul w15, w15, w8
+; CHECK-NEON-NEXT:    mul w11, w11, w13
+; CHECK-NEON-NEXT:    eor w17, w17, w5
+; CHECK-NEON-NEXT:    mul w13, w18, w13
+; CHECK-NEON-NEXT:    mul w8, w18, w8
+; CHECK-NEON-NEXT:    eor w9, w11, w9
+; CHECK-NEON-NEXT:    eor w13, w15, w13
+; CHECK-NEON-NEXT:    eor w8, w10, w8
+; CHECK-NEON-NEXT:    eor w10, w14, w16
+; CHECK-NEON-NEXT:    eor w11, w17, w13
+; CHECK-NEON-NEXT:    eor w8, w9, w8
+; CHECK-NEON-NEXT:    and w9, w10, #0x11111111
+; CHECK-NEON-NEXT:    and w10, w11, #0x44444444
+; CHECK-NEON-NEXT:    and w8, w8, #0x88888888
+; CHECK-NEON-NEXT:    orr w9, w9, w12
+; CHECK-NEON-NEXT:    orr w8, w10, w8
+; CHECK-NEON-NEXT:    orr w0, w9, w8
 ; CHECK-NEON-NEXT:    ret
 ;
 ; CHECK-AES-LABEL: clmul_i32:
@@ -191,274 +139,49 @@ define i32 @clmul_i32(i32 %x, i32 %y) {
 define i64 @clmul_i64(i64 %x, i64 %y) {
 ; CHECK-NEON-LABEL: clmul_i64:
 ; CHECK-NEON:       // %bb.0:
-; CHECK-NEON-NEXT:    sub sp, sp, #304
-; CHECK-NEON-NEXT:    stp x29, x30, [sp, #208] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    stp x28, x27, [sp, #224] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    stp x26, x25, [sp, #240] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    stp x24, x23, [sp, #256] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    stp x22, x21, [sp, #272] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    stp x20, x19, [sp, #288] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    .cfi_def_cfa_offset 304
-; CHECK-NEON-NEXT:    .cfi_offset w19, -8
-; CHECK-NEON-NEXT:    .cfi_offset w20, -16
-; CHECK-NEON-NEXT:    .cfi_offset w21, -24
-; CHECK-NEON-NEXT:    .cfi_offset w22, -32
-; CHECK-NEON-NEXT:    .cfi_offset w23, -40
-; CHECK-NEON-NEXT:    .cfi_offset w24, -48
-; CHECK-NEON-NEXT:    .cfi_offset w25, -56
-; CHECK-NEON-NEXT:    .cfi_offset w26, -64
-; CHECK-NEON-NEXT:    .cfi_offset w27, -72
-; CHECK-NEON-NEXT:    .cfi_offset w28, -80
-; CHECK-NEON-NEXT:    .cfi_offset w30, -88
-; CHECK-NEON-NEXT:    .cfi_offset w29, -96
-; CHECK-NEON-NEXT:    and x8, x1, #0x2
-; CHECK-NEON-NEXT:    mul x9, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x1
-; CHECK-NEON-NEXT:    mul x10, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x4
-; CHECK-NEON-NEXT:    mul x11, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x8
-; CHECK-NEON-NEXT:    mul x13, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x10
-; CHECK-NEON-NEXT:    eor x9, x10, x9
-; CHECK-NEON-NEXT:    mul x12, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x20
-; CHECK-NEON-NEXT:    mul x14, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x40
-; CHECK-NEON-NEXT:    eor x10, x11, x13
-; CHECK-NEON-NEXT:    and x11, x1, #0x10000000000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #200] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x80
-; CHECK-NEON-NEXT:    mul x15, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x100
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #160] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x200
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #152] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x400
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #184] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x800
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #192] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x1000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #144] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x2000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #136] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x4000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #176] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x8000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #168] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x10000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #120] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x20000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #80] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x40000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #72] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x80000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #104] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x100000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #96] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x200000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #128] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x400000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #112] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x800000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #64] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x1000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #40] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x2000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    ldr x30, [sp, #40] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x8, [sp, #32] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x4000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #56] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x8000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #48] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x10000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #88] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x20000000
-; CHECK-NEON-NEXT:    mul x26, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x40000000
-; CHECK-NEON-NEXT:    mul x22, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x80000000
-; CHECK-NEON-NEXT:    mul x23, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x100000000
-; CHECK-NEON-NEXT:    mul x24, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x200000000
-; CHECK-NEON-NEXT:    eor x22, x26, x22
-; CHECK-NEON-NEXT:    ldr x26, [sp, #32] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x25, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x400000000
-; CHECK-NEON-NEXT:    eor x22, x22, x23
-; CHECK-NEON-NEXT:    and x23, x1, #0x400000000000000
-; CHECK-NEON-NEXT:    mul x27, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x800000000
-; CHECK-NEON-NEXT:    eor x22, x22, x24
-; CHECK-NEON-NEXT:    ldr x24, [sp, #48] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x28, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x1000000000
-; CHECK-NEON-NEXT:    eor x22, x22, x25
-; CHECK-NEON-NEXT:    ldr x25, [sp, #88] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x29, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x2000000000
-; CHECK-NEON-NEXT:    eor x22, x22, x27
-; CHECK-NEON-NEXT:    mul x21, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x4000000000
-; CHECK-NEON-NEXT:    mul x7, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x8000000000
-; CHECK-NEON-NEXT:    mul x19, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x10000000000
-; CHECK-NEON-NEXT:    mul x5, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x20000000000
-; CHECK-NEON-NEXT:    eor x7, x21, x7
-; CHECK-NEON-NEXT:    mul x6, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x40000000000
-; CHECK-NEON-NEXT:    mul x20, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x80000000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    mul x23, x0, x23
-; CHECK-NEON-NEXT:    str x8, [sp, #24] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x100000000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #16] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x200000000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #8] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x400000000000
-; CHECK-NEON-NEXT:    mul x4, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x800000000000
-; CHECK-NEON-NEXT:    mul x17, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x1000000000000
-; CHECK-NEON-NEXT:    mul x18, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x2000000000000
-; CHECK-NEON-NEXT:    mul x3, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x4000000000000
-; CHECK-NEON-NEXT:    eor x17, x4, x17
-; CHECK-NEON-NEXT:    mul x2, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x8000000000000
-; CHECK-NEON-NEXT:    eor x17, x17, x18
-; CHECK-NEON-NEXT:    and x18, x1, #0x4000000000000000
-; CHECK-NEON-NEXT:    mul x16, x0, x8
-; CHECK-NEON-NEXT:    eor x8, x9, x10
-; CHECK-NEON-NEXT:    ldr x9, [sp, #160] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x12, x14
-; CHECK-NEON-NEXT:    ldr x12, [sp, #80] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x17, x17, x3
-; CHECK-NEON-NEXT:    eor x9, x15, x9
-; CHECK-NEON-NEXT:    mul x15, x0, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #200] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x17, x17, x2
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #152] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x18, x0, x18
-; CHECK-NEON-NEXT:    eor x8, x8, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #184] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x16, x17, x16
-; CHECK-NEON-NEXT:    eor x9, x9, x11
-; CHECK-NEON-NEXT:    and x11, x1, #0x20000000000000
-; CHECK-NEON-NEXT:    ldr x17, [sp, #24] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x10
-; CHECK-NEON-NEXT:    mul x14, x0, x11
-; CHECK-NEON-NEXT:    and x10, x1, #0x40000000000000
-; CHECK-NEON-NEXT:    eor x11, x8, x9
-; CHECK-NEON-NEXT:    ldr x8, [sp, #192] // 8-byte Reload
-; CHECK-NEON-NEXT:    ldr x9, [sp, #144] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x13, x0, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #136] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x15, x16, x15
-; CHECK-NEON-NEXT:    eor x8, x8, x9
-; CHECK-NEON-NEXT:    ldr x9, [sp, #120] // 8-byte Reload
-; CHECK-NEON-NEXT:    ldr x16, [sp, #16] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x8, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #72] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #176] // 8-byte Reload
+; CHECK-NEON-NEXT:    and x8, x1, #0x1111111111111111
+; CHECK-NEON-NEXT:    and x9, x0, #0x2222222222222222
+; CHECK-NEON-NEXT:    and x11, x1, #0x2222222222222222
+; CHECK-NEON-NEXT:    and x12, x0, #0x1111111111111111
+; CHECK-NEON-NEXT:    and x14, x1, #0x8888888888888888
+; CHECK-NEON-NEXT:    and x15, x0, #0x4444444444444444
+; CHECK-NEON-NEXT:    and x17, x1, #0x4444444444444444
+; CHECK-NEON-NEXT:    and x18, x0, #0x8888888888888888
+; CHECK-NEON-NEXT:    mul x10, x9, x8
+; CHECK-NEON-NEXT:    mul x13, x12, x11
+; CHECK-NEON-NEXT:    mul x16, x15, x14
+; CHECK-NEON-NEXT:    mul x0, x18, x17
+; CHECK-NEON-NEXT:    eor x10, x13, x10
+; CHECK-NEON-NEXT:    mul x1, x9, x14
+; CHECK-NEON-NEXT:    mul x2, x12, x8
+; CHECK-NEON-NEXT:    eor x13, x16, x0
+; CHECK-NEON-NEXT:    mul x3, x15, x17
+; CHECK-NEON-NEXT:    eor x10, x10, x13
+; CHECK-NEON-NEXT:    and x10, x10, #0x2222222222222222
+; CHECK-NEON-NEXT:    mul x4, x18, x11
+; CHECK-NEON-NEXT:    eor x13, x2, x1
+; CHECK-NEON-NEXT:    mul x5, x9, x11
+; CHECK-NEON-NEXT:    mul x9, x9, x17
+; CHECK-NEON-NEXT:    eor x16, x3, x4
+; CHECK-NEON-NEXT:    mul x17, x12, x17
+; CHECK-NEON-NEXT:    mul x11, x15, x11
+; CHECK-NEON-NEXT:    mul x15, x15, x8
+; CHECK-NEON-NEXT:    eor x17, x17, x5
+; CHECK-NEON-NEXT:    mul x12, x12, x14
+; CHECK-NEON-NEXT:    mul x14, x18, x14
+; CHECK-NEON-NEXT:    mul x8, x18, x8
+; CHECK-NEON-NEXT:    eor x9, x12, x9
 ; CHECK-NEON-NEXT:    eor x14, x15, x14
-; CHECK-NEON-NEXT:    eor x9, x9, x10
-; CHECK-NEON-NEXT:    and x10, x1, #0x80000000000000
-; CHECK-NEON-NEXT:    ldr x15, [sp, #8] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x8, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #104] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x13, x14, x13
-; CHECK-NEON-NEXT:    eor x9, x9, x12
-; CHECK-NEON-NEXT:    mul x12, x0, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #168] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x8, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #96] // 8-byte Reload
+; CHECK-NEON-NEXT:    eor x12, x17, x14
 ; CHECK-NEON-NEXT:    eor x8, x11, x8
-; CHECK-NEON-NEXT:    ldr x11, [sp, #128] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x10
-; CHECK-NEON-NEXT:    and x10, x1, #0x100000000000000
-; CHECK-NEON-NEXT:    eor x9, x9, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #64] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x10, x0, x10
-; CHECK-NEON-NEXT:    eor x8, x8, x9
-; CHECK-NEON-NEXT:    ldr x9, [sp, #112] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x11
-; CHECK-NEON-NEXT:    and x11, x1, #0x200000000000000
-; CHECK-NEON-NEXT:    eor x9, x9, x30
-; CHECK-NEON-NEXT:    mul x11, x0, x11
-; CHECK-NEON-NEXT:    eor x9, x9, x26
-; CHECK-NEON-NEXT:    ldr x26, [sp, #56] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x26
-; CHECK-NEON-NEXT:    eor x9, x9, x24
-; CHECK-NEON-NEXT:    and x24, x1, #0x800000000000000
-; CHECK-NEON-NEXT:    eor x9, x9, x25
-; CHECK-NEON-NEXT:    mul x24, x0, x24
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    eor x8, x8, x9
-; CHECK-NEON-NEXT:    eor x9, x22, x28
-; CHECK-NEON-NEXT:    and x22, x1, #0x1000000000000000
-; CHECK-NEON-NEXT:    eor x9, x9, x29
-; CHECK-NEON-NEXT:    mul x21, x0, x22
-; CHECK-NEON-NEXT:    and x11, x1, #0x8000000000000000
-; CHECK-NEON-NEXT:    eor x8, x8, x9
-; CHECK-NEON-NEXT:    eor x9, x7, x19
-; CHECK-NEON-NEXT:    and x7, x1, #0x2000000000000000
-; CHECK-NEON-NEXT:    eor x9, x9, x5
-; CHECK-NEON-NEXT:    mul x4, x0, x7
-; CHECK-NEON-NEXT:    eor x10, x10, x23
-; CHECK-NEON-NEXT:    eor x9, x9, x6
-; CHECK-NEON-NEXT:    eor x10, x10, x24
-; CHECK-NEON-NEXT:    eor x9, x9, x20
-; CHECK-NEON-NEXT:    mul x11, x0, x11
-; CHECK-NEON-NEXT:    eor x9, x9, x17
-; CHECK-NEON-NEXT:    eor x10, x10, x21
-; CHECK-NEON-NEXT:    eor x9, x9, x16
-; CHECK-NEON-NEXT:    ldp x20, x19, [sp, #288] // 16-byte Folded Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x15
-; CHECK-NEON-NEXT:    eor x10, x10, x4
-; CHECK-NEON-NEXT:    eor x8, x8, x9
-; CHECK-NEON-NEXT:    eor x9, x13, x12
-; CHECK-NEON-NEXT:    eor x10, x10, x18
-; CHECK-NEON-NEXT:    ldp x22, x21, [sp, #272] // 16-byte Folded Reload
-; CHECK-NEON-NEXT:    eor x8, x8, x9
-; CHECK-NEON-NEXT:    ldp x24, x23, [sp, #256] // 16-byte Folded Reload
-; CHECK-NEON-NEXT:    eor x9, x10, x11
-; CHECK-NEON-NEXT:    ldp x26, x25, [sp, #240] // 16-byte Folded Reload
-; CHECK-NEON-NEXT:    eor x0, x8, x9
-; CHECK-NEON-NEXT:    ldp x28, x27, [sp, #224] // 16-byte Folded Reload
-; CHECK-NEON-NEXT:    ldp x29, x30, [sp, #208] // 16-byte Folded Reload
-; CHECK-NEON-NEXT:    add sp, sp, #304
+; CHECK-NEON-NEXT:    eor x11, x13, x16
+; CHECK-NEON-NEXT:    eor x8, x9, x8
+; CHECK-NEON-NEXT:    and x9, x11, #0x1111111111111111
+; CHECK-NEON-NEXT:    and x11, x12, #0x4444444444444444
+; CHECK-NEON-NEXT:    and x8, x8, #0x8888888888888888
+; CHECK-NEON-NEXT:    orr x9, x9, x10
+; CHECK-NEON-NEXT:    orr x8, x11, x8
+; CHECK-NEON-NEXT:    orr x0, x9, x8
 ; CHECK-NEON-NEXT:    ret
 ;
 ; CHECK-AES-LABEL: clmul_i64:
@@ -476,14 +199,14 @@ define i64 @clmul_i64(i64 %x, i64 %y) {
 define i128 @clmul_i128(i128 %x, i128 %y) {
 ; CHECK-NEON-LABEL: clmul_i128:
 ; CHECK-NEON:       // %bb.0:
-; CHECK-NEON-NEXT:    stp x29, x30, [sp, #-96]! // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    stp x28, x27, [sp, #16] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    stp x26, x25, [sp, #32] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    stp x24, x23, [sp, #48] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    stp x22, x21, [sp, #64] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    stp x20, x19, [sp, #80] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    sub sp, sp, #1952
-; CHECK-NEON-NEXT:    .cfi_def_cfa_offset 2048
+; CHECK-NEON-NEXT:    sub sp, sp, #352
+; CHECK-NEON-NEXT:    stp x29, x30, [sp, #256] // 16-byte Folded Spill
+; CHECK-NEON-NEXT:    stp x28, x27, [sp, #272] // 16-byte Folded Spill
+; CHECK-NEON-NEXT:    stp x26, x25, [sp, #288] // 16-byte Folded Spill
+; CHECK-NEON-NEXT:    stp x24, x23, [sp, #304] // 16-byte Folded Spill
+; CHECK-NEON-NEXT:    stp x22, x21, [sp, #320] // 16-byte Folded Spill
+; CHECK-NEON-NEXT:    stp x20, x19, [sp, #336] // 16-byte Folded Spill
+; CHECK-NEON-NEXT:    .cfi_def_cfa_offset 352
 ; CHECK-NEON-NEXT:    .cfi_offset w19, -8
 ; CHECK-NEON-NEXT:    .cfi_offset w20, -16
 ; CHECK-NEON-NEXT:    .cfi_offset w21, -24
@@ -496,1262 +219,229 @@ define i128 @clmul_i128(i128 %x, i128 %y) {
 ; CHECK-NEON-NEXT:    .cfi_offset w28, -80
 ; CHECK-NEON-NEXT:    .cfi_offset w30, -88
 ; CHECK-NEON-NEXT:    .cfi_offset w29, -96
+; CHECK-NEON-NEXT:    and x10, x2, #0x2222222222222222
+; CHECK-NEON-NEXT:    and x12, x1, #0x1111111111111111
+; CHECK-NEON-NEXT:    and x11, x2, #0x8888888888888888
+; CHECK-NEON-NEXT:    mul x8, x12, x10
+; CHECK-NEON-NEXT:    and x5, x1, #0x4444444444444444
+; CHECK-NEON-NEXT:    and x21, x1, #0x2222222222222222
+; CHECK-NEON-NEXT:    and x6, x2, #0x1111111111111111
+; CHECK-NEON-NEXT:    rbit x13, x0
+; CHECK-NEON-NEXT:    and x29, x0, #0x2222222222222222
+; CHECK-NEON-NEXT:    mul x14, x21, x10
+; CHECK-NEON-NEXT:    and x15, x13, #0x2222222222222222
+; CHECK-NEON-NEXT:    and x17, x13, #0x1111111111111111
+; CHECK-NEON-NEXT:    and x4, x13, #0x4444444444444444
+; CHECK-NEON-NEXT:    and x13, x13, #0x8888888888888888
+; CHECK-NEON-NEXT:    mul x26, x12, x11
+; CHECK-NEON-NEXT:    str x8, [sp, #248] // 8-byte Spill
+; CHECK-NEON-NEXT:    mul x8, x5, x11
+; CHECK-NEON-NEXT:    str x14, [sp, #88] // 8-byte Spill
+; CHECK-NEON-NEXT:    mul x9, x21, x6
+; CHECK-NEON-NEXT:    mul x20, x5, x6
+; CHECK-NEON-NEXT:    str x8, [sp, #240] // 8-byte Spill
+; CHECK-NEON-NEXT:    mul x8, x21, x11
+; CHECK-NEON-NEXT:    mul x22, x5, x10
+; CHECK-NEON-NEXT:    str x8, [sp, #232] // 8-byte Spill
+; CHECK-NEON-NEXT:    mul x8, x12, x6
+; CHECK-NEON-NEXT:    str x8, [sp, #224] // 8-byte Spill
 ; CHECK-NEON-NEXT:    rbit x8, x2
-; CHECK-NEON-NEXT:    rbit x9, x0
-; CHECK-NEON-NEXT:    and x7, x2, #0x2
-; CHECK-NEON-NEXT:    and x18, x2, #0x1
-; CHECK-NEON-NEXT:    and x4, x2, #0x4
-; CHECK-NEON-NEXT:    and x5, x2, #0x10
-; CHECK-NEON-NEXT:    and x10, x8, #0x2
-; CHECK-NEON-NEXT:    and x6, x2, #0x80
-; CHECK-NEON-NEXT:    and x17, x2, #0x800
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1944] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x1
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1936] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x4
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1928] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x8
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1920] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x10
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1912] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x20
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1904] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x40
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1888] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x80
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1896] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x100
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1880] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x200
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1864] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x400
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1872] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x800
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1856] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x1000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1832] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x2000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1840] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x4000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1824] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x8000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1848] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x10000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1816] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x20000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1792] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x40000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1784] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x80000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1808] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x100000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1776] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x200000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1800] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x400000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1768] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x800000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1728] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x1000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1760] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x2000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1720] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x4000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1752] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x8000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1736] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x10000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1744] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x20000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1696] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x40000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1656] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x80000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1688] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x100000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1672] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x200000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1680] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x400000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1664] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x800000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1704] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x1000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1712] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x2000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1640] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x4000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1608] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x8000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1584] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x10000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1632] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x20000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1600] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x40000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1616] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x80000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1592] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x100000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1624] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x200000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1648] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x400000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1568] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x800000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1520] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x1000000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1536] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x2000000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1512] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x4000000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1560] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x8000000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1504] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x10000000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1552] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x20000000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1528] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x40000000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1576] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x80000000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1544] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x100000000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1480] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x200000000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1448] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x400000000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1472] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x800000000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1456] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x1000000000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #1488] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x2000000000000000
-; CHECK-NEON-NEXT:    and x8, x8, #0x4000000000000000
-; CHECK-NEON-NEXT:    mul x8, x9, x8
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x8, [sp, #1496] // 8-byte Spill
-; CHECK-NEON-NEXT:    mul x8, x1, x7
-; CHECK-NEON-NEXT:    str x10, [sp, #1464] // 8-byte Spill
-; CHECK-NEON-NEXT:    str x8, [sp, #1032] // 8-byte Spill
-; CHECK-NEON-NEXT:    mul x8, x1, x18
-; CHECK-NEON-NEXT:    mul x18, x0, x18
-; CHECK-NEON-NEXT:    str x8, [sp, #1008] // 8-byte Spill
-; CHECK-NEON-NEXT:    mul x8, x1, x4
-; CHECK-NEON-NEXT:    str x18, [sp, #1128] // 8-byte Spill
-; CHECK-NEON-NEXT:    mul x18, x0, x4
-; CHECK-NEON-NEXT:    str x8, [sp, #992] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x2, #0x8
-; CHECK-NEON-NEXT:    mul x9, x1, x8
-; CHECK-NEON-NEXT:    str x18, [sp, #1120] // 8-byte Spill
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x9, [sp, #976] // 8-byte Spill
-; CHECK-NEON-NEXT:    mul x9, x1, x5
-; CHECK-NEON-NEXT:    str x8, [sp, #1112] // 8-byte Spill
-; CHECK-NEON-NEXT:    mul x8, x0, x5
-; CHECK-NEON-NEXT:    str x9, [sp, #952] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x9, x2, #0x20
-; CHECK-NEON-NEXT:    mul x10, x1, x9
-; CHECK-NEON-NEXT:    str x8, [sp, #1096] // 8-byte Spill
-; CHECK-NEON-NEXT:    mul x8, x0, x9
-; CHECK-NEON-NEXT:    ldr x9, [sp, #1008] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x10, [sp, #928] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x2, #0x40
-; CHECK-NEON-NEXT:    mul x11, x1, x10
-; CHECK-NEON-NEXT:    str x8, [sp, #1072] // 8-byte Spill
-; CHECK-NEON-NEXT:    mul x8, x0, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #976] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x11, [sp, #944] // 8-byte Spill
-; CHECK-NEON-NEXT:    mul x11, x1, x6
-; CHECK-NEON-NEXT:    str x8, [sp, #1104] // 8-byte Spill
-; CHECK-NEON-NEXT:    mul x8, x0, x6
-; CHECK-NEON-NEXT:    str x11, [sp, #936] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x11, x2, #0x100
-; CHECK-NEON-NEXT:    mul x12, x1, x11
-; CHECK-NEON-NEXT:    str x8, [sp, #1064] // 8-byte Spill
-; CHECK-NEON-NEXT:    mul x8, x0, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #928] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x12, [sp, #920] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x12, x2, #0x200
+; CHECK-NEON-NEXT:    and x14, x8, #0x1111111111111111
+; CHECK-NEON-NEXT:    mul x16, x15, x14
+; CHECK-NEON-NEXT:    str x16, [sp, #216] // 8-byte Spill
+; CHECK-NEON-NEXT:    and x16, x8, #0x2222222222222222
+; CHECK-NEON-NEXT:    mul x18, x17, x16
+; CHECK-NEON-NEXT:    str x18, [sp, #208] // 8-byte Spill
+; CHECK-NEON-NEXT:    and x18, x8, #0x8888888888888888
+; CHECK-NEON-NEXT:    and x8, x8, #0x4444444444444444
+; CHECK-NEON-NEXT:    mul x7, x4, x18
+; CHECK-NEON-NEXT:    str x7, [sp, #200] // 8-byte Spill
+; CHECK-NEON-NEXT:    mul x7, x13, x8
+; CHECK-NEON-NEXT:    str x7, [sp, #192] // 8-byte Spill
+; CHECK-NEON-NEXT:    mul x7, x15, x18
+; CHECK-NEON-NEXT:    str x7, [sp, #184] // 8-byte Spill
+; CHECK-NEON-NEXT:    mul x7, x17, x14
+; CHECK-NEON-NEXT:    str x7, [sp, #176] // 8-byte Spill
+; CHECK-NEON-NEXT:    mul x7, x4, x8
+; CHECK-NEON-NEXT:    str x7, [sp, #168] // 8-byte Spill
+; CHECK-NEON-NEXT:    mul x7, x13, x16
+; CHECK-NEON-NEXT:    str x7, [sp, #160] // 8-byte Spill
+; CHECK-NEON-NEXT:    mul x7, x15, x16
+; CHECK-NEON-NEXT:    mul x15, x15, x8
+; CHECK-NEON-NEXT:    mul x8, x17, x8
+; CHECK-NEON-NEXT:    stp x15, x7, [sp, #144] // 16-byte Folded Spill
+; CHECK-NEON-NEXT:    str x8, [sp, #128] // 8-byte Spill
+; CHECK-NEON-NEXT:    mul x8, x4, x16
+; CHECK-NEON-NEXT:    str x8, [sp, #136] // 8-byte Spill
+; CHECK-NEON-NEXT:    mul x8, x4, x14
+; CHECK-NEON-NEXT:    str x8, [sp, #104] // 8-byte Spill
+; CHECK-NEON-NEXT:    mul x8, x17, x18
+; CHECK-NEON-NEXT:    str x8, [sp, #120] // 8-byte Spill
+; CHECK-NEON-NEXT:    mul x8, x13, x18
+; CHECK-NEON-NEXT:    str x8, [sp, #96] // 8-byte Spill
+; CHECK-NEON-NEXT:    mul x8, x13, x14
+; CHECK-NEON-NEXT:    str x8, [sp, #112] // 8-byte Spill
+; CHECK-NEON-NEXT:    and x8, x2, #0x4444444444444444
+; CHECK-NEON-NEXT:    and x2, x0, #0x1111111111111111
+; CHECK-NEON-NEXT:    mul x24, x12, x8
+; CHECK-NEON-NEXT:    and x12, x1, #0x8888888888888888
+; CHECK-NEON-NEXT:    and x1, x0, #0x8888888888888888
+; CHECK-NEON-NEXT:    mul x27, x21, x8
+; CHECK-NEON-NEXT:    and x21, x0, #0x4444444444444444
+; CHECK-NEON-NEXT:    mul x14, x12, x8
+; CHECK-NEON-NEXT:    mul x15, x12, x10
+; CHECK-NEON-NEXT:    mul x18, x12, x11
+; CHECK-NEON-NEXT:    mul x4, x12, x6
+; CHECK-NEON-NEXT:    and x12, x3, #0x4444444444444444
+; CHECK-NEON-NEXT:    mul x30, x1, x12
+; CHECK-NEON-NEXT:    eor x18, x20, x18
+; CHECK-NEON-NEXT:    eor x20, x26, x27
+; CHECK-NEON-NEXT:    mul x28, x21, x12
+; CHECK-NEON-NEXT:    eor x4, x22, x4
+; CHECK-NEON-NEXT:    mul x13, x2, x12
+; CHECK-NEON-NEXT:    mul x12, x29, x12
+; CHECK-NEON-NEXT:    mul x19, x5, x8
+; CHECK-NEON-NEXT:    and x5, x3, #0x2222222222222222
+; CHECK-NEON-NEXT:    mul x16, x2, x5
+; CHECK-NEON-NEXT:    stp x13, x12, [sp, #48] // 16-byte Folded Spill
+; CHECK-NEON-NEXT:    and x12, x3, #0x8888888888888888
+; CHECK-NEON-NEXT:    and x3, x3, #0x1111111111111111
+; CHECK-NEON-NEXT:    mul x0, x21, x12
+; CHECK-NEON-NEXT:    eor x15, x19, x15
+; CHECK-NEON-NEXT:    mul x7, x29, x12
 ; CHECK-NEON-NEXT:    mul x13, x1, x12
-; CHECK-NEON-NEXT:    str x8, [sp, #1048] // 8-byte Spill
-; CHECK-NEON-NEXT:    mul x8, x0, x12
-; CHECK-NEON-NEXT:    str x13, [sp, #968] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x13, x2, #0x400
-; CHECK-NEON-NEXT:    mul x14, x1, x13
-; CHECK-NEON-NEXT:    str x8, [sp, #1040] // 8-byte Spill
-; CHECK-NEON-NEXT:    mul x8, x0, x13
-; CHECK-NEON-NEXT:    str x14, [sp, #960] // 8-byte Spill
-; CHECK-NEON-NEXT:    mul x14, x1, x17
-; CHECK-NEON-NEXT:    str x8, [sp, #1088] // 8-byte Spill
-; CHECK-NEON-NEXT:    mul x8, x0, x17
-; CHECK-NEON-NEXT:    str x14, [sp, #848] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x14, x2, #0x1000
-; CHECK-NEON-NEXT:    mul x15, x1, x14
-; CHECK-NEON-NEXT:    str x8, [sp, #1000] // 8-byte Spill
-; CHECK-NEON-NEXT:    mul x8, x0, x14
-; CHECK-NEON-NEXT:    str x15, [sp, #824] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x15, x2, #0x2000
-; CHECK-NEON-NEXT:    mul x16, x1, x15
-; CHECK-NEON-NEXT:    ldr x12, [sp, #824] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x8, [sp, #984] // 8-byte Spill
-; CHECK-NEON-NEXT:    mul x8, x0, x15
-; CHECK-NEON-NEXT:    str x16, [sp, #896] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x16, x2, #0x4000
-; CHECK-NEON-NEXT:    mul x19, x1, x16
-; CHECK-NEON-NEXT:    str x8, [sp, #1024] // 8-byte Spill
-; CHECK-NEON-NEXT:    mul x8, x0, x16
-; CHECK-NEON-NEXT:    str x19, [sp, #888] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x19, x2, #0x8000
-; CHECK-NEON-NEXT:    mul x20, x1, x19
-; CHECK-NEON-NEXT:    str x8, [sp, #1016] // 8-byte Spill
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1032] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x9, x8
-; CHECK-NEON-NEXT:    ldr x9, [sp, #992] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x20, [sp, #904] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x20, x2, #0x10000
-; CHECK-NEON-NEXT:    eor x9, x9, x10
-; CHECK-NEON-NEXT:    mul x21, x1, x20
-; CHECK-NEON-NEXT:    ldr x10, [sp, #952] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x8, x9
-; CHECK-NEON-NEXT:    ldr x9, [sp, #944] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    mul x11, x0, x19
-; CHECK-NEON-NEXT:    eor x9, x10, x9
-; CHECK-NEON-NEXT:    ldr x10, [sp, #936] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x8, x9
-; CHECK-NEON-NEXT:    ldr x9, [sp, #968] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x21, [sp, #880] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x21, x2, #0x20000
-; CHECK-NEON-NEXT:    mul x22, x1, x21
-; CHECK-NEON-NEXT:    str x11, [sp, #1032] // 8-byte Spill
-; CHECK-NEON-NEXT:    ldr x11, [sp, #920] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #848] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x10, x9
-; CHECK-NEON-NEXT:    ldr x10, [sp, #896] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x22, [sp, #840] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x22, x2, #0x40000
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    mul x23, x1, x22
-; CHECK-NEON-NEXT:    eor x10, x11, x10
-; CHECK-NEON-NEXT:    ldr x11, [sp, #960] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #888] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x12, x0, x20
-; CHECK-NEON-NEXT:    eor x8, x8, x9
-; CHECK-NEON-NEXT:    ldr x9, [sp, #904] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    mul x11, x0, x21
-; CHECK-NEON-NEXT:    str x23, [sp, #832] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x23, x2, #0x80000
-; CHECK-NEON-NEXT:    eor x9, x10, x9
-; CHECK-NEON-NEXT:    mul x24, x1, x23
-; CHECK-NEON-NEXT:    ldr x10, [sp, #880] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x8, x9
-; CHECK-NEON-NEXT:    str x12, [sp, #1008] // 8-byte Spill
-; CHECK-NEON-NEXT:    str x11, [sp, #992] // 8-byte Spill
-; CHECK-NEON-NEXT:    ldr x11, [sp, #840] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x24, [sp, #872] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x24, x2, #0x100000
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    mul x25, x1, x24
-; CHECK-NEON-NEXT:    ldr x11, [sp, #832] // 8-byte Reload
-; CHECK-NEON-NEXT:    ldr x9, [sp, #872] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    eor x9, x10, x9
-; CHECK-NEON-NEXT:    str x25, [sp, #864] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x25, x2, #0x200000
-; CHECK-NEON-NEXT:    mul x26, x1, x25
-; CHECK-NEON-NEXT:    str x26, [sp, #912] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x26, x2, #0x400000
-; CHECK-NEON-NEXT:    mul x27, x1, x26
-; CHECK-NEON-NEXT:    str x27, [sp, #760] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x27, x2, #0x800000
-; CHECK-NEON-NEXT:    mul x28, x1, x27
-; CHECK-NEON-NEXT:    ldr x11, [sp, #760] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x28, [sp, #736] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x28, x2, #0x1000000
-; CHECK-NEON-NEXT:    mul x29, x1, x28
-; CHECK-NEON-NEXT:    ldr x12, [sp, #736] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    mul x12, x0, x22
-; CHECK-NEON-NEXT:    str x29, [sp, #784] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x29, x2, #0x2000000
-; CHECK-NEON-NEXT:    mul x30, x1, x29
-; CHECK-NEON-NEXT:    ldr x10, [sp, #784] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x12, [sp, #976] // 8-byte Spill
-; CHECK-NEON-NEXT:    eor x10, x11, x10
-; CHECK-NEON-NEXT:    ldr x11, [sp, #864] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x11
-; CHECK-NEON-NEXT:    str x30, [sp, #776] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x4000000
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    ldr x11, [sp, #776] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    mul x11, x0, x23
-; CHECK-NEON-NEXT:    str x30, [sp, #800] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x4000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1144] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x8000000
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    str x11, [sp, #968] // 8-byte Spill
-; CHECK-NEON-NEXT:    ldr x11, [sp, #912] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #800] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x8, x9
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    str x30, [sp, #792] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x8000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1152] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x10000000
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    str x30, [sp, #816] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x10000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1160] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x20000000
-; CHECK-NEON-NEXT:    ldr x9, [sp, #816] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    str x30, [sp, #728] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x20000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1168] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x40000000
-; CHECK-NEON-NEXT:    ldr x11, [sp, #728] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    str x30, [sp, #696] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x40000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1176] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x80000000
-; CHECK-NEON-NEXT:    ldr x12, [sp, #696] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #792] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x12
-; CHECK-NEON-NEXT:    eor x9, x10, x9
-; CHECK-NEON-NEXT:    str x30, [sp, #688] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x80000000
-; CHECK-NEON-NEXT:    eor x8, x8, x9
-; CHECK-NEON-NEXT:    str x30, [sp, #1184] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x100000000
-; CHECK-NEON-NEXT:    ldr x12, [sp, #688] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    mul x12, x0, x24
-; CHECK-NEON-NEXT:    str x30, [sp, #744] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x100000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1192] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x200000000
-; CHECK-NEON-NEXT:    ldr x10, [sp, #744] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    str x12, [sp, #960] // 8-byte Spill
-; CHECK-NEON-NEXT:    eor x10, x11, x10
-; CHECK-NEON-NEXT:    str x30, [sp, #720] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x200000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1200] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x400000000
-; CHECK-NEON-NEXT:    ldr x11, [sp, #720] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    str x30, [sp, #768] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x400000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1208] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x800000000
-; CHECK-NEON-NEXT:    ldr x9, [sp, #768] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    eor x9, x10, x9
-; CHECK-NEON-NEXT:    str x30, [sp, #808] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x800000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1216] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x1000000000
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    str x30, [sp, #856] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x1000000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1224] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x2000000000
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    str x30, [sp, #648] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x2000000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1232] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x4000000000
-; CHECK-NEON-NEXT:    ldr x11, [sp, #648] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    str x30, [sp, #632] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x4000000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1240] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x8000000000
-; CHECK-NEON-NEXT:    ldr x12, [sp, #632] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    mul x12, x0, x25
-; CHECK-NEON-NEXT:    str x30, [sp, #664] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x8000000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1248] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x10000000000
-; CHECK-NEON-NEXT:    ldr x10, [sp, #664] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    str x12, [sp, #952] // 8-byte Spill
-; CHECK-NEON-NEXT:    eor x10, x11, x10
-; CHECK-NEON-NEXT:    str x30, [sp, #640] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x10000000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1256] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x20000000000
-; CHECK-NEON-NEXT:    ldr x11, [sp, #640] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    str x30, [sp, #680] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x20000000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1264] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x40000000000
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    str x30, [sp, #672] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x40000000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1272] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x80000000000
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    str x30, [sp, #712] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x80000000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1280] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x100000000000
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    str x30, [sp, #704] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x100000000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1288] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x200000000000
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    str x30, [sp, #752] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x200000000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1296] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x400000000000
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    str x30, [sp, #520] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x400000000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1304] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x800000000000
-; CHECK-NEON-NEXT:    ldr x11, [sp, #520] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    str x30, [sp, #504] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x800000000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1312] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x1000000000000
-; CHECK-NEON-NEXT:    ldr x12, [sp, #504] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    mul x12, x0, x26
-; CHECK-NEON-NEXT:    str x30, [sp, #560] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x1000000000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1320] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x2000000000000
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    str x12, [sp, #944] // 8-byte Spill
-; CHECK-NEON-NEXT:    ldr x12, [sp, #808] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #680] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #560] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x30, [sp, #552] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x2000000000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1328] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x4000000000000
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    ldr x12, [sp, #672] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #552] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    mul x12, x0, x27
-; CHECK-NEON-NEXT:    str x30, [sp, #584] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x4000000000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1336] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x8000000000000
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    str x12, [sp, #936] // 8-byte Spill
-; CHECK-NEON-NEXT:    ldr x12, [sp, #856] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #712] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x30, [sp, #576] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x8000000000000
-; CHECK-NEON-NEXT:    eor x8, x8, x9
-; CHECK-NEON-NEXT:    str x30, [sp, #1344] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x10000000000000
-; CHECK-NEON-NEXT:    eor x10, x10, x12
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    ldr x12, [sp, #584] // 8-byte Reload
-; CHECK-NEON-NEXT:    ldr x9, [sp, #752] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #704] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #576] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x30, [sp, #608] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x10000000000000
-; CHECK-NEON-NEXT:    eor x9, x10, x9
-; CHECK-NEON-NEXT:    str x30, [sp, #1352] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x20000000000000
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    ldr x10, [sp, #608] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x8, x9
-; CHECK-NEON-NEXT:    mul x12, x0, x28
-; CHECK-NEON-NEXT:    eor x10, x11, x10
-; CHECK-NEON-NEXT:    str x30, [sp, #592] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x20000000000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1360] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x40000000000000
-; CHECK-NEON-NEXT:    ldr x11, [sp, #592] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    str x12, [sp, #928] // 8-byte Spill
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    str x30, [sp, #624] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x40000000000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1368] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x80000000000000
-; CHECK-NEON-NEXT:    ldr x9, [sp, #624] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    eor x9, x10, x9
-; CHECK-NEON-NEXT:    str x30, [sp, #616] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x80000000000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1376] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x100000000000000
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    str x30, [sp, #528] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x100000000000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1384] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x200000000000000
-; CHECK-NEON-NEXT:    ldr x11, [sp, #528] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    str x30, [sp, #512] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x200000000000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1392] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x400000000000000
-; CHECK-NEON-NEXT:    ldr x12, [sp, #512] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    mul x12, x0, x29
-; CHECK-NEON-NEXT:    str x30, [sp, #544] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x400000000000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1400] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x800000000000000
-; CHECK-NEON-NEXT:    ldr x10, [sp, #544] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    str x12, [sp, #920] // 8-byte Spill
-; CHECK-NEON-NEXT:    eor x10, x11, x10
-; CHECK-NEON-NEXT:    ldr x11, [sp, #616] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x11
-; CHECK-NEON-NEXT:    eor x21, x8, x9
-; CHECK-NEON-NEXT:    str x30, [sp, #536] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x800000000000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1408] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x1000000000000000
-; CHECK-NEON-NEXT:    ldr x11, [sp, #536] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #1144] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x11, x0, x11
-; CHECK-NEON-NEXT:    str x30, [sp, #568] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x1000000000000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1416] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x2000000000000000
-; CHECK-NEON-NEXT:    ldr x8, [sp, #568] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    eor x8, x10, x8
-; CHECK-NEON-NEXT:    str x11, [sp, #1144] // 8-byte Spill
-; CHECK-NEON-NEXT:    str x30, [sp, #600] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x2000000000000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1424] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x4000000000000000
-; CHECK-NEON-NEXT:    mul x30, x1, x30
-; CHECK-NEON-NEXT:    str x30, [sp, #656] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x30, x2, #0x4000000000000000
-; CHECK-NEON-NEXT:    and x2, x2, #0x8000000000000000
-; CHECK-NEON-NEXT:    str x30, [sp, #1432] // 8-byte Spill
-; CHECK-NEON-NEXT:    mul x30, x1, x2
-; CHECK-NEON-NEXT:    and x1, x3, #0x2
-; CHECK-NEON-NEXT:    str x2, [sp, #1440] // 8-byte Spill
-; CHECK-NEON-NEXT:    mul x2, x0, x1
-; CHECK-NEON-NEXT:    and x1, x3, #0x1
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    stp x1, x2, [sp, #488] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x4
-; CHECK-NEON-NEXT:    mul x2, x0, x1
-; CHECK-NEON-NEXT:    and x1, x3, #0x8
-; CHECK-NEON-NEXT:    ldp x10, x9, [sp, #488] // 16-byte Folded Reload
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    eor x9, x10, x9
-; CHECK-NEON-NEXT:    str x1, [sp, #456] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x10
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    ldr x11, [sp, #456] // 8-byte Reload
-; CHECK-NEON-NEXT:    stp x2, x1, [sp, #472] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x20
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    ldr x10, [sp, #472] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #1152] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #480] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x11, x0, x11
-; CHECK-NEON-NEXT:    str x1, [sp, #464] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x40
-; CHECK-NEON-NEXT:    mul x2, x0, x1
-; CHECK-NEON-NEXT:    and x1, x3, #0x80
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    str x11, [sp, #1152] // 8-byte Spill
-; CHECK-NEON-NEXT:    ldr x11, [sp, #600] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x8, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #464] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x1, [sp, #424] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x100
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    str x1, [sp, #384] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x200
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    ldr x12, [sp, #384] // 8-byte Reload
-; CHECK-NEON-NEXT:    stp x1, x2, [sp, #440] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x400
-; CHECK-NEON-NEXT:    mul x2, x0, x1
-; CHECK-NEON-NEXT:    and x1, x3, #0x800
-; CHECK-NEON-NEXT:    ldr x11, [sp, #448] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #424] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #440] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #1160] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x11, x10
-; CHECK-NEON-NEXT:    mul x12, x0, x12
-; CHECK-NEON-NEXT:    str x1, [sp, #392] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x1000
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    str x12, [sp, #1160] // 8-byte Spill
-; CHECK-NEON-NEXT:    ldr x12, [sp, #656] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x1, [sp, #376] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x2000
-; CHECK-NEON-NEXT:    eor x8, x8, x12
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    ldr x12, [sp, #376] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x16, x8, x30
-; CHECK-NEON-NEXT:    stp x1, x2, [sp, #408] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x4000
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    ldr x11, [sp, #416] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #392] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #408] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #1168] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x1, [sp, #400] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x8000
-; CHECK-NEON-NEXT:    eor x10, x11, x10
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    mul x12, x0, x12
-; CHECK-NEON-NEXT:    str x1, [sp, #432] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x10000
-; CHECK-NEON-NEXT:    mul x2, x0, x1
-; CHECK-NEON-NEXT:    and x1, x3, #0x20000
-; CHECK-NEON-NEXT:    str x12, [sp, #1168] // 8-byte Spill
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    str x1, [sp, #328] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x40000
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    ldr x12, [sp, #328] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x1, [sp, #320] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x80000
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    stp x1, x2, [sp, #352] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x100000
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    ldr x11, [sp, #360] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #400] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #320] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x1, [sp, #344] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x200000
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #1176] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x12, x0, x12
-; CHECK-NEON-NEXT:    str x1, [sp, #368] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x400000
-; CHECK-NEON-NEXT:    mul x2, x0, x1
-; CHECK-NEON-NEXT:    and x1, x3, #0x800000
-; CHECK-NEON-NEXT:    str x12, [sp, #1176] // 8-byte Spill
-; CHECK-NEON-NEXT:    ldr x12, [sp, #432] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    eor x10, x10, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #352] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #368] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    str x1, [sp, #232] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x1000000
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    ldr x13, [sp, #232] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x1, [sp, #224] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x2000000
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    stp x2, x1, [sp, #280] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x4000000
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    ldr x12, [sp, #280] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x12, x12, x13
-; CHECK-NEON-NEXT:    ldr x13, [sp, #344] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x13
-; CHECK-NEON-NEXT:    ldr x13, [sp, #224] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x1, [sp, #272] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x8000000
-; CHECK-NEON-NEXT:    eor x10, x11, x10
-; CHECK-NEON-NEXT:    mul x2, x0, x1
-; CHECK-NEON-NEXT:    and x1, x3, #0x10000000
-; CHECK-NEON-NEXT:    eor x12, x12, x13
-; CHECK-NEON-NEXT:    ldr x13, [sp, #1184] // 8-byte Reload
-; CHECK-NEON-NEXT:    ldr x11, [sp, #288] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x10
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    eor x11, x12, x11
-; CHECK-NEON-NEXT:    ldr x12, [sp, #272] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x13, x0, x13
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    stp x1, x2, [sp, #304] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x20000000
-; CHECK-NEON-NEXT:    mul x2, x0, x1
-; CHECK-NEON-NEXT:    and x1, x3, #0x40000000
-; CHECK-NEON-NEXT:    ldr x10, [sp, #312] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x13, [sp, #1184] // 8-byte Spill
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    eor x10, x11, x10
-; CHECK-NEON-NEXT:    stp x1, x2, [sp, #168] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x80000000
-; CHECK-NEON-NEXT:    mul x2, x0, x1
-; CHECK-NEON-NEXT:    and x1, x3, #0x100000000
-; CHECK-NEON-NEXT:    ldp x13, x12, [sp, #168] // 16-byte Folded Reload
-; CHECK-NEON-NEXT:    mul x1, x0, x1
+; CHECK-NEON-NEXT:    mul x12, x2, x12
+; CHECK-NEON-NEXT:    mul x17, x1, x5
+; CHECK-NEON-NEXT:    mul x25, x29, x5
+; CHECK-NEON-NEXT:    str x12, [sp, #40] // 8-byte Spill
+; CHECK-NEON-NEXT:    mul x12, x21, x5
+; CHECK-NEON-NEXT:    eor x17, x28, x17
+; CHECK-NEON-NEXT:    ldp x28, x27, [sp, #272] // 16-byte Folded Reload
+; CHECK-NEON-NEXT:    mul x5, x21, x3
+; CHECK-NEON-NEXT:    mul x23, x1, x3
+; CHECK-NEON-NEXT:    stp x12, x13, [sp, #16] // 16-byte Folded Spill
+; CHECK-NEON-NEXT:    mul x12, x29, x3
+; CHECK-NEON-NEXT:    mul x13, x2, x3
+; CHECK-NEON-NEXT:    mul x3, x29, x6
+; CHECK-NEON-NEXT:    eor x12, x16, x12
+; CHECK-NEON-NEXT:    eor x16, x0, x30
+; CHECK-NEON-NEXT:    eor x12, x12, x16
+; CHECK-NEON-NEXT:    mul x0, x1, x10
+; CHECK-NEON-NEXT:    eor x13, x13, x7
+; CHECK-NEON-NEXT:    and x12, x12, #0x2222222222222222
+; CHECK-NEON-NEXT:    eor x13, x13, x17
+; CHECK-NEON-NEXT:    str x3, [sp, #80] // 8-byte Spill
+; CHECK-NEON-NEXT:    mul x3, x29, x11
+; CHECK-NEON-NEXT:    and x13, x13, #0x1111111111111111
+; CHECK-NEON-NEXT:    orr x12, x13, x12
+; CHECK-NEON-NEXT:    ldp x17, x13, [sp, #40] // 16-byte Folded Reload
+; CHECK-NEON-NEXT:    eor x13, x13, x25
+; CHECK-NEON-NEXT:    ldp x26, x25, [sp, #288] // 16-byte Folded Reload
+; CHECK-NEON-NEXT:    str x3, [sp, #72] // 8-byte Spill
+; CHECK-NEON-NEXT:    mul x3, x29, x10
+; CHECK-NEON-NEXT:    str x3, [sp, #64] // 8-byte Spill
+; CHECK-NEON-NEXT:    mul x3, x29, x8
+; CHECK-NEON-NEXT:    str x3, [sp, #32] // 8-byte Spill
+; CHECK-NEON-NEXT:    ldr x3, [sp, #248] // 8-byte Reload
+; CHECK-NEON-NEXT:    eor x3, x3, x9
+; CHECK-NEON-NEXT:    ldr x9, [sp, #240] // 8-byte Reload
+; CHECK-NEON-NEXT:    eor x9, x9, x14
+; CHECK-NEON-NEXT:    mul x14, x1, x8
+; CHECK-NEON-NEXT:    eor x9, x3, x9
+; CHECK-NEON-NEXT:    ldp x29, x3, [sp, #224] // 16-byte Folded Reload
+; CHECK-NEON-NEXT:    and x9, x9, #0x2222222222222222
+; CHECK-NEON-NEXT:    eor x3, x29, x3
+; CHECK-NEON-NEXT:    ldp x29, x30, [sp, #256] // 16-byte Folded Reload
+; CHECK-NEON-NEXT:    eor x3, x3, x15
+; CHECK-NEON-NEXT:    ldr x15, [sp, #88] // 8-byte Reload
+; CHECK-NEON-NEXT:    and x3, x3, #0x1111111111111111
+; CHECK-NEON-NEXT:    eor x19, x24, x15
+; CHECK-NEON-NEXT:    orr x9, x3, x9
+; CHECK-NEON-NEXT:    eor x3, x20, x4
+; CHECK-NEON-NEXT:    eor x18, x19, x18
+; CHECK-NEON-NEXT:    and x3, x3, #0x8888888888888888
+; CHECK-NEON-NEXT:    mul x15, x21, x8
+; CHECK-NEON-NEXT:    and x18, x18, #0x4444444444444444
+; CHECK-NEON-NEXT:    orr x18, x18, x3
+; CHECK-NEON-NEXT:    mul x4, x2, x10
+; CHECK-NEON-NEXT:    orr x9, x9, x18
+; CHECK-NEON-NEXT:    ldp x18, x16, [sp, #16] // 16-byte Folded Reload
+; CHECK-NEON-NEXT:    mul x10, x21, x10
+; CHECK-NEON-NEXT:    eor x16, x5, x16
+; CHECK-NEON-NEXT:    ldp x5, x3, [sp, #208] // 16-byte Folded Reload
+; CHECK-NEON-NEXT:    eor x13, x13, x16
+; CHECK-NEON-NEXT:    ldr x16, [sp, #56] // 8-byte Reload
+; CHECK-NEON-NEXT:    eor x18, x18, x23
+; CHECK-NEON-NEXT:    and x13, x13, #0x4444444444444444
+; CHECK-NEON-NEXT:    mul x8, x2, x8
+; CHECK-NEON-NEXT:    eor x3, x5, x3
+; CHECK-NEON-NEXT:    ldp x7, x5, [sp, #192] // 16-byte Folded Reload
+; CHECK-NEON-NEXT:    eor x16, x17, x16
+; CHECK-NEON-NEXT:    mul x17, x21, x11
+; CHECK-NEON-NEXT:    eor x16, x16, x18
+; CHECK-NEON-NEXT:    ldp x24, x23, [sp, #304] // 16-byte Folded Reload
+; CHECK-NEON-NEXT:    eor x5, x5, x7
+; CHECK-NEON-NEXT:    ldp x19, x7, [sp, #176] // 16-byte Folded Reload
+; CHECK-NEON-NEXT:    eor x18, x3, x5
+; CHECK-NEON-NEXT:    ldr x5, [sp, #152] // 8-byte Reload
+; CHECK-NEON-NEXT:    and x16, x16, #0x8888888888888888
+; CHECK-NEON-NEXT:    and x18, x18, #0x2222222222222222
+; CHECK-NEON-NEXT:    orr x13, x13, x16
+; CHECK-NEON-NEXT:    eor x7, x19, x7
+; CHECK-NEON-NEXT:    ldp x20, x19, [sp, #160] // 16-byte Folded Reload
+; CHECK-NEON-NEXT:    orr x12, x12, x13
+; CHECK-NEON-NEXT:    eor x14, x17, x14
+; CHECK-NEON-NEXT:    eor x9, x12, x9
+; CHECK-NEON-NEXT:    eor x19, x19, x20
+; CHECK-NEON-NEXT:    mul x20, x2, x6
+; CHECK-NEON-NEXT:    eor x3, x7, x19
+; CHECK-NEON-NEXT:    ldr x7, [sp, #128] // 8-byte Reload
+; CHECK-NEON-NEXT:    and x3, x3, #0x1111111111111111
+; CHECK-NEON-NEXT:    mul x2, x2, x11
+; CHECK-NEON-NEXT:    eor x5, x7, x5
+; CHECK-NEON-NEXT:    ldp x19, x7, [sp, #96] // 16-byte Folded Reload
+; CHECK-NEON-NEXT:    orr x16, x3, x18
+; CHECK-NEON-NEXT:    mul x11, x1, x11
+; CHECK-NEON-NEXT:    eor x7, x7, x19
+; CHECK-NEON-NEXT:    mul x19, x21, x6
+; CHECK-NEON-NEXT:    ldp x22, x21, [sp, #112] // 16-byte Folded Reload
+; CHECK-NEON-NEXT:    eor x5, x5, x7
+; CHECK-NEON-NEXT:    ldr x7, [sp, #144] // 8-byte Reload
+; CHECK-NEON-NEXT:    and x5, x5, #0x4444444444444444
+; CHECK-NEON-NEXT:    eor x7, x21, x7
+; CHECK-NEON-NEXT:    ldr x21, [sp, #136] // 8-byte Reload
+; CHECK-NEON-NEXT:    eor x11, x19, x11
+; CHECK-NEON-NEXT:    eor x21, x21, x22
+; CHECK-NEON-NEXT:    eor x7, x7, x21
+; CHECK-NEON-NEXT:    ldp x22, x21, [sp, #320] // 16-byte Folded Reload
+; CHECK-NEON-NEXT:    and x7, x7, #0x8888888888888888
+; CHECK-NEON-NEXT:    orr x18, x5, x7
+; CHECK-NEON-NEXT:    orr x13, x16, x18
+; CHECK-NEON-NEXT:    ldr x18, [sp, #80] // 8-byte Reload
+; CHECK-NEON-NEXT:    mul x16, x1, x6
+; CHECK-NEON-NEXT:    rbit x13, x13
+; CHECK-NEON-NEXT:    eor x18, x4, x18
+; CHECK-NEON-NEXT:    eor x12, x18, x14
+; CHECK-NEON-NEXT:    eor x1, x9, x13, lsr #1
+; CHECK-NEON-NEXT:    eor x13, x15, x0
+; CHECK-NEON-NEXT:    and x9, x12, #0x2222222222222222
+; CHECK-NEON-NEXT:    ldp x14, x12, [sp, #64] // 16-byte Folded Reload
+; CHECK-NEON-NEXT:    eor x10, x10, x16
+; CHECK-NEON-NEXT:    eor x8, x8, x14
+; CHECK-NEON-NEXT:    ldr x14, [sp, #32] // 8-byte Reload
+; CHECK-NEON-NEXT:    eor x12, x20, x12
 ; CHECK-NEON-NEXT:    eor x12, x12, x13
-; CHECK-NEON-NEXT:    ldr x13, [sp, #1192] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x13, x0, x13
-; CHECK-NEON-NEXT:    stp x1, x2, [sp, #200] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x200000000
-; CHECK-NEON-NEXT:    mul x2, x0, x1
-; CHECK-NEON-NEXT:    and x1, x3, #0x400000000
-; CHECK-NEON-NEXT:    ldr x11, [sp, #208] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    eor x11, x12, x11
-; CHECK-NEON-NEXT:    ldr x12, [sp, #304] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x13, [sp, #1192] // 8-byte Spill
-; CHECK-NEON-NEXT:    eor x10, x10, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #200] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x9, x10
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #1200] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x1, [sp, #216] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x800000000
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    ldr x10, [sp, #216] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x12, x0, x12
-; CHECK-NEON-NEXT:    stp x2, x1, [sp, #256] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x1000000000
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    ldr x9, [sp, #256] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x12, [sp, #1200] // 8-byte Spill
-; CHECK-NEON-NEXT:    eor x9, x11, x9
-; CHECK-NEON-NEXT:    eor x9, x9, x10
-; CHECK-NEON-NEXT:    str x1, [sp, #336] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x2000000000
-; CHECK-NEON-NEXT:    mul x2, x0, x1
-; CHECK-NEON-NEXT:    and x1, x3, #0x4000000000
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    str x1, [sp, #136] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x8000000000
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    ldr x11, [sp, #136] // 8-byte Reload
-; CHECK-NEON-NEXT:    stp x1, x2, [sp, #152] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x10000000000
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    ldr x10, [sp, #160] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #1208] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x11, x0, x11
-; CHECK-NEON-NEXT:    str x1, [sp, #144] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x20000000000
-; CHECK-NEON-NEXT:    mul x2, x0, x1
-; CHECK-NEON-NEXT:    and x1, x3, #0x40000000000
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    str x11, [sp, #1208] // 8-byte Spill
-; CHECK-NEON-NEXT:    ldr x11, [sp, #264] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #152] // 8-byte Reload
-; CHECK-NEON-NEXT:    stp x1, x2, [sp, #184] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x80000000000
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    mul x2, x0, x1
-; CHECK-NEON-NEXT:    and x1, x3, #0x100000000000
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    stp x1, x2, [sp, #240] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x200000000000
-; CHECK-NEON-NEXT:    and x2, x3, #0x1000000000000
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    mul x2, x0, x2
-; CHECK-NEON-NEXT:    str x1, [sp, #296] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x400000000000
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    str x1, [sp, #40] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x800000000000
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    ldr x11, [sp, #40] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x1, [sp, #16] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x2000000000000
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    ldr x12, [sp, #16] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #144] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x2
-; CHECK-NEON-NEXT:    eor x10, x10, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #1216] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x1, [sp, #56] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x4000000000000
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    mul x23, x0, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #336] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #192] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x1, [sp, #48] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x8000000000000
-; CHECK-NEON-NEXT:    eor x8, x8, x9
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    eor x10, x10, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #56] // 8-byte Reload
-; CHECK-NEON-NEXT:    ldr x9, [sp, #248] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #184] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #48] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x1, [sp, #96] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x10000000000000
-; CHECK-NEON-NEXT:    eor x9, x10, x9
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    ldr x10, [sp, #96] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #1224] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x11, x10
-; CHECK-NEON-NEXT:    ldr x11, [sp, #240] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x24, x0, x12
-; CHECK-NEON-NEXT:    eor x9, x9, x11
-; CHECK-NEON-NEXT:    str x1, [sp, #88] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x20000000000000
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    ldr x11, [sp, #88] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #1232] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x25, x0, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #296] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x1, [sp, #112] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x40000000000000
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    eor x9, x9, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #112] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x8, x9
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    str x1, [sp, #104] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x80000000000000
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    str x1, [sp, #120] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x100000000000000
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    ldr x9, [sp, #120] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x1, [sp, #80] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x200000000000000
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    ldr x11, [sp, #80] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x1, [sp, #32] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x400000000000000
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    ldr x12, [sp, #32] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #104] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x12
-; CHECK-NEON-NEXT:    str x1, [sp, #24] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x800000000000000
-; CHECK-NEON-NEXT:    eor x9, x10, x9
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    ldr x12, [sp, #24] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x15, x8, x9
-; CHECK-NEON-NEXT:    ldr x9, [sp, #1248] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #1240] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x27, x0, x9
-; CHECK-NEON-NEXT:    ldr x9, [sp, #1944] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x1, [sp, #72] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x1000000000000000
-; CHECK-NEON-NEXT:    mul x26, x0, x12
-; CHECK-NEON-NEXT:    ldr x10, [sp, #72] // 8-byte Reload
-; CHECK-NEON-NEXT:    ldr x12, [sp, #1904] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    eor x10, x11, x10
-; CHECK-NEON-NEXT:    ldr x11, [sp, #1920] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x1, [sp, #64] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x2000000000000000
-; CHECK-NEON-NEXT:    ldr x8, [sp, #64] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    eor x8, x10, x8
-; CHECK-NEON-NEXT:    ldr x10, [sp, #1936] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x10, x9
-; CHECK-NEON-NEXT:    ldr x10, [sp, #1928] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x1, [sp, #128] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x4000000000000000
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #1912] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    eor x9, x9, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #1888] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #1792] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x11, x10
-; CHECK-NEON-NEXT:    ldr x11, [sp, #1256] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #1880] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x28, x0, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #128] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x1, [sp, #1056] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x1, x3, #0x8000000000000000
-; CHECK-NEON-NEXT:    eor x14, x8, x11
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1896] // 8-byte Reload
-; CHECK-NEON-NEXT:    ldr x11, [sp, #1832] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x1, x0, x1
-; CHECK-NEON-NEXT:    eor x10, x8, x10
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1864] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1856] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x8, x11
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1264] // 8-byte Reload
-; CHECK-NEON-NEXT:    str x1, [sp, #1080] // 8-byte Spill
-; CHECK-NEON-NEXT:    mul x1, x0, x7
-; CHECK-NEON-NEXT:    mul x29, x0, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1872] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1840] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x10
-; CHECK-NEON-NEXT:    str x1, [sp, #1136] // 8-byte Spill
-; CHECK-NEON-NEXT:    eor x11, x11, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1816] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x12, x8, x12
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1824] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1784] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x12, x12, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1272] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x30, x0, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1848] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x11, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1808] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x10
-; CHECK-NEON-NEXT:    eor x11, x12, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1776] // 8-byte Reload
-; CHECK-NEON-NEXT:    ldr x12, [sp, #1728] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1768] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x12, x8, x12
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1280] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x22, x0, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1800] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x11, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1760] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x10
-; CHECK-NEON-NEXT:    eor x11, x12, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1720] // 8-byte Reload
-; CHECK-NEON-NEXT:    ldr x12, [sp, #1656] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1696] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x12, x8, x12
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1288] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x20, x0, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1752] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x11, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1688] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x12, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1736] // 8-byte Reload
-; CHECK-NEON-NEXT:    ldr x12, [sp, #1608] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1672] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1296] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x19, x0, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1744] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1680] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x10
-; CHECK-NEON-NEXT:    eor x11, x11, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1640] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x12, x8, x12
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1664] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1584] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x12, x12, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1304] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x7, x0, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1704] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x11, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1632] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x12, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1600] // 8-byte Reload
-; CHECK-NEON-NEXT:    ldr x12, [sp, #1520] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1568] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x12, x8, x12
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1312] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x6, x0, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1712] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1616] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x13, x9, x10
-; CHECK-NEON-NEXT:    ldr x9, [sp, #1448] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1536] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x6, x7, x6
-; CHECK-NEON-NEXT:    ldr x7, [sp, #1424] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x12, x12, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1592] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x7, x0, x7
-; CHECK-NEON-NEXT:    eor x11, x11, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1512] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x12, x12, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1320] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x5, x0, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1624] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x11, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1560] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x12, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1504] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1480] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x8, x9
-; CHECK-NEON-NEXT:    ldr x9, [sp, #1328] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x4, x0, x9
-; CHECK-NEON-NEXT:    ldr x9, [sp, #1648] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x9
-; CHECK-NEON-NEXT:    ldr x9, [sp, #1552] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x13, x13, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #1576] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x9
-; CHECK-NEON-NEXT:    ldr x9, [sp, #1472] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x8, x9
-; CHECK-NEON-NEXT:    ldr x9, [sp, #1528] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x11, x9
-; CHECK-NEON-NEXT:    ldr x11, [sp, #1456] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #1488] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x8, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #1336] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x8, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #1544] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x3, x0, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #992] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #1464] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x13, x9
-; CHECK-NEON-NEXT:    eor x8, x8, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #1344] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x1, x0, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #1056] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x14, x14, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #1496] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x8, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #1080] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x9, x8
-; CHECK-NEON-NEXT:    ldr x9, [sp, #1352] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x13, x14, x10
-; CHECK-NEON-NEXT:    eor x14, x21, x16
-; CHECK-NEON-NEXT:    rbit x8, x8
-; CHECK-NEON-NEXT:    mul x18, x0, x9
-; CHECK-NEON-NEXT:    eor x13, x15, x13
-; CHECK-NEON-NEXT:    ldr x9, [sp, #1360] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x13, x13, x14
-; CHECK-NEON-NEXT:    ldr x10, [sp, #1112] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x17, x0, x9
-; CHECK-NEON-NEXT:    eor x2, x13, x8, lsr #1
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1136] // 8-byte Reload
-; CHECK-NEON-NEXT:    ldr x9, [sp, #1128] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x9, x8
-; CHECK-NEON-NEXT:    ldr x9, [sp, #1120] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x13, x9, x10
-; CHECK-NEON-NEXT:    ldr x9, [sp, #1096] // 8-byte Reload
-; CHECK-NEON-NEXT:    ldr x10, [sp, #1072] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x13, x8, x13
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1104] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x14, x9, x10
-; CHECK-NEON-NEXT:    ldr x9, [sp, #1368] // 8-byte Reload
-; CHECK-NEON-NEXT:    ldr x10, [sp, #984] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x14, x14, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1064] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x21, x0, x9
-; CHECK-NEON-NEXT:    ldr x9, [sp, #1048] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x13, x13, x14
-; CHECK-NEON-NEXT:    eor x8, x8, x9
-; CHECK-NEON-NEXT:    ldr x9, [sp, #1040] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x8, x9
-; CHECK-NEON-NEXT:    ldr x9, [sp, #1000] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #1376] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x16, x0, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #1088] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x8, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #1024] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x13, x8
-; CHECK-NEON-NEXT:    eor x9, x9, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #1016] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #1008] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x14, x10, x11
-; CHECK-NEON-NEXT:    ldr x10, [sp, #1384] // 8-byte Reload
-; CHECK-NEON-NEXT:    ldr x11, [sp, #936] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x15, x0, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #1032] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #976] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x13, x14, x10
-; CHECK-NEON-NEXT:    eor x10, x8, x9
-; CHECK-NEON-NEXT:    ldr x8, [sp, #968] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x13, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1392] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x13, x0, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #960] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #944] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x14, x8, x11
-; CHECK-NEON-NEXT:    ldr x8, [sp, #1168] // 8-byte Reload
-; CHECK-NEON-NEXT:    ldr x11, [sp, #1176] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x13, x15, x13
-; CHECK-NEON-NEXT:    eor x8, x8, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #928] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x14, x14, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #1184] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x8, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #1400] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x12, x0, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #952] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #920] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x10, x9
-; CHECK-NEON-NEXT:    ldr x10, [sp, #1152] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x14, x14, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #1192] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x12, x13, x12
-; CHECK-NEON-NEXT:    ldr x13, [sp, #1440] // 8-byte Reload
 ; CHECK-NEON-NEXT:    eor x8, x8, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #1144] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x13, x0, x13
-; CHECK-NEON-NEXT:    eor x14, x14, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #1200] // 8-byte Reload
+; CHECK-NEON-NEXT:    eor x14, x2, x14
+; CHECK-NEON-NEXT:    and x11, x12, #0x1111111111111111
+; CHECK-NEON-NEXT:    and x8, x8, #0x4444444444444444
 ; CHECK-NEON-NEXT:    eor x10, x14, x10
-; CHECK-NEON-NEXT:    ldr x14, [sp, #1208] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x8, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #1408] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x8, x14
-; CHECK-NEON-NEXT:    ldr x14, [sp, #1160] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x8, x23
-; CHECK-NEON-NEXT:    mul x11, x0, x11
-; CHECK-NEON-NEXT:    eor x10, x10, x14
-; CHECK-NEON-NEXT:    ldr x14, [sp, #1416] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x8, x24
-; CHECK-NEON-NEXT:    eor x9, x9, x10
-; CHECK-NEON-NEXT:    eor x10, x25, x26
-; CHECK-NEON-NEXT:    eor x10, x10, x27
-; CHECK-NEON-NEXT:    eor x8, x9, x8
-; CHECK-NEON-NEXT:    mul x14, x0, x14
-; CHECK-NEON-NEXT:    eor x9, x10, x28
-; CHECK-NEON-NEXT:    eor x10, x6, x5
-; CHECK-NEON-NEXT:    eor x10, x10, x4
-; CHECK-NEON-NEXT:    ldr x4, [sp, #1432] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x29
-; CHECK-NEON-NEXT:    eor x10, x10, x3
-; CHECK-NEON-NEXT:    eor x9, x9, x30
-; CHECK-NEON-NEXT:    eor x11, x12, x11
-; CHECK-NEON-NEXT:    mul x4, x0, x4
-; CHECK-NEON-NEXT:    eor x10, x10, x1
-; CHECK-NEON-NEXT:    eor x9, x9, x22
-; CHECK-NEON-NEXT:    eor x10, x10, x18
-; CHECK-NEON-NEXT:    eor x9, x9, x20
-; CHECK-NEON-NEXT:    eor x11, x11, x14
-; CHECK-NEON-NEXT:    eor x10, x10, x17
-; CHECK-NEON-NEXT:    eor x9, x9, x19
-; CHECK-NEON-NEXT:    eor x11, x11, x7
-; CHECK-NEON-NEXT:    eor x10, x10, x21
-; CHECK-NEON-NEXT:    eor x8, x8, x9
-; CHECK-NEON-NEXT:    mov x1, x2
-; CHECK-NEON-NEXT:    eor x9, x10, x16
-; CHECK-NEON-NEXT:    eor x10, x11, x4
-; CHECK-NEON-NEXT:    eor x8, x8, x9
-; CHECK-NEON-NEXT:    eor x9, x10, x13
-; CHECK-NEON-NEXT:    eor x0, x8, x9
-; CHECK-NEON-NEXT:    add sp, sp, #1952
-; CHECK-NEON-NEXT:    ldp x20, x19, [sp, #80] // 16-byte Folded Reload
-; CHECK-NEON-NEXT:    ldp x22, x21, [sp, #64] // 16-byte Folded Reload
-; CHECK-NEON-NEXT:    ldp x24, x23, [sp, #48] // 16-byte Folded Reload
-; CHECK-NEON-NEXT:    ldp x26, x25, [sp, #32] // 16-byte Folded Reload
-; CHECK-NEON-NEXT:    ldp x28, x27, [sp, #16] // 16-byte Folded Reload
-; CHECK-NEON-NEXT:    ldp x29, x30, [sp], #96 // 16-byte Folded Reload
+; CHECK-NEON-NEXT:    ldp x20, x19, [sp, #336] // 16-byte Folded Reload
+; CHECK-NEON-NEXT:    and x10, x10, #0x8888888888888888
+; CHECK-NEON-NEXT:    orr x9, x11, x9
+; CHECK-NEON-NEXT:    orr x8, x8, x10
+; CHECK-NEON-NEXT:    orr x0, x9, x8
+; CHECK-NEON-NEXT:    add sp, sp, #352
 ; CHECK-NEON-NEXT:    ret
 ;
 ; CHECK-AES-LABEL: clmul_i128:
@@ -1827,54 +517,53 @@ define i16 @clmul_i16_zext(i8 %x, i8 %y) {
 define i32 @clmul_i32_zext(i16 %x, i16 %y) {
 ; CHECK-NEON-LABEL: clmul_i32_zext:
 ; CHECK-NEON:       // %bb.0:
-; CHECK-NEON-NEXT:    and w8, w0, #0xffff
-; CHECK-NEON-NEXT:    and w9, w1, #0x2
-; CHECK-NEON-NEXT:    and w10, w1, #0x1
-; CHECK-NEON-NEXT:    mul w9, w8, w9
-; CHECK-NEON-NEXT:    and w11, w1, #0x4
-; CHECK-NEON-NEXT:    and w12, w1, #0x8
-; CHECK-NEON-NEXT:    mul w10, w8, w10
-; CHECK-NEON-NEXT:    and w13, w1, #0x10
-; CHECK-NEON-NEXT:    and w14, w1, #0x20
-; CHECK-NEON-NEXT:    mul w11, w8, w11
-; CHECK-NEON-NEXT:    and w16, w1, #0x80
-; CHECK-NEON-NEXT:    and w17, w1, #0x100
-; CHECK-NEON-NEXT:    mul w12, w8, w12
-; CHECK-NEON-NEXT:    and w2, w1, #0x800
-; CHECK-NEON-NEXT:    and w15, w1, #0x40
-; CHECK-NEON-NEXT:    mul w13, w8, w13
-; CHECK-NEON-NEXT:    eor w9, w10, w9
-; CHECK-NEON-NEXT:    and w10, w1, #0x1000
+; CHECK-NEON-NEXT:    mov w8, #4369 // =0x1111
+; CHECK-NEON-NEXT:    mov w9, #8738 // =0x2222
+; CHECK-NEON-NEXT:    mov w13, #34952 // =0x8888
+; CHECK-NEON-NEXT:    mov w14, #17476 // =0x4444
+; CHECK-NEON-NEXT:    and w10, w1, w8
+; CHECK-NEON-NEXT:    and w11, w0, w9
+; CHECK-NEON-NEXT:    and w9, w1, w9
+; CHECK-NEON-NEXT:    and w8, w0, w8
+; CHECK-NEON-NEXT:    and w16, w1, w13
+; CHECK-NEON-NEXT:    and w17, w0, w14
+; CHECK-NEON-NEXT:    and w14, w1, w14
+; CHECK-NEON-NEXT:    and w13, w0, w13
+; CHECK-NEON-NEXT:    mul w12, w11, w10
+; CHECK-NEON-NEXT:    mul w15, w8, w9
+; CHECK-NEON-NEXT:    mul w18, w17, w16
+; CHECK-NEON-NEXT:    mul w0, w13, w14
+; CHECK-NEON-NEXT:    mul w1, w11, w16
+; CHECK-NEON-NEXT:    eor w12, w15, w12
+; CHECK-NEON-NEXT:    mul w2, w8, w10
+; CHECK-NEON-NEXT:    mul w3, w17, w14
+; CHECK-NEON-NEXT:    eor w15, w18, w0
+; CHECK-NEON-NEXT:    mul w4, w13, w9
+; CHECK-NEON-NEXT:    eor w12, w12, w15
+; CHECK-NEON-NEXT:    mul w5, w11, w9
+; CHECK-NEON-NEXT:    and w12, w12, #0x22222222
+; CHECK-NEON-NEXT:    mul w11, w11, w14
 ; CHECK-NEON-NEXT:    mul w14, w8, w14
-; CHECK-NEON-NEXT:    and w18, w1, #0x200
-; CHECK-NEON-NEXT:    and w0, w1, #0x400
-; CHECK-NEON-NEXT:    mul w16, w8, w16
-; CHECK-NEON-NEXT:    eor w11, w11, w12
-; CHECK-NEON-NEXT:    and w12, w1, #0x2000
-; CHECK-NEON-NEXT:    mul w17, w8, w17
-; CHECK-NEON-NEXT:    eor w9, w9, w11
-; CHECK-NEON-NEXT:    and w11, w1, #0x4000
-; CHECK-NEON-NEXT:    mul w2, w8, w2
-; CHECK-NEON-NEXT:    eor w13, w13, w14
-; CHECK-NEON-NEXT:    and w14, w1, #0x8000
-; CHECK-NEON-NEXT:    mul w10, w8, w10
-; CHECK-NEON-NEXT:    mul w15, w8, w15
-; CHECK-NEON-NEXT:    eor w16, w16, w17
-; CHECK-NEON-NEXT:    mul w18, w8, w18
-; CHECK-NEON-NEXT:    mul w12, w8, w12
-; CHECK-NEON-NEXT:    eor w10, w2, w10
-; CHECK-NEON-NEXT:    mul w0, w8, w0
-; CHECK-NEON-NEXT:    eor w13, w13, w15
-; CHECK-NEON-NEXT:    mul w11, w8, w11
-; CHECK-NEON-NEXT:    eor w9, w9, w13
-; CHECK-NEON-NEXT:    mul w8, w8, w14
-; CHECK-NEON-NEXT:    eor w14, w16, w18
-; CHECK-NEON-NEXT:    eor w10, w10, w12
-; CHECK-NEON-NEXT:    eor w12, w14, w0
-; CHECK-NEON-NEXT:    eor w10, w10, w11
-; CHECK-NEON-NEXT:    eor w9, w9, w12
-; CHECK-NEON-NEXT:    eor w8, w10, w8
-; CHECK-NEON-NEXT:    eor w0, w9, w8
+; CHECK-NEON-NEXT:    eor w15, w3, w4
+; CHECK-NEON-NEXT:    mul w9, w17, w9
+; CHECK-NEON-NEXT:    mul w17, w17, w10
+; CHECK-NEON-NEXT:    mul w8, w8, w16
+; CHECK-NEON-NEXT:    eor w14, w14, w5
+; CHECK-NEON-NEXT:    mul w16, w13, w16
+; CHECK-NEON-NEXT:    mul w10, w13, w10
+; CHECK-NEON-NEXT:    eor w13, w2, w1
+; CHECK-NEON-NEXT:    eor w8, w8, w11
+; CHECK-NEON-NEXT:    eor w16, w17, w16
+; CHECK-NEON-NEXT:    eor w9, w9, w10
+; CHECK-NEON-NEXT:    eor w10, w13, w15
+; CHECK-NEON-NEXT:    eor w11, w14, w16
+; CHECK-NEON-NEXT:    eor w8, w8, w9
+; CHECK-NEON-NEXT:    and w9, w10, #0x11111111
+; CHECK-NEON-NEXT:    and w10, w11, #0x44444444
+; CHECK-NEON-NEXT:    and w8, w8, #0x88888888
+; CHECK-NEON-NEXT:    orr w9, w9, w12
+; CHECK-NEON-NEXT:    orr w8, w10, w8
+; CHECK-NEON-NEXT:    orr w0, w9, w8
 ; CHECK-NEON-NEXT:    ret
 ;
 ; CHECK-AES-LABEL: clmul_i32_zext:
@@ -1895,102 +584,51 @@ define i32 @clmul_i32_zext(i16 %x, i16 %y) {
 define i64 @clmul_i64_zext(i32 %x, i32 %y) {
 ; CHECK-NEON-LABEL: clmul_i64_zext:
 ; CHECK-NEON:       // %bb.0:
-; CHECK-NEON-NEXT:    mov w8, w1
-; CHECK-NEON-NEXT:    and x9, x8, #0x2
-; CHECK-NEON-NEXT:    and x10, x8, #0x1
-; CHECK-NEON-NEXT:    and x11, x8, #0x4
-; CHECK-NEON-NEXT:    umull x9, w9, w0
-; CHECK-NEON-NEXT:    and x12, x8, #0x8
-; CHECK-NEON-NEXT:    and x13, x8, #0x10
-; CHECK-NEON-NEXT:    umull x10, w10, w0
-; CHECK-NEON-NEXT:    and x14, x8, #0x20
-; CHECK-NEON-NEXT:    and x15, x8, #0x40
-; CHECK-NEON-NEXT:    umull x11, w11, w0
-; CHECK-NEON-NEXT:    and x2, x8, #0x800
-; CHECK-NEON-NEXT:    and x16, x8, #0x80
-; CHECK-NEON-NEXT:    umull x12, w12, w0
-; CHECK-NEON-NEXT:    and x17, x8, #0x100
-; CHECK-NEON-NEXT:    and x18, x8, #0x200
-; CHECK-NEON-NEXT:    umull x13, w13, w0
-; CHECK-NEON-NEXT:    eor x9, x10, x9
-; CHECK-NEON-NEXT:    and x10, x8, #0x1000
-; CHECK-NEON-NEXT:    umull x14, w14, w0
-; CHECK-NEON-NEXT:    and x1, x8, #0x400
-; CHECK-NEON-NEXT:    umull x15, w15, w0
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    and x12, x8, #0x2000
-; CHECK-NEON-NEXT:    umull x2, w2, w0
-; CHECK-NEON-NEXT:    eor x9, x9, x11
-; CHECK-NEON-NEXT:    and x11, x8, #0x4000
-; CHECK-NEON-NEXT:    umull x10, w10, w0
-; CHECK-NEON-NEXT:    eor x13, x13, x14
-; CHECK-NEON-NEXT:    and x14, x8, #0x8000
-; CHECK-NEON-NEXT:    umull x16, w16, w0
-; CHECK-NEON-NEXT:    eor x13, x13, x15
-; CHECK-NEON-NEXT:    and x15, x8, #0x10000
-; CHECK-NEON-NEXT:    umull x17, w17, w0
-; CHECK-NEON-NEXT:    eor x9, x9, x13
-; CHECK-NEON-NEXT:    and x13, x8, #0x20000
-; CHECK-NEON-NEXT:    umull x12, w12, w0
-; CHECK-NEON-NEXT:    eor x10, x2, x10
-; CHECK-NEON-NEXT:    and x2, x8, #0x400000
-; CHECK-NEON-NEXT:    umull x18, w18, w0
-; CHECK-NEON-NEXT:    umull x11, w11, w0
-; CHECK-NEON-NEXT:    eor x16, x16, x17
-; CHECK-NEON-NEXT:    and x17, x8, #0x40000
-; CHECK-NEON-NEXT:    umull x14, w14, w0
+; CHECK-NEON-NEXT:    mov w8, w0
+; CHECK-NEON-NEXT:    mov w9, w1
+; CHECK-NEON-NEXT:    and x10, x9, #0x1111111111111111
+; CHECK-NEON-NEXT:    and x11, x8, #0x2222222222222222
+; CHECK-NEON-NEXT:    and x13, x9, #0x2222222222222222
+; CHECK-NEON-NEXT:    and x14, x8, #0x1111111111111111
+; CHECK-NEON-NEXT:    and x16, x9, #0x8888888888888888
+; CHECK-NEON-NEXT:    and x17, x8, #0x4444444444444444
+; CHECK-NEON-NEXT:    and x9, x9, #0x4444444444444444
+; CHECK-NEON-NEXT:    and x8, x8, #0x8888888888888888
+; CHECK-NEON-NEXT:    mul x12, x11, x10
+; CHECK-NEON-NEXT:    mul x15, x14, x13
+; CHECK-NEON-NEXT:    mul x18, x17, x16
+; CHECK-NEON-NEXT:    mul x0, x8, x9
+; CHECK-NEON-NEXT:    mul x1, x11, x16
+; CHECK-NEON-NEXT:    mul x2, x14, x10
+; CHECK-NEON-NEXT:    mul x3, x17, x9
+; CHECK-NEON-NEXT:    mul x4, x8, x13
+; CHECK-NEON-NEXT:    mul x5, x11, x13
+; CHECK-NEON-NEXT:    mul x11, x11, x9
+; CHECK-NEON-NEXT:    mul x9, x14, x9
+; CHECK-NEON-NEXT:    mul x13, x17, x13
+; CHECK-NEON-NEXT:    mul x17, x17, x10
+; CHECK-NEON-NEXT:    eor x9, x9, x5
+; CHECK-NEON-NEXT:    mul x14, x14, x16
+; CHECK-NEON-NEXT:    mul x16, x8, x16
+; CHECK-NEON-NEXT:    mul x8, x8, x10
+; CHECK-NEON-NEXT:    eor x10, x15, x12
+; CHECK-NEON-NEXT:    eor x12, x18, x0
 ; CHECK-NEON-NEXT:    eor x10, x10, x12
-; CHECK-NEON-NEXT:    and x12, x8, #0x800000
-; CHECK-NEON-NEXT:    umull x1, w1, w0
-; CHECK-NEON-NEXT:    eor x16, x16, x18
-; CHECK-NEON-NEXT:    and x18, x8, #0x80000
-; CHECK-NEON-NEXT:    umull x15, w15, w0
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    and x11, x8, #0x1000000
-; CHECK-NEON-NEXT:    umull x13, w13, w0
-; CHECK-NEON-NEXT:    eor x10, x10, x14
-; CHECK-NEON-NEXT:    and x14, x8, #0x2000000
-; CHECK-NEON-NEXT:    umull x17, w17, w0
-; CHECK-NEON-NEXT:    eor x16, x16, x1
-; CHECK-NEON-NEXT:    and x1, x8, #0x100000
-; CHECK-NEON-NEXT:    umull x2, w2, w0
+; CHECK-NEON-NEXT:    eor x12, x2, x1
+; CHECK-NEON-NEXT:    eor x15, x3, x4
+; CHECK-NEON-NEXT:    eor x11, x14, x11
+; CHECK-NEON-NEXT:    eor x12, x12, x15
+; CHECK-NEON-NEXT:    and x10, x10, #0x2222222222222222
+; CHECK-NEON-NEXT:    eor x16, x17, x16
 ; CHECK-NEON-NEXT:    eor x9, x9, x16
-; CHECK-NEON-NEXT:    and x16, x8, #0x200000
-; CHECK-NEON-NEXT:    umull x12, w12, w0
-; CHECK-NEON-NEXT:    eor x13, x15, x13
-; CHECK-NEON-NEXT:    and x15, x8, #0x4000000
-; CHECK-NEON-NEXT:    umull x18, w18, w0
-; CHECK-NEON-NEXT:    eor x13, x13, x17
-; CHECK-NEON-NEXT:    and x17, x8, #0x8000000
-; CHECK-NEON-NEXT:    umull x11, w11, w0
-; CHECK-NEON-NEXT:    eor x9, x9, x10
-; CHECK-NEON-NEXT:    umull x14, w14, w0
-; CHECK-NEON-NEXT:    eor x12, x2, x12
-; CHECK-NEON-NEXT:    and x2, x8, #0x20000000
-; CHECK-NEON-NEXT:    umull x1, w1, w0
-; CHECK-NEON-NEXT:    eor x13, x13, x18
-; CHECK-NEON-NEXT:    and x18, x8, #0x10000000
-; CHECK-NEON-NEXT:    umull x15, w15, w0
-; CHECK-NEON-NEXT:    eor x11, x12, x11
-; CHECK-NEON-NEXT:    and x12, x8, #0x40000000
-; CHECK-NEON-NEXT:    umull x16, w16, w0
-; CHECK-NEON-NEXT:    eor x11, x11, x14
-; CHECK-NEON-NEXT:    and x8, x8, #0x80000000
-; CHECK-NEON-NEXT:    umull x17, w17, w0
-; CHECK-NEON-NEXT:    eor x13, x13, x1
-; CHECK-NEON-NEXT:    umull x18, w18, w0
-; CHECK-NEON-NEXT:    eor x11, x11, x15
-; CHECK-NEON-NEXT:    umull x2, w2, w0
-; CHECK-NEON-NEXT:    eor x10, x13, x16
-; CHECK-NEON-NEXT:    umull x12, w12, w0
-; CHECK-NEON-NEXT:    eor x11, x11, x17
-; CHECK-NEON-NEXT:    eor x9, x9, x10
-; CHECK-NEON-NEXT:    umull x8, w8, w0
-; CHECK-NEON-NEXT:    eor x10, x11, x18
-; CHECK-NEON-NEXT:    eor x9, x9, x10
-; CHECK-NEON-NEXT:    eor x11, x2, x12
+; CHECK-NEON-NEXT:    eor x8, x13, x8
+; CHECK-NEON-NEXT:    and x9, x9, #0x4444444444444444
 ; CHECK-NEON-NEXT:    eor x8, x11, x8
-; CHECK-NEON-NEXT:    eor x0, x9, x8
+; CHECK-NEON-NEXT:    and x11, x12, #0x1111111111111111
+; CHECK-NEON-NEXT:    and x8, x8, #0x8888888888888888
+; CHECK-NEON-NEXT:    orr x10, x11, x10
+; CHECK-NEON-NEXT:    orr x8, x9, x8
+; CHECK-NEON-NEXT:    orr x0, x10, x8
 ; CHECK-NEON-NEXT:    ret
 ;
 ; CHECK-AES-LABEL: clmul_i64_zext:
@@ -2011,14 +649,14 @@ define i64 @clmul_i64_zext(i32 %x, i32 %y) {
 define i128 @clmul_i128_zext(i64 %x, i64 %y) {
 ; CHECK-NEON-LABEL: clmul_i128_zext:
 ; CHECK-NEON:       // %bb.0:
-; CHECK-NEON-NEXT:    stp x29, x30, [sp, #-96]! // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    stp x28, x27, [sp, #16] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    stp x26, x25, [sp, #32] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    stp x24, x23, [sp, #48] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    stp x22, x21, [sp, #64] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    stp x20, x19, [sp, #80] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    sub sp, sp, #624
-; CHECK-NEON-NEXT:    .cfi_def_cfa_offset 720
+; CHECK-NEON-NEXT:    sub sp, sp, #144
+; CHECK-NEON-NEXT:    stp x29, x30, [sp, #48] // 16-byte Folded Spill
+; CHECK-NEON-NEXT:    stp x28, x27, [sp, #64] // 16-byte Folded Spill
+; CHECK-NEON-NEXT:    stp x26, x25, [sp, #80] // 16-byte Folded Spill
+; CHECK-NEON-NEXT:    stp x24, x23, [sp, #96] // 16-byte Folded Spill
+; CHECK-NEON-NEXT:    stp x22, x21, [sp, #112] // 16-byte Folded Spill
+; CHECK-NEON-NEXT:    stp x20, x19, [sp, #128] // 16-byte Folded Spill
+; CHECK-NEON-NEXT:    .cfi_def_cfa_offset 144
 ; CHECK-NEON-NEXT:    .cfi_offset w19, -8
 ; CHECK-NEON-NEXT:    .cfi_offset w20, -16
 ; CHECK-NEON-NEXT:    .cfi_offset w21, -24
@@ -2031,534 +669,110 @@ define i128 @clmul_i128_zext(i64 %x, i64 %y) {
 ; CHECK-NEON-NEXT:    .cfi_offset w28, -80
 ; CHECK-NEON-NEXT:    .cfi_offset w30, -88
 ; CHECK-NEON-NEXT:    .cfi_offset w29, -96
-; CHECK-NEON-NEXT:    and x8, x1, #0x2
-; CHECK-NEON-NEXT:    mul x11, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x1
-; CHECK-NEON-NEXT:    mul x12, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x4
-; CHECK-NEON-NEXT:    mul x13, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x8
-; CHECK-NEON-NEXT:    mul x14, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x10
-; CHECK-NEON-NEXT:    mul x15, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x20
-; CHECK-NEON-NEXT:    mul x16, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x40
-; CHECK-NEON-NEXT:    mul x17, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x80
-; CHECK-NEON-NEXT:    mul x18, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x100
-; CHECK-NEON-NEXT:    mul x3, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x200
-; CHECK-NEON-NEXT:    mul x2, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x400
-; CHECK-NEON-NEXT:    mul x4, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x800
-; CHECK-NEON-NEXT:    mul x5, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x1000
-; CHECK-NEON-NEXT:    mul x20, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x2000
-; CHECK-NEON-NEXT:    mul x6, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x4000
-; CHECK-NEON-NEXT:    mul x7, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x8000
-; CHECK-NEON-NEXT:    mul x19, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x10000
-; CHECK-NEON-NEXT:    mul x21, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x20000
-; CHECK-NEON-NEXT:    mul x22, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x40000
-; CHECK-NEON-NEXT:    mul x23, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x80000
-; CHECK-NEON-NEXT:    mul x24, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x100000
-; CHECK-NEON-NEXT:    mul x25, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x200000
-; CHECK-NEON-NEXT:    mul x26, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x400000
-; CHECK-NEON-NEXT:    mul x27, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x800000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #592] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x1000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #584] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x2000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #616] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x4000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #576] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x8000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #608] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x10000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #600] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x20000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #568] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x40000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #512] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x80000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #536] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x100000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #528] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x200000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #560] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x400000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #520] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x800000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #552] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x1000000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #544] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x2000000000
-; CHECK-NEON-NEXT:    mul x9, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x4000000000
-; CHECK-NEON-NEXT:    mul x10, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x8000000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #464] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x10000000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #456] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x20000000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    stp x8, x9, [sp, #488] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x40000000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    stp x10, x8, [sp, #440] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x80000000000
-; CHECK-NEON-NEXT:    mul x9, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x100000000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    stp x8, x9, [sp, #472] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x200000000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #504] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x400000000000
-; CHECK-NEON-NEXT:    mul x9, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x800000000000
-; CHECK-NEON-NEXT:    mul x10, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x1000000000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #416] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x2000000000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    str x8, [sp, #408] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x4000000000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    stp x8, x9, [sp, #424] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x8000000000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    stp x10, x8, [sp, #392] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    and x8, x1, #0x100000000000000
-; CHECK-NEON-NEXT:    mul x9, x0, x8
-; CHECK-NEON-NEXT:    and x8, x1, #0x200000000000000
-; CHECK-NEON-NEXT:    mul x8, x0, x8
-; CHECK-NEON-NEXT:    stp x8, x9, [sp, #376] // 16-byte Folded Spill
-; CHECK-NEON-NEXT:    and x9, x1, #0x400000000000000
-; CHECK-NEON-NEXT:    rbit x8, x1
-; CHECK-NEON-NEXT:    mul x9, x0, x9
-; CHECK-NEON-NEXT:    and x10, x8, #0x2
-; CHECK-NEON-NEXT:    str x9, [sp, #368] // 8-byte Spill
-; CHECK-NEON-NEXT:    rbit x9, x0
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #360] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x1
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #352] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x4
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #344] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x8
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #336] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x10
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #328] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x20
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #320] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x40
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #312] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x80
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #304] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x100
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #296] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x200
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #288] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x400
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #280] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x800
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #272] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x1000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #256] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x2000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #248] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x4000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #264] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x8000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #240] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x10000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #232] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x20000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #200] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x40000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #224] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x80000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #192] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x100000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #216] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x200000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #208] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x400000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #184] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x800000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #136] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x1000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #168] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x2000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #160] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x4000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #176] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x8000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #152] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x10000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #144] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x20000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #128] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x40000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #120] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x80000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #112] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x100000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #104] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x200000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #96] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x400000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #88] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x800000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #80] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x1000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #72] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x2000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #64] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x4000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #56] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x8000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #48] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x10000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #40] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x20000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #32] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x40000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #24] // 8-byte Spill
-; CHECK-NEON-NEXT:    and x10, x8, #0x80000000000
-; CHECK-NEON-NEXT:    mul x10, x9, x10
-; CHECK-NEON-NEXT:    str x10, [sp, #16] // 8-byte Spill
-; CHECK-NEON-NEXT:    eor x10, x12, x11
-; CHECK-NEON-NEXT:    and x12, x8, #0x100000000000
-; CHECK-NEON-NEXT:    mul x12, x9, x12
-; CHECK-NEON-NEXT:    eor x11, x13, x14
-; CHECK-NEON-NEXT:    and x13, x8, #0x200000000000
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    eor x11, x15, x16
-; CHECK-NEON-NEXT:    ldr x14, [sp, #608] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x17
-; CHECK-NEON-NEXT:    mul x30, x9, x13
-; CHECK-NEON-NEXT:    and x13, x8, #0x400000000000
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    mul x29, x9, x13
-; CHECK-NEON-NEXT:    and x13, x8, #0x800000000000
-; CHECK-NEON-NEXT:    str x12, [sp, #8] // 8-byte Spill
-; CHECK-NEON-NEXT:    eor x12, x18, x3
-; CHECK-NEON-NEXT:    eor x11, x12, x2
-; CHECK-NEON-NEXT:    eor x12, x5, x20
-; CHECK-NEON-NEXT:    mul x28, x9, x13
-; CHECK-NEON-NEXT:    eor x11, x11, x4
-; CHECK-NEON-NEXT:    ldr x13, [sp, #592] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    eor x11, x12, x6
-; CHECK-NEON-NEXT:    eor x12, x21, x22
-; CHECK-NEON-NEXT:    eor x11, x11, x7
-; CHECK-NEON-NEXT:    eor x12, x12, x23
-; CHECK-NEON-NEXT:    eor x13, x27, x13
-; CHECK-NEON-NEXT:    eor x11, x11, x19
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    eor x11, x12, x24
-; CHECK-NEON-NEXT:    and x12, x8, #0x1000000000000
-; CHECK-NEON-NEXT:    eor x11, x11, x25
-; CHECK-NEON-NEXT:    mul x27, x9, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #584] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x26
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #616] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x12, x13, x12
-; CHECK-NEON-NEXT:    and x13, x8, #0x2000000000000
-; CHECK-NEON-NEXT:    eor x11, x12, x11
-; CHECK-NEON-NEXT:    ldr x12, [sp, #576] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x25, x9, x13
-; CHECK-NEON-NEXT:    ldr x13, [sp, #512] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #568] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x14
-; CHECK-NEON-NEXT:    ldr x14, [sp, #536] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x12, x12, x13
-; CHECK-NEON-NEXT:    and x13, x8, #0x4000000000000
-; CHECK-NEON-NEXT:    mul x24, x9, x13
-; CHECK-NEON-NEXT:    ldr x13, [sp, #600] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x12, x12, x14
-; CHECK-NEON-NEXT:    ldr x14, [sp, #552] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x13
-; CHECK-NEON-NEXT:    ldr x13, [sp, #528] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #560] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x12, x12, x13
-; CHECK-NEON-NEXT:    and x13, x8, #0x8000000000000
-; CHECK-NEON-NEXT:    eor x11, x12, x11
-; CHECK-NEON-NEXT:    ldr x12, [sp, #520] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x23, x9, x13
-; CHECK-NEON-NEXT:    ldr x13, [sp, #440] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #496] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x14
-; CHECK-NEON-NEXT:    ldr x14, [sp, #464] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x12, x12, x13
-; CHECK-NEON-NEXT:    and x13, x8, #0x10000000000000
-; CHECK-NEON-NEXT:    mul x21, x9, x13
-; CHECK-NEON-NEXT:    ldr x13, [sp, #544] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x12, x12, x14
-; CHECK-NEON-NEXT:    ldr x14, [sp, #256] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x13
-; CHECK-NEON-NEXT:    ldr x13, [sp, #456] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x26, x10, x11
-; CHECK-NEON-NEXT:    ldr x10, [sp, #488] // 8-byte Reload
-; CHECK-NEON-NEXT:    ldr x11, [sp, #448] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x12, x12, x13
-; CHECK-NEON-NEXT:    and x13, x8, #0x20000000000000
-; CHECK-NEON-NEXT:    eor x10, x12, x10
-; CHECK-NEON-NEXT:    mul x20, x9, x13
-; CHECK-NEON-NEXT:    ldr x12, [sp, #392] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #432] // 8-byte Reload
-; CHECK-NEON-NEXT:    ldr x13, [sp, #480] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    and x12, x8, #0x40000000000000
-; CHECK-NEON-NEXT:    eor x10, x10, x13
-; CHECK-NEON-NEXT:    ldr x13, [sp, #416] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x7, x9, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #472] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x13
-; CHECK-NEON-NEXT:    ldr x13, [sp, #504] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #408] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x22, x10, x13
-; CHECK-NEON-NEXT:    ldr x10, [sp, #424] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    and x12, x8, #0x80000000000000
-; CHECK-NEON-NEXT:    eor x10, x11, x10
-; CHECK-NEON-NEXT:    ldr x11, [sp, #400] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x5, x9, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #368] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x19, x10, x11
-; CHECK-NEON-NEXT:    ldp x11, x10, [sp, #376] // 16-byte Folded Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    and x11, x8, #0x100000000000000
-; CHECK-NEON-NEXT:    eor x6, x10, x12
-; CHECK-NEON-NEXT:    ldp x12, x10, [sp, #352] // 16-byte Folded Reload
-; CHECK-NEON-NEXT:    mul x4, x9, x11
-; CHECK-NEON-NEXT:    eor x10, x12, x10
-; CHECK-NEON-NEXT:    ldp x12, x11, [sp, #336] // 16-byte Folded Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    ldp x13, x12, [sp, #320] // 16-byte Folded Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #312] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x12, x12, x13
-; CHECK-NEON-NEXT:    and x13, x8, #0x200000000000000
-; CHECK-NEON-NEXT:    eor x11, x12, x11
-; CHECK-NEON-NEXT:    mul x3, x9, x13
-; CHECK-NEON-NEXT:    and x13, x8, #0x400000000000000
-; CHECK-NEON-NEXT:    eor x12, x10, x11
-; CHECK-NEON-NEXT:    ldp x11, x10, [sp, #296] // 16-byte Folded Reload
-; CHECK-NEON-NEXT:    mul x2, x9, x13
-; CHECK-NEON-NEXT:    ldr x13, [sp, #280] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x10, x11
-; CHECK-NEON-NEXT:    ldr x10, [sp, #288] // 8-byte Reload
+; CHECK-NEON-NEXT:    and x12, x1, #0x1111111111111111
+; CHECK-NEON-NEXT:    and x13, x0, #0x2222222222222222
+; CHECK-NEON-NEXT:    and x14, x1, #0x2222222222222222
+; CHECK-NEON-NEXT:    and x11, x0, #0x1111111111111111
+; CHECK-NEON-NEXT:    mul x8, x13, x12
+; CHECK-NEON-NEXT:    and x10, x1, #0x8888888888888888
+; CHECK-NEON-NEXT:    and x16, x0, #0x4444444444444444
+; CHECK-NEON-NEXT:    and x2, x1, #0x4444444444444444
+; CHECK-NEON-NEXT:    rbit x1, x1
+; CHECK-NEON-NEXT:    mul x9, x11, x14
+; CHECK-NEON-NEXT:    and x7, x1, #0x1111111111111111
+; CHECK-NEON-NEXT:    and x21, x1, #0x2222222222222222
+; CHECK-NEON-NEXT:    and x24, x1, #0x8888888888888888
+; CHECK-NEON-NEXT:    mul x4, x11, x12
+; CHECK-NEON-NEXT:    and x1, x1, #0x4444444444444444
+; CHECK-NEON-NEXT:    mul x5, x16, x2
+; CHECK-NEON-NEXT:    stp x8, x9, [sp, #32] // 16-byte Folded Spill
+; CHECK-NEON-NEXT:    and x9, x0, #0x8888888888888888
+; CHECK-NEON-NEXT:    rbit x0, x0
+; CHECK-NEON-NEXT:    mul x8, x16, x10
+; CHECK-NEON-NEXT:    and x19, x0, #0x2222222222222222
+; CHECK-NEON-NEXT:    and x22, x0, #0x1111111111111111
+; CHECK-NEON-NEXT:    and x25, x0, #0x4444444444444444
+; CHECK-NEON-NEXT:    mul x6, x9, x14
+; CHECK-NEON-NEXT:    and x0, x0, #0x8888888888888888
+; CHECK-NEON-NEXT:    mul x28, x19, x24
+; CHECK-NEON-NEXT:    str x8, [sp, #24] // 8-byte Spill
+; CHECK-NEON-NEXT:    mul x8, x9, x2
+; CHECK-NEON-NEXT:    mul x29, x22, x7
+; CHECK-NEON-NEXT:    mul x30, x25, x1
+; CHECK-NEON-NEXT:    str x8, [sp, #16] // 8-byte Spill
+; CHECK-NEON-NEXT:    mul x8, x13, x10
+; CHECK-NEON-NEXT:    mul x17, x0, x21
+; CHECK-NEON-NEXT:    mul x20, x19, x7
+; CHECK-NEON-NEXT:    str x8, [sp, #8] // 8-byte Spill
+; CHECK-NEON-NEXT:    ldp x18, x8, [sp, #32] // 16-byte Folded Reload
+; CHECK-NEON-NEXT:    mul x23, x22, x21
+; CHECK-NEON-NEXT:    eor x17, x30, x17
+; CHECK-NEON-NEXT:    eor x8, x8, x18
+; CHECK-NEON-NEXT:    ldp x3, x18, [sp, #16] // 16-byte Folded Reload
+; CHECK-NEON-NEXT:    mul x26, x25, x24
+; CHECK-NEON-NEXT:    mul x27, x0, x1
+; CHECK-NEON-NEXT:    eor x18, x18, x3
+; CHECK-NEON-NEXT:    ldr x3, [sp, #8] // 8-byte Reload
+; CHECK-NEON-NEXT:    eor x8, x8, x18
+; CHECK-NEON-NEXT:    mul x15, x19, x21
 ; CHECK-NEON-NEXT:    eor x3, x4, x3
-; CHECK-NEON-NEXT:    and x4, x1, #0x2000000000000000
-; CHECK-NEON-NEXT:    eor x11, x11, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #272] // 8-byte Reload
+; CHECK-NEON-NEXT:    eor x4, x5, x6
+; CHECK-NEON-NEXT:    eor x18, x3, x4
+; CHECK-NEON-NEXT:    eor x4, x29, x28
+; CHECK-NEON-NEXT:    eor x5, x23, x20
+; CHECK-NEON-NEXT:    mul x19, x19, x1
+; CHECK-NEON-NEXT:    eor x17, x4, x17
+; CHECK-NEON-NEXT:    and x8, x8, #0x2222222222222222
+; CHECK-NEON-NEXT:    eor x6, x26, x27
+; CHECK-NEON-NEXT:    and x17, x17, #0x1111111111111111
+; CHECK-NEON-NEXT:    mul x1, x22, x1
+; CHECK-NEON-NEXT:    eor x3, x5, x6
+; CHECK-NEON-NEXT:    ldp x28, x27, [sp, #64] // 16-byte Folded Reload
+; CHECK-NEON-NEXT:    mul x21, x25, x21
+; CHECK-NEON-NEXT:    ldp x29, x30, [sp, #48] // 16-byte Folded Reload
+; CHECK-NEON-NEXT:    mul x25, x25, x7
+; CHECK-NEON-NEXT:    eor x15, x1, x15
+; CHECK-NEON-NEXT:    mul x22, x22, x24
+; CHECK-NEON-NEXT:    mul x24, x0, x24
+; CHECK-NEON-NEXT:    mul x0, x0, x7
+; CHECK-NEON-NEXT:    eor x4, x22, x19
+; CHECK-NEON-NEXT:    ldp x20, x19, [sp, #128] // 16-byte Folded Reload
+; CHECK-NEON-NEXT:    mul x7, x13, x14
+; CHECK-NEON-NEXT:    eor x1, x25, x24
+; CHECK-NEON-NEXT:    ldp x24, x23, [sp, #96] // 16-byte Folded Reload
+; CHECK-NEON-NEXT:    mul x13, x13, x2
+; CHECK-NEON-NEXT:    ldp x26, x25, [sp, #80] // 16-byte Folded Reload
+; CHECK-NEON-NEXT:    eor x0, x21, x0
+; CHECK-NEON-NEXT:    ldp x22, x21, [sp, #112] // 16-byte Folded Reload
+; CHECK-NEON-NEXT:    mul x2, x11, x2
+; CHECK-NEON-NEXT:    mul x14, x16, x14
+; CHECK-NEON-NEXT:    mul x16, x16, x12
+; CHECK-NEON-NEXT:    mul x11, x11, x10
+; CHECK-NEON-NEXT:    mul x10, x9, x10
+; CHECK-NEON-NEXT:    mul x9, x9, x12
+; CHECK-NEON-NEXT:    eor x12, x15, x1
+; CHECK-NEON-NEXT:    eor x15, x4, x0
+; CHECK-NEON-NEXT:    and x0, x3, #0x2222222222222222
+; CHECK-NEON-NEXT:    and x12, x12, #0x4444444444444444
+; CHECK-NEON-NEXT:    and x15, x15, #0x8888888888888888
+; CHECK-NEON-NEXT:    orr x17, x17, x0
+; CHECK-NEON-NEXT:    orr x12, x12, x15
+; CHECK-NEON-NEXT:    eor x15, x2, x7
+; CHECK-NEON-NEXT:    eor x10, x16, x10
 ; CHECK-NEON-NEXT:    eor x11, x11, x13
-; CHECK-NEON-NEXT:    ldr x13, [sp, #248] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x2, x3, x2
-; CHECK-NEON-NEXT:    eor x10, x10, x14
-; CHECK-NEON-NEXT:    eor x11, x12, x11
-; CHECK-NEON-NEXT:    ldr x12, [sp, #264] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x10, x10, x13
-; CHECK-NEON-NEXT:    and x13, x8, #0x800000000000000
-; CHECK-NEON-NEXT:    mul x3, x0, x4
-; CHECK-NEON-NEXT:    eor x12, x10, x12
-; CHECK-NEON-NEXT:    ldr x10, [sp, #240] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x18, x9, x13
-; CHECK-NEON-NEXT:    ldr x13, [sp, #200] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x12, x12, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #232] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    ldp x14, x12, [sp, #216] // 16-byte Folded Reload
-; CHECK-NEON-NEXT:    eor x13, x10, x13
-; CHECK-NEON-NEXT:    and x10, x8, #0x1000000000000000
-; CHECK-NEON-NEXT:    mul x17, x9, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #192] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x12, x13, x12
-; CHECK-NEON-NEXT:    ldr x13, [sp, #136] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x18, x2, x18
-; CHECK-NEON-NEXT:    eor x12, x12, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #184] // 8-byte Reload
-; CHECK-NEON-NEXT:    and x2, x1, #0x4000000000000000
-; CHECK-NEON-NEXT:    eor x12, x12, x14
-; CHECK-NEON-NEXT:    ldr x14, [sp, #168] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x13, x10, x13
-; CHECK-NEON-NEXT:    and x10, x8, #0x2000000000000000
-; CHECK-NEON-NEXT:    and x8, x8, #0x4000000000000000
-; CHECK-NEON-NEXT:    mul x16, x9, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #208] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x13, x13, x14
-; CHECK-NEON-NEXT:    eor x17, x18, x17
-; CHECK-NEON-NEXT:    eor x12, x12, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #160] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x15, x9, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #152] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    eor x13, x13, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #176] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x18, x0, x2
-; CHECK-NEON-NEXT:    eor x16, x17, x16
-; CHECK-NEON-NEXT:    and x17, x1, #0x8000000000000000
-; CHECK-NEON-NEXT:    eor x12, x13, x10
-; CHECK-NEON-NEXT:    and x13, x1, #0x10000000000000
-; CHECK-NEON-NEXT:    eor x9, x12, x8
-; CHECK-NEON-NEXT:    ldp x10, x8, [sp, #120] // 16-byte Folded Reload
-; CHECK-NEON-NEXT:    mul x14, x0, x13
-; CHECK-NEON-NEXT:    eor x12, x8, x10
-; CHECK-NEON-NEXT:    ldr x8, [sp, #144] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x8
-; CHECK-NEON-NEXT:    ldp x8, x10, [sp, #104] // 16-byte Folded Reload
+; CHECK-NEON-NEXT:    orr x12, x17, x12
+; CHECK-NEON-NEXT:    eor x10, x15, x10
+; CHECK-NEON-NEXT:    rbit x12, x12
+; CHECK-NEON-NEXT:    and x13, x18, #0x1111111111111111
+; CHECK-NEON-NEXT:    eor x9, x14, x9
+; CHECK-NEON-NEXT:    and x10, x10, #0x4444444444444444
+; CHECK-NEON-NEXT:    orr x8, x13, x8
 ; CHECK-NEON-NEXT:    eor x9, x11, x9
-; CHECK-NEON-NEXT:    eor x14, x19, x14
-; CHECK-NEON-NEXT:    eor x12, x12, x10
-; CHECK-NEON-NEXT:    eor x11, x12, x8
-; CHECK-NEON-NEXT:    ldr x8, [sp, #96] // 8-byte Reload
-; CHECK-NEON-NEXT:    and x12, x1, #0x20000000000000
-; CHECK-NEON-NEXT:    mul x13, x0, x12
-; CHECK-NEON-NEXT:    eor x11, x11, x8
-; CHECK-NEON-NEXT:    ldp x10, x8, [sp, #56] // 16-byte Folded Reload
-; CHECK-NEON-NEXT:    eor x8, x8, x10
-; CHECK-NEON-NEXT:    ldp x12, x10, [sp, #80] // 16-byte Folded Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #48] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x12
-; CHECK-NEON-NEXT:    ldr x12, [sp, #40] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x8, x10
-; CHECK-NEON-NEXT:    and x10, x1, #0x40000000000000
-; CHECK-NEON-NEXT:    eor x8, x8, x12
-; CHECK-NEON-NEXT:    mul x12, x0, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #72] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x10
-; CHECK-NEON-NEXT:    ldr x10, [sp, #32] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x9, x9, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #24] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x8, x8, x10
-; CHECK-NEON-NEXT:    and x10, x1, #0x80000000000000
-; CHECK-NEON-NEXT:    eor x8, x8, x11
-; CHECK-NEON-NEXT:    ldr x11, [sp, #16] // 8-byte Reload
-; CHECK-NEON-NEXT:    mul x10, x0, x10
-; CHECK-NEON-NEXT:    eor x8, x8, x11
-; CHECK-NEON-NEXT:    eor x11, x29, x28
-; CHECK-NEON-NEXT:    ldr x29, [sp, #8] // 8-byte Reload
-; CHECK-NEON-NEXT:    eor x11, x11, x27
-; CHECK-NEON-NEXT:    and x28, x1, #0x800000000000000
-; CHECK-NEON-NEXT:    eor x8, x8, x29
-; CHECK-NEON-NEXT:    eor x11, x11, x25
-; CHECK-NEON-NEXT:    mul x27, x0, x28
-; CHECK-NEON-NEXT:    eor x8, x8, x30
-; CHECK-NEON-NEXT:    and x25, x1, #0x1000000000000000
-; CHECK-NEON-NEXT:    eor x8, x9, x8
-; CHECK-NEON-NEXT:    eor x9, x11, x24
-; CHECK-NEON-NEXT:    mul x11, x0, x25
-; CHECK-NEON-NEXT:    eor x9, x9, x23
-; CHECK-NEON-NEXT:    eor x9, x9, x21
-; CHECK-NEON-NEXT:    eor x9, x9, x20
-; CHECK-NEON-NEXT:    eor x9, x9, x7
-; CHECK-NEON-NEXT:    eor x9, x9, x5
-; CHECK-NEON-NEXT:    eor x8, x8, x9
-; CHECK-NEON-NEXT:    eor x9, x16, x15
-; CHECK-NEON-NEXT:    mul x15, x0, x17
-; CHECK-NEON-NEXT:    eor x16, x6, x27
-; CHECK-NEON-NEXT:    eor x8, x8, x9
-; CHECK-NEON-NEXT:    eor x9, x14, x13
-; CHECK-NEON-NEXT:    eor x11, x16, x11
-; CHECK-NEON-NEXT:    rbit x8, x8
-; CHECK-NEON-NEXT:    eor x9, x9, x12
-; CHECK-NEON-NEXT:    eor x11, x11, x3
-; CHECK-NEON-NEXT:    eor x13, x26, x22
-; CHECK-NEON-NEXT:    eor x9, x9, x10
-; CHECK-NEON-NEXT:    eor x10, x11, x18
-; CHECK-NEON-NEXT:    lsr x1, x8, #1
-; CHECK-NEON-NEXT:    eor x8, x13, x9
-; CHECK-NEON-NEXT:    eor x9, x10, x15
-; CHECK-NEON-NEXT:    eor x0, x8, x9
-; CHECK-NEON-NEXT:    add sp, sp, #624
-; CHECK-NEON-NEXT:    ldp x20, x19, [sp, #80] // 16-byte Folded Reload
-; CHECK-NEON-NEXT:    ldp x22, x21, [sp, #64] // 16-byte Folded Reload
-; CHECK-NEON-NEXT:    ldp x24, x23, [sp, #48] // 16-byte Folded Reload
-; CHECK-NEON-NEXT:    ldp x26, x25, [sp, #32] // 16-byte Folded Reload
-; CHECK-NEON-NEXT:    ldp x28, x27, [sp, #16] // 16-byte Folded Reload
-; CHECK-NEON-NEXT:    ldp x29, x30, [sp], #96 // 16-byte Folded Reload
+; CHECK-NEON-NEXT:    lsr x1, x12, #1
+; CHECK-NEON-NEXT:    and x9, x9, #0x8888888888888888
+; CHECK-NEON-NEXT:    orr x9, x10, x9
+; CHECK-NEON-NEXT:    orr x0, x8, x9
+; CHECK-NEON-NEXT:    add sp, sp, #144
 ; CHECK-NEON-NEXT:    ret
 ;
 ; CHECK-AES-LABEL: clmul_i128_zext:
diff --git a/llvm/test/CodeGen/X86/clmul.ll b/llvm/test/CodeGen/X86/clmul.ll
index 638243ff309fe..fe0762e188bbe 100644
--- a/llvm/test/CodeGen/X86/clmul.ll
+++ b/llvm/test/CodeGen/X86/clmul.ll
@@ -79,96 +79,87 @@ define i8 @clmul_i8(i8 %a, i8 %b) nounwind {
 define i16 @clmul_i16(i16 %a, i16 %b) nounwind {
 ; SCALAR-LABEL: clmul_i16:
 ; SCALAR:       # %bb.0:
-; SCALAR-NEXT:    # kill: def $edi killed $edi def $rdi
-; SCALAR-NEXT:    leal (%rdi,%rdi), %eax
-; SCALAR-NEXT:    movl %esi, %ecx
-; SCALAR-NEXT:    andl $2, %ecx
-; SCALAR-NEXT:    cmovnel %eax, %ecx
-; SCALAR-NEXT:    movl %esi, %eax
-; SCALAR-NEXT:    andl $1, %eax
-; SCALAR-NEXT:    cmovnel %edi, %eax
-; SCALAR-NEXT:    xorl %ecx, %eax
-; SCALAR-NEXT:    leal (,%rdi,4), %ecx
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $4, %edx
-; SCALAR-NEXT:    cmovnel %ecx, %edx
-; SCALAR-NEXT:    leal (,%rdi,8), %r8d
-; SCALAR-NEXT:    movl %esi, %ecx
-; SCALAR-NEXT:    andl $8, %ecx
-; SCALAR-NEXT:    cmovnel %r8d, %ecx
-; SCALAR-NEXT:    xorl %edx, %ecx
-; SCALAR-NEXT:    xorl %eax, %ecx
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    shll $4, %eax
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $16, %edx
-; SCALAR-NEXT:    cmovnel %eax, %edx
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    shll $5, %eax
-; SCALAR-NEXT:    movl %esi, %r8d
-; SCALAR-NEXT:    andl $32, %r8d
-; SCALAR-NEXT:    cmovnel %eax, %r8d
-; SCALAR-NEXT:    xorl %edx, %r8d
-; SCALAR-NEXT:    movl %edi, %edx
-; SCALAR-NEXT:    shll $6, %edx
-; SCALAR-NEXT:    movl %esi, %eax
-; SCALAR-NEXT:    andl $64, %eax
-; SCALAR-NEXT:    cmovnel %edx, %eax
-; SCALAR-NEXT:    xorl %r8d, %eax
-; SCALAR-NEXT:    xorl %ecx, %eax
-; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    shll $7, %ecx
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $128, %edx
-; SCALAR-NEXT:    cmovnel %ecx, %edx
-; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    shll $8, %ecx
-; SCALAR-NEXT:    movl %esi, %r8d
-; SCALAR-NEXT:    andl $256, %r8d # imm = 0x100
-; SCALAR-NEXT:    cmovnel %ecx, %r8d
-; SCALAR-NEXT:    xorl %edx, %r8d
-; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    shll $9, %ecx
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $512, %edx # imm = 0x200
-; SCALAR-NEXT:    cmovnel %ecx, %edx
-; SCALAR-NEXT:    xorl %r8d, %edx
-; SCALAR-NEXT:    movl %edi, %r8d
-; SCALAR-NEXT:    shll $10, %r8d
+; SCALAR-NEXT:    pushq %rbp
+; SCALAR-NEXT:    pushq %r15
+; SCALAR-NEXT:    pushq %r14
+; SCALAR-NEXT:    pushq %r13
+; SCALAR-NEXT:    pushq %r12
+; SCALAR-NEXT:    pushq %rbx
 ; SCALAR-NEXT:    movl %esi, %ecx
-; SCALAR-NEXT:    andl $1024, %ecx # imm = 0x400
-; SCALAR-NEXT:    cmovnel %r8d, %ecx
-; SCALAR-NEXT:    xorl %edx, %ecx
-; SCALAR-NEXT:    xorl %eax, %ecx
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    shll $11, %eax
+; SCALAR-NEXT:    andl $4369, %ecx # imm = 0x1111
+; SCALAR-NEXT:    movl %edi, %r10d
+; SCALAR-NEXT:    andl $8738, %r10d # imm = 0x2222
+; SCALAR-NEXT:    movl %r10d, %r11d
+; SCALAR-NEXT:    imull %ecx, %r11d
 ; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $2048, %edx # imm = 0x800
-; SCALAR-NEXT:    cmovnel %eax, %edx
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    shll $12, %eax
+; SCALAR-NEXT:    andl $8738, %edx # imm = 0x2222
+; SCALAR-NEXT:    movl %edi, %ebx
+; SCALAR-NEXT:    andl $4369, %ebx # imm = 0x1111
 ; SCALAR-NEXT:    movl %esi, %r8d
-; SCALAR-NEXT:    andl $4096, %r8d # imm = 0x1000
-; SCALAR-NEXT:    cmovnel %eax, %r8d
-; SCALAR-NEXT:    xorl %edx, %r8d
+; SCALAR-NEXT:    andl $34952, %r8d # imm = 0x8888
+; SCALAR-NEXT:    movl %edi, %r9d
+; SCALAR-NEXT:    andl $17476, %r9d # imm = 0x4444
+; SCALAR-NEXT:    movl %r9d, %ebp
+; SCALAR-NEXT:    imull %r8d, %ebp
 ; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    shll $13, %eax
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $8192, %edx # imm = 0x2000
-; SCALAR-NEXT:    cmovnel %eax, %edx
-; SCALAR-NEXT:    xorl %r8d, %edx
-; SCALAR-NEXT:    movl %edi, %r8d
-; SCALAR-NEXT:    shll $14, %r8d
-; SCALAR-NEXT:    movl %esi, %eax
-; SCALAR-NEXT:    andl $16384, %eax # imm = 0x4000
-; SCALAR-NEXT:    cmovnel %r8d, %eax
-; SCALAR-NEXT:    xorl %edx, %eax
-; SCALAR-NEXT:    shll $15, %edi
-; SCALAR-NEXT:    andl $32768, %esi # imm = 0x8000
-; SCALAR-NEXT:    cmovnel %edi, %esi
-; SCALAR-NEXT:    xorl %esi, %eax
-; SCALAR-NEXT:    xorl %ecx, %eax
+; SCALAR-NEXT:    andl $34952, %eax # imm = 0x8888
+; SCALAR-NEXT:    movl %ebx, %r14d
+; SCALAR-NEXT:    movl %ebx, %r12d
+; SCALAR-NEXT:    movl %eax, %r15d
+; SCALAR-NEXT:    imull %r8d, %r15d
+; SCALAR-NEXT:    imull %ebx, %r8d
+; SCALAR-NEXT:    imull %edx, %ebx
+; SCALAR-NEXT:    xorl %r11d, %ebx
+; SCALAR-NEXT:    movl %esi, %r11d
+; SCALAR-NEXT:    andl $17476, %r11d # imm = 0x4444
+; SCALAR-NEXT:    movl %eax, %r13d
+; SCALAR-NEXT:    imull %r11d, %r13d
+; SCALAR-NEXT:    imull %r11d, %r12d
+; SCALAR-NEXT:    imull %r10d, %r11d
+; SCALAR-NEXT:    xorl %ebp, %r13d
+; SCALAR-NEXT:    xorl %ebx, %r13d
+; SCALAR-NEXT:    movl %esi, %ebx
+; SCALAR-NEXT:    andl $2184, %ebx # imm = 0x888
+; SCALAR-NEXT:    movl %edi, %ebp
+; SCALAR-NEXT:    andl $546, %ebp # imm = 0x222
+; SCALAR-NEXT:    imull %ebx, %ebp
+; SCALAR-NEXT:    imull %ecx, %r14d
+; SCALAR-NEXT:    xorl %ebp, %r14d
+; SCALAR-NEXT:    movl %esi, %ebx
+; SCALAR-NEXT:    andl $1092, %ebx # imm = 0x444
+; SCALAR-NEXT:    movl %edi, %ebp
+; SCALAR-NEXT:    andl $1092, %ebp # imm = 0x444
+; SCALAR-NEXT:    imull %ebx, %ebp
+; SCALAR-NEXT:    andl $8738, %r13d # imm = 0x2222
+; SCALAR-NEXT:    andl $546, %esi # imm = 0x222
+; SCALAR-NEXT:    andl $2184, %edi # imm = 0x888
+; SCALAR-NEXT:    imull %esi, %edi
+; SCALAR-NEXT:    xorl %edi, %ebp
+; SCALAR-NEXT:    xorl %r14d, %ebp
+; SCALAR-NEXT:    andl $4369, %ebp # imm = 0x1111
+; SCALAR-NEXT:    orl %r13d, %ebp
+; SCALAR-NEXT:    imull %edx, %r10d
+; SCALAR-NEXT:    xorl %r10d, %r12d
+; SCALAR-NEXT:    movl %r9d, %esi
+; SCALAR-NEXT:    imull %ecx, %esi
+; SCALAR-NEXT:    xorl %esi, %r15d
+; SCALAR-NEXT:    xorl %r12d, %r15d
+; SCALAR-NEXT:    andl $17476, %r15d # imm = 0x4444
+; SCALAR-NEXT:    orl %ebp, %r15d
+; SCALAR-NEXT:    xorl %r11d, %r8d
+; SCALAR-NEXT:    imull %edx, %r9d
+; SCALAR-NEXT:    imull %ecx, %eax
+; SCALAR-NEXT:    xorl %r9d, %eax
+; SCALAR-NEXT:    xorl %r8d, %eax
+; SCALAR-NEXT:    andl $268404872, %eax # imm = 0xFFF8888
+; SCALAR-NEXT:    orl %r15d, %eax
 ; SCALAR-NEXT:    # kill: def $ax killed $ax killed $eax
+; SCALAR-NEXT:    popq %rbx
+; SCALAR-NEXT:    popq %r12
+; SCALAR-NEXT:    popq %r13
+; SCALAR-NEXT:    popq %r14
+; SCALAR-NEXT:    popq %r15
+; SCALAR-NEXT:    popq %rbp
 ; SCALAR-NEXT:    retq
 ;
 ; SSE-PCLMUL-LABEL: clmul_i16:
@@ -195,191 +186,77 @@ define i16 @clmul_i16(i16 %a, i16 %b) nounwind {
 define i32 @clmul_i32(i32 %a, i32 %b) nounwind {
 ; SCALAR-LABEL: clmul_i32:
 ; SCALAR:       # %bb.0:
-; SCALAR-NEXT:    # kill: def $edi killed $edi def $rdi
-; SCALAR-NEXT:    leal (%rdi,%rdi), %eax
-; SCALAR-NEXT:    movl %esi, %ecx
-; SCALAR-NEXT:    andl $2, %ecx
-; SCALAR-NEXT:    cmovnel %eax, %ecx
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $1, %edx
-; SCALAR-NEXT:    cmovnel %edi, %edx
-; SCALAR-NEXT:    xorl %ecx, %edx
-; SCALAR-NEXT:    leal (,%rdi,4), %eax
-; SCALAR-NEXT:    movl %esi, %ecx
-; SCALAR-NEXT:    andl $4, %ecx
-; SCALAR-NEXT:    cmovnel %eax, %ecx
-; SCALAR-NEXT:    leal (,%rdi,8), %r8d
-; SCALAR-NEXT:    movl %esi, %eax
-; SCALAR-NEXT:    andl $8, %eax
-; SCALAR-NEXT:    cmovnel %r8d, %eax
-; SCALAR-NEXT:    xorl %ecx, %eax
-; SCALAR-NEXT:    xorl %edx, %eax
-; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    shll $4, %ecx
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $16, %edx
-; SCALAR-NEXT:    cmovnel %ecx, %edx
-; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    shll $5, %ecx
-; SCALAR-NEXT:    movl %esi, %r8d
-; SCALAR-NEXT:    andl $32, %r8d
-; SCALAR-NEXT:    cmovnel %ecx, %r8d
-; SCALAR-NEXT:    xorl %edx, %r8d
-; SCALAR-NEXT:    movl %edi, %edx
-; SCALAR-NEXT:    shll $6, %edx
-; SCALAR-NEXT:    movl %esi, %ecx
-; SCALAR-NEXT:    andl $64, %ecx
-; SCALAR-NEXT:    cmovnel %edx, %ecx
-; SCALAR-NEXT:    xorl %r8d, %ecx
-; SCALAR-NEXT:    xorl %eax, %ecx
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    shll $7, %eax
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $128, %edx
-; SCALAR-NEXT:    cmovnel %eax, %edx
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    shll $8, %eax
-; SCALAR-NEXT:    movl %esi, %r8d
-; SCALAR-NEXT:    andl $256, %r8d # imm = 0x100
-; SCALAR-NEXT:    cmovnel %eax, %r8d
-; SCALAR-NEXT:    xorl %edx, %r8d
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    shll $9, %eax
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $512, %edx # imm = 0x200
-; SCALAR-NEXT:    cmovnel %eax, %edx
-; SCALAR-NEXT:    xorl %r8d, %edx
-; SCALAR-NEXT:    movl %edi, %r8d
-; SCALAR-NEXT:    shll $10, %r8d
-; SCALAR-NEXT:    movl %esi, %eax
-; SCALAR-NEXT:    andl $1024, %eax # imm = 0x400
-; SCALAR-NEXT:    cmovnel %r8d, %eax
-; SCALAR-NEXT:    xorl %edx, %eax
-; SCALAR-NEXT:    xorl %ecx, %eax
-; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    shll $11, %ecx
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $2048, %edx # imm = 0x800
-; SCALAR-NEXT:    cmovnel %ecx, %edx
-; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    shll $12, %ecx
-; SCALAR-NEXT:    movl %esi, %r8d
-; SCALAR-NEXT:    andl $4096, %r8d # imm = 0x1000
-; SCALAR-NEXT:    cmovnel %ecx, %r8d
-; SCALAR-NEXT:    xorl %edx, %r8d
-; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    shll $13, %ecx
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $8192, %edx # imm = 0x2000
-; SCALAR-NEXT:    cmovnel %ecx, %edx
-; SCALAR-NEXT:    xorl %r8d, %edx
-; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    shll $14, %ecx
-; SCALAR-NEXT:    movl %esi, %r8d
-; SCALAR-NEXT:    andl $16384, %r8d # imm = 0x4000
-; SCALAR-NEXT:    cmovnel %ecx, %r8d
-; SCALAR-NEXT:    xorl %edx, %r8d
-; SCALAR-NEXT:    movl %edi, %edx
-; SCALAR-NEXT:    shll $15, %edx
-; SCALAR-NEXT:    movl %esi, %ecx
-; SCALAR-NEXT:    andl $32768, %ecx # imm = 0x8000
-; SCALAR-NEXT:    cmovnel %edx, %ecx
-; SCALAR-NEXT:    xorl %r8d, %ecx
-; SCALAR-NEXT:    xorl %eax, %ecx
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    shll $16, %eax
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $65536, %edx # imm = 0x10000
-; SCALAR-NEXT:    cmovnel %eax, %edx
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    shll $17, %eax
-; SCALAR-NEXT:    movl %esi, %r8d
-; SCALAR-NEXT:    andl $131072, %r8d # imm = 0x20000
-; SCALAR-NEXT:    cmovnel %eax, %r8d
-; SCALAR-NEXT:    xorl %edx, %r8d
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    shll $18, %eax
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $262144, %edx # imm = 0x40000
-; SCALAR-NEXT:    cmovnel %eax, %edx
-; SCALAR-NEXT:    xorl %r8d, %edx
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    shll $19, %eax
-; SCALAR-NEXT:    movl %esi, %r8d
-; SCALAR-NEXT:    andl $524288, %r8d # imm = 0x80000
-; SCALAR-NEXT:    cmovnel %eax, %r8d
-; SCALAR-NEXT:    xorl %edx, %r8d
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    shll $20, %eax
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $1048576, %edx # imm = 0x100000
-; SCALAR-NEXT:    cmovnel %eax, %edx
-; SCALAR-NEXT:    xorl %r8d, %edx
-; SCALAR-NEXT:    movl %edi, %r8d
-; SCALAR-NEXT:    shll $21, %r8d
+; SCALAR-NEXT:    pushq %rbp
+; SCALAR-NEXT:    pushq %r15
+; SCALAR-NEXT:    pushq %r14
+; SCALAR-NEXT:    pushq %r13
+; SCALAR-NEXT:    pushq %r12
+; SCALAR-NEXT:    pushq %rbx
 ; SCALAR-NEXT:    movl %esi, %eax
-; SCALAR-NEXT:    andl $2097152, %eax # imm = 0x200000
-; SCALAR-NEXT:    cmovnel %r8d, %eax
-; SCALAR-NEXT:    xorl %edx, %eax
-; SCALAR-NEXT:    xorl %ecx, %eax
-; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    shll $22, %ecx
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $4194304, %edx # imm = 0x400000
-; SCALAR-NEXT:    cmovnel %ecx, %edx
-; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    shll $23, %ecx
-; SCALAR-NEXT:    movl %esi, %r8d
-; SCALAR-NEXT:    andl $8388608, %r8d # imm = 0x800000
-; SCALAR-NEXT:    cmovnel %ecx, %r8d
-; SCALAR-NEXT:    xorl %edx, %r8d
-; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    shll $24, %ecx
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $16777216, %edx # imm = 0x1000000
-; SCALAR-NEXT:    cmovnel %ecx, %edx
-; SCALAR-NEXT:    xorl %r8d, %edx
+; SCALAR-NEXT:    andl $286331153, %eax # imm = 0x11111111
 ; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    shll $25, %ecx
-; SCALAR-NEXT:    movl %esi, %r8d
-; SCALAR-NEXT:    andl $33554432, %r8d # imm = 0x2000000
-; SCALAR-NEXT:    cmovnel %ecx, %r8d
-; SCALAR-NEXT:    xorl %edx, %r8d
-; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    shll $26, %ecx
+; SCALAR-NEXT:    andl $572662306, %ecx # imm = 0x22222222
+; SCALAR-NEXT:    movl %ecx, %r10d
+; SCALAR-NEXT:    imull %eax, %r10d
 ; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $67108864, %edx # imm = 0x4000000
-; SCALAR-NEXT:    cmovnel %ecx, %edx
-; SCALAR-NEXT:    xorl %r8d, %edx
-; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    shll $27, %ecx
+; SCALAR-NEXT:    andl $572662306, %edx # imm = 0x22222222
+; SCALAR-NEXT:    movl %edi, %ebx
+; SCALAR-NEXT:    andl $286331153, %ebx # imm = 0x11111111
 ; SCALAR-NEXT:    movl %esi, %r8d
-; SCALAR-NEXT:    andl $134217728, %r8d # imm = 0x8000000
-; SCALAR-NEXT:    cmovnel %ecx, %r8d
-; SCALAR-NEXT:    xorl %edx, %r8d
-; SCALAR-NEXT:    movl %edi, %edx
-; SCALAR-NEXT:    shll $28, %edx
-; SCALAR-NEXT:    movl %esi, %ecx
-; SCALAR-NEXT:    andl $268435456, %ecx # imm = 0x10000000
-; SCALAR-NEXT:    cmovnel %edx, %ecx
-; SCALAR-NEXT:    xorl %r8d, %ecx
-; SCALAR-NEXT:    xorl %eax, %ecx
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    shll $29, %eax
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $536870912, %edx # imm = 0x20000000
-; SCALAR-NEXT:    cmovnel %eax, %edx
-; SCALAR-NEXT:    movl %edi, %r8d
-; SCALAR-NEXT:    shll $30, %r8d
-; SCALAR-NEXT:    movl %esi, %eax
-; SCALAR-NEXT:    andl $1073741824, %eax # imm = 0x40000000
-; SCALAR-NEXT:    cmovnel %r8d, %eax
-; SCALAR-NEXT:    xorl %edx, %eax
-; SCALAR-NEXT:    shll $31, %edi
-; SCALAR-NEXT:    andl $-2147483648, %esi # imm = 0x80000000
-; SCALAR-NEXT:    cmovnel %edi, %esi
-; SCALAR-NEXT:    xorl %esi, %eax
-; SCALAR-NEXT:    xorl %ecx, %eax
+; SCALAR-NEXT:    andl $-2004318072, %r8d # imm = 0x88888888
+; SCALAR-NEXT:    movl %edi, %r9d
+; SCALAR-NEXT:    andl $1145324612, %r9d # imm = 0x44444444
+; SCALAR-NEXT:    movl %r9d, %r11d
+; SCALAR-NEXT:    imull %r8d, %r11d
+; SCALAR-NEXT:    andl $-2004318072, %edi # imm = 0x88888888
+; SCALAR-NEXT:    movl %ecx, %r15d
+; SCALAR-NEXT:    imull %r8d, %r15d
+; SCALAR-NEXT:    movl %ebx, %ebp
+; SCALAR-NEXT:    movl %ebx, %r12d
+; SCALAR-NEXT:    movl %edi, %r14d
+; SCALAR-NEXT:    imull %r8d, %r14d
+; SCALAR-NEXT:    imull %ebx, %r8d
+; SCALAR-NEXT:    imull %edx, %ebx
+; SCALAR-NEXT:    xorl %r10d, %ebx
+; SCALAR-NEXT:    andl $1145324612, %esi # imm = 0x44444444
+; SCALAR-NEXT:    imull %eax, %ebp
+; SCALAR-NEXT:    movl %edi, %r10d
+; SCALAR-NEXT:    movl %r9d, %r13d
+; SCALAR-NEXT:    imull %eax, %r13d
+; SCALAR-NEXT:    imull %edi, %eax
+; SCALAR-NEXT:    imull %esi, %edi
+; SCALAR-NEXT:    xorl %r11d, %edi
+; SCALAR-NEXT:    xorl %ebx, %edi
+; SCALAR-NEXT:    xorl %r15d, %ebp
+; SCALAR-NEXT:    movl %r9d, %r11d
+; SCALAR-NEXT:    imull %esi, %r11d
+; SCALAR-NEXT:    imull %edx, %r10d
+; SCALAR-NEXT:    xorl %r11d, %r10d
+; SCALAR-NEXT:    andl $572662306, %edi # imm = 0x22222222
+; SCALAR-NEXT:    xorl %ebp, %r10d
+; SCALAR-NEXT:    andl $286331153, %r10d # imm = 0x11111111
+; SCALAR-NEXT:    orl %edi, %r10d
+; SCALAR-NEXT:    movl %ecx, %edi
+; SCALAR-NEXT:    imull %edx, %edi
+; SCALAR-NEXT:    imull %esi, %r12d
+; SCALAR-NEXT:    xorl %edi, %r12d
+; SCALAR-NEXT:    xorl %r13d, %r14d
+; SCALAR-NEXT:    xorl %r12d, %r14d
+; SCALAR-NEXT:    andl $1145324612, %r14d # imm = 0x44444444
+; SCALAR-NEXT:    orl %r10d, %r14d
+; SCALAR-NEXT:    imull %esi, %ecx
+; SCALAR-NEXT:    xorl %ecx, %r8d
+; SCALAR-NEXT:    imull %edx, %r9d
+; SCALAR-NEXT:    xorl %r9d, %eax
+; SCALAR-NEXT:    xorl %r8d, %eax
+; SCALAR-NEXT:    andl $-2004318072, %eax # imm = 0x88888888
+; SCALAR-NEXT:    orl %r14d, %eax
+; SCALAR-NEXT:    popq %rbx
+; SCALAR-NEXT:    popq %r12
+; SCALAR-NEXT:    popq %r13
+; SCALAR-NEXT:    popq %r14
+; SCALAR-NEXT:    popq %r15
+; SCALAR-NEXT:    popq %rbp
 ; SCALAR-NEXT:    retq
 ;
 ; SSE-PCLMUL-LABEL: clmul_i32:
@@ -406,352 +283,81 @@ define i32 @clmul_i32(i32 %a, i32 %b) nounwind {
 define i64 @clmul_i64(i64 %a, i64 %b) nounwind {
 ; SCALAR-LABEL: clmul_i64:
 ; SCALAR:       # %bb.0:
-; SCALAR-NEXT:    leaq (%rdi,%rdi), %rax
-; SCALAR-NEXT:    movl %esi, %ecx
-; SCALAR-NEXT:    andl $2, %ecx
-; SCALAR-NEXT:    cmovneq %rax, %rcx
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $1, %edx
-; SCALAR-NEXT:    cmovneq %rdi, %rdx
-; SCALAR-NEXT:    xorq %rcx, %rdx
-; SCALAR-NEXT:    leaq (,%rdi,4), %rax
-; SCALAR-NEXT:    movl %esi, %ecx
-; SCALAR-NEXT:    andl $4, %ecx
-; SCALAR-NEXT:    cmovneq %rax, %rcx
-; SCALAR-NEXT:    leaq (,%rdi,8), %r8
-; SCALAR-NEXT:    movl %esi, %eax
-; SCALAR-NEXT:    andl $8, %eax
-; SCALAR-NEXT:    cmovneq %r8, %rax
-; SCALAR-NEXT:    xorq %rcx, %rax
-; SCALAR-NEXT:    xorq %rdx, %rax
-; SCALAR-NEXT:    movq %rdi, %rcx
-; SCALAR-NEXT:    shlq $4, %rcx
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $16, %edx
-; SCALAR-NEXT:    cmovneq %rcx, %rdx
-; SCALAR-NEXT:    movq %rdi, %rcx
-; SCALAR-NEXT:    shlq $5, %rcx
-; SCALAR-NEXT:    movl %esi, %r8d
-; SCALAR-NEXT:    andl $32, %r8d
-; SCALAR-NEXT:    cmovneq %rcx, %r8
-; SCALAR-NEXT:    xorq %rdx, %r8
-; SCALAR-NEXT:    movq %rdi, %rdx
-; SCALAR-NEXT:    shlq $6, %rdx
-; SCALAR-NEXT:    movl %esi, %ecx
-; SCALAR-NEXT:    andl $64, %ecx
-; SCALAR-NEXT:    cmovneq %rdx, %rcx
-; SCALAR-NEXT:    xorq %r8, %rcx
-; SCALAR-NEXT:    xorq %rax, %rcx
-; SCALAR-NEXT:    movq %rdi, %rax
-; SCALAR-NEXT:    shlq $7, %rax
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $128, %edx
-; SCALAR-NEXT:    cmovneq %rax, %rdx
-; SCALAR-NEXT:    movq %rdi, %rax
-; SCALAR-NEXT:    shlq $8, %rax
-; SCALAR-NEXT:    movl %esi, %r8d
-; SCALAR-NEXT:    andl $256, %r8d # imm = 0x100
-; SCALAR-NEXT:    cmovneq %rax, %r8
-; SCALAR-NEXT:    xorq %rdx, %r8
-; SCALAR-NEXT:    movq %rdi, %rax
-; SCALAR-NEXT:    shlq $9, %rax
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $512, %edx # imm = 0x200
-; SCALAR-NEXT:    cmovneq %rax, %rdx
-; SCALAR-NEXT:    xorq %r8, %rdx
-; SCALAR-NEXT:    movq %rdi, %r8
-; SCALAR-NEXT:    shlq $10, %r8
-; SCALAR-NEXT:    movl %esi, %eax
-; SCALAR-NEXT:    andl $1024, %eax # imm = 0x400
-; SCALAR-NEXT:    cmovneq %r8, %rax
-; SCALAR-NEXT:    xorq %rdx, %rax
-; SCALAR-NEXT:    xorq %rcx, %rax
-; SCALAR-NEXT:    movq %rdi, %rcx
-; SCALAR-NEXT:    shlq $11, %rcx
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $2048, %edx # imm = 0x800
-; SCALAR-NEXT:    cmovneq %rcx, %rdx
-; SCALAR-NEXT:    movq %rdi, %rcx
-; SCALAR-NEXT:    shlq $12, %rcx
-; SCALAR-NEXT:    movl %esi, %r8d
-; SCALAR-NEXT:    andl $4096, %r8d # imm = 0x1000
-; SCALAR-NEXT:    cmovneq %rcx, %r8
-; SCALAR-NEXT:    xorq %rdx, %r8
-; SCALAR-NEXT:    movq %rdi, %rcx
-; SCALAR-NEXT:    shlq $13, %rcx
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $8192, %edx # imm = 0x2000
-; SCALAR-NEXT:    cmovneq %rcx, %rdx
-; SCALAR-NEXT:    xorq %r8, %rdx
-; SCALAR-NEXT:    movq %rdi, %rcx
-; SCALAR-NEXT:    shlq $14, %rcx
-; SCALAR-NEXT:    movl %esi, %r8d
-; SCALAR-NEXT:    andl $16384, %r8d # imm = 0x4000
-; SCALAR-NEXT:    cmovneq %rcx, %r8
-; SCALAR-NEXT:    xorq %rdx, %r8
-; SCALAR-NEXT:    movq %rdi, %rdx
-; SCALAR-NEXT:    shlq $15, %rdx
-; SCALAR-NEXT:    movl %esi, %ecx
-; SCALAR-NEXT:    andl $32768, %ecx # imm = 0x8000
-; SCALAR-NEXT:    cmovneq %rdx, %rcx
-; SCALAR-NEXT:    xorq %r8, %rcx
-; SCALAR-NEXT:    xorq %rax, %rcx
-; SCALAR-NEXT:    movq %rdi, %rax
-; SCALAR-NEXT:    shlq $16, %rax
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $65536, %edx # imm = 0x10000
-; SCALAR-NEXT:    cmovneq %rax, %rdx
-; SCALAR-NEXT:    movq %rdi, %rax
-; SCALAR-NEXT:    shlq $17, %rax
-; SCALAR-NEXT:    movl %esi, %r8d
-; SCALAR-NEXT:    andl $131072, %r8d # imm = 0x20000
-; SCALAR-NEXT:    cmovneq %rax, %r8
-; SCALAR-NEXT:    xorq %rdx, %r8
-; SCALAR-NEXT:    movq %rdi, %rax
-; SCALAR-NEXT:    shlq $18, %rax
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $262144, %edx # imm = 0x40000
-; SCALAR-NEXT:    cmovneq %rax, %rdx
-; SCALAR-NEXT:    xorq %r8, %rdx
-; SCALAR-NEXT:    movq %rdi, %rax
-; SCALAR-NEXT:    shlq $19, %rax
-; SCALAR-NEXT:    movl %esi, %r8d
-; SCALAR-NEXT:    andl $524288, %r8d # imm = 0x80000
-; SCALAR-NEXT:    cmovneq %rax, %r8
-; SCALAR-NEXT:    xorq %rdx, %r8
-; SCALAR-NEXT:    movq %rdi, %rax
-; SCALAR-NEXT:    shlq $20, %rax
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $1048576, %edx # imm = 0x100000
-; SCALAR-NEXT:    cmovneq %rax, %rdx
-; SCALAR-NEXT:    xorq %r8, %rdx
-; SCALAR-NEXT:    movq %rdi, %r8
-; SCALAR-NEXT:    shlq $21, %r8
-; SCALAR-NEXT:    movl %esi, %eax
-; SCALAR-NEXT:    andl $2097152, %eax # imm = 0x200000
-; SCALAR-NEXT:    cmovneq %r8, %rax
-; SCALAR-NEXT:    xorq %rdx, %rax
-; SCALAR-NEXT:    xorq %rcx, %rax
-; SCALAR-NEXT:    movq %rdi, %rcx
-; SCALAR-NEXT:    shlq $22, %rcx
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $4194304, %edx # imm = 0x400000
-; SCALAR-NEXT:    cmovneq %rcx, %rdx
-; SCALAR-NEXT:    movq %rdi, %rcx
-; SCALAR-NEXT:    shlq $23, %rcx
-; SCALAR-NEXT:    movl %esi, %r8d
-; SCALAR-NEXT:    andl $8388608, %r8d # imm = 0x800000
-; SCALAR-NEXT:    cmovneq %rcx, %r8
-; SCALAR-NEXT:    xorq %rdx, %r8
-; SCALAR-NEXT:    movq %rdi, %rcx
-; SCALAR-NEXT:    shlq $24, %rcx
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $16777216, %edx # imm = 0x1000000
-; SCALAR-NEXT:    cmovneq %rcx, %rdx
-; SCALAR-NEXT:    xorq %r8, %rdx
-; SCALAR-NEXT:    movq %rdi, %rcx
-; SCALAR-NEXT:    shlq $25, %rcx
-; SCALAR-NEXT:    movl %esi, %r8d
-; SCALAR-NEXT:    andl $33554432, %r8d # imm = 0x2000000
-; SCALAR-NEXT:    cmovneq %rcx, %r8
-; SCALAR-NEXT:    xorq %rdx, %r8
-; SCALAR-NEXT:    movq %rdi, %rcx
-; SCALAR-NEXT:    shlq $26, %rcx
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $67108864, %edx # imm = 0x4000000
-; SCALAR-NEXT:    cmovneq %rcx, %rdx
-; SCALAR-NEXT:    xorq %r8, %rdx
-; SCALAR-NEXT:    movq %rdi, %rcx
-; SCALAR-NEXT:    shlq $27, %rcx
-; SCALAR-NEXT:    movl %esi, %r8d
-; SCALAR-NEXT:    andl $134217728, %r8d # imm = 0x8000000
-; SCALAR-NEXT:    cmovneq %rcx, %r8
-; SCALAR-NEXT:    xorq %rdx, %r8
-; SCALAR-NEXT:    movq %rdi, %rdx
-; SCALAR-NEXT:    shlq $28, %rdx
-; SCALAR-NEXT:    movl %esi, %ecx
-; SCALAR-NEXT:    andl $268435456, %ecx # imm = 0x10000000
-; SCALAR-NEXT:    cmovneq %rdx, %rcx
-; SCALAR-NEXT:    xorq %r8, %rcx
-; SCALAR-NEXT:    xorq %rax, %rcx
-; SCALAR-NEXT:    movq %rdi, %rax
-; SCALAR-NEXT:    shlq $29, %rax
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $536870912, %edx # imm = 0x20000000
-; SCALAR-NEXT:    cmovneq %rax, %rdx
-; SCALAR-NEXT:    movq %rdi, %rax
-; SCALAR-NEXT:    shlq $30, %rax
-; SCALAR-NEXT:    movl %esi, %r8d
-; SCALAR-NEXT:    andl $1073741824, %r8d # imm = 0x40000000
-; SCALAR-NEXT:    cmovneq %rax, %r8
-; SCALAR-NEXT:    xorq %rdx, %r8
-; SCALAR-NEXT:    movq %rdi, %rax
-; SCALAR-NEXT:    shlq $31, %rax
-; SCALAR-NEXT:    movl %esi, %edx
-; SCALAR-NEXT:    andl $-2147483648, %edx # imm = 0x80000000
-; SCALAR-NEXT:    cmovneq %rax, %rdx
-; SCALAR-NEXT:    xorq %r8, %rdx
-; SCALAR-NEXT:    movq %rdi, %r8
-; SCALAR-NEXT:    shlq $32, %r8
-; SCALAR-NEXT:    xorl %eax, %eax
-; SCALAR-NEXT:    btq $32, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %r8
-; SCALAR-NEXT:    xorq %rdx, %r8
-; SCALAR-NEXT:    movq %rdi, %rdx
-; SCALAR-NEXT:    shlq $33, %rdx
-; SCALAR-NEXT:    btq $33, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %rdx
-; SCALAR-NEXT:    xorq %r8, %rdx
-; SCALAR-NEXT:    movq %rdi, %r8
-; SCALAR-NEXT:    shlq $34, %r8
-; SCALAR-NEXT:    btq $34, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %r8
-; SCALAR-NEXT:    xorq %rdx, %r8
-; SCALAR-NEXT:    movq %rdi, %r9
-; SCALAR-NEXT:    shlq $35, %r9
-; SCALAR-NEXT:    btq $35, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %r9
-; SCALAR-NEXT:    xorq %r8, %r9
-; SCALAR-NEXT:    movq %rdi, %rdx
-; SCALAR-NEXT:    shlq $36, %rdx
-; SCALAR-NEXT:    btq $36, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %rdx
-; SCALAR-NEXT:    xorq %r9, %rdx
-; SCALAR-NEXT:    xorq %rcx, %rdx
-; SCALAR-NEXT:    movq %rdi, %rcx
-; SCALAR-NEXT:    shlq $37, %rcx
-; SCALAR-NEXT:    btq $37, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %rcx
-; SCALAR-NEXT:    movq %rdi, %r8
-; SCALAR-NEXT:    shlq $38, %r8
-; SCALAR-NEXT:    btq $38, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %r8
-; SCALAR-NEXT:    xorq %rcx, %r8
-; SCALAR-NEXT:    movq %rdi, %rcx
-; SCALAR-NEXT:    shlq $39, %rcx
-; SCALAR-NEXT:    btq $39, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %rcx
-; SCALAR-NEXT:    xorq %r8, %rcx
-; SCALAR-NEXT:    movq %rdi, %r8
-; SCALAR-NEXT:    shlq $40, %r8
-; SCALAR-NEXT:    btq $40, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %r8
-; SCALAR-NEXT:    xorq %rcx, %r8
-; SCALAR-NEXT:    movq %rdi, %rcx
-; SCALAR-NEXT:    shlq $41, %rcx
-; SCALAR-NEXT:    btq $41, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %rcx
-; SCALAR-NEXT:    xorq %r8, %rcx
-; SCALAR-NEXT:    movq %rdi, %r8
-; SCALAR-NEXT:    shlq $42, %r8
-; SCALAR-NEXT:    btq $42, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %r8
-; SCALAR-NEXT:    xorq %rcx, %r8
-; SCALAR-NEXT:    movq %rdi, %rcx
-; SCALAR-NEXT:    shlq $43, %rcx
-; SCALAR-NEXT:    btq $43, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %rcx
-; SCALAR-NEXT:    xorq %r8, %rcx
-; SCALAR-NEXT:    movq %rdi, %r8
-; SCALAR-NEXT:    shlq $44, %r8
-; SCALAR-NEXT:    btq $44, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %r8
-; SCALAR-NEXT:    xorq %rcx, %r8
-; SCALAR-NEXT:    movq %rdi, %r9
-; SCALAR-NEXT:    shlq $45, %r9
-; SCALAR-NEXT:    btq $45, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %r9
-; SCALAR-NEXT:    xorq %r8, %r9
-; SCALAR-NEXT:    movq %rdi, %rcx
-; SCALAR-NEXT:    shlq $46, %rcx
-; SCALAR-NEXT:    btq $46, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %rcx
-; SCALAR-NEXT:    xorq %r9, %rcx
-; SCALAR-NEXT:    xorq %rdx, %rcx
-; SCALAR-NEXT:    movq %rdi, %rdx
-; SCALAR-NEXT:    shlq $47, %rdx
-; SCALAR-NEXT:    btq $47, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %rdx
-; SCALAR-NEXT:    movq %rdi, %r8
-; SCALAR-NEXT:    shlq $48, %r8
-; SCALAR-NEXT:    btq $48, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %r8
-; SCALAR-NEXT:    xorq %rdx, %r8
-; SCALAR-NEXT:    movq %rdi, %rdx
-; SCALAR-NEXT:    shlq $49, %rdx
-; SCALAR-NEXT:    btq $49, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %rdx
-; SCALAR-NEXT:    xorq %r8, %rdx
-; SCALAR-NEXT:    movq %rdi, %r8
-; SCALAR-NEXT:    shlq $50, %r8
-; SCALAR-NEXT:    btq $50, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %r8
-; SCALAR-NEXT:    xorq %rdx, %r8
-; SCALAR-NEXT:    movq %rdi, %rdx
-; SCALAR-NEXT:    shlq $51, %rdx
-; SCALAR-NEXT:    btq $51, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %rdx
-; SCALAR-NEXT:    xorq %r8, %rdx
-; SCALAR-NEXT:    movq %rdi, %r8
-; SCALAR-NEXT:    shlq $52, %r8
-; SCALAR-NEXT:    btq $52, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %r8
-; SCALAR-NEXT:    xorq %rdx, %r8
-; SCALAR-NEXT:    movq %rdi, %rdx
-; SCALAR-NEXT:    shlq $53, %rdx
-; SCALAR-NEXT:    btq $53, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %rdx
-; SCALAR-NEXT:    xorq %r8, %rdx
-; SCALAR-NEXT:    movq %rdi, %r8
-; SCALAR-NEXT:    shlq $54, %r8
-; SCALAR-NEXT:    btq $54, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %r8
-; SCALAR-NEXT:    xorq %rdx, %r8
-; SCALAR-NEXT:    movq %rdi, %rdx
-; SCALAR-NEXT:    shlq $55, %rdx
-; SCALAR-NEXT:    btq $55, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %rdx
-; SCALAR-NEXT:    xorq %r8, %rdx
-; SCALAR-NEXT:    movq %rdi, %r8
-; SCALAR-NEXT:    shlq $56, %r8
-; SCALAR-NEXT:    btq $56, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %r8
-; SCALAR-NEXT:    xorq %rdx, %r8
-; SCALAR-NEXT:    movq %rdi, %rdx
-; SCALAR-NEXT:    shlq $57, %rdx
-; SCALAR-NEXT:    btq $57, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %rdx
-; SCALAR-NEXT:    xorq %r8, %rdx
-; SCALAR-NEXT:    xorq %rcx, %rdx
-; SCALAR-NEXT:    movq %rdi, %rcx
-; SCALAR-NEXT:    shlq $58, %rcx
-; SCALAR-NEXT:    btq $58, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %rcx
-; SCALAR-NEXT:    movq %rdi, %r8
-; SCALAR-NEXT:    shlq $59, %r8
-; SCALAR-NEXT:    btq $59, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %r8
-; SCALAR-NEXT:    xorq %rcx, %r8
-; SCALAR-NEXT:    movq %rdi, %rcx
-; SCALAR-NEXT:    shlq $60, %rcx
-; SCALAR-NEXT:    btq $60, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %rcx
-; SCALAR-NEXT:    xorq %r8, %rcx
-; SCALAR-NEXT:    movq %rdi, %r8
-; SCALAR-NEXT:    shlq $61, %r8
-; SCALAR-NEXT:    btq $61, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %r8
-; SCALAR-NEXT:    xorq %rcx, %r8
+; SCALAR-NEXT:    pushq %rbp
+; SCALAR-NEXT:    pushq %r15
+; SCALAR-NEXT:    pushq %r14
+; SCALAR-NEXT:    pushq %r13
+; SCALAR-NEXT:    pushq %r12
+; SCALAR-NEXT:    pushq %rbx
+; SCALAR-NEXT:    movabsq $1229782938247303441, %r9 # imm = 0x1111111111111111
+; SCALAR-NEXT:    movq %rsi, %rax
+; SCALAR-NEXT:    andq %r9, %rax
+; SCALAR-NEXT:    movabsq $2459565876494606882, %rbp # imm = 0x2222222222222222
 ; SCALAR-NEXT:    movq %rdi, %rcx
-; SCALAR-NEXT:    shlq $62, %rcx
-; SCALAR-NEXT:    btq $62, %rsi
-; SCALAR-NEXT:    cmovaeq %rax, %rcx
-; SCALAR-NEXT:    xorq %r8, %rcx
-; SCALAR-NEXT:    shlq $63, %rdi
-; SCALAR-NEXT:    btq $63, %rsi
-; SCALAR-NEXT:    cmovbq %rdi, %rax
-; SCALAR-NEXT:    xorq %rcx, %rax
-; SCALAR-NEXT:    xorq %rdx, %rax
+; SCALAR-NEXT:    andq %rbp, %rcx
+; SCALAR-NEXT:    movq %rcx, %r8
+; SCALAR-NEXT:    imulq %rax, %r8
+; SCALAR-NEXT:    movq %rsi, %rdx
+; SCALAR-NEXT:    andq %rbp, %rdx
+; SCALAR-NEXT:    movq %rdi, %r14
+; SCALAR-NEXT:    andq %r9, %r14
+; SCALAR-NEXT:    movq %r14, %r12
+; SCALAR-NEXT:    imulq %rdx, %r12
+; SCALAR-NEXT:    xorq %r8, %r12
+; SCALAR-NEXT:    movabsq $-8608480567731124088, %r8 # imm = 0x8888888888888888
+; SCALAR-NEXT:    movq %rsi, %r10
+; SCALAR-NEXT:    andq %r8, %r10
+; SCALAR-NEXT:    movabsq $4919131752989213764, %rbx # imm = 0x4444444444444444
+; SCALAR-NEXT:    movq %rdi, %r11
+; SCALAR-NEXT:    andq %rbx, %r11
+; SCALAR-NEXT:    movq %r11, %r13
+; SCALAR-NEXT:    imulq %r10, %r13
+; SCALAR-NEXT:    andq %rbx, %rsi
+; SCALAR-NEXT:    andq %r8, %rdi
+; SCALAR-NEXT:    movq %rdi, %r15
+; SCALAR-NEXT:    imulq %rsi, %r15
+; SCALAR-NEXT:    xorq %r13, %r15
+; SCALAR-NEXT:    movq %r14, %r13
+; SCALAR-NEXT:    xorq %r12, %r15
+; SCALAR-NEXT:    movq %rdi, %r12
+; SCALAR-NEXT:    andq %rbp, %r15
+; SCALAR-NEXT:    movq %rcx, %rbp
+; SCALAR-NEXT:    imulq %r10, %rbp
+; SCALAR-NEXT:    imulq %rax, %r13
+; SCALAR-NEXT:    xorq %rbp, %r13
+; SCALAR-NEXT:    movq %r11, %rbp
+; SCALAR-NEXT:    imulq %rsi, %rbp
+; SCALAR-NEXT:    imulq %rdx, %r12
+; SCALAR-NEXT:    xorq %rbp, %r12
+; SCALAR-NEXT:    movq %r11, %rbp
+; SCALAR-NEXT:    imulq %rax, %rbp
+; SCALAR-NEXT:    imulq %rdi, %rax
+; SCALAR-NEXT:    imulq %r10, %rdi
+; SCALAR-NEXT:    imulq %r14, %r10
+; SCALAR-NEXT:    xorq %r13, %r12
+; SCALAR-NEXT:    andq %r9, %r12
+; SCALAR-NEXT:    orq %r15, %r12
+; SCALAR-NEXT:    movq %rcx, %r9
+; SCALAR-NEXT:    imulq %rdx, %r9
+; SCALAR-NEXT:    imulq %rsi, %r14
+; SCALAR-NEXT:    xorq %r9, %r14
+; SCALAR-NEXT:    xorq %rbp, %rdi
+; SCALAR-NEXT:    xorq %r14, %rdi
+; SCALAR-NEXT:    andq %rbx, %rdi
+; SCALAR-NEXT:    orq %r12, %rdi
+; SCALAR-NEXT:    imulq %rsi, %rcx
+; SCALAR-NEXT:    xorq %rcx, %r10
+; SCALAR-NEXT:    imulq %rdx, %r11
+; SCALAR-NEXT:    xorq %r11, %rax
+; SCALAR-NEXT:    xorq %r10, %rax
+; SCALAR-NEXT:    andq %r8, %rax
+; SCALAR-NEXT:    orq %rdi, %rax
+; SCALAR-NEXT:    popq %rbx
+; SCALAR-NEXT:    popq %r12
+; SCALAR-NEXT:    popq %r13
+; SCALAR-NEXT:    popq %r14
+; SCALAR-NEXT:    popq %r15
+; SCALAR-NEXT:    popq %rbp
 ; SCALAR-NEXT:    retq
 ;
 ; SSE-PCLMUL-LABEL: clmul_i64:
@@ -779,81 +385,73 @@ define i8 @clmulr_i8(i8 %a, i8 %b) nounwind {
 ; SCALAR-NEXT:    pushq %rbp
 ; SCALAR-NEXT:    pushq %r15
 ; SCALAR-NEXT:    pushq %r14
+; SCALAR-NEXT:    pushq %r13
+; SCALAR-NEXT:    pushq %r12
 ; SCALAR-NEXT:    pushq %rbx
-; SCALAR-NEXT:    movzbl %dil, %ecx
-; SCALAR-NEXT:    movl %ecx, %r11d
-; SCALAR-NEXT:    shll $8, %r11d
+; SCALAR-NEXT:    movl %esi, %eax
+; SCALAR-NEXT:    andl $17, %eax
+; SCALAR-NEXT:    movl %edi, %ecx
+; SCALAR-NEXT:    andl $34, %ecx
 ; SCALAR-NEXT:    movl %ecx, %r10d
-; SCALAR-NEXT:    shll $9, %r10d
-; SCALAR-NEXT:    movl %ecx, %r9d
-; SCALAR-NEXT:    shll $10, %r9d
-; SCALAR-NEXT:    movl %ecx, %eax
-; SCALAR-NEXT:    shll $11, %eax
-; SCALAR-NEXT:    movl %ecx, %r8d
-; SCALAR-NEXT:    shll $12, %r8d
-; SCALAR-NEXT:    movl %edi, %edx
-; SCALAR-NEXT:    shll $13, %edx
-; SCALAR-NEXT:    xorl %ebx, %ebx
-; SCALAR-NEXT:    testw %bx, %bx
-; SCALAR-NEXT:    cmovel %ebx, %edx
-; SCALAR-NEXT:    cmovel %ebx, %r8d
-; SCALAR-NEXT:    cmovel %ebx, %eax
-; SCALAR-NEXT:    cmovel %ebx, %r9d
-; SCALAR-NEXT:    cmovel %ebx, %r10d
-; SCALAR-NEXT:    cmovel %ebx, %r11d
-; SCALAR-NEXT:    shll $14, %edi
-; SCALAR-NEXT:    testw %bx, %bx
-; SCALAR-NEXT:    cmovnel %edi, %ebx
-; SCALAR-NEXT:    movl %esi, %edi
-; SCALAR-NEXT:    andl $1, %edi
-; SCALAR-NEXT:    cmovnel %ecx, %edi
-; SCALAR-NEXT:    leal (%rcx,%rcx), %ebp
-; SCALAR-NEXT:    movl %esi, %r14d
-; SCALAR-NEXT:    andl $2, %r14d
-; SCALAR-NEXT:    cmovnel %ebp, %r14d
-; SCALAR-NEXT:    xorl %edi, %r14d
-; SCALAR-NEXT:    leal (,%rcx,4), %edi
-; SCALAR-NEXT:    movl %esi, %ebp
-; SCALAR-NEXT:    andl $4, %ebp
-; SCALAR-NEXT:    cmovnel %edi, %ebp
-; SCALAR-NEXT:    leal (,%rcx,8), %r15d
-; SCALAR-NEXT:    movl %esi, %edi
-; SCALAR-NEXT:    andl $8, %edi
-; SCALAR-NEXT:    cmovnel %r15d, %edi
-; SCALAR-NEXT:    xorl %ebp, %edi
-; SCALAR-NEXT:    xorl %r14d, %edi
-; SCALAR-NEXT:    movl %ecx, %ebp
-; SCALAR-NEXT:    shll $4, %ebp
-; SCALAR-NEXT:    movl %esi, %r14d
-; SCALAR-NEXT:    andl $16, %r14d
-; SCALAR-NEXT:    cmovnel %ebp, %r14d
-; SCALAR-NEXT:    movl %ecx, %ebp
-; SCALAR-NEXT:    shll $5, %ebp
-; SCALAR-NEXT:    movl %esi, %r15d
-; SCALAR-NEXT:    andl $32, %r15d
-; SCALAR-NEXT:    cmovnel %ebp, %r15d
-; SCALAR-NEXT:    xorl %r14d, %r15d
-; SCALAR-NEXT:    movl %ecx, %ebp
-; SCALAR-NEXT:    shll $6, %ebp
-; SCALAR-NEXT:    movl %esi, %r14d
-; SCALAR-NEXT:    andl $64, %r14d
-; SCALAR-NEXT:    cmovnel %ebp, %r14d
-; SCALAR-NEXT:    xorl %r15d, %r14d
-; SCALAR-NEXT:    xorl %edi, %r14d
-; SCALAR-NEXT:    shll $7, %ecx
-; SCALAR-NEXT:    andl $128, %esi
-; SCALAR-NEXT:    cmovel %esi, %ecx
-; SCALAR-NEXT:    xorl %r11d, %ecx
-; SCALAR-NEXT:    xorl %r10d, %ecx
-; SCALAR-NEXT:    xorl %r9d, %ecx
-; SCALAR-NEXT:    xorl %r14d, %ecx
+; SCALAR-NEXT:    imull %eax, %r10d
+; SCALAR-NEXT:    movl %esi, %edx
+; SCALAR-NEXT:    andl $34, %edx
+; SCALAR-NEXT:    movl %edi, %ebx
+; SCALAR-NEXT:    andl $17, %ebx
+; SCALAR-NEXT:    movl %esi, %r8d
+; SCALAR-NEXT:    andl $136, %r8d
+; SCALAR-NEXT:    movl %edi, %r9d
+; SCALAR-NEXT:    andl $68, %r9d
+; SCALAR-NEXT:    movl %r9d, %r11d
+; SCALAR-NEXT:    imull %r8d, %r11d
+; SCALAR-NEXT:    andl $136, %edi
+; SCALAR-NEXT:    movl %ecx, %r15d
+; SCALAR-NEXT:    imull %r8d, %r15d
+; SCALAR-NEXT:    movl %ebx, %ebp
+; SCALAR-NEXT:    movl %ebx, %r12d
+; SCALAR-NEXT:    movl %edi, %r14d
+; SCALAR-NEXT:    imull %r8d, %r14d
+; SCALAR-NEXT:    imull %ebx, %r8d
+; SCALAR-NEXT:    imull %edx, %ebx
+; SCALAR-NEXT:    xorl %r10d, %ebx
+; SCALAR-NEXT:    andl $68, %esi
+; SCALAR-NEXT:    imull %eax, %ebp
+; SCALAR-NEXT:    movl %edi, %r10d
+; SCALAR-NEXT:    movl %r9d, %r13d
+; SCALAR-NEXT:    imull %eax, %r13d
+; SCALAR-NEXT:    imull %edi, %eax
+; SCALAR-NEXT:    imull %esi, %edi
+; SCALAR-NEXT:    xorl %r11d, %edi
+; SCALAR-NEXT:    xorl %ebx, %edi
+; SCALAR-NEXT:    xorl %r15d, %ebp
+; SCALAR-NEXT:    movl %r9d, %r11d
+; SCALAR-NEXT:    imull %esi, %r11d
+; SCALAR-NEXT:    imull %edx, %r10d
+; SCALAR-NEXT:    xorl %r11d, %r10d
+; SCALAR-NEXT:    andl $8704, %edi # imm = 0x2200
+; SCALAR-NEXT:    xorl %ebp, %r10d
+; SCALAR-NEXT:    andl $4352, %r10d # imm = 0x1100
+; SCALAR-NEXT:    orl %edi, %r10d
+; SCALAR-NEXT:    movl %ecx, %edi
+; SCALAR-NEXT:    imull %edx, %edi
+; SCALAR-NEXT:    imull %esi, %r12d
+; SCALAR-NEXT:    xorl %edi, %r12d
+; SCALAR-NEXT:    xorl %r13d, %r14d
+; SCALAR-NEXT:    xorl %r12d, %r14d
+; SCALAR-NEXT:    andl $17408, %r14d # imm = 0x4400
+; SCALAR-NEXT:    orl %r10d, %r14d
+; SCALAR-NEXT:    imull %esi, %ecx
+; SCALAR-NEXT:    xorl %ecx, %r8d
+; SCALAR-NEXT:    imull %edx, %r9d
+; SCALAR-NEXT:    xorl %r9d, %eax
 ; SCALAR-NEXT:    xorl %r8d, %eax
-; SCALAR-NEXT:    xorl %edx, %eax
-; SCALAR-NEXT:    xorl %ebx, %eax
-; SCALAR-NEXT:    xorl %ecx, %eax
+; SCALAR-NEXT:    andl $2176, %eax # imm = 0x880
+; SCALAR-NEXT:    orl %r14d, %eax
 ; SCALAR-NEXT:    shrl $7, %eax
 ; SCALAR-NEXT:    # kill: def $al killed $al killed $eax
 ; SCALAR-NEXT:    popq %rbx
+; SCALAR-NEXT:    popq %r12
+; SCALAR-NEXT:    popq %r13
 ; SCALAR-NEXT:    popq %r14
 ; SCALAR-NEXT:    popq %r15
 ; SCALAR-NEXT:    popq %rbp
@@ -899,569 +497,184 @@ define i16 @clmulr_i16(i16 %a, i16 %b) nounwind {
 ; SCALAR-NEXT:    pushq %r13
 ; SCALAR-NEXT:    pushq %r12
 ; SCALAR-NEXT:    pushq %rbx
-; SCALAR-NEXT:    movl %esi, %r10d
+; SCALAR-NEXT:    movl %esi, %eax
+; SCALAR-NEXT:    andl $4369, %eax # imm = 0x1111
 ; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    movl %edi, %r13d
-; SCALAR-NEXT:    movl %edi, %r12d
-; SCALAR-NEXT:    movl %edi, %r15d
-; SCALAR-NEXT:    movl %edi, %r9d
-; SCALAR-NEXT:    movl %edi, %ebp
+; SCALAR-NEXT:    andl $8738, %ecx # imm = 0x2222
+; SCALAR-NEXT:    movl %ecx, %r10d
+; SCALAR-NEXT:    imull %eax, %r10d
+; SCALAR-NEXT:    movl %esi, %edx
+; SCALAR-NEXT:    andl $8738, %edx # imm = 0x2222
 ; SCALAR-NEXT:    movl %edi, %ebx
-; SCALAR-NEXT:    movl %edi, %r11d
-; SCALAR-NEXT:    movl %edi, %r8d
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    movl %edi, %esi
-; SCALAR-NEXT:    movl %edi, %edx
+; SCALAR-NEXT:    andl $4369, %ebx # imm = 0x1111
+; SCALAR-NEXT:    movl %esi, %r8d
+; SCALAR-NEXT:    andl $34952, %r8d # imm = 0x8888
+; SCALAR-NEXT:    movl %edi, %r9d
+; SCALAR-NEXT:    andl $17476, %r9d # imm = 0x4444
+; SCALAR-NEXT:    movl %r9d, %r11d
+; SCALAR-NEXT:    imull %r8d, %r11d
+; SCALAR-NEXT:    andl $34952, %edi # imm = 0x8888
+; SCALAR-NEXT:    movl %ecx, %r15d
+; SCALAR-NEXT:    imull %r8d, %r15d
+; SCALAR-NEXT:    movl %ebx, %ebp
+; SCALAR-NEXT:    movl %ebx, %r12d
 ; SCALAR-NEXT:    movl %edi, %r14d
-; SCALAR-NEXT:    shll $16, %r14d
-; SCALAR-NEXT:    shll $17, %ecx
-; SCALAR-NEXT:    movl %ecx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SCALAR-NEXT:    shll $18, %r13d
-; SCALAR-NEXT:    movl %r13d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SCALAR-NEXT:    shll $19, %r12d
-; SCALAR-NEXT:    movl %r12d, %r13d
-; SCALAR-NEXT:    shll $20, %r15d
-; SCALAR-NEXT:    movl %r15d, %r12d
-; SCALAR-NEXT:    shll $21, %r9d
-; SCALAR-NEXT:    movl %r9d, %r15d
-; SCALAR-NEXT:    shll $22, %ebp
-; SCALAR-NEXT:    shll $23, %ebx
-; SCALAR-NEXT:    movl %ebx, %r9d
-; SCALAR-NEXT:    shll $24, %r11d
-; SCALAR-NEXT:    movl %r11d, %ebx
-; SCALAR-NEXT:    shll $25, %r8d
-; SCALAR-NEXT:    movl %r8d, %r11d
-; SCALAR-NEXT:    shll $26, %eax
-; SCALAR-NEXT:    movl %eax, %r8d
-; SCALAR-NEXT:    shll $27, %esi
-; SCALAR-NEXT:    shll $28, %edx
-; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    shll $29, %ecx
-; SCALAR-NEXT:    xorl %eax, %eax
-; SCALAR-NEXT:    testl $0, %eax
-; SCALAR-NEXT:    cmovel %eax, %ecx
-; SCALAR-NEXT:    movl %ecx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SCALAR-NEXT:    cmovel %eax, %edx
-; SCALAR-NEXT:    movl %edx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SCALAR-NEXT:    cmovel %eax, %esi
-; SCALAR-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SCALAR-NEXT:    cmovel %eax, %r8d
-; SCALAR-NEXT:    movl %r8d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SCALAR-NEXT:    cmovel %eax, %r11d
-; SCALAR-NEXT:    movl %r11d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SCALAR-NEXT:    cmovel %eax, %ebx
-; SCALAR-NEXT:    movl %ebx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SCALAR-NEXT:    cmovel %eax, %r9d
-; SCALAR-NEXT:    cmovel %eax, %ebp
-; SCALAR-NEXT:    cmovel %eax, %r15d
-; SCALAR-NEXT:    movl %r15d, %r8d
-; SCALAR-NEXT:    cmovel %eax, %r12d
-; SCALAR-NEXT:    movl %r12d, %r15d
-; SCALAR-NEXT:    cmovel %eax, %r13d
-; SCALAR-NEXT:    movl %r13d, %r12d
-; SCALAR-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %ecx # 4-byte Reload
-; SCALAR-NEXT:    cmovel %eax, %ecx
-; SCALAR-NEXT:    movl %ecx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SCALAR-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %r13d # 4-byte Reload
-; SCALAR-NEXT:    cmovel %eax, %r13d
-; SCALAR-NEXT:    cmovel %eax, %r14d
-; SCALAR-NEXT:    movl %edi, %r11d
-; SCALAR-NEXT:    shll $30, %r11d
-; SCALAR-NEXT:    testl $0, %eax
-; SCALAR-NEXT:    cmovel %eax, %r11d
-; SCALAR-NEXT:    movzwl %di, %edi
-; SCALAR-NEXT:    movl %r10d, %eax
-; SCALAR-NEXT:    andl $1, %eax
-; SCALAR-NEXT:    cmovnel %edi, %eax
-; SCALAR-NEXT:    movl %r10d, %ecx
-; SCALAR-NEXT:    andl $2, %ecx
-; SCALAR-NEXT:    leal (%rdi,%rdi), %esi
-; SCALAR-NEXT:    cmovnel %esi, %ecx
-; SCALAR-NEXT:    xorl %eax, %ecx
-; SCALAR-NEXT:    movl %r10d, %eax
-; SCALAR-NEXT:    andl $4, %eax
-; SCALAR-NEXT:    leal (,%rdi,4), %esi
-; SCALAR-NEXT:    cmovnel %esi, %eax
-; SCALAR-NEXT:    movl %r10d, %esi
-; SCALAR-NEXT:    andl $8, %esi
-; SCALAR-NEXT:    leal (,%rdi,8), %ebx
-; SCALAR-NEXT:    cmovnel %ebx, %esi
-; SCALAR-NEXT:    xorl %eax, %esi
-; SCALAR-NEXT:    xorl %ecx, %esi
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    shll $4, %eax
-; SCALAR-NEXT:    movl %r10d, %ecx
-; SCALAR-NEXT:    andl $16, %ecx
-; SCALAR-NEXT:    cmovnel %eax, %ecx
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    shll $5, %eax
-; SCALAR-NEXT:    movl %r10d, %ebx
-; SCALAR-NEXT:    andl $32, %ebx
-; SCALAR-NEXT:    cmovnel %eax, %ebx
-; SCALAR-NEXT:    xorl %ecx, %ebx
-; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    shll $6, %ecx
-; SCALAR-NEXT:    movl %r10d, %eax
-; SCALAR-NEXT:    andl $64, %eax
-; SCALAR-NEXT:    cmovnel %ecx, %eax
-; SCALAR-NEXT:    xorl %ebx, %eax
-; SCALAR-NEXT:    xorl %esi, %eax
-; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    shll $7, %ecx
-; SCALAR-NEXT:    movl %r10d, %esi
-; SCALAR-NEXT:    andl $128, %esi
-; SCALAR-NEXT:    cmovnel %ecx, %esi
-; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    shll $8, %ecx
-; SCALAR-NEXT:    movl %r10d, %ebx
-; SCALAR-NEXT:    andl $256, %ebx # imm = 0x100
-; SCALAR-NEXT:    cmovnel %ecx, %ebx
-; SCALAR-NEXT:    xorl %esi, %ebx
-; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    shll $9, %ecx
-; SCALAR-NEXT:    movl %r10d, %edx
-; SCALAR-NEXT:    andl $512, %edx # imm = 0x200
-; SCALAR-NEXT:    cmovnel %ecx, %edx
-; SCALAR-NEXT:    xorl %ebx, %edx
-; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    shll $10, %ecx
-; SCALAR-NEXT:    movl %r10d, %esi
-; SCALAR-NEXT:    andl $1024, %esi # imm = 0x400
-; SCALAR-NEXT:    cmovnel %ecx, %esi
-; SCALAR-NEXT:    xorl %edx, %esi
-; SCALAR-NEXT:    xorl %eax, %esi
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    shll $11, %eax
-; SCALAR-NEXT:    movl %r10d, %ecx
-; SCALAR-NEXT:    andl $2048, %ecx # imm = 0x800
-; SCALAR-NEXT:    cmovnel %eax, %ecx
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    shll $12, %eax
-; SCALAR-NEXT:    movl %r10d, %edx
-; SCALAR-NEXT:    andl $4096, %edx # imm = 0x1000
-; SCALAR-NEXT:    cmovnel %eax, %edx
-; SCALAR-NEXT:    xorl %ecx, %edx
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    shll $13, %eax
-; SCALAR-NEXT:    movl %r10d, %ecx
-; SCALAR-NEXT:    andl $8192, %ecx # imm = 0x2000
-; SCALAR-NEXT:    cmovnel %eax, %ecx
-; SCALAR-NEXT:    xorl %edx, %ecx
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    shll $14, %eax
-; SCALAR-NEXT:    movl %r10d, %edx
-; SCALAR-NEXT:    andl $16384, %edx # imm = 0x4000
-; SCALAR-NEXT:    cmovnel %eax, %edx
-; SCALAR-NEXT:    xorl %ecx, %edx
-; SCALAR-NEXT:    shll $15, %edi
-; SCALAR-NEXT:    andl $32768, %r10d # imm = 0x8000
-; SCALAR-NEXT:    cmovel %r10d, %edi
-; SCALAR-NEXT:    xorl %edx, %edi
-; SCALAR-NEXT:    xorl %esi, %edi
+; SCALAR-NEXT:    imull %r8d, %r14d
+; SCALAR-NEXT:    imull %ebx, %r8d
+; SCALAR-NEXT:    imull %edx, %ebx
+; SCALAR-NEXT:    xorl %r10d, %ebx
+; SCALAR-NEXT:    andl $17476, %esi # imm = 0x4444
+; SCALAR-NEXT:    imull %eax, %ebp
+; SCALAR-NEXT:    movl %edi, %r10d
+; SCALAR-NEXT:    movl %r9d, %r13d
+; SCALAR-NEXT:    imull %eax, %r13d
+; SCALAR-NEXT:    imull %edi, %eax
+; SCALAR-NEXT:    imull %esi, %edi
+; SCALAR-NEXT:    xorl %r11d, %edi
+; SCALAR-NEXT:    xorl %ebx, %edi
+; SCALAR-NEXT:    xorl %r15d, %ebp
+; SCALAR-NEXT:    movl %r9d, %r11d
+; SCALAR-NEXT:    imull %esi, %r11d
+; SCALAR-NEXT:    imull %edx, %r10d
+; SCALAR-NEXT:    xorl %r11d, %r10d
+; SCALAR-NEXT:    andl $572653568, %edi # imm = 0x22220000
+; SCALAR-NEXT:    xorl %ebp, %r10d
+; SCALAR-NEXT:    andl $286326784, %r10d # imm = 0x11110000
+; SCALAR-NEXT:    orl %edi, %r10d
+; SCALAR-NEXT:    movl %ecx, %edi
+; SCALAR-NEXT:    imull %edx, %edi
+; SCALAR-NEXT:    imull %esi, %r12d
+; SCALAR-NEXT:    xorl %edi, %r12d
 ; SCALAR-NEXT:    xorl %r13d, %r14d
-; SCALAR-NEXT:    xorl {{[-0-9]+}}(%r{{[sb]}}p), %r14d # 4-byte Folded Reload
 ; SCALAR-NEXT:    xorl %r12d, %r14d
-; SCALAR-NEXT:    xorl %r15d, %r14d
-; SCALAR-NEXT:    xorl %r8d, %r14d
-; SCALAR-NEXT:    xorl %edi, %r14d
-; SCALAR-NEXT:    xorl %r9d, %ebp
-; SCALAR-NEXT:    xorl {{[-0-9]+}}(%r{{[sb]}}p), %ebp # 4-byte Folded Reload
-; SCALAR-NEXT:    xorl {{[-0-9]+}}(%r{{[sb]}}p), %ebp # 4-byte Folded Reload
-; SCALAR-NEXT:    xorl {{[-0-9]+}}(%r{{[sb]}}p), %ebp # 4-byte Folded Reload
-; SCALAR-NEXT:    xorl {{[-0-9]+}}(%r{{[sb]}}p), %ebp # 4-byte Folded Reload
-; SCALAR-NEXT:    xorl {{[-0-9]+}}(%r{{[sb]}}p), %ebp # 4-byte Folded Reload
-; SCALAR-NEXT:    xorl %r14d, %ebp
-; SCALAR-NEXT:    xorl {{[-0-9]+}}(%r{{[sb]}}p), %r11d # 4-byte Folded Reload
-; SCALAR-NEXT:    xorl %ebp, %r11d
-; SCALAR-NEXT:    shrl $15, %r11d
-; SCALAR-NEXT:    movl %r11d, %eax
+; SCALAR-NEXT:    andl $1145307136, %r14d # imm = 0x44440000
+; SCALAR-NEXT:    orl %r10d, %r14d
+; SCALAR-NEXT:    imull %esi, %ecx
+; SCALAR-NEXT:    xorl %ecx, %r8d
+; SCALAR-NEXT:    imull %edx, %r9d
+; SCALAR-NEXT:    xorl %r9d, %eax
+; SCALAR-NEXT:    xorl %r8d, %eax
+; SCALAR-NEXT:    andl $143163392, %eax # imm = 0x8888000
+; SCALAR-NEXT:    orl %r14d, %eax
+; SCALAR-NEXT:    shrl $15, %eax
+; SCALAR-NEXT:    # kill: def $ax killed $ax killed $eax
 ; SCALAR-NEXT:    popq %rbx
 ; SCALAR-NEXT:    popq %r12
 ; SCALAR-NEXT:    popq %r13
 ; SCALAR-NEXT:    popq %r14
 ; SCALAR-NEXT:    popq %r15
-; SCALAR-NEXT:    popq %rbp
-; SCALAR-NEXT:    retq
-;
-; SSE-PCLMUL-LABEL: clmulr_i16:
-; SSE-PCLMUL:       # %bb.0:
-; SSE-PCLMUL-NEXT:    movzwl %si, %eax
-; SSE-PCLMUL-NEXT:    movd %eax, %xmm0
-; SSE-PCLMUL-NEXT:    movzwl %di, %eax
-; SSE-PCLMUL-NEXT:    movd %eax, %xmm1
-; SSE-PCLMUL-NEXT:    pclmulqdq $0, %xmm0, %xmm1
-; SSE-PCLMUL-NEXT:    movq %xmm1, %rax
-; SSE-PCLMUL-NEXT:    shrl $15, %eax
-; SSE-PCLMUL-NEXT:    # kill: def $ax killed $ax killed $rax
-; SSE-PCLMUL-NEXT:    retq
-;
-; AVX-LABEL: clmulr_i16:
-; AVX:       # %bb.0:
-; AVX-NEXT:    movzwl %si, %eax
-; AVX-NEXT:    vmovd %eax, %xmm0
-; AVX-NEXT:    movzwl %di, %eax
-; AVX-NEXT:    vmovd %eax, %xmm1
-; AVX-NEXT:    vpclmulqdq $0, %xmm0, %xmm1, %xmm0
-; AVX-NEXT:    vmovq %xmm0, %rax
-; AVX-NEXT:    shrl $15, %eax
-; AVX-NEXT:    # kill: def $ax killed $ax killed $rax
-; AVX-NEXT:    retq
-  %a.ext = zext i16 %a to i32
-  %b.ext = zext i16 %b to i32
-  %clmul = call i32 @llvm.clmul.i32(i32 %a.ext, i32 %b.ext)
-  %res.ext = lshr i32 %clmul, 15
-  %res = trunc i32 %res.ext to i16
-  ret i16 %res
-}
-
-define i32 @clmulr_i32(i32 %a, i32 %b) nounwind {
-; SCALAR-LABEL: clmulr_i32:
-; SCALAR:       # %bb.0:
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    leaq (%rax,%rax), %rdx
-; SCALAR-NEXT:    movl %esi, %ecx
-; SCALAR-NEXT:    movl %ecx, %esi
-; SCALAR-NEXT:    andl $2, %esi
-; SCALAR-NEXT:    cmovneq %rdx, %rsi
-; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $1, %edi
-; SCALAR-NEXT:    cmovneq %rax, %rdi
-; SCALAR-NEXT:    xorq %rsi, %rdi
-; SCALAR-NEXT:    leaq (,%rax,4), %rdx
-; SCALAR-NEXT:    movl %ecx, %esi
-; SCALAR-NEXT:    andl $4, %esi
-; SCALAR-NEXT:    cmovneq %rdx, %rsi
-; SCALAR-NEXT:    leaq (,%rax,8), %r8
-; SCALAR-NEXT:    movl %ecx, %edx
-; SCALAR-NEXT:    andl $8, %edx
-; SCALAR-NEXT:    cmovneq %r8, %rdx
-; SCALAR-NEXT:    xorq %rsi, %rdx
-; SCALAR-NEXT:    xorq %rdi, %rdx
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $4, %rsi
-; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $16, %edi
-; SCALAR-NEXT:    cmovneq %rsi, %rdi
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $5, %rsi
-; SCALAR-NEXT:    movl %ecx, %r8d
-; SCALAR-NEXT:    andl $32, %r8d
-; SCALAR-NEXT:    cmovneq %rsi, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rdi
-; SCALAR-NEXT:    shlq $6, %rdi
-; SCALAR-NEXT:    movl %ecx, %esi
-; SCALAR-NEXT:    andl $64, %esi
-; SCALAR-NEXT:    cmovneq %rdi, %rsi
-; SCALAR-NEXT:    xorq %r8, %rsi
-; SCALAR-NEXT:    xorq %rdx, %rsi
-; SCALAR-NEXT:    movq %rax, %rdx
-; SCALAR-NEXT:    shlq $7, %rdx
-; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $128, %edi
-; SCALAR-NEXT:    cmovneq %rdx, %rdi
-; SCALAR-NEXT:    movq %rax, %rdx
-; SCALAR-NEXT:    shlq $8, %rdx
-; SCALAR-NEXT:    movl %ecx, %r8d
-; SCALAR-NEXT:    andl $256, %r8d # imm = 0x100
-; SCALAR-NEXT:    cmovneq %rdx, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rdx
-; SCALAR-NEXT:    shlq $9, %rdx
-; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $512, %edi # imm = 0x200
-; SCALAR-NEXT:    cmovneq %rdx, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $10, %r8
-; SCALAR-NEXT:    movl %ecx, %edx
-; SCALAR-NEXT:    andl $1024, %edx # imm = 0x400
-; SCALAR-NEXT:    cmovneq %r8, %rdx
-; SCALAR-NEXT:    xorq %rdi, %rdx
-; SCALAR-NEXT:    xorq %rsi, %rdx
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $11, %rsi
-; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $2048, %edi # imm = 0x800
-; SCALAR-NEXT:    cmovneq %rsi, %rdi
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $12, %rsi
-; SCALAR-NEXT:    movl %ecx, %r8d
-; SCALAR-NEXT:    andl $4096, %r8d # imm = 0x1000
-; SCALAR-NEXT:    cmovneq %rsi, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $13, %rsi
-; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $8192, %edi # imm = 0x2000
-; SCALAR-NEXT:    cmovneq %rsi, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $14, %rsi
-; SCALAR-NEXT:    movl %ecx, %r8d
-; SCALAR-NEXT:    andl $16384, %r8d # imm = 0x4000
-; SCALAR-NEXT:    cmovneq %rsi, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rdi
-; SCALAR-NEXT:    shlq $15, %rdi
-; SCALAR-NEXT:    movl %ecx, %esi
-; SCALAR-NEXT:    andl $32768, %esi # imm = 0x8000
-; SCALAR-NEXT:    cmovneq %rdi, %rsi
-; SCALAR-NEXT:    xorq %r8, %rsi
-; SCALAR-NEXT:    xorq %rdx, %rsi
-; SCALAR-NEXT:    movq %rax, %rdx
-; SCALAR-NEXT:    shlq $16, %rdx
-; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $65536, %edi # imm = 0x10000
-; SCALAR-NEXT:    cmovneq %rdx, %rdi
-; SCALAR-NEXT:    movq %rax, %rdx
-; SCALAR-NEXT:    shlq $17, %rdx
-; SCALAR-NEXT:    movl %ecx, %r8d
-; SCALAR-NEXT:    andl $131072, %r8d # imm = 0x20000
-; SCALAR-NEXT:    cmovneq %rdx, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rdx
-; SCALAR-NEXT:    shlq $18, %rdx
-; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $262144, %edi # imm = 0x40000
-; SCALAR-NEXT:    cmovneq %rdx, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
-; SCALAR-NEXT:    movq %rax, %rdx
-; SCALAR-NEXT:    shlq $19, %rdx
-; SCALAR-NEXT:    movl %ecx, %r8d
-; SCALAR-NEXT:    andl $524288, %r8d # imm = 0x80000
-; SCALAR-NEXT:    cmovneq %rdx, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rdx
-; SCALAR-NEXT:    shlq $20, %rdx
-; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $1048576, %edi # imm = 0x100000
-; SCALAR-NEXT:    cmovneq %rdx, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $21, %r8
-; SCALAR-NEXT:    movl %ecx, %edx
-; SCALAR-NEXT:    andl $2097152, %edx # imm = 0x200000
-; SCALAR-NEXT:    cmovneq %r8, %rdx
-; SCALAR-NEXT:    xorq %rdi, %rdx
-; SCALAR-NEXT:    xorq %rsi, %rdx
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $22, %rsi
-; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $4194304, %edi # imm = 0x400000
-; SCALAR-NEXT:    cmovneq %rsi, %rdi
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $23, %rsi
-; SCALAR-NEXT:    movl %ecx, %r8d
-; SCALAR-NEXT:    andl $8388608, %r8d # imm = 0x800000
-; SCALAR-NEXT:    cmovneq %rsi, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $24, %rsi
-; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $16777216, %edi # imm = 0x1000000
-; SCALAR-NEXT:    cmovneq %rsi, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $25, %rsi
-; SCALAR-NEXT:    movl %ecx, %r8d
-; SCALAR-NEXT:    andl $33554432, %r8d # imm = 0x2000000
-; SCALAR-NEXT:    cmovneq %rsi, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $26, %rsi
-; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $67108864, %edi # imm = 0x4000000
-; SCALAR-NEXT:    cmovneq %rsi, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $27, %rsi
-; SCALAR-NEXT:    movl %ecx, %r8d
-; SCALAR-NEXT:    andl $134217728, %r8d # imm = 0x8000000
-; SCALAR-NEXT:    cmovneq %rsi, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rdi
-; SCALAR-NEXT:    shlq $28, %rdi
+; SCALAR-NEXT:    popq %rbp
+; SCALAR-NEXT:    retq
+;
+; SSE-PCLMUL-LABEL: clmulr_i16:
+; SSE-PCLMUL:       # %bb.0:
+; SSE-PCLMUL-NEXT:    movzwl %si, %eax
+; SSE-PCLMUL-NEXT:    movd %eax, %xmm0
+; SSE-PCLMUL-NEXT:    movzwl %di, %eax
+; SSE-PCLMUL-NEXT:    movd %eax, %xmm1
+; SSE-PCLMUL-NEXT:    pclmulqdq $0, %xmm0, %xmm1
+; SSE-PCLMUL-NEXT:    movq %xmm1, %rax
+; SSE-PCLMUL-NEXT:    shrl $15, %eax
+; SSE-PCLMUL-NEXT:    # kill: def $ax killed $ax killed $rax
+; SSE-PCLMUL-NEXT:    retq
+;
+; AVX-LABEL: clmulr_i16:
+; AVX:       # %bb.0:
+; AVX-NEXT:    movzwl %si, %eax
+; AVX-NEXT:    vmovd %eax, %xmm0
+; AVX-NEXT:    movzwl %di, %eax
+; AVX-NEXT:    vmovd %eax, %xmm1
+; AVX-NEXT:    vpclmulqdq $0, %xmm0, %xmm1, %xmm0
+; AVX-NEXT:    vmovq %xmm0, %rax
+; AVX-NEXT:    shrl $15, %eax
+; AVX-NEXT:    # kill: def $ax killed $ax killed $rax
+; AVX-NEXT:    retq
+  %a.ext = zext i16 %a to i32
+  %b.ext = zext i16 %b to i32
+  %clmul = call i32 @llvm.clmul.i32(i32 %a.ext, i32 %b.ext)
+  %res.ext = lshr i32 %clmul, 15
+  %res = trunc i32 %res.ext to i16
+  ret i16 %res
+}
+
+define i32 @clmulr_i32(i32 %a, i32 %b) nounwind {
+; SCALAR-LABEL: clmulr_i32:
+; SCALAR:       # %bb.0:
+; SCALAR-NEXT:    pushq %r15
+; SCALAR-NEXT:    pushq %r14
+; SCALAR-NEXT:    pushq %rbx
+; SCALAR-NEXT:    movl %esi, %edx
+; SCALAR-NEXT:    movl %edx, %eax
+; SCALAR-NEXT:    andl $286331153, %eax # imm = 0x11111111
+; SCALAR-NEXT:    movl %edi, %ecx
 ; SCALAR-NEXT:    movl %ecx, %esi
-; SCALAR-NEXT:    andl $268435456, %esi # imm = 0x10000000
-; SCALAR-NEXT:    cmovneq %rdi, %rsi
-; SCALAR-NEXT:    xorq %r8, %rsi
-; SCALAR-NEXT:    xorq %rdx, %rsi
-; SCALAR-NEXT:    movq %rax, %rdx
-; SCALAR-NEXT:    shlq $29, %rdx
-; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $536870912, %edi # imm = 0x20000000
-; SCALAR-NEXT:    cmovneq %rdx, %rdi
-; SCALAR-NEXT:    movq %rax, %rdx
-; SCALAR-NEXT:    shlq $30, %rdx
-; SCALAR-NEXT:    movl %ecx, %r8d
-; SCALAR-NEXT:    andl $1073741824, %r8d # imm = 0x40000000
-; SCALAR-NEXT:    cmovneq %rdx, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rdx
-; SCALAR-NEXT:    shlq $31, %rdx
+; SCALAR-NEXT:    andl $572662306, %esi # imm = 0x22222222
+; SCALAR-NEXT:    movq %rsi, %r9
+; SCALAR-NEXT:    imulq %rax, %r9
+; SCALAR-NEXT:    movl %edx, %r8d
+; SCALAR-NEXT:    andl $572662306, %r8d # imm = 0x22222222
 ; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $-2147483648, %edi # imm = 0x80000000
-; SCALAR-NEXT:    cmovneq %rdx, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $32, %r8
-; SCALAR-NEXT:    xorl %edx, %edx
-; SCALAR-NEXT:    btq $32, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rdi
-; SCALAR-NEXT:    shlq $33, %rdi
-; SCALAR-NEXT:    btq $33, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $34, %r8
-; SCALAR-NEXT:    btq $34, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %r9
-; SCALAR-NEXT:    shlq $35, %r9
-; SCALAR-NEXT:    btq $35, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r9
-; SCALAR-NEXT:    xorq %r8, %r9
-; SCALAR-NEXT:    movq %rax, %rdi
-; SCALAR-NEXT:    shlq $36, %rdi
-; SCALAR-NEXT:    btq $36, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rdi
-; SCALAR-NEXT:    xorq %r9, %rdi
-; SCALAR-NEXT:    xorq %rsi, %rdi
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $37, %rsi
-; SCALAR-NEXT:    btq $37, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rsi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $38, %r8
-; SCALAR-NEXT:    btq $38, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r8
-; SCALAR-NEXT:    xorq %rsi, %r8
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $39, %rsi
-; SCALAR-NEXT:    btq $39, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rsi
-; SCALAR-NEXT:    xorq %r8, %rsi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $40, %r8
-; SCALAR-NEXT:    btq $40, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r8
-; SCALAR-NEXT:    xorq %rsi, %r8
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $41, %rsi
-; SCALAR-NEXT:    btq $41, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rsi
-; SCALAR-NEXT:    xorq %r8, %rsi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $42, %r8
-; SCALAR-NEXT:    btq $42, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r8
-; SCALAR-NEXT:    xorq %rsi, %r8
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $43, %rsi
-; SCALAR-NEXT:    btq $43, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rsi
-; SCALAR-NEXT:    xorq %r8, %rsi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $44, %r8
-; SCALAR-NEXT:    btq $44, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r8
-; SCALAR-NEXT:    xorq %rsi, %r8
-; SCALAR-NEXT:    movq %rax, %r9
-; SCALAR-NEXT:    shlq $45, %r9
-; SCALAR-NEXT:    btq $45, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r9
-; SCALAR-NEXT:    xorq %r8, %r9
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $46, %rsi
-; SCALAR-NEXT:    btq $46, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rsi
-; SCALAR-NEXT:    xorq %r9, %rsi
-; SCALAR-NEXT:    xorq %rdi, %rsi
-; SCALAR-NEXT:    movq %rax, %rdi
-; SCALAR-NEXT:    shlq $47, %rdi
-; SCALAR-NEXT:    btq $47, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rdi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $48, %r8
-; SCALAR-NEXT:    btq $48, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rdi
-; SCALAR-NEXT:    shlq $49, %rdi
-; SCALAR-NEXT:    btq $49, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $50, %r8
-; SCALAR-NEXT:    btq $50, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rdi
-; SCALAR-NEXT:    shlq $51, %rdi
-; SCALAR-NEXT:    btq $51, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $52, %r8
-; SCALAR-NEXT:    btq $52, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rdi
-; SCALAR-NEXT:    shlq $53, %rdi
-; SCALAR-NEXT:    btq $53, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $54, %r8
-; SCALAR-NEXT:    btq $54, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rdi
-; SCALAR-NEXT:    shlq $55, %rdi
-; SCALAR-NEXT:    btq $55, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $56, %r8
-; SCALAR-NEXT:    btq $56, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rdi
-; SCALAR-NEXT:    shlq $57, %rdi
-; SCALAR-NEXT:    btq $57, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
+; SCALAR-NEXT:    andl $286331153, %edi # imm = 0x11111111
+; SCALAR-NEXT:    movq %rdi, %r11
+; SCALAR-NEXT:    imulq %r8, %r11
+; SCALAR-NEXT:    xorq %r9, %r11
+; SCALAR-NEXT:    movl %edx, %r10d
+; SCALAR-NEXT:    andl $-2004318072, %r10d # imm = 0x88888888
+; SCALAR-NEXT:    movl %ecx, %r9d
+; SCALAR-NEXT:    andl $1145324612, %r9d # imm = 0x44444444
+; SCALAR-NEXT:    movq %r9, %rbx
+; SCALAR-NEXT:    imulq %r10, %rbx
+; SCALAR-NEXT:    andl $1145324612, %edx # imm = 0x44444444
+; SCALAR-NEXT:    andl $-2004318072, %ecx # imm = 0x88888888
+; SCALAR-NEXT:    movq %rcx, %r14
+; SCALAR-NEXT:    imulq %rdx, %r14
+; SCALAR-NEXT:    xorq %rbx, %r14
+; SCALAR-NEXT:    xorq %r11, %r14
+; SCALAR-NEXT:    movabsq $2459565875921944576, %r11 # imm = 0x2222222200000000
+; SCALAR-NEXT:    andq %r14, %r11
+; SCALAR-NEXT:    movq %rsi, %rbx
+; SCALAR-NEXT:    imulq %r10, %rbx
+; SCALAR-NEXT:    movq %rdi, %r14
+; SCALAR-NEXT:    imulq %rax, %r14
+; SCALAR-NEXT:    xorq %rbx, %r14
+; SCALAR-NEXT:    movq %r9, %rbx
+; SCALAR-NEXT:    imulq %rdx, %rbx
+; SCALAR-NEXT:    movq %rcx, %r15
+; SCALAR-NEXT:    imulq %r8, %r15
+; SCALAR-NEXT:    xorq %rbx, %r15
+; SCALAR-NEXT:    xorq %r14, %r15
+; SCALAR-NEXT:    movabsq $1229782937960972288, %rbx # imm = 0x1111111100000000
+; SCALAR-NEXT:    andq %r15, %rbx
+; SCALAR-NEXT:    orq %r11, %rbx
+; SCALAR-NEXT:    movq %rsi, %r11
+; SCALAR-NEXT:    imulq %r8, %r11
+; SCALAR-NEXT:    movq %rdi, %r14
+; SCALAR-NEXT:    imulq %rdx, %r14
+; SCALAR-NEXT:    xorq %r11, %r14
+; SCALAR-NEXT:    movq %r9, %r11
+; SCALAR-NEXT:    imulq %rax, %r11
+; SCALAR-NEXT:    movq %rcx, %r15
+; SCALAR-NEXT:    imulq %r10, %r15
+; SCALAR-NEXT:    xorq %r11, %r15
+; SCALAR-NEXT:    xorq %r14, %r15
+; SCALAR-NEXT:    movabsq $4919131751843889152, %r11 # imm = 0x4444444400000000
+; SCALAR-NEXT:    andq %r15, %r11
+; SCALAR-NEXT:    orq %rbx, %r11
+; SCALAR-NEXT:    imulq %rdx, %rsi
+; SCALAR-NEXT:    imulq %r10, %rdi
 ; SCALAR-NEXT:    xorq %rsi, %rdi
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $58, %rsi
-; SCALAR-NEXT:    btq $58, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rsi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $59, %r8
-; SCALAR-NEXT:    btq $59, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r8
-; SCALAR-NEXT:    xorq %rsi, %r8
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $60, %rsi
-; SCALAR-NEXT:    btq $60, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rsi
-; SCALAR-NEXT:    xorq %r8, %rsi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $61, %r8
-; SCALAR-NEXT:    btq $61, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r8
-; SCALAR-NEXT:    xorq %rsi, %r8
-; SCALAR-NEXT:    shlq $62, %rax
-; SCALAR-NEXT:    btq $62, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rax
-; SCALAR-NEXT:    xorq %r8, %rax
-; SCALAR-NEXT:    xorq %rdi, %rax
+; SCALAR-NEXT:    imulq %r8, %r9
+; SCALAR-NEXT:    imulq %rax, %rcx
+; SCALAR-NEXT:    xorq %r9, %rcx
+; SCALAR-NEXT:    xorq %rdi, %rcx
+; SCALAR-NEXT:    movabsq $614891468980486144, %rax # imm = 0x888888880000000
+; SCALAR-NEXT:    andq %rcx, %rax
+; SCALAR-NEXT:    orq %r11, %rax
 ; SCALAR-NEXT:    shrq $31, %rax
 ; SCALAR-NEXT:    # kill: def $eax killed $eax killed $rax
+; SCALAR-NEXT:    popq %rbx
+; SCALAR-NEXT:    popq %r14
+; SCALAR-NEXT:    popq %r15
 ; SCALAR-NEXT:    retq
 ;
 ; SSE-PCLMUL-LABEL: clmulr_i32:
@@ -1494,410 +707,138 @@ define i32 @clmulr_i32(i32 %a, i32 %b) nounwind {
 define i64 @clmulr_i64(i64 %a, i64 %b) nounwind {
 ; SCALAR-LABEL: clmulr_i64:
 ; SCALAR:       # %bb.0:
+; SCALAR-NEXT:    pushq %rbp
+; SCALAR-NEXT:    pushq %r15
 ; SCALAR-NEXT:    pushq %r14
+; SCALAR-NEXT:    pushq %r13
+; SCALAR-NEXT:    pushq %r12
 ; SCALAR-NEXT:    pushq %rbx
-; SCALAR-NEXT:    bswapq %rdi
-; SCALAR-NEXT:    movq %rdi, %rax
+; SCALAR-NEXT:    bswapq %rsi
+; SCALAR-NEXT:    movq %rsi, %rax
 ; SCALAR-NEXT:    shrq $4, %rax
 ; SCALAR-NEXT:    movabsq $1085102592571150095, %rcx # imm = 0xF0F0F0F0F0F0F0F
 ; SCALAR-NEXT:    andq %rcx, %rax
-; SCALAR-NEXT:    andq %rcx, %rdi
-; SCALAR-NEXT:    shlq $4, %rdi
-; SCALAR-NEXT:    orq %rax, %rdi
-; SCALAR-NEXT:    movabsq $3689348814741910323, %rax # imm = 0x3333333333333333
-; SCALAR-NEXT:    movq %rdi, %rdx
-; SCALAR-NEXT:    andq %rax, %rdx
-; SCALAR-NEXT:    shrq $2, %rdi
-; SCALAR-NEXT:    andq %rax, %rdi
-; SCALAR-NEXT:    leaq (%rdi,%rdx,4), %rdi
-; SCALAR-NEXT:    movabsq $6148914691236517205, %rdx # imm = 0x5555555555555555
-; SCALAR-NEXT:    movq %rdi, %r8
-; SCALAR-NEXT:    andq %rdx, %r8
-; SCALAR-NEXT:    shrq %rdi
-; SCALAR-NEXT:    movq %rdi, %r9
-; SCALAR-NEXT:    andq %rdx, %r9
-; SCALAR-NEXT:    leaq (%r9,%r8,2), %r8
-; SCALAR-NEXT:    leaq (%r8,%r8), %r9
-; SCALAR-NEXT:    bswapq %rsi
-; SCALAR-NEXT:    movq %rsi, %r10
-; SCALAR-NEXT:    shrq $4, %r10
-; SCALAR-NEXT:    andq %rcx, %r10
 ; SCALAR-NEXT:    andq %rcx, %rsi
 ; SCALAR-NEXT:    shlq $4, %rsi
-; SCALAR-NEXT:    orq %r10, %rsi
-; SCALAR-NEXT:    movq %rsi, %r10
-; SCALAR-NEXT:    andq %rax, %r10
+; SCALAR-NEXT:    orq %rax, %rsi
+; SCALAR-NEXT:    movabsq $3689348814741910323, %rax # imm = 0x3333333333333333
+; SCALAR-NEXT:    movq %rsi, %rdx
+; SCALAR-NEXT:    andq %rax, %rdx
 ; SCALAR-NEXT:    shrq $2, %rsi
 ; SCALAR-NEXT:    andq %rax, %rsi
-; SCALAR-NEXT:    leaq (%rsi,%r10,4), %rsi
-; SCALAR-NEXT:    movq %rsi, %r10
-; SCALAR-NEXT:    andq %rdx, %r10
+; SCALAR-NEXT:    leaq (%rsi,%rdx,4), %rsi
+; SCALAR-NEXT:    movabsq $6148914691236517205, %rdx # imm = 0x5555555555555555
+; SCALAR-NEXT:    movq %rsi, %r8
+; SCALAR-NEXT:    andq %rdx, %r8
 ; SCALAR-NEXT:    shrq %rsi
 ; SCALAR-NEXT:    andq %rdx, %rsi
-; SCALAR-NEXT:    leaq (%rsi,%r10,2), %rsi
-; SCALAR-NEXT:    movl %esi, %r10d
-; SCALAR-NEXT:    andl $2, %r10d
-; SCALAR-NEXT:    cmovneq %r9, %r10
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $1, %r11d
-; SCALAR-NEXT:    cmovneq %r8, %r11
-; SCALAR-NEXT:    xorq %r10, %r11
-; SCALAR-NEXT:    leaq (,%r8,4), %r9
-; SCALAR-NEXT:    movl %esi, %r10d
-; SCALAR-NEXT:    andl $4, %r10d
-; SCALAR-NEXT:    cmovneq %r9, %r10
-; SCALAR-NEXT:    leaq (,%r8,8), %rbx
-; SCALAR-NEXT:    movl %esi, %r9d
-; SCALAR-NEXT:    andl $8, %r9d
-; SCALAR-NEXT:    cmovneq %rbx, %r9
-; SCALAR-NEXT:    xorq %r10, %r9
-; SCALAR-NEXT:    xorq %r11, %r9
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $4, %r10
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $16, %r11d
-; SCALAR-NEXT:    cmovneq %r10, %r11
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $5, %r10
-; SCALAR-NEXT:    movl %esi, %ebx
-; SCALAR-NEXT:    andl $32, %ebx
-; SCALAR-NEXT:    cmovneq %r10, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r11
-; SCALAR-NEXT:    shlq $6, %r11
-; SCALAR-NEXT:    movl %esi, %r10d
-; SCALAR-NEXT:    andl $64, %r10d
-; SCALAR-NEXT:    cmovneq %r11, %r10
-; SCALAR-NEXT:    xorq %rbx, %r10
-; SCALAR-NEXT:    xorq %r9, %r10
-; SCALAR-NEXT:    movq %r8, %r9
-; SCALAR-NEXT:    shlq $7, %r9
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $128, %r11d
-; SCALAR-NEXT:    cmovneq %r9, %r11
-; SCALAR-NEXT:    movq %r8, %r9
-; SCALAR-NEXT:    shlq $8, %r9
-; SCALAR-NEXT:    movl %esi, %ebx
-; SCALAR-NEXT:    andl $256, %ebx # imm = 0x100
-; SCALAR-NEXT:    cmovneq %r9, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r9
-; SCALAR-NEXT:    shlq $9, %r9
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $512, %r11d # imm = 0x200
-; SCALAR-NEXT:    cmovneq %r9, %r11
-; SCALAR-NEXT:    xorq %rbx, %r11
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $10, %rbx
-; SCALAR-NEXT:    movl %esi, %r9d
-; SCALAR-NEXT:    andl $1024, %r9d # imm = 0x400
-; SCALAR-NEXT:    cmovneq %rbx, %r9
-; SCALAR-NEXT:    xorq %r11, %r9
-; SCALAR-NEXT:    xorq %r10, %r9
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $11, %r10
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $2048, %r11d # imm = 0x800
-; SCALAR-NEXT:    cmovneq %r10, %r11
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $12, %r10
-; SCALAR-NEXT:    movl %esi, %ebx
-; SCALAR-NEXT:    andl $4096, %ebx # imm = 0x1000
-; SCALAR-NEXT:    cmovneq %r10, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $13, %r10
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $8192, %r11d # imm = 0x2000
-; SCALAR-NEXT:    cmovneq %r10, %r11
-; SCALAR-NEXT:    xorq %rbx, %r11
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $14, %r10
-; SCALAR-NEXT:    movl %esi, %ebx
-; SCALAR-NEXT:    andl $16384, %ebx # imm = 0x4000
-; SCALAR-NEXT:    cmovneq %r10, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r11
-; SCALAR-NEXT:    shlq $15, %r11
-; SCALAR-NEXT:    movl %esi, %r10d
-; SCALAR-NEXT:    andl $32768, %r10d # imm = 0x8000
-; SCALAR-NEXT:    cmovneq %r11, %r10
-; SCALAR-NEXT:    xorq %rbx, %r10
-; SCALAR-NEXT:    xorq %r9, %r10
-; SCALAR-NEXT:    movq %r8, %r9
-; SCALAR-NEXT:    shlq $16, %r9
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $65536, %r11d # imm = 0x10000
-; SCALAR-NEXT:    cmovneq %r9, %r11
-; SCALAR-NEXT:    movq %r8, %r9
-; SCALAR-NEXT:    shlq $17, %r9
-; SCALAR-NEXT:    movl %esi, %ebx
-; SCALAR-NEXT:    andl $131072, %ebx # imm = 0x20000
-; SCALAR-NEXT:    cmovneq %r9, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r9
-; SCALAR-NEXT:    shlq $18, %r9
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $262144, %r11d # imm = 0x40000
-; SCALAR-NEXT:    cmovneq %r9, %r11
-; SCALAR-NEXT:    xorq %rbx, %r11
-; SCALAR-NEXT:    movq %r8, %r9
-; SCALAR-NEXT:    shlq $19, %r9
-; SCALAR-NEXT:    movl %esi, %ebx
-; SCALAR-NEXT:    andl $524288, %ebx # imm = 0x80000
-; SCALAR-NEXT:    cmovneq %r9, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r9
-; SCALAR-NEXT:    shlq $20, %r9
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $1048576, %r11d # imm = 0x100000
-; SCALAR-NEXT:    cmovneq %r9, %r11
-; SCALAR-NEXT:    xorq %rbx, %r11
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $21, %rbx
-; SCALAR-NEXT:    movl %esi, %r9d
-; SCALAR-NEXT:    andl $2097152, %r9d # imm = 0x200000
-; SCALAR-NEXT:    cmovneq %rbx, %r9
-; SCALAR-NEXT:    xorq %r11, %r9
-; SCALAR-NEXT:    xorq %r10, %r9
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $22, %r10
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $4194304, %r11d # imm = 0x400000
-; SCALAR-NEXT:    cmovneq %r10, %r11
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $23, %r10
-; SCALAR-NEXT:    movl %esi, %ebx
-; SCALAR-NEXT:    andl $8388608, %ebx # imm = 0x800000
-; SCALAR-NEXT:    cmovneq %r10, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $24, %r10
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $16777216, %r11d # imm = 0x1000000
-; SCALAR-NEXT:    cmovneq %r10, %r11
-; SCALAR-NEXT:    xorq %rbx, %r11
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $25, %r10
-; SCALAR-NEXT:    movl %esi, %ebx
-; SCALAR-NEXT:    andl $33554432, %ebx # imm = 0x2000000
-; SCALAR-NEXT:    cmovneq %r10, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $26, %r10
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $67108864, %r11d # imm = 0x4000000
-; SCALAR-NEXT:    cmovneq %r10, %r11
-; SCALAR-NEXT:    xorq %rbx, %r11
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $27, %r10
-; SCALAR-NEXT:    movl %esi, %ebx
-; SCALAR-NEXT:    andl $134217728, %ebx # imm = 0x8000000
-; SCALAR-NEXT:    cmovneq %r10, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r11
-; SCALAR-NEXT:    shlq $28, %r11
-; SCALAR-NEXT:    movl %esi, %r10d
-; SCALAR-NEXT:    andl $268435456, %r10d # imm = 0x10000000
-; SCALAR-NEXT:    cmovneq %r11, %r10
-; SCALAR-NEXT:    xorq %rbx, %r10
-; SCALAR-NEXT:    xorq %r9, %r10
-; SCALAR-NEXT:    movq %r8, %r9
-; SCALAR-NEXT:    shlq $29, %r9
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $536870912, %r11d # imm = 0x20000000
-; SCALAR-NEXT:    cmovneq %r9, %r11
-; SCALAR-NEXT:    movq %r8, %r9
-; SCALAR-NEXT:    shlq $30, %r9
-; SCALAR-NEXT:    movl %esi, %ebx
-; SCALAR-NEXT:    andl $1073741824, %ebx # imm = 0x40000000
-; SCALAR-NEXT:    cmovneq %r9, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
+; SCALAR-NEXT:    leaq (%rsi,%r8,2), %r8
+; SCALAR-NEXT:    movabsq $1229782938247303441, %rbx # imm = 0x1111111111111111
+; SCALAR-NEXT:    movq %r8, %rsi
+; SCALAR-NEXT:    andq %rbx, %rsi
+; SCALAR-NEXT:    bswapq %rdi
+; SCALAR-NEXT:    movq %rdi, %r9
+; SCALAR-NEXT:    shrq $4, %r9
+; SCALAR-NEXT:    andq %rcx, %r9
+; SCALAR-NEXT:    andq %rcx, %rdi
+; SCALAR-NEXT:    shlq $4, %rdi
+; SCALAR-NEXT:    orq %r9, %rdi
+; SCALAR-NEXT:    movq %rdi, %r9
+; SCALAR-NEXT:    andq %rax, %r9
+; SCALAR-NEXT:    shrq $2, %rdi
+; SCALAR-NEXT:    andq %rax, %rdi
+; SCALAR-NEXT:    leaq (%rdi,%r9,4), %rdi
+; SCALAR-NEXT:    movq %rdi, %r9
+; SCALAR-NEXT:    andq %rdx, %r9
+; SCALAR-NEXT:    shrq %rdi
+; SCALAR-NEXT:    andq %rdx, %rdi
+; SCALAR-NEXT:    leaq (%rdi,%r9,2), %rdi
+; SCALAR-NEXT:    movabsq $2459565876494606882, %rbp # imm = 0x2222222222222222
+; SCALAR-NEXT:    movq %rdi, %r14
+; SCALAR-NEXT:    andq %rbp, %r14
+; SCALAR-NEXT:    movq %r14, %r10
+; SCALAR-NEXT:    imulq %rsi, %r10
 ; SCALAR-NEXT:    movq %r8, %r9
-; SCALAR-NEXT:    shlq $31, %r9
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $-2147483648, %r11d # imm = 0x80000000
-; SCALAR-NEXT:    cmovneq %r9, %r11
-; SCALAR-NEXT:    xorq %rbx, %r11
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $32, %rbx
-; SCALAR-NEXT:    xorl %r9d, %r9d
-; SCALAR-NEXT:    btq $32, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r11
-; SCALAR-NEXT:    shlq $33, %r11
-; SCALAR-NEXT:    btq $33, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r11
-; SCALAR-NEXT:    xorq %rbx, %r11
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $34, %rbx
-; SCALAR-NEXT:    btq $34, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r14
-; SCALAR-NEXT:    shlq $35, %r14
-; SCALAR-NEXT:    btq $35, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r14
-; SCALAR-NEXT:    xorq %rbx, %r14
-; SCALAR-NEXT:    movq %r8, %r11
-; SCALAR-NEXT:    shlq $36, %r11
-; SCALAR-NEXT:    btq $36, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r11
-; SCALAR-NEXT:    xorq %r14, %r11
-; SCALAR-NEXT:    xorq %r10, %r11
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $37, %r10
-; SCALAR-NEXT:    btq $37, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r10
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $38, %rbx
-; SCALAR-NEXT:    btq $38, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rbx
-; SCALAR-NEXT:    xorq %r10, %rbx
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $39, %r10
-; SCALAR-NEXT:    btq $39, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r10
-; SCALAR-NEXT:    xorq %rbx, %r10
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $40, %rbx
-; SCALAR-NEXT:    btq $40, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rbx
-; SCALAR-NEXT:    xorq %r10, %rbx
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $41, %r10
-; SCALAR-NEXT:    btq $41, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r10
-; SCALAR-NEXT:    xorq %rbx, %r10
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $42, %rbx
-; SCALAR-NEXT:    btq $42, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rbx
-; SCALAR-NEXT:    xorq %r10, %rbx
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $43, %r10
-; SCALAR-NEXT:    btq $43, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r10
-; SCALAR-NEXT:    xorq %rbx, %r10
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $44, %rbx
-; SCALAR-NEXT:    btq $44, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rbx
-; SCALAR-NEXT:    xorq %r10, %rbx
-; SCALAR-NEXT:    movq %r8, %r14
-; SCALAR-NEXT:    shlq $45, %r14
-; SCALAR-NEXT:    btq $45, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r14
-; SCALAR-NEXT:    xorq %rbx, %r14
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $46, %r10
-; SCALAR-NEXT:    btq $46, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r10
-; SCALAR-NEXT:    xorq %r14, %r10
-; SCALAR-NEXT:    xorq %r11, %r10
-; SCALAR-NEXT:    movq %r8, %r11
-; SCALAR-NEXT:    shlq $47, %r11
-; SCALAR-NEXT:    btq $47, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r11
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $48, %rbx
-; SCALAR-NEXT:    btq $48, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r11
-; SCALAR-NEXT:    shlq $49, %r11
-; SCALAR-NEXT:    btq $49, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r11
-; SCALAR-NEXT:    xorq %rbx, %r11
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $50, %rbx
-; SCALAR-NEXT:    btq $50, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r11
-; SCALAR-NEXT:    shlq $51, %r11
-; SCALAR-NEXT:    btq $51, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r11
-; SCALAR-NEXT:    xorq %rbx, %r11
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $52, %rbx
-; SCALAR-NEXT:    btq $52, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r11
-; SCALAR-NEXT:    shlq $53, %r11
-; SCALAR-NEXT:    btq $53, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r11
-; SCALAR-NEXT:    xorq %rbx, %r11
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $54, %rbx
-; SCALAR-NEXT:    btq $54, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r11
-; SCALAR-NEXT:    shlq $55, %r11
-; SCALAR-NEXT:    btq $55, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r11
-; SCALAR-NEXT:    xorq %rbx, %r11
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $56, %rbx
-; SCALAR-NEXT:    btq $56, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
+; SCALAR-NEXT:    andq %rbp, %r9
+; SCALAR-NEXT:    movq %rdi, %rdx
+; SCALAR-NEXT:    andq %rbx, %rdx
+; SCALAR-NEXT:    movq %rdx, %r13
+; SCALAR-NEXT:    imulq %r9, %r13
+; SCALAR-NEXT:    xorq %r10, %r13
+; SCALAR-NEXT:    movabsq $-8608480567731124088, %r10 # imm = 0x8888888888888888
 ; SCALAR-NEXT:    movq %r8, %r11
-; SCALAR-NEXT:    shlq $57, %r11
-; SCALAR-NEXT:    btq $57, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r11
-; SCALAR-NEXT:    xorq %rbx, %r11
-; SCALAR-NEXT:    xorq %r10, %r11
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $58, %r10
-; SCALAR-NEXT:    btq $58, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r10
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $59, %rbx
-; SCALAR-NEXT:    btq $59, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rbx
-; SCALAR-NEXT:    xorq %r10, %rbx
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $60, %r10
-; SCALAR-NEXT:    btq $60, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r10
-; SCALAR-NEXT:    xorq %rbx, %r10
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $61, %rbx
-; SCALAR-NEXT:    btq $61, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rbx
-; SCALAR-NEXT:    xorq %r10, %rbx
-; SCALAR-NEXT:    shlq $62, %r8
-; SCALAR-NEXT:    btq $62, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r8
-; SCALAR-NEXT:    xorq %rbx, %r8
-; SCALAR-NEXT:    shlq $63, %rdi
-; SCALAR-NEXT:    btq $63, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
+; SCALAR-NEXT:    andq %r10, %r11
+; SCALAR-NEXT:    movabsq $4919131752989213764, %r12 # imm = 0x4444444444444444
+; SCALAR-NEXT:    movq %rdi, %rcx
+; SCALAR-NEXT:    andq %r12, %rcx
+; SCALAR-NEXT:    movq %rcx, %r15
+; SCALAR-NEXT:    imulq %r11, %r15
+; SCALAR-NEXT:    andq %r12, %r8
+; SCALAR-NEXT:    andq %r10, %rdi
+; SCALAR-NEXT:    movq %rdi, %rax
+; SCALAR-NEXT:    imulq %r8, %rax
+; SCALAR-NEXT:    xorq %r15, %rax
+; SCALAR-NEXT:    xorq %r13, %rax
+; SCALAR-NEXT:    andq %rbp, %rax
+; SCALAR-NEXT:    movq %r14, %r15
+; SCALAR-NEXT:    imulq %r11, %r15
+; SCALAR-NEXT:    movq %rdx, %r13
+; SCALAR-NEXT:    imulq %rsi, %r13
+; SCALAR-NEXT:    xorq %r15, %r13
+; SCALAR-NEXT:    movq %rcx, %r15
+; SCALAR-NEXT:    imulq %r8, %r15
+; SCALAR-NEXT:    movq %rdi, %rbp
+; SCALAR-NEXT:    imulq %r9, %rbp
+; SCALAR-NEXT:    xorq %r15, %rbp
+; SCALAR-NEXT:    xorq %r13, %rbp
+; SCALAR-NEXT:    andq %rbx, %rbp
+; SCALAR-NEXT:    orq %rax, %rbp
+; SCALAR-NEXT:    movq %rdi, %rbx
+; SCALAR-NEXT:    imulq %r11, %rbx
+; SCALAR-NEXT:    imulq %rdx, %r11
+; SCALAR-NEXT:    imulq %r8, %rdx
+; SCALAR-NEXT:    imulq %r14, %r8
+; SCALAR-NEXT:    imulq %r9, %r14
+; SCALAR-NEXT:    xorq %r14, %rdx
+; SCALAR-NEXT:    movq %rcx, %rax
+; SCALAR-NEXT:    imulq %rsi, %rax
+; SCALAR-NEXT:    xorq %rax, %rbx
+; SCALAR-NEXT:    xorq %rdx, %rbx
+; SCALAR-NEXT:    andq %r12, %rbx
+; SCALAR-NEXT:    orq %rbp, %rbx
+; SCALAR-NEXT:    xorq %r8, %r11
+; SCALAR-NEXT:    imulq %r9, %rcx
+; SCALAR-NEXT:    imulq %rsi, %rdi
+; SCALAR-NEXT:    xorq %rcx, %rdi
 ; SCALAR-NEXT:    xorq %r11, %rdi
+; SCALAR-NEXT:    andq %r10, %rdi
+; SCALAR-NEXT:    orq %rbx, %rdi
 ; SCALAR-NEXT:    bswapq %rdi
-; SCALAR-NEXT:    movq %rdi, %rsi
-; SCALAR-NEXT:    shrq $4, %rsi
-; SCALAR-NEXT:    andq %rcx, %rsi
+; SCALAR-NEXT:    movq %rdi, %rax
+; SCALAR-NEXT:    shrq $4, %rax
+; SCALAR-NEXT:    movabsq $1085102592571150095, %rcx # imm = 0xF0F0F0F0F0F0F0F
+; SCALAR-NEXT:    andq %rcx, %rax
 ; SCALAR-NEXT:    andq %rcx, %rdi
 ; SCALAR-NEXT:    shlq $4, %rdi
-; SCALAR-NEXT:    orq %rsi, %rdi
-; SCALAR-NEXT:    movq %rdi, %rcx
-; SCALAR-NEXT:    andq %rax, %rcx
+; SCALAR-NEXT:    orq %rax, %rdi
+; SCALAR-NEXT:    movq %rdi, %rax
+; SCALAR-NEXT:    movabsq $3689348814741910323, %rcx # imm = 0x3333333333333333
+; SCALAR-NEXT:    andq %rcx, %rax
 ; SCALAR-NEXT:    shrq $2, %rdi
-; SCALAR-NEXT:    andq %rax, %rdi
-; SCALAR-NEXT:    leaq (%rdi,%rcx,4), %rax
+; SCALAR-NEXT:    andq %rcx, %rdi
+; SCALAR-NEXT:    leaq (%rdi,%rax,4), %rax
 ; SCALAR-NEXT:    movq %rax, %rcx
+; SCALAR-NEXT:    movabsq $6148914691236517205, %rdx # imm = 0x5555555555555555
 ; SCALAR-NEXT:    andq %rdx, %rcx
 ; SCALAR-NEXT:    shrq %rax
 ; SCALAR-NEXT:    andq %rdx, %rax
 ; SCALAR-NEXT:    leaq (%rax,%rcx,2), %rax
 ; SCALAR-NEXT:    popq %rbx
+; SCALAR-NEXT:    popq %r12
+; SCALAR-NEXT:    popq %r13
 ; SCALAR-NEXT:    popq %r14
+; SCALAR-NEXT:    popq %r15
+; SCALAR-NEXT:    popq %rbp
 ; SCALAR-NEXT:    retq
 ;
 ; SSE2-PCLMUL-LABEL: clmulr_i64:
@@ -1944,87 +885,73 @@ define i8 @clmulh_i8(i8 %a, i8 %b) nounwind {
 ; SCALAR-NEXT:    pushq %rbp
 ; SCALAR-NEXT:    pushq %r15
 ; SCALAR-NEXT:    pushq %r14
+; SCALAR-NEXT:    pushq %r13
 ; SCALAR-NEXT:    pushq %r12
 ; SCALAR-NEXT:    pushq %rbx
-; SCALAR-NEXT:    movzbl %dil, %ecx
-; SCALAR-NEXT:    movl %ecx, %ebx
-; SCALAR-NEXT:    shll $8, %ebx
-; SCALAR-NEXT:    movl %ecx, %r11d
-; SCALAR-NEXT:    shll $9, %r11d
+; SCALAR-NEXT:    movl %esi, %eax
+; SCALAR-NEXT:    andl $17, %eax
+; SCALAR-NEXT:    movl %edi, %ecx
+; SCALAR-NEXT:    andl $34, %ecx
 ; SCALAR-NEXT:    movl %ecx, %r10d
-; SCALAR-NEXT:    shll $10, %r10d
-; SCALAR-NEXT:    movl %ecx, %eax
-; SCALAR-NEXT:    shll $11, %eax
-; SCALAR-NEXT:    movl %ecx, %r9d
-; SCALAR-NEXT:    shll $12, %r9d
-; SCALAR-NEXT:    movl %ecx, %r8d
-; SCALAR-NEXT:    shll $13, %r8d
-; SCALAR-NEXT:    movl %edi, %edx
-; SCALAR-NEXT:    shll $14, %edx
-; SCALAR-NEXT:    xorl %ebp, %ebp
-; SCALAR-NEXT:    testw %bp, %bp
-; SCALAR-NEXT:    cmovel %ebp, %edx
-; SCALAR-NEXT:    cmovel %ebp, %r8d
-; SCALAR-NEXT:    cmovel %ebp, %r9d
-; SCALAR-NEXT:    cmovel %ebp, %eax
-; SCALAR-NEXT:    cmovel %ebp, %r10d
-; SCALAR-NEXT:    cmovel %ebp, %r11d
-; SCALAR-NEXT:    cmovel %ebp, %ebx
-; SCALAR-NEXT:    shll $15, %edi
-; SCALAR-NEXT:    testw %bp, %bp
-; SCALAR-NEXT:    cmovnel %edi, %ebp
-; SCALAR-NEXT:    movl %esi, %edi
-; SCALAR-NEXT:    andl $1, %edi
-; SCALAR-NEXT:    cmovnel %ecx, %edi
-; SCALAR-NEXT:    leal (%rcx,%rcx), %r14d
-; SCALAR-NEXT:    movl %esi, %r15d
-; SCALAR-NEXT:    andl $2, %r15d
-; SCALAR-NEXT:    cmovnel %r14d, %r15d
-; SCALAR-NEXT:    xorl %edi, %r15d
-; SCALAR-NEXT:    leal (,%rcx,4), %edi
-; SCALAR-NEXT:    movl %esi, %r14d
-; SCALAR-NEXT:    andl $4, %r14d
-; SCALAR-NEXT:    cmovnel %edi, %r14d
-; SCALAR-NEXT:    leal (,%rcx,8), %r12d
-; SCALAR-NEXT:    movl %esi, %edi
-; SCALAR-NEXT:    andl $8, %edi
-; SCALAR-NEXT:    cmovnel %r12d, %edi
-; SCALAR-NEXT:    xorl %r14d, %edi
-; SCALAR-NEXT:    xorl %r15d, %edi
-; SCALAR-NEXT:    movl %ecx, %r14d
-; SCALAR-NEXT:    shll $4, %r14d
-; SCALAR-NEXT:    movl %esi, %r15d
-; SCALAR-NEXT:    andl $16, %r15d
-; SCALAR-NEXT:    cmovnel %r14d, %r15d
-; SCALAR-NEXT:    movl %ecx, %r14d
-; SCALAR-NEXT:    shll $5, %r14d
-; SCALAR-NEXT:    movl %esi, %r12d
-; SCALAR-NEXT:    andl $32, %r12d
-; SCALAR-NEXT:    cmovnel %r14d, %r12d
-; SCALAR-NEXT:    xorl %r15d, %r12d
-; SCALAR-NEXT:    movl %ecx, %r14d
-; SCALAR-NEXT:    shll $6, %r14d
-; SCALAR-NEXT:    movl %esi, %r15d
-; SCALAR-NEXT:    andl $64, %r15d
-; SCALAR-NEXT:    cmovnel %r14d, %r15d
-; SCALAR-NEXT:    xorl %r12d, %r15d
-; SCALAR-NEXT:    xorl %edi, %r15d
-; SCALAR-NEXT:    shll $7, %ecx
-; SCALAR-NEXT:    andl $128, %esi
-; SCALAR-NEXT:    cmovel %esi, %ecx
-; SCALAR-NEXT:    xorl %ebx, %ecx
-; SCALAR-NEXT:    xorl %r11d, %ecx
-; SCALAR-NEXT:    xorl %r10d, %ecx
-; SCALAR-NEXT:    xorl %r15d, %ecx
+; SCALAR-NEXT:    imull %eax, %r10d
+; SCALAR-NEXT:    movl %esi, %edx
+; SCALAR-NEXT:    andl $34, %edx
+; SCALAR-NEXT:    movl %edi, %ebx
+; SCALAR-NEXT:    andl $17, %ebx
+; SCALAR-NEXT:    movl %esi, %r8d
+; SCALAR-NEXT:    andl $136, %r8d
+; SCALAR-NEXT:    movl %edi, %r9d
+; SCALAR-NEXT:    andl $68, %r9d
+; SCALAR-NEXT:    movl %r9d, %r11d
+; SCALAR-NEXT:    imull %r8d, %r11d
+; SCALAR-NEXT:    andl $136, %edi
+; SCALAR-NEXT:    movl %ecx, %r15d
+; SCALAR-NEXT:    imull %r8d, %r15d
+; SCALAR-NEXT:    movl %ebx, %ebp
+; SCALAR-NEXT:    movl %ebx, %r12d
+; SCALAR-NEXT:    movl %edi, %r14d
+; SCALAR-NEXT:    imull %r8d, %r14d
+; SCALAR-NEXT:    imull %ebx, %r8d
+; SCALAR-NEXT:    imull %edx, %ebx
+; SCALAR-NEXT:    xorl %r10d, %ebx
+; SCALAR-NEXT:    andl $68, %esi
+; SCALAR-NEXT:    imull %eax, %ebp
+; SCALAR-NEXT:    movl %edi, %r10d
+; SCALAR-NEXT:    movl %r9d, %r13d
+; SCALAR-NEXT:    imull %eax, %r13d
+; SCALAR-NEXT:    imull %edi, %eax
+; SCALAR-NEXT:    imull %esi, %edi
+; SCALAR-NEXT:    xorl %r11d, %edi
+; SCALAR-NEXT:    xorl %ebx, %edi
+; SCALAR-NEXT:    xorl %r15d, %ebp
+; SCALAR-NEXT:    movl %r9d, %r11d
+; SCALAR-NEXT:    imull %esi, %r11d
+; SCALAR-NEXT:    imull %edx, %r10d
+; SCALAR-NEXT:    xorl %r11d, %r10d
+; SCALAR-NEXT:    andl $8704, %edi # imm = 0x2200
+; SCALAR-NEXT:    xorl %ebp, %r10d
+; SCALAR-NEXT:    andl $4352, %r10d # imm = 0x1100
+; SCALAR-NEXT:    orl %edi, %r10d
+; SCALAR-NEXT:    movl %ecx, %edi
+; SCALAR-NEXT:    imull %edx, %edi
+; SCALAR-NEXT:    imull %esi, %r12d
+; SCALAR-NEXT:    xorl %edi, %r12d
+; SCALAR-NEXT:    xorl %r13d, %r14d
+; SCALAR-NEXT:    xorl %r12d, %r14d
+; SCALAR-NEXT:    andl $17408, %r14d # imm = 0x4400
+; SCALAR-NEXT:    orl %r10d, %r14d
+; SCALAR-NEXT:    imull %esi, %ecx
+; SCALAR-NEXT:    xorl %ecx, %r8d
+; SCALAR-NEXT:    imull %edx, %r9d
 ; SCALAR-NEXT:    xorl %r9d, %eax
 ; SCALAR-NEXT:    xorl %r8d, %eax
-; SCALAR-NEXT:    xorl %edx, %eax
-; SCALAR-NEXT:    xorl %ebp, %eax
-; SCALAR-NEXT:    xorl %ecx, %eax
+; SCALAR-NEXT:    andl $2048, %eax # imm = 0x800
+; SCALAR-NEXT:    orl %r14d, %eax
 ; SCALAR-NEXT:    shrl $8, %eax
 ; SCALAR-NEXT:    # kill: def $al killed $al killed $eax
 ; SCALAR-NEXT:    popq %rbx
 ; SCALAR-NEXT:    popq %r12
+; SCALAR-NEXT:    popq %r13
 ; SCALAR-NEXT:    popq %r14
 ; SCALAR-NEXT:    popq %r15
 ; SCALAR-NEXT:    popq %rbp
@@ -2070,187 +997,67 @@ define i16 @clmulh_i16(i16 %a, i16 %b) nounwind {
 ; SCALAR-NEXT:    pushq %r13
 ; SCALAR-NEXT:    pushq %r12
 ; SCALAR-NEXT:    pushq %rbx
-; SCALAR-NEXT:    movl %esi, %r9d
-; SCALAR-NEXT:    movl %edi, %r14d
-; SCALAR-NEXT:    movl %edi, %r13d
-; SCALAR-NEXT:    movl %edi, %r12d
-; SCALAR-NEXT:    movl %edi, %r15d
-; SCALAR-NEXT:    movl %edi, %ebp
-; SCALAR-NEXT:    movl %edi, %r10d
-; SCALAR-NEXT:    movl %edi, %ebx
-; SCALAR-NEXT:    movl %edi, %r11d
-; SCALAR-NEXT:    movl %edi, %esi
-; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    movl %edi, %edx
-; SCALAR-NEXT:    movl %edi, %r8d
-; SCALAR-NEXT:    shll $16, %r8d
-; SCALAR-NEXT:    movl %r8d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SCALAR-NEXT:    shll $17, %r14d
-; SCALAR-NEXT:    movl %r14d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SCALAR-NEXT:    shll $18, %r13d
-; SCALAR-NEXT:    movl %r13d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SCALAR-NEXT:    shll $19, %r12d
-; SCALAR-NEXT:    movl %r12d, %r13d
-; SCALAR-NEXT:    shll $20, %r15d
-; SCALAR-NEXT:    movl %r15d, %r12d
-; SCALAR-NEXT:    shll $21, %ebp
-; SCALAR-NEXT:    movl %ebp, %r15d
-; SCALAR-NEXT:    shll $22, %r10d
-; SCALAR-NEXT:    shll $23, %ebx
-; SCALAR-NEXT:    movl %ebx, %r14d
-; SCALAR-NEXT:    shll $24, %r11d
-; SCALAR-NEXT:    movl %r11d, %ebp
-; SCALAR-NEXT:    shll $25, %esi
-; SCALAR-NEXT:    movl %esi, %ebx
-; SCALAR-NEXT:    shll $26, %ecx
-; SCALAR-NEXT:    movl %ecx, %r8d
-; SCALAR-NEXT:    shll $27, %eax
-; SCALAR-NEXT:    movl %eax, %esi
-; SCALAR-NEXT:    shll $28, %edx
-; SCALAR-NEXT:    movl %edi, %r11d
-; SCALAR-NEXT:    shll $29, %r11d
+; SCALAR-NEXT:    movl %esi, %eax
+; SCALAR-NEXT:    andl $4369, %eax # imm = 0x1111
 ; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    shll $30, %ecx
-; SCALAR-NEXT:    xorl %eax, %eax
-; SCALAR-NEXT:    testl $0, %eax
-; SCALAR-NEXT:    cmovel %eax, %ecx
-; SCALAR-NEXT:    movl %ecx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SCALAR-NEXT:    cmovel %eax, %r11d
-; SCALAR-NEXT:    cmovel %eax, %edx
-; SCALAR-NEXT:    movl %edx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SCALAR-NEXT:    cmovel %eax, %esi
-; SCALAR-NEXT:    movl %esi, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SCALAR-NEXT:    cmovel %eax, %r8d
-; SCALAR-NEXT:    movl %r8d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SCALAR-NEXT:    cmovel %eax, %ebx
-; SCALAR-NEXT:    movl %ebx, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SCALAR-NEXT:    cmovel %eax, %ebp
-; SCALAR-NEXT:    movl %ebp, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SCALAR-NEXT:    cmovel %eax, %r14d
-; SCALAR-NEXT:    movl %r14d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SCALAR-NEXT:    cmovel %eax, %r10d
-; SCALAR-NEXT:    cmovel %eax, %r15d
-; SCALAR-NEXT:    movl %r15d, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SCALAR-NEXT:    cmovel %eax, %r12d
-; SCALAR-NEXT:    cmovel %eax, %r13d
-; SCALAR-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %r15d # 4-byte Reload
-; SCALAR-NEXT:    cmovel %eax, %r15d
-; SCALAR-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %r14d # 4-byte Reload
-; SCALAR-NEXT:    cmovel %eax, %r14d
-; SCALAR-NEXT:    movl {{[-0-9]+}}(%r{{[sb]}}p), %r8d # 4-byte Reload
-; SCALAR-NEXT:    cmovel %eax, %r8d
+; SCALAR-NEXT:    andl $8738, %ecx # imm = 0x2222
+; SCALAR-NEXT:    movl %ecx, %r10d
+; SCALAR-NEXT:    imull %eax, %r10d
+; SCALAR-NEXT:    movl %esi, %edx
+; SCALAR-NEXT:    andl $8738, %edx # imm = 0x2222
 ; SCALAR-NEXT:    movl %edi, %ebx
-; SCALAR-NEXT:    shll $31, %ebx
-; SCALAR-NEXT:    testl $0, %eax
-; SCALAR-NEXT:    cmovel %eax, %ebx
-; SCALAR-NEXT:    movzwl %di, %edi
-; SCALAR-NEXT:    movl %r9d, %eax
-; SCALAR-NEXT:    andl $1, %eax
-; SCALAR-NEXT:    cmovnel %edi, %eax
-; SCALAR-NEXT:    movl %r9d, %ecx
-; SCALAR-NEXT:    andl $2, %ecx
-; SCALAR-NEXT:    leal (%rdi,%rdi), %esi
-; SCALAR-NEXT:    cmovnel %esi, %ecx
-; SCALAR-NEXT:    xorl %eax, %ecx
-; SCALAR-NEXT:    movl %r9d, %eax
-; SCALAR-NEXT:    andl $4, %eax
-; SCALAR-NEXT:    leal (,%rdi,4), %esi
-; SCALAR-NEXT:    cmovnel %esi, %eax
-; SCALAR-NEXT:    movl %r9d, %esi
-; SCALAR-NEXT:    andl $8, %esi
-; SCALAR-NEXT:    leal (,%rdi,8), %ebp
-; SCALAR-NEXT:    cmovnel %ebp, %esi
-; SCALAR-NEXT:    xorl %eax, %esi
-; SCALAR-NEXT:    xorl %ecx, %esi
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    shll $4, %eax
-; SCALAR-NEXT:    movl %r9d, %ecx
-; SCALAR-NEXT:    andl $16, %ecx
-; SCALAR-NEXT:    cmovnel %eax, %ecx
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    shll $5, %eax
-; SCALAR-NEXT:    movl %r9d, %ebp
-; SCALAR-NEXT:    andl $32, %ebp
-; SCALAR-NEXT:    cmovnel %eax, %ebp
-; SCALAR-NEXT:    xorl %ecx, %ebp
-; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    shll $6, %ecx
-; SCALAR-NEXT:    movl %r9d, %eax
-; SCALAR-NEXT:    andl $64, %eax
-; SCALAR-NEXT:    cmovnel %ecx, %eax
-; SCALAR-NEXT:    xorl %ebp, %eax
-; SCALAR-NEXT:    xorl %esi, %eax
-; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    shll $7, %ecx
-; SCALAR-NEXT:    movl %r9d, %esi
-; SCALAR-NEXT:    andl $128, %esi
-; SCALAR-NEXT:    cmovnel %ecx, %esi
-; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    shll $8, %ecx
-; SCALAR-NEXT:    movl %r9d, %ebp
-; SCALAR-NEXT:    andl $256, %ebp # imm = 0x100
-; SCALAR-NEXT:    cmovnel %ecx, %ebp
-; SCALAR-NEXT:    xorl %esi, %ebp
-; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    shll $9, %ecx
-; SCALAR-NEXT:    movl %r9d, %edx
-; SCALAR-NEXT:    andl $512, %edx # imm = 0x200
-; SCALAR-NEXT:    cmovnel %ecx, %edx
-; SCALAR-NEXT:    xorl %ebp, %edx
-; SCALAR-NEXT:    movl %edi, %ecx
-; SCALAR-NEXT:    shll $10, %ecx
-; SCALAR-NEXT:    movl %r9d, %esi
-; SCALAR-NEXT:    andl $1024, %esi # imm = 0x400
-; SCALAR-NEXT:    cmovnel %ecx, %esi
-; SCALAR-NEXT:    xorl %edx, %esi
-; SCALAR-NEXT:    xorl %eax, %esi
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    shll $11, %eax
-; SCALAR-NEXT:    movl %r9d, %ecx
-; SCALAR-NEXT:    andl $2048, %ecx # imm = 0x800
-; SCALAR-NEXT:    cmovnel %eax, %ecx
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    shll $12, %eax
-; SCALAR-NEXT:    movl %r9d, %edx
-; SCALAR-NEXT:    andl $4096, %edx # imm = 0x1000
-; SCALAR-NEXT:    cmovnel %eax, %edx
-; SCALAR-NEXT:    xorl %ecx, %edx
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    shll $13, %eax
-; SCALAR-NEXT:    movl %r9d, %ecx
-; SCALAR-NEXT:    andl $8192, %ecx # imm = 0x2000
-; SCALAR-NEXT:    cmovnel %eax, %ecx
-; SCALAR-NEXT:    xorl %edx, %ecx
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    shll $14, %eax
-; SCALAR-NEXT:    movl %r9d, %edx
-; SCALAR-NEXT:    andl $16384, %edx # imm = 0x4000
-; SCALAR-NEXT:    cmovnel %eax, %edx
-; SCALAR-NEXT:    xorl %ecx, %edx
-; SCALAR-NEXT:    shll $15, %edi
-; SCALAR-NEXT:    andl $32768, %r9d # imm = 0x8000
-; SCALAR-NEXT:    cmovel %r9d, %edi
-; SCALAR-NEXT:    xorl %edx, %edi
-; SCALAR-NEXT:    xorl %esi, %edi
-; SCALAR-NEXT:    xorl %r14d, %r8d
-; SCALAR-NEXT:    xorl %r15d, %r8d
-; SCALAR-NEXT:    xorl %r13d, %r8d
-; SCALAR-NEXT:    xorl %r12d, %r8d
-; SCALAR-NEXT:    xorl {{[-0-9]+}}(%r{{[sb]}}p), %r8d # 4-byte Folded Reload
-; SCALAR-NEXT:    xorl %edi, %r8d
-; SCALAR-NEXT:    xorl {{[-0-9]+}}(%r{{[sb]}}p), %r10d # 4-byte Folded Reload
-; SCALAR-NEXT:    xorl {{[-0-9]+}}(%r{{[sb]}}p), %r10d # 4-byte Folded Reload
-; SCALAR-NEXT:    xorl {{[-0-9]+}}(%r{{[sb]}}p), %r10d # 4-byte Folded Reload
-; SCALAR-NEXT:    xorl {{[-0-9]+}}(%r{{[sb]}}p), %r10d # 4-byte Folded Reload
-; SCALAR-NEXT:    xorl {{[-0-9]+}}(%r{{[sb]}}p), %r10d # 4-byte Folded Reload
-; SCALAR-NEXT:    xorl {{[-0-9]+}}(%r{{[sb]}}p), %r10d # 4-byte Folded Reload
-; SCALAR-NEXT:    xorl %r8d, %r10d
-; SCALAR-NEXT:    xorl {{[-0-9]+}}(%r{{[sb]}}p), %r11d # 4-byte Folded Reload
-; SCALAR-NEXT:    xorl %ebx, %r11d
-; SCALAR-NEXT:    xorl %r10d, %r11d
-; SCALAR-NEXT:    shrl $16, %r11d
-; SCALAR-NEXT:    movl %r11d, %eax
+; SCALAR-NEXT:    andl $4369, %ebx # imm = 0x1111
+; SCALAR-NEXT:    movl %esi, %r8d
+; SCALAR-NEXT:    andl $34952, %r8d # imm = 0x8888
+; SCALAR-NEXT:    movl %edi, %r9d
+; SCALAR-NEXT:    andl $17476, %r9d # imm = 0x4444
+; SCALAR-NEXT:    movl %r9d, %r11d
+; SCALAR-NEXT:    imull %r8d, %r11d
+; SCALAR-NEXT:    andl $34952, %edi # imm = 0x8888
+; SCALAR-NEXT:    movl %ecx, %r15d
+; SCALAR-NEXT:    imull %r8d, %r15d
+; SCALAR-NEXT:    movl %ebx, %ebp
+; SCALAR-NEXT:    movl %ebx, %r12d
+; SCALAR-NEXT:    movl %edi, %r14d
+; SCALAR-NEXT:    imull %r8d, %r14d
+; SCALAR-NEXT:    imull %ebx, %r8d
+; SCALAR-NEXT:    imull %edx, %ebx
+; SCALAR-NEXT:    xorl %r10d, %ebx
+; SCALAR-NEXT:    andl $17476, %esi # imm = 0x4444
+; SCALAR-NEXT:    imull %eax, %ebp
+; SCALAR-NEXT:    movl %edi, %r10d
+; SCALAR-NEXT:    movl %r9d, %r13d
+; SCALAR-NEXT:    imull %eax, %r13d
+; SCALAR-NEXT:    imull %edi, %eax
+; SCALAR-NEXT:    imull %esi, %edi
+; SCALAR-NEXT:    xorl %r11d, %edi
+; SCALAR-NEXT:    xorl %ebx, %edi
+; SCALAR-NEXT:    xorl %r15d, %ebp
+; SCALAR-NEXT:    movl %r9d, %r11d
+; SCALAR-NEXT:    imull %esi, %r11d
+; SCALAR-NEXT:    imull %edx, %r10d
+; SCALAR-NEXT:    xorl %r11d, %r10d
+; SCALAR-NEXT:    andl $572653568, %edi # imm = 0x22220000
+; SCALAR-NEXT:    xorl %ebp, %r10d
+; SCALAR-NEXT:    andl $286326784, %r10d # imm = 0x11110000
+; SCALAR-NEXT:    orl %edi, %r10d
+; SCALAR-NEXT:    movl %ecx, %edi
+; SCALAR-NEXT:    imull %edx, %edi
+; SCALAR-NEXT:    imull %esi, %r12d
+; SCALAR-NEXT:    xorl %edi, %r12d
+; SCALAR-NEXT:    xorl %r13d, %r14d
+; SCALAR-NEXT:    xorl %r12d, %r14d
+; SCALAR-NEXT:    andl $1145307136, %r14d # imm = 0x44440000
+; SCALAR-NEXT:    orl %r10d, %r14d
+; SCALAR-NEXT:    imull %esi, %ecx
+; SCALAR-NEXT:    xorl %ecx, %r8d
+; SCALAR-NEXT:    imull %edx, %r9d
+; SCALAR-NEXT:    xorl %r9d, %eax
+; SCALAR-NEXT:    xorl %r8d, %eax
+; SCALAR-NEXT:    andl $143130624, %eax # imm = 0x8880000
+; SCALAR-NEXT:    orl %r14d, %eax
+; SCALAR-NEXT:    shrl $16, %eax
+; SCALAR-NEXT:    # kill: def $ax killed $ax killed $eax
 ; SCALAR-NEXT:    popq %rbx
 ; SCALAR-NEXT:    popq %r12
 ; SCALAR-NEXT:    popq %r13
@@ -2293,356 +1100,81 @@ define i16 @clmulh_i16(i16 %a, i16 %b) nounwind {
 define i32 @clmulh_i32(i32 %a, i32 %b) nounwind {
 ; SCALAR-LABEL: clmulh_i32:
 ; SCALAR:       # %bb.0:
-; SCALAR-NEXT:    movl %edi, %eax
-; SCALAR-NEXT:    leaq (%rax,%rax), %rdx
-; SCALAR-NEXT:    movl %esi, %ecx
-; SCALAR-NEXT:    movl %ecx, %esi
-; SCALAR-NEXT:    andl $2, %esi
-; SCALAR-NEXT:    cmovneq %rdx, %rsi
-; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $1, %edi
-; SCALAR-NEXT:    cmovneq %rax, %rdi
-; SCALAR-NEXT:    xorq %rsi, %rdi
-; SCALAR-NEXT:    leaq (,%rax,4), %rdx
-; SCALAR-NEXT:    movl %ecx, %esi
-; SCALAR-NEXT:    andl $4, %esi
-; SCALAR-NEXT:    cmovneq %rdx, %rsi
-; SCALAR-NEXT:    leaq (,%rax,8), %r8
-; SCALAR-NEXT:    movl %ecx, %edx
-; SCALAR-NEXT:    andl $8, %edx
-; SCALAR-NEXT:    cmovneq %r8, %rdx
-; SCALAR-NEXT:    xorq %rsi, %rdx
-; SCALAR-NEXT:    xorq %rdi, %rdx
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $4, %rsi
-; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $16, %edi
-; SCALAR-NEXT:    cmovneq %rsi, %rdi
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $5, %rsi
-; SCALAR-NEXT:    movl %ecx, %r8d
-; SCALAR-NEXT:    andl $32, %r8d
-; SCALAR-NEXT:    cmovneq %rsi, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rdi
-; SCALAR-NEXT:    shlq $6, %rdi
-; SCALAR-NEXT:    movl %ecx, %esi
-; SCALAR-NEXT:    andl $64, %esi
-; SCALAR-NEXT:    cmovneq %rdi, %rsi
-; SCALAR-NEXT:    xorq %r8, %rsi
-; SCALAR-NEXT:    xorq %rdx, %rsi
-; SCALAR-NEXT:    movq %rax, %rdx
-; SCALAR-NEXT:    shlq $7, %rdx
-; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $128, %edi
-; SCALAR-NEXT:    cmovneq %rdx, %rdi
-; SCALAR-NEXT:    movq %rax, %rdx
-; SCALAR-NEXT:    shlq $8, %rdx
-; SCALAR-NEXT:    movl %ecx, %r8d
-; SCALAR-NEXT:    andl $256, %r8d # imm = 0x100
-; SCALAR-NEXT:    cmovneq %rdx, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rdx
-; SCALAR-NEXT:    shlq $9, %rdx
-; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $512, %edi # imm = 0x200
-; SCALAR-NEXT:    cmovneq %rdx, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $10, %r8
-; SCALAR-NEXT:    movl %ecx, %edx
-; SCALAR-NEXT:    andl $1024, %edx # imm = 0x400
-; SCALAR-NEXT:    cmovneq %r8, %rdx
-; SCALAR-NEXT:    xorq %rdi, %rdx
-; SCALAR-NEXT:    xorq %rsi, %rdx
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $11, %rsi
-; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $2048, %edi # imm = 0x800
-; SCALAR-NEXT:    cmovneq %rsi, %rdi
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $12, %rsi
-; SCALAR-NEXT:    movl %ecx, %r8d
-; SCALAR-NEXT:    andl $4096, %r8d # imm = 0x1000
-; SCALAR-NEXT:    cmovneq %rsi, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $13, %rsi
-; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $8192, %edi # imm = 0x2000
-; SCALAR-NEXT:    cmovneq %rsi, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $14, %rsi
-; SCALAR-NEXT:    movl %ecx, %r8d
-; SCALAR-NEXT:    andl $16384, %r8d # imm = 0x4000
-; SCALAR-NEXT:    cmovneq %rsi, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rdi
-; SCALAR-NEXT:    shlq $15, %rdi
-; SCALAR-NEXT:    movl %ecx, %esi
-; SCALAR-NEXT:    andl $32768, %esi # imm = 0x8000
-; SCALAR-NEXT:    cmovneq %rdi, %rsi
-; SCALAR-NEXT:    xorq %r8, %rsi
-; SCALAR-NEXT:    xorq %rdx, %rsi
-; SCALAR-NEXT:    movq %rax, %rdx
-; SCALAR-NEXT:    shlq $16, %rdx
-; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $65536, %edi # imm = 0x10000
-; SCALAR-NEXT:    cmovneq %rdx, %rdi
-; SCALAR-NEXT:    movq %rax, %rdx
-; SCALAR-NEXT:    shlq $17, %rdx
-; SCALAR-NEXT:    movl %ecx, %r8d
-; SCALAR-NEXT:    andl $131072, %r8d # imm = 0x20000
-; SCALAR-NEXT:    cmovneq %rdx, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rdx
-; SCALAR-NEXT:    shlq $18, %rdx
-; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $262144, %edi # imm = 0x40000
-; SCALAR-NEXT:    cmovneq %rdx, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
-; SCALAR-NEXT:    movq %rax, %rdx
-; SCALAR-NEXT:    shlq $19, %rdx
-; SCALAR-NEXT:    movl %ecx, %r8d
-; SCALAR-NEXT:    andl $524288, %r8d # imm = 0x80000
-; SCALAR-NEXT:    cmovneq %rdx, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rdx
-; SCALAR-NEXT:    shlq $20, %rdx
-; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $1048576, %edi # imm = 0x100000
-; SCALAR-NEXT:    cmovneq %rdx, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $21, %r8
-; SCALAR-NEXT:    movl %ecx, %edx
-; SCALAR-NEXT:    andl $2097152, %edx # imm = 0x200000
-; SCALAR-NEXT:    cmovneq %r8, %rdx
-; SCALAR-NEXT:    xorq %rdi, %rdx
-; SCALAR-NEXT:    xorq %rsi, %rdx
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $22, %rsi
-; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $4194304, %edi # imm = 0x400000
-; SCALAR-NEXT:    cmovneq %rsi, %rdi
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $23, %rsi
-; SCALAR-NEXT:    movl %ecx, %r8d
-; SCALAR-NEXT:    andl $8388608, %r8d # imm = 0x800000
-; SCALAR-NEXT:    cmovneq %rsi, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $24, %rsi
-; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $16777216, %edi # imm = 0x1000000
-; SCALAR-NEXT:    cmovneq %rsi, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $25, %rsi
-; SCALAR-NEXT:    movl %ecx, %r8d
-; SCALAR-NEXT:    andl $33554432, %r8d # imm = 0x2000000
-; SCALAR-NEXT:    cmovneq %rsi, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $26, %rsi
-; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $67108864, %edi # imm = 0x4000000
-; SCALAR-NEXT:    cmovneq %rsi, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $27, %rsi
-; SCALAR-NEXT:    movl %ecx, %r8d
-; SCALAR-NEXT:    andl $134217728, %r8d # imm = 0x8000000
-; SCALAR-NEXT:    cmovneq %rsi, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rdi
-; SCALAR-NEXT:    shlq $28, %rdi
+; SCALAR-NEXT:    pushq %r15
+; SCALAR-NEXT:    pushq %r14
+; SCALAR-NEXT:    pushq %rbx
+; SCALAR-NEXT:    movl %esi, %edx
+; SCALAR-NEXT:    movl %edx, %eax
+; SCALAR-NEXT:    andl $286331153, %eax # imm = 0x11111111
+; SCALAR-NEXT:    movl %edi, %ecx
 ; SCALAR-NEXT:    movl %ecx, %esi
-; SCALAR-NEXT:    andl $268435456, %esi # imm = 0x10000000
-; SCALAR-NEXT:    cmovneq %rdi, %rsi
-; SCALAR-NEXT:    xorq %r8, %rsi
-; SCALAR-NEXT:    xorq %rdx, %rsi
-; SCALAR-NEXT:    movq %rax, %rdx
-; SCALAR-NEXT:    shlq $29, %rdx
-; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $536870912, %edi # imm = 0x20000000
-; SCALAR-NEXT:    cmovneq %rdx, %rdi
-; SCALAR-NEXT:    movq %rax, %rdx
-; SCALAR-NEXT:    shlq $30, %rdx
-; SCALAR-NEXT:    movl %ecx, %r8d
-; SCALAR-NEXT:    andl $1073741824, %r8d # imm = 0x40000000
-; SCALAR-NEXT:    cmovneq %rdx, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rdx
-; SCALAR-NEXT:    shlq $31, %rdx
+; SCALAR-NEXT:    andl $572662306, %esi # imm = 0x22222222
+; SCALAR-NEXT:    movq %rsi, %r9
+; SCALAR-NEXT:    imulq %rax, %r9
+; SCALAR-NEXT:    movl %edx, %r8d
+; SCALAR-NEXT:    andl $572662306, %r8d # imm = 0x22222222
 ; SCALAR-NEXT:    movl %ecx, %edi
-; SCALAR-NEXT:    andl $-2147483648, %edi # imm = 0x80000000
-; SCALAR-NEXT:    cmovneq %rdx, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $32, %r8
-; SCALAR-NEXT:    xorl %edx, %edx
-; SCALAR-NEXT:    btq $32, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rdi
-; SCALAR-NEXT:    shlq $33, %rdi
-; SCALAR-NEXT:    btq $33, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $34, %r8
-; SCALAR-NEXT:    btq $34, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %r9
-; SCALAR-NEXT:    shlq $35, %r9
-; SCALAR-NEXT:    btq $35, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r9
-; SCALAR-NEXT:    xorq %r8, %r9
-; SCALAR-NEXT:    movq %rax, %rdi
-; SCALAR-NEXT:    shlq $36, %rdi
-; SCALAR-NEXT:    btq $36, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rdi
-; SCALAR-NEXT:    xorq %r9, %rdi
-; SCALAR-NEXT:    xorq %rsi, %rdi
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $37, %rsi
-; SCALAR-NEXT:    btq $37, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rsi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $38, %r8
-; SCALAR-NEXT:    btq $38, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r8
-; SCALAR-NEXT:    xorq %rsi, %r8
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $39, %rsi
-; SCALAR-NEXT:    btq $39, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rsi
-; SCALAR-NEXT:    xorq %r8, %rsi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $40, %r8
-; SCALAR-NEXT:    btq $40, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r8
-; SCALAR-NEXT:    xorq %rsi, %r8
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $41, %rsi
-; SCALAR-NEXT:    btq $41, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rsi
-; SCALAR-NEXT:    xorq %r8, %rsi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $42, %r8
-; SCALAR-NEXT:    btq $42, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r8
-; SCALAR-NEXT:    xorq %rsi, %r8
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $43, %rsi
-; SCALAR-NEXT:    btq $43, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rsi
-; SCALAR-NEXT:    xorq %r8, %rsi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $44, %r8
-; SCALAR-NEXT:    btq $44, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r8
-; SCALAR-NEXT:    xorq %rsi, %r8
-; SCALAR-NEXT:    movq %rax, %r9
-; SCALAR-NEXT:    shlq $45, %r9
-; SCALAR-NEXT:    btq $45, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r9
-; SCALAR-NEXT:    xorq %r8, %r9
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $46, %rsi
-; SCALAR-NEXT:    btq $46, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rsi
-; SCALAR-NEXT:    xorq %r9, %rsi
-; SCALAR-NEXT:    xorq %rdi, %rsi
-; SCALAR-NEXT:    movq %rax, %rdi
-; SCALAR-NEXT:    shlq $47, %rdi
-; SCALAR-NEXT:    btq $47, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rdi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $48, %r8
-; SCALAR-NEXT:    btq $48, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rdi
-; SCALAR-NEXT:    shlq $49, %rdi
-; SCALAR-NEXT:    btq $49, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $50, %r8
-; SCALAR-NEXT:    btq $50, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rdi
-; SCALAR-NEXT:    shlq $51, %rdi
-; SCALAR-NEXT:    btq $51, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $52, %r8
-; SCALAR-NEXT:    btq $52, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rdi
-; SCALAR-NEXT:    shlq $53, %rdi
-; SCALAR-NEXT:    btq $53, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $54, %r8
-; SCALAR-NEXT:    btq $54, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rdi
-; SCALAR-NEXT:    shlq $55, %rdi
-; SCALAR-NEXT:    btq $55, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $56, %r8
-; SCALAR-NEXT:    btq $56, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r8
-; SCALAR-NEXT:    xorq %rdi, %r8
-; SCALAR-NEXT:    movq %rax, %rdi
-; SCALAR-NEXT:    shlq $57, %rdi
-; SCALAR-NEXT:    btq $57, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
+; SCALAR-NEXT:    andl $286331153, %edi # imm = 0x11111111
+; SCALAR-NEXT:    movq %rdi, %r11
+; SCALAR-NEXT:    imulq %r8, %r11
+; SCALAR-NEXT:    xorq %r9, %r11
+; SCALAR-NEXT:    movl %edx, %r10d
+; SCALAR-NEXT:    andl $-2004318072, %r10d # imm = 0x88888888
+; SCALAR-NEXT:    movl %ecx, %r9d
+; SCALAR-NEXT:    andl $1145324612, %r9d # imm = 0x44444444
+; SCALAR-NEXT:    movq %r9, %rbx
+; SCALAR-NEXT:    imulq %r10, %rbx
+; SCALAR-NEXT:    andl $1145324612, %edx # imm = 0x44444444
+; SCALAR-NEXT:    andl $-2004318072, %ecx # imm = 0x88888888
+; SCALAR-NEXT:    movq %rcx, %r14
+; SCALAR-NEXT:    imulq %rdx, %r14
+; SCALAR-NEXT:    xorq %rbx, %r14
+; SCALAR-NEXT:    xorq %r11, %r14
+; SCALAR-NEXT:    movabsq $2459565875921944576, %r11 # imm = 0x2222222200000000
+; SCALAR-NEXT:    andq %r14, %r11
+; SCALAR-NEXT:    movq %rsi, %rbx
+; SCALAR-NEXT:    imulq %r10, %rbx
+; SCALAR-NEXT:    movq %rdi, %r14
+; SCALAR-NEXT:    imulq %rax, %r14
+; SCALAR-NEXT:    xorq %rbx, %r14
+; SCALAR-NEXT:    movq %r9, %rbx
+; SCALAR-NEXT:    imulq %rdx, %rbx
+; SCALAR-NEXT:    movq %rcx, %r15
+; SCALAR-NEXT:    imulq %r8, %r15
+; SCALAR-NEXT:    xorq %rbx, %r15
+; SCALAR-NEXT:    xorq %r14, %r15
+; SCALAR-NEXT:    movabsq $1229782937960972288, %rbx # imm = 0x1111111100000000
+; SCALAR-NEXT:    andq %r15, %rbx
+; SCALAR-NEXT:    orq %r11, %rbx
+; SCALAR-NEXT:    movq %rsi, %r11
+; SCALAR-NEXT:    imulq %r8, %r11
+; SCALAR-NEXT:    movq %rdi, %r14
+; SCALAR-NEXT:    imulq %rdx, %r14
+; SCALAR-NEXT:    xorq %r11, %r14
+; SCALAR-NEXT:    movq %r9, %r11
+; SCALAR-NEXT:    imulq %rax, %r11
+; SCALAR-NEXT:    movq %rcx, %r15
+; SCALAR-NEXT:    imulq %r10, %r15
+; SCALAR-NEXT:    xorq %r11, %r15
+; SCALAR-NEXT:    xorq %r14, %r15
+; SCALAR-NEXT:    movabsq $4919131751843889152, %r11 # imm = 0x4444444400000000
+; SCALAR-NEXT:    andq %r15, %r11
+; SCALAR-NEXT:    orq %rbx, %r11
+; SCALAR-NEXT:    imulq %rdx, %rsi
+; SCALAR-NEXT:    imulq %r10, %rdi
 ; SCALAR-NEXT:    xorq %rsi, %rdi
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $58, %rsi
-; SCALAR-NEXT:    btq $58, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rsi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $59, %r8
-; SCALAR-NEXT:    btq $59, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r8
-; SCALAR-NEXT:    xorq %rsi, %r8
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $60, %rsi
-; SCALAR-NEXT:    btq $60, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rsi
-; SCALAR-NEXT:    xorq %r8, %rsi
-; SCALAR-NEXT:    movq %rax, %r8
-; SCALAR-NEXT:    shlq $61, %r8
-; SCALAR-NEXT:    btq $61, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %r8
-; SCALAR-NEXT:    xorq %rsi, %r8
-; SCALAR-NEXT:    movq %rax, %rsi
-; SCALAR-NEXT:    shlq $62, %rsi
-; SCALAR-NEXT:    btq $62, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rsi
-; SCALAR-NEXT:    xorq %r8, %rsi
-; SCALAR-NEXT:    shlq $63, %rax
-; SCALAR-NEXT:    btq $63, %rcx
-; SCALAR-NEXT:    cmovaeq %rdx, %rax
-; SCALAR-NEXT:    xorq %rsi, %rax
-; SCALAR-NEXT:    xorq %rdi, %rax
+; SCALAR-NEXT:    imulq %r8, %r9
+; SCALAR-NEXT:    imulq %rax, %rcx
+; SCALAR-NEXT:    xorq %r9, %rcx
+; SCALAR-NEXT:    xorq %rdi, %rcx
+; SCALAR-NEXT:    movabsq $614891466833002496, %rax # imm = 0x888888800000000
+; SCALAR-NEXT:    andq %rcx, %rax
+; SCALAR-NEXT:    orq %r11, %rax
 ; SCALAR-NEXT:    shrq $32, %rax
 ; SCALAR-NEXT:    # kill: def $eax killed $eax killed $rax
+; SCALAR-NEXT:    popq %rbx
+; SCALAR-NEXT:    popq %r14
+; SCALAR-NEXT:    popq %r15
 ; SCALAR-NEXT:    retq
 ;
 ; SSE-PCLMUL-LABEL: clmulh_i32:
@@ -2675,411 +1207,139 @@ define i32 @clmulh_i32(i32 %a, i32 %b) nounwind {
 define i64 @clmulh_i64(i64 %a, i64 %b) nounwind {
 ; SCALAR-LABEL: clmulh_i64:
 ; SCALAR:       # %bb.0:
+; SCALAR-NEXT:    pushq %rbp
+; SCALAR-NEXT:    pushq %r15
 ; SCALAR-NEXT:    pushq %r14
+; SCALAR-NEXT:    pushq %r13
+; SCALAR-NEXT:    pushq %r12
 ; SCALAR-NEXT:    pushq %rbx
-; SCALAR-NEXT:    bswapq %rdi
-; SCALAR-NEXT:    movq %rdi, %rax
+; SCALAR-NEXT:    bswapq %rsi
+; SCALAR-NEXT:    movq %rsi, %rax
 ; SCALAR-NEXT:    shrq $4, %rax
 ; SCALAR-NEXT:    movabsq $1085102592571150095, %rdx # imm = 0xF0F0F0F0F0F0F0F
 ; SCALAR-NEXT:    andq %rdx, %rax
-; SCALAR-NEXT:    andq %rdx, %rdi
-; SCALAR-NEXT:    shlq $4, %rdi
-; SCALAR-NEXT:    orq %rax, %rdi
-; SCALAR-NEXT:    movabsq $3689348814741910323, %rcx # imm = 0x3333333333333333
-; SCALAR-NEXT:    movq %rdi, %rax
-; SCALAR-NEXT:    andq %rcx, %rax
-; SCALAR-NEXT:    shrq $2, %rdi
-; SCALAR-NEXT:    andq %rcx, %rdi
-; SCALAR-NEXT:    leaq (%rdi,%rax,4), %rdi
-; SCALAR-NEXT:    movabsq $6148914691236517205, %rax # imm = 0x5555555555555555
-; SCALAR-NEXT:    movq %rdi, %r8
-; SCALAR-NEXT:    andq %rax, %r8
-; SCALAR-NEXT:    shrq %rdi
-; SCALAR-NEXT:    movq %rdi, %r9
-; SCALAR-NEXT:    andq %rax, %r9
-; SCALAR-NEXT:    leaq (%r9,%r8,2), %r8
-; SCALAR-NEXT:    leaq (%r8,%r8), %r9
-; SCALAR-NEXT:    bswapq %rsi
-; SCALAR-NEXT:    movq %rsi, %r10
-; SCALAR-NEXT:    shrq $4, %r10
-; SCALAR-NEXT:    andq %rdx, %r10
 ; SCALAR-NEXT:    andq %rdx, %rsi
 ; SCALAR-NEXT:    shlq $4, %rsi
-; SCALAR-NEXT:    orq %r10, %rsi
-; SCALAR-NEXT:    movq %rsi, %r10
-; SCALAR-NEXT:    andq %rcx, %r10
+; SCALAR-NEXT:    orq %rax, %rsi
+; SCALAR-NEXT:    movabsq $3689348814741910323, %rcx # imm = 0x3333333333333333
+; SCALAR-NEXT:    movq %rsi, %rax
+; SCALAR-NEXT:    andq %rcx, %rax
 ; SCALAR-NEXT:    shrq $2, %rsi
 ; SCALAR-NEXT:    andq %rcx, %rsi
-; SCALAR-NEXT:    leaq (%rsi,%r10,4), %rsi
-; SCALAR-NEXT:    movq %rsi, %r10
-; SCALAR-NEXT:    andq %rax, %r10
+; SCALAR-NEXT:    leaq (%rsi,%rax,4), %rsi
+; SCALAR-NEXT:    movabsq $6148914691236517205, %r12 # imm = 0x5555555555555555
+; SCALAR-NEXT:    movq %rsi, %r8
+; SCALAR-NEXT:    andq %r12, %r8
 ; SCALAR-NEXT:    shrq %rsi
-; SCALAR-NEXT:    andq %rax, %rsi
-; SCALAR-NEXT:    leaq (%rsi,%r10,2), %rsi
-; SCALAR-NEXT:    movl %esi, %r10d
-; SCALAR-NEXT:    andl $2, %r10d
-; SCALAR-NEXT:    cmovneq %r9, %r10
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $1, %r11d
-; SCALAR-NEXT:    cmovneq %r8, %r11
-; SCALAR-NEXT:    xorq %r10, %r11
-; SCALAR-NEXT:    leaq (,%r8,4), %r9
-; SCALAR-NEXT:    movl %esi, %r10d
-; SCALAR-NEXT:    andl $4, %r10d
-; SCALAR-NEXT:    cmovneq %r9, %r10
-; SCALAR-NEXT:    leaq (,%r8,8), %rbx
-; SCALAR-NEXT:    movl %esi, %r9d
-; SCALAR-NEXT:    andl $8, %r9d
-; SCALAR-NEXT:    cmovneq %rbx, %r9
-; SCALAR-NEXT:    xorq %r10, %r9
-; SCALAR-NEXT:    xorq %r11, %r9
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $4, %r10
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $16, %r11d
-; SCALAR-NEXT:    cmovneq %r10, %r11
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $5, %r10
-; SCALAR-NEXT:    movl %esi, %ebx
-; SCALAR-NEXT:    andl $32, %ebx
-; SCALAR-NEXT:    cmovneq %r10, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r11
-; SCALAR-NEXT:    shlq $6, %r11
-; SCALAR-NEXT:    movl %esi, %r10d
-; SCALAR-NEXT:    andl $64, %r10d
-; SCALAR-NEXT:    cmovneq %r11, %r10
-; SCALAR-NEXT:    xorq %rbx, %r10
-; SCALAR-NEXT:    xorq %r9, %r10
-; SCALAR-NEXT:    movq %r8, %r9
-; SCALAR-NEXT:    shlq $7, %r9
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $128, %r11d
-; SCALAR-NEXT:    cmovneq %r9, %r11
-; SCALAR-NEXT:    movq %r8, %r9
-; SCALAR-NEXT:    shlq $8, %r9
-; SCALAR-NEXT:    movl %esi, %ebx
-; SCALAR-NEXT:    andl $256, %ebx # imm = 0x100
-; SCALAR-NEXT:    cmovneq %r9, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r9
-; SCALAR-NEXT:    shlq $9, %r9
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $512, %r11d # imm = 0x200
-; SCALAR-NEXT:    cmovneq %r9, %r11
-; SCALAR-NEXT:    xorq %rbx, %r11
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $10, %rbx
-; SCALAR-NEXT:    movl %esi, %r9d
-; SCALAR-NEXT:    andl $1024, %r9d # imm = 0x400
-; SCALAR-NEXT:    cmovneq %rbx, %r9
-; SCALAR-NEXT:    xorq %r11, %r9
-; SCALAR-NEXT:    xorq %r10, %r9
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $11, %r10
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $2048, %r11d # imm = 0x800
-; SCALAR-NEXT:    cmovneq %r10, %r11
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $12, %r10
-; SCALAR-NEXT:    movl %esi, %ebx
-; SCALAR-NEXT:    andl $4096, %ebx # imm = 0x1000
-; SCALAR-NEXT:    cmovneq %r10, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $13, %r10
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $8192, %r11d # imm = 0x2000
-; SCALAR-NEXT:    cmovneq %r10, %r11
-; SCALAR-NEXT:    xorq %rbx, %r11
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $14, %r10
-; SCALAR-NEXT:    movl %esi, %ebx
-; SCALAR-NEXT:    andl $16384, %ebx # imm = 0x4000
-; SCALAR-NEXT:    cmovneq %r10, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r11
-; SCALAR-NEXT:    shlq $15, %r11
-; SCALAR-NEXT:    movl %esi, %r10d
-; SCALAR-NEXT:    andl $32768, %r10d # imm = 0x8000
-; SCALAR-NEXT:    cmovneq %r11, %r10
-; SCALAR-NEXT:    xorq %rbx, %r10
-; SCALAR-NEXT:    xorq %r9, %r10
-; SCALAR-NEXT:    movq %r8, %r9
-; SCALAR-NEXT:    shlq $16, %r9
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $65536, %r11d # imm = 0x10000
-; SCALAR-NEXT:    cmovneq %r9, %r11
-; SCALAR-NEXT:    movq %r8, %r9
-; SCALAR-NEXT:    shlq $17, %r9
-; SCALAR-NEXT:    movl %esi, %ebx
-; SCALAR-NEXT:    andl $131072, %ebx # imm = 0x20000
-; SCALAR-NEXT:    cmovneq %r9, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r9
-; SCALAR-NEXT:    shlq $18, %r9
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $262144, %r11d # imm = 0x40000
-; SCALAR-NEXT:    cmovneq %r9, %r11
-; SCALAR-NEXT:    xorq %rbx, %r11
-; SCALAR-NEXT:    movq %r8, %r9
-; SCALAR-NEXT:    shlq $19, %r9
-; SCALAR-NEXT:    movl %esi, %ebx
-; SCALAR-NEXT:    andl $524288, %ebx # imm = 0x80000
-; SCALAR-NEXT:    cmovneq %r9, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r9
-; SCALAR-NEXT:    shlq $20, %r9
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $1048576, %r11d # imm = 0x100000
-; SCALAR-NEXT:    cmovneq %r9, %r11
-; SCALAR-NEXT:    xorq %rbx, %r11
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $21, %rbx
-; SCALAR-NEXT:    movl %esi, %r9d
-; SCALAR-NEXT:    andl $2097152, %r9d # imm = 0x200000
-; SCALAR-NEXT:    cmovneq %rbx, %r9
-; SCALAR-NEXT:    xorq %r11, %r9
-; SCALAR-NEXT:    xorq %r10, %r9
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $22, %r10
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $4194304, %r11d # imm = 0x400000
-; SCALAR-NEXT:    cmovneq %r10, %r11
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $23, %r10
-; SCALAR-NEXT:    movl %esi, %ebx
-; SCALAR-NEXT:    andl $8388608, %ebx # imm = 0x800000
-; SCALAR-NEXT:    cmovneq %r10, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $24, %r10
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $16777216, %r11d # imm = 0x1000000
-; SCALAR-NEXT:    cmovneq %r10, %r11
-; SCALAR-NEXT:    xorq %rbx, %r11
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $25, %r10
-; SCALAR-NEXT:    movl %esi, %ebx
-; SCALAR-NEXT:    andl $33554432, %ebx # imm = 0x2000000
-; SCALAR-NEXT:    cmovneq %r10, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $26, %r10
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $67108864, %r11d # imm = 0x4000000
-; SCALAR-NEXT:    cmovneq %r10, %r11
-; SCALAR-NEXT:    xorq %rbx, %r11
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $27, %r10
-; SCALAR-NEXT:    movl %esi, %ebx
-; SCALAR-NEXT:    andl $134217728, %ebx # imm = 0x8000000
-; SCALAR-NEXT:    cmovneq %r10, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r11
-; SCALAR-NEXT:    shlq $28, %r11
-; SCALAR-NEXT:    movl %esi, %r10d
-; SCALAR-NEXT:    andl $268435456, %r10d # imm = 0x10000000
-; SCALAR-NEXT:    cmovneq %r11, %r10
-; SCALAR-NEXT:    xorq %rbx, %r10
-; SCALAR-NEXT:    xorq %r9, %r10
-; SCALAR-NEXT:    movq %r8, %r9
-; SCALAR-NEXT:    shlq $29, %r9
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $536870912, %r11d # imm = 0x20000000
-; SCALAR-NEXT:    cmovneq %r9, %r11
-; SCALAR-NEXT:    movq %r8, %r9
-; SCALAR-NEXT:    shlq $30, %r9
-; SCALAR-NEXT:    movl %esi, %ebx
-; SCALAR-NEXT:    andl $1073741824, %ebx # imm = 0x40000000
-; SCALAR-NEXT:    cmovneq %r9, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
+; SCALAR-NEXT:    andq %r12, %rsi
+; SCALAR-NEXT:    leaq (%rsi,%r8,2), %r8
+; SCALAR-NEXT:    movabsq $1229782938247303441, %rbx # imm = 0x1111111111111111
+; SCALAR-NEXT:    movq %r8, %rsi
+; SCALAR-NEXT:    andq %rbx, %rsi
+; SCALAR-NEXT:    bswapq %rdi
+; SCALAR-NEXT:    movq %rdi, %r9
+; SCALAR-NEXT:    shrq $4, %r9
+; SCALAR-NEXT:    andq %rdx, %r9
+; SCALAR-NEXT:    andq %rdx, %rdi
+; SCALAR-NEXT:    shlq $4, %rdi
+; SCALAR-NEXT:    orq %r9, %rdi
+; SCALAR-NEXT:    movq %rdi, %r9
+; SCALAR-NEXT:    andq %rcx, %r9
+; SCALAR-NEXT:    shrq $2, %rdi
+; SCALAR-NEXT:    andq %rcx, %rdi
+; SCALAR-NEXT:    leaq (%rdi,%r9,4), %rdi
+; SCALAR-NEXT:    movq %rdi, %r9
+; SCALAR-NEXT:    andq %r12, %r9
+; SCALAR-NEXT:    shrq %rdi
+; SCALAR-NEXT:    andq %r12, %rdi
+; SCALAR-NEXT:    leaq (%rdi,%r9,2), %rdi
+; SCALAR-NEXT:    movabsq $2459565876494606882, %rbp # imm = 0x2222222222222222
+; SCALAR-NEXT:    movq %rdi, %r14
+; SCALAR-NEXT:    andq %rbp, %r14
+; SCALAR-NEXT:    movq %r14, %r10
+; SCALAR-NEXT:    imulq %rsi, %r10
 ; SCALAR-NEXT:    movq %r8, %r9
-; SCALAR-NEXT:    shlq $31, %r9
-; SCALAR-NEXT:    movl %esi, %r11d
-; SCALAR-NEXT:    andl $-2147483648, %r11d # imm = 0x80000000
-; SCALAR-NEXT:    cmovneq %r9, %r11
-; SCALAR-NEXT:    xorq %rbx, %r11
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $32, %rbx
-; SCALAR-NEXT:    xorl %r9d, %r9d
-; SCALAR-NEXT:    btq $32, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r11
-; SCALAR-NEXT:    shlq $33, %r11
-; SCALAR-NEXT:    btq $33, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r11
-; SCALAR-NEXT:    xorq %rbx, %r11
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $34, %rbx
-; SCALAR-NEXT:    btq $34, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r14
-; SCALAR-NEXT:    shlq $35, %r14
-; SCALAR-NEXT:    btq $35, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r14
-; SCALAR-NEXT:    xorq %rbx, %r14
-; SCALAR-NEXT:    movq %r8, %r11
-; SCALAR-NEXT:    shlq $36, %r11
-; SCALAR-NEXT:    btq $36, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r11
-; SCALAR-NEXT:    xorq %r14, %r11
-; SCALAR-NEXT:    xorq %r10, %r11
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $37, %r10
-; SCALAR-NEXT:    btq $37, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r10
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $38, %rbx
-; SCALAR-NEXT:    btq $38, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rbx
-; SCALAR-NEXT:    xorq %r10, %rbx
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $39, %r10
-; SCALAR-NEXT:    btq $39, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r10
-; SCALAR-NEXT:    xorq %rbx, %r10
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $40, %rbx
-; SCALAR-NEXT:    btq $40, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rbx
-; SCALAR-NEXT:    xorq %r10, %rbx
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $41, %r10
-; SCALAR-NEXT:    btq $41, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r10
-; SCALAR-NEXT:    xorq %rbx, %r10
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $42, %rbx
-; SCALAR-NEXT:    btq $42, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rbx
-; SCALAR-NEXT:    xorq %r10, %rbx
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $43, %r10
-; SCALAR-NEXT:    btq $43, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r10
-; SCALAR-NEXT:    xorq %rbx, %r10
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $44, %rbx
-; SCALAR-NEXT:    btq $44, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rbx
-; SCALAR-NEXT:    xorq %r10, %rbx
-; SCALAR-NEXT:    movq %r8, %r14
-; SCALAR-NEXT:    shlq $45, %r14
-; SCALAR-NEXT:    btq $45, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r14
-; SCALAR-NEXT:    xorq %rbx, %r14
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $46, %r10
-; SCALAR-NEXT:    btq $46, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r10
-; SCALAR-NEXT:    xorq %r14, %r10
-; SCALAR-NEXT:    xorq %r11, %r10
-; SCALAR-NEXT:    movq %r8, %r11
-; SCALAR-NEXT:    shlq $47, %r11
-; SCALAR-NEXT:    btq $47, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r11
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $48, %rbx
-; SCALAR-NEXT:    btq $48, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r11
-; SCALAR-NEXT:    shlq $49, %r11
-; SCALAR-NEXT:    btq $49, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r11
-; SCALAR-NEXT:    xorq %rbx, %r11
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $50, %rbx
-; SCALAR-NEXT:    btq $50, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r11
-; SCALAR-NEXT:    shlq $51, %r11
-; SCALAR-NEXT:    btq $51, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r11
-; SCALAR-NEXT:    xorq %rbx, %r11
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $52, %rbx
-; SCALAR-NEXT:    btq $52, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r11
-; SCALAR-NEXT:    shlq $53, %r11
-; SCALAR-NEXT:    btq $53, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r11
-; SCALAR-NEXT:    xorq %rbx, %r11
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $54, %rbx
-; SCALAR-NEXT:    btq $54, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
-; SCALAR-NEXT:    movq %r8, %r11
-; SCALAR-NEXT:    shlq $55, %r11
-; SCALAR-NEXT:    btq $55, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r11
-; SCALAR-NEXT:    xorq %rbx, %r11
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $56, %rbx
-; SCALAR-NEXT:    btq $56, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rbx
-; SCALAR-NEXT:    xorq %r11, %rbx
+; SCALAR-NEXT:    andq %rbp, %r9
+; SCALAR-NEXT:    movq %rdi, %rdx
+; SCALAR-NEXT:    andq %rbx, %rdx
+; SCALAR-NEXT:    movq %rdx, %r13
+; SCALAR-NEXT:    imulq %r9, %r13
+; SCALAR-NEXT:    xorq %r10, %r13
+; SCALAR-NEXT:    movabsq $-8608480567731124088, %rax # imm = 0x8888888888888888
 ; SCALAR-NEXT:    movq %r8, %r11
-; SCALAR-NEXT:    shlq $57, %r11
-; SCALAR-NEXT:    btq $57, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r11
-; SCALAR-NEXT:    xorq %rbx, %r11
-; SCALAR-NEXT:    xorq %r10, %r11
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $58, %r10
-; SCALAR-NEXT:    btq $58, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r10
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $59, %rbx
-; SCALAR-NEXT:    btq $59, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rbx
-; SCALAR-NEXT:    xorq %r10, %rbx
-; SCALAR-NEXT:    movq %r8, %r10
-; SCALAR-NEXT:    shlq $60, %r10
-; SCALAR-NEXT:    btq $60, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r10
-; SCALAR-NEXT:    xorq %rbx, %r10
-; SCALAR-NEXT:    movq %r8, %rbx
-; SCALAR-NEXT:    shlq $61, %rbx
-; SCALAR-NEXT:    btq $61, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rbx
-; SCALAR-NEXT:    xorq %r10, %rbx
-; SCALAR-NEXT:    shlq $62, %r8
-; SCALAR-NEXT:    btq $62, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %r8
-; SCALAR-NEXT:    xorq %rbx, %r8
-; SCALAR-NEXT:    shlq $63, %rdi
-; SCALAR-NEXT:    btq $63, %rsi
-; SCALAR-NEXT:    cmovaeq %r9, %rdi
-; SCALAR-NEXT:    xorq %r8, %rdi
+; SCALAR-NEXT:    andq %rax, %r11
+; SCALAR-NEXT:    movabsq $4919131752989213764, %r10 # imm = 0x4444444444444444
+; SCALAR-NEXT:    movq %rdi, %rcx
+; SCALAR-NEXT:    andq %r10, %rcx
+; SCALAR-NEXT:    movq %rcx, %r15
+; SCALAR-NEXT:    imulq %r11, %r15
+; SCALAR-NEXT:    andq %r10, %r8
+; SCALAR-NEXT:    andq %rax, %rdi
+; SCALAR-NEXT:    movq %rdi, %rax
+; SCALAR-NEXT:    imulq %r8, %rax
+; SCALAR-NEXT:    xorq %r15, %rax
+; SCALAR-NEXT:    xorq %r13, %rax
+; SCALAR-NEXT:    andq %rbp, %rax
+; SCALAR-NEXT:    movq %r14, %r15
+; SCALAR-NEXT:    imulq %r11, %r15
+; SCALAR-NEXT:    movq %rdx, %r13
+; SCALAR-NEXT:    imulq %rsi, %r13
+; SCALAR-NEXT:    xorq %r15, %r13
+; SCALAR-NEXT:    movq %rcx, %r15
+; SCALAR-NEXT:    imulq %r8, %r15
+; SCALAR-NEXT:    movq %rdi, %rbp
+; SCALAR-NEXT:    imulq %r9, %rbp
+; SCALAR-NEXT:    xorq %r15, %rbp
+; SCALAR-NEXT:    xorq %r13, %rbp
+; SCALAR-NEXT:    andq %rbx, %rbp
+; SCALAR-NEXT:    orq %rax, %rbp
+; SCALAR-NEXT:    movq %rdi, %rbx
+; SCALAR-NEXT:    imulq %r11, %rbx
+; SCALAR-NEXT:    imulq %rdx, %r11
+; SCALAR-NEXT:    imulq %r8, %rdx
+; SCALAR-NEXT:    imulq %r14, %r8
+; SCALAR-NEXT:    imulq %r9, %r14
+; SCALAR-NEXT:    xorq %r14, %rdx
+; SCALAR-NEXT:    movq %rcx, %rax
+; SCALAR-NEXT:    imulq %rsi, %rax
+; SCALAR-NEXT:    xorq %rax, %rbx
+; SCALAR-NEXT:    xorq %rdx, %rbx
+; SCALAR-NEXT:    andq %r10, %rbx
+; SCALAR-NEXT:    orq %rbp, %rbx
+; SCALAR-NEXT:    xorq %r8, %r11
+; SCALAR-NEXT:    imulq %r9, %rcx
+; SCALAR-NEXT:    imulq %rsi, %rdi
+; SCALAR-NEXT:    xorq %rcx, %rdi
 ; SCALAR-NEXT:    xorq %r11, %rdi
+; SCALAR-NEXT:    movabsq $-8608480567731124088, %rax # imm = 0x8888888888888888
+; SCALAR-NEXT:    andq %rax, %rdi
+; SCALAR-NEXT:    orq %rbx, %rdi
 ; SCALAR-NEXT:    bswapq %rdi
-; SCALAR-NEXT:    movq %rdi, %rsi
-; SCALAR-NEXT:    shrq $4, %rsi
-; SCALAR-NEXT:    andq %rdx, %rsi
-; SCALAR-NEXT:    andq %rdx, %rdi
+; SCALAR-NEXT:    movq %rdi, %rax
+; SCALAR-NEXT:    shrq $4, %rax
+; SCALAR-NEXT:    movabsq $1085102592571150095, %rcx # imm = 0xF0F0F0F0F0F0F0F
+; SCALAR-NEXT:    andq %rcx, %rax
+; SCALAR-NEXT:    andq %rcx, %rdi
 ; SCALAR-NEXT:    shlq $4, %rdi
-; SCALAR-NEXT:    orq %rsi, %rdi
-; SCALAR-NEXT:    movq %rdi, %rdx
-; SCALAR-NEXT:    andq %rcx, %rdx
+; SCALAR-NEXT:    orq %rax, %rdi
+; SCALAR-NEXT:    movq %rdi, %rax
+; SCALAR-NEXT:    movabsq $3689348814741910323, %rcx # imm = 0x3333333333333333
+; SCALAR-NEXT:    andq %rcx, %rax
 ; SCALAR-NEXT:    shrq $2, %rdi
 ; SCALAR-NEXT:    andq %rcx, %rdi
-; SCALAR-NEXT:    leaq (%rdi,%rdx,4), %rcx
-; SCALAR-NEXT:    andq %rcx, %rax
-; SCALAR-NEXT:    shrq %rcx
-; SCALAR-NEXT:    movabsq $6148914691236517204, %rdx # imm = 0x5555555555555554
-; SCALAR-NEXT:    andq %rcx, %rdx
-; SCALAR-NEXT:    leaq (%rdx,%rax,2), %rax
+; SCALAR-NEXT:    leaq (%rdi,%rax,4), %rax
+; SCALAR-NEXT:    andq %rax, %r12
+; SCALAR-NEXT:    shrq %rax
+; SCALAR-NEXT:    movabsq $6148914691236517204, %rcx # imm = 0x5555555555555554
+; SCALAR-NEXT:    andq %rax, %rcx
+; SCALAR-NEXT:    leaq (%rcx,%r12,2), %rax
 ; SCALAR-NEXT:    shrq %rax
 ; SCALAR-NEXT:    popq %rbx
+; SCALAR-NEXT:    popq %r12
+; SCALAR-NEXT:    popq %r13
 ; SCALAR-NEXT:    popq %r14
+; SCALAR-NEXT:    popq %r15
+; SCALAR-NEXT:    popq %rbp
 ; SCALAR-NEXT:    retq
 ;
 ; SSE2-PCLMUL-LABEL: clmulh_i64:
@@ -3253,86 +1513,68 @@ define void @commutative_clmulh_i8(i8 %x, i8 %y, ptr %p0, ptr %p1) nounwind {
 ; SCALAR-NEXT:    pushq %r13
 ; SCALAR-NEXT:    pushq %r12
 ; SCALAR-NEXT:    pushq %rbx
-; SCALAR-NEXT:    movq %rcx, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
-; SCALAR-NEXT:    movzbl %sil, %r14d
-; SCALAR-NEXT:    movl %r14d, %ebp
-; SCALAR-NEXT:    shll $8, %ebp
-; SCALAR-NEXT:    movl %r14d, %ebx
-; SCALAR-NEXT:    shll $9, %ebx
-; SCALAR-NEXT:    movl %r14d, %r11d
-; SCALAR-NEXT:    shll $10, %r11d
-; SCALAR-NEXT:    movl %r14d, %eax
-; SCALAR-NEXT:    shll $11, %eax
-; SCALAR-NEXT:    movl %r14d, %r10d
-; SCALAR-NEXT:    shll $12, %r10d
-; SCALAR-NEXT:    movl %r14d, %ecx
-; SCALAR-NEXT:    shll $13, %ecx
+; SCALAR-NEXT:    movl %edi, %eax
+; SCALAR-NEXT:    andl $17, %eax
 ; SCALAR-NEXT:    movl %esi, %r8d
-; SCALAR-NEXT:    shll $14, %r8d
-; SCALAR-NEXT:    xorl %r15d, %r15d
-; SCALAR-NEXT:    testw %r15w, %r15w
-; SCALAR-NEXT:    cmovel %r15d, %r8d
-; SCALAR-NEXT:    cmovel %r15d, %ecx
-; SCALAR-NEXT:    cmovel %r15d, %r10d
-; SCALAR-NEXT:    cmovel %r15d, %eax
-; SCALAR-NEXT:    cmovel %r15d, %r11d
-; SCALAR-NEXT:    cmovel %r15d, %ebx
-; SCALAR-NEXT:    cmovel %r15d, %ebp
-; SCALAR-NEXT:    shll $15, %esi
-; SCALAR-NEXT:    testw %r15w, %r15w
-; SCALAR-NEXT:    cmovel %r15d, %esi
-; SCALAR-NEXT:    movl %edi, %r15d
-; SCALAR-NEXT:    andl $1, %r15d
-; SCALAR-NEXT:    cmovnel %r14d, %r15d
-; SCALAR-NEXT:    leal (%r14,%r14), %r12d
-; SCALAR-NEXT:    movl %edi, %r13d
-; SCALAR-NEXT:    andl $2, %r13d
-; SCALAR-NEXT:    cmovnel %r12d, %r13d
-; SCALAR-NEXT:    xorl %r15d, %r13d
-; SCALAR-NEXT:    leal (,%r14,4), %r15d
-; SCALAR-NEXT:    movl %edi, %r12d
-; SCALAR-NEXT:    andl $4, %r12d
-; SCALAR-NEXT:    cmovnel %r15d, %r12d
-; SCALAR-NEXT:    movl %edi, %r15d
-; SCALAR-NEXT:    andl $8, %r15d
-; SCALAR-NEXT:    leal (,%r14,8), %r9d
-; SCALAR-NEXT:    cmovnel %r9d, %r15d
-; SCALAR-NEXT:    xorl %r12d, %r15d
+; SCALAR-NEXT:    andl $34, %r8d
+; SCALAR-NEXT:    movl %r8d, %ebx
+; SCALAR-NEXT:    movl %r8d, %r10d
+; SCALAR-NEXT:    imull %eax, %r10d
+; SCALAR-NEXT:    movl %edi, %r9d
+; SCALAR-NEXT:    andl $34, %r9d
+; SCALAR-NEXT:    movl %esi, %r14d
+; SCALAR-NEXT:    andl $17, %r14d
+; SCALAR-NEXT:    movl %r14d, %ebp
+; SCALAR-NEXT:    movl %r14d, %r12d
+; SCALAR-NEXT:    imull %r9d, %r12d
+; SCALAR-NEXT:    xorl %r10d, %r12d
+; SCALAR-NEXT:    movl %edi, %r10d
+; SCALAR-NEXT:    andl $136, %r10d
+; SCALAR-NEXT:    movl %esi, %r11d
+; SCALAR-NEXT:    andl $68, %r11d
+; SCALAR-NEXT:    movl %r11d, %r13d
+; SCALAR-NEXT:    imull %r10d, %r13d
+; SCALAR-NEXT:    andl $68, %edi
+; SCALAR-NEXT:    andl $136, %esi
+; SCALAR-NEXT:    movl %esi, %r15d
+; SCALAR-NEXT:    imull %edi, %r15d
 ; SCALAR-NEXT:    xorl %r13d, %r15d
-; SCALAR-NEXT:    movl %r14d, %r9d
-; SCALAR-NEXT:    shll $4, %r9d
-; SCALAR-NEXT:    movl %edi, %r12d
-; SCALAR-NEXT:    andl $16, %r12d
-; SCALAR-NEXT:    cmovnel %r9d, %r12d
-; SCALAR-NEXT:    movl %r14d, %r9d
-; SCALAR-NEXT:    shll $5, %r9d
-; SCALAR-NEXT:    movl %edi, %r13d
-; SCALAR-NEXT:    andl $32, %r13d
-; SCALAR-NEXT:    cmovnel %r9d, %r13d
-; SCALAR-NEXT:    xorl %r12d, %r13d
-; SCALAR-NEXT:    movl %r14d, %r9d
-; SCALAR-NEXT:    shll $6, %r9d
-; SCALAR-NEXT:    movl %edi, %r12d
-; SCALAR-NEXT:    andl $64, %r12d
-; SCALAR-NEXT:    cmovnel %r9d, %r12d
-; SCALAR-NEXT:    xorl %r13d, %r12d
-; SCALAR-NEXT:    xorl %r15d, %r12d
-; SCALAR-NEXT:    shll $7, %r14d
-; SCALAR-NEXT:    andl $128, %edi
-; SCALAR-NEXT:    cmovnel %r14d, %edi
-; SCALAR-NEXT:    xorl %ebp, %edi
-; SCALAR-NEXT:    xorl %ebx, %edi
-; SCALAR-NEXT:    xorl %r11d, %edi
-; SCALAR-NEXT:    xorl %r12d, %edi
-; SCALAR-NEXT:    xorl %r10d, %eax
-; SCALAR-NEXT:    xorl %ecx, %eax
-; SCALAR-NEXT:    xorl %r8d, %eax
-; SCALAR-NEXT:    xorl %esi, %eax
-; SCALAR-NEXT:    xorl %edi, %eax
-; SCALAR-NEXT:    shrl $8, %eax
-; SCALAR-NEXT:    movb %al, (%rdx)
-; SCALAR-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %rcx # 8-byte Reload
-; SCALAR-NEXT:    movb %al, (%rcx)
+; SCALAR-NEXT:    imull %r10d, %ebx
+; SCALAR-NEXT:    movl %esi, %r13d
+; SCALAR-NEXT:    imull %r10d, %r13d
+; SCALAR-NEXT:    imull %r14d, %r10d
+; SCALAR-NEXT:    xorl %r12d, %r15d
+; SCALAR-NEXT:    movl %r11d, %r12d
+; SCALAR-NEXT:    imull %edi, %r12d
+; SCALAR-NEXT:    imull %edi, %r14d
+; SCALAR-NEXT:    imull %r8d, %edi
+; SCALAR-NEXT:    imull %eax, %ebp
+; SCALAR-NEXT:    xorl %ebx, %ebp
+; SCALAR-NEXT:    movl %esi, %ebx
+; SCALAR-NEXT:    imull %r9d, %ebx
+; SCALAR-NEXT:    xorl %r12d, %ebx
+; SCALAR-NEXT:    andl $8704, %r15d # imm = 0x2200
+; SCALAR-NEXT:    xorl %ebp, %ebx
+; SCALAR-NEXT:    andl $4352, %ebx # imm = 0x1100
+; SCALAR-NEXT:    orl %r15d, %ebx
+; SCALAR-NEXT:    imull %r9d, %r8d
+; SCALAR-NEXT:    xorl %r8d, %r14d
+; SCALAR-NEXT:    movl %r11d, %r8d
+; SCALAR-NEXT:    imull %eax, %r8d
+; SCALAR-NEXT:    xorl %r8d, %r13d
+; SCALAR-NEXT:    xorl %r14d, %r13d
+; SCALAR-NEXT:    andl $17408, %r13d # imm = 0x4400
+; SCALAR-NEXT:    orl %ebx, %r13d
+; SCALAR-NEXT:    xorl %edi, %r10d
+; SCALAR-NEXT:    imull %r9d, %r11d
+; SCALAR-NEXT:    imull %eax, %esi
+; SCALAR-NEXT:    xorl %r11d, %esi
+; SCALAR-NEXT:    xorl %r10d, %esi
+; SCALAR-NEXT:    andl $2048, %esi # imm = 0x800
+; SCALAR-NEXT:    orl %r13d, %esi
+; SCALAR-NEXT:    shrl $8, %esi
+; SCALAR-NEXT:    movb %sil, (%rdx)
+; SCALAR-NEXT:    movb %sil, (%rcx)
 ; SCALAR-NEXT:    popq %rbx
 ; SCALAR-NEXT:    popq %r12
 ; SCALAR-NEXT:    popq %r13
@@ -3388,80 +1630,68 @@ define void @commutative_clmulr_i8(i8 %x, i8 %y, ptr %p0, ptr %p1) nounwind {
 ; SCALAR-NEXT:    pushq %r13
 ; SCALAR-NEXT:    pushq %r12
 ; SCALAR-NEXT:    pushq %rbx
-; SCALAR-NEXT:    movzbl %sil, %r14d
-; SCALAR-NEXT:    movl %r14d, %ebx
-; SCALAR-NEXT:    shll $8, %ebx
-; SCALAR-NEXT:    movl %r14d, %r11d
-; SCALAR-NEXT:    shll $9, %r11d
-; SCALAR-NEXT:    movl %r14d, %r10d
-; SCALAR-NEXT:    shll $10, %r10d
-; SCALAR-NEXT:    movl %r14d, %eax
-; SCALAR-NEXT:    shll $11, %eax
-; SCALAR-NEXT:    movl %r14d, %r9d
-; SCALAR-NEXT:    shll $12, %r9d
+; SCALAR-NEXT:    movl %edi, %eax
+; SCALAR-NEXT:    andl $17, %eax
 ; SCALAR-NEXT:    movl %esi, %r8d
-; SCALAR-NEXT:    shll $13, %r8d
-; SCALAR-NEXT:    xorl %ebp, %ebp
-; SCALAR-NEXT:    testw %bp, %bp
-; SCALAR-NEXT:    cmovel %ebp, %r8d
-; SCALAR-NEXT:    cmovel %ebp, %r9d
-; SCALAR-NEXT:    cmovel %ebp, %eax
-; SCALAR-NEXT:    cmovel %ebp, %r10d
-; SCALAR-NEXT:    cmovel %ebp, %r11d
-; SCALAR-NEXT:    cmovel %ebp, %ebx
-; SCALAR-NEXT:    shll $14, %esi
-; SCALAR-NEXT:    testw %bp, %bp
-; SCALAR-NEXT:    cmovel %ebp, %esi
-; SCALAR-NEXT:    movl %edi, %ebp
-; SCALAR-NEXT:    andl $1, %ebp
-; SCALAR-NEXT:    cmovnel %r14d, %ebp
-; SCALAR-NEXT:    leal (%r14,%r14), %r15d
-; SCALAR-NEXT:    movl %edi, %r12d
-; SCALAR-NEXT:    andl $2, %r12d
-; SCALAR-NEXT:    cmovnel %r15d, %r12d
-; SCALAR-NEXT:    xorl %ebp, %r12d
-; SCALAR-NEXT:    leal (,%r14,4), %ebp
-; SCALAR-NEXT:    movl %edi, %r15d
-; SCALAR-NEXT:    andl $4, %r15d
-; SCALAR-NEXT:    cmovnel %ebp, %r15d
-; SCALAR-NEXT:    leal (,%r14,8), %r13d
-; SCALAR-NEXT:    movl %edi, %ebp
-; SCALAR-NEXT:    andl $8, %ebp
-; SCALAR-NEXT:    cmovnel %r13d, %ebp
-; SCALAR-NEXT:    xorl %r15d, %ebp
-; SCALAR-NEXT:    xorl %r12d, %ebp
-; SCALAR-NEXT:    movl %r14d, %r15d
-; SCALAR-NEXT:    shll $4, %r15d
-; SCALAR-NEXT:    movl %edi, %r12d
-; SCALAR-NEXT:    andl $16, %r12d
-; SCALAR-NEXT:    cmovnel %r15d, %r12d
-; SCALAR-NEXT:    movl %r14d, %r15d
-; SCALAR-NEXT:    shll $5, %r15d
-; SCALAR-NEXT:    movl %edi, %r13d
-; SCALAR-NEXT:    andl $32, %r13d
-; SCALAR-NEXT:    cmovnel %r15d, %r13d
-; SCALAR-NEXT:    xorl %r12d, %r13d
-; SCALAR-NEXT:    movl %r14d, %r15d
-; SCALAR-NEXT:    shll $6, %r15d
-; SCALAR-NEXT:    movl %edi, %r12d
-; SCALAR-NEXT:    andl $64, %r12d
-; SCALAR-NEXT:    cmovnel %r15d, %r12d
-; SCALAR-NEXT:    xorl %r13d, %r12d
-; SCALAR-NEXT:    xorl %ebp, %r12d
-; SCALAR-NEXT:    shll $7, %r14d
-; SCALAR-NEXT:    andl $128, %edi
-; SCALAR-NEXT:    cmovnel %r14d, %edi
-; SCALAR-NEXT:    xorl %ebx, %edi
-; SCALAR-NEXT:    xorl %r11d, %edi
-; SCALAR-NEXT:    xorl %r10d, %edi
-; SCALAR-NEXT:    xorl %r12d, %edi
-; SCALAR-NEXT:    xorl %r9d, %eax
-; SCALAR-NEXT:    xorl %r8d, %eax
-; SCALAR-NEXT:    xorl %esi, %eax
-; SCALAR-NEXT:    xorl %edi, %eax
-; SCALAR-NEXT:    shrl $7, %eax
-; SCALAR-NEXT:    movb %al, (%rdx)
-; SCALAR-NEXT:    movb %al, (%rcx)
+; SCALAR-NEXT:    andl $34, %r8d
+; SCALAR-NEXT:    movl %r8d, %ebx
+; SCALAR-NEXT:    movl %r8d, %r10d
+; SCALAR-NEXT:    imull %eax, %r10d
+; SCALAR-NEXT:    movl %edi, %r9d
+; SCALAR-NEXT:    andl $34, %r9d
+; SCALAR-NEXT:    movl %esi, %r14d
+; SCALAR-NEXT:    andl $17, %r14d
+; SCALAR-NEXT:    movl %r14d, %ebp
+; SCALAR-NEXT:    movl %r14d, %r12d
+; SCALAR-NEXT:    imull %r9d, %r12d
+; SCALAR-NEXT:    xorl %r10d, %r12d
+; SCALAR-NEXT:    movl %edi, %r10d
+; SCALAR-NEXT:    andl $136, %r10d
+; SCALAR-NEXT:    movl %esi, %r11d
+; SCALAR-NEXT:    andl $68, %r11d
+; SCALAR-NEXT:    movl %r11d, %r13d
+; SCALAR-NEXT:    imull %r10d, %r13d
+; SCALAR-NEXT:    andl $68, %edi
+; SCALAR-NEXT:    andl $136, %esi
+; SCALAR-NEXT:    movl %esi, %r15d
+; SCALAR-NEXT:    imull %edi, %r15d
+; SCALAR-NEXT:    xorl %r13d, %r15d
+; SCALAR-NEXT:    imull %r10d, %ebx
+; SCALAR-NEXT:    movl %esi, %r13d
+; SCALAR-NEXT:    imull %r10d, %r13d
+; SCALAR-NEXT:    imull %r14d, %r10d
+; SCALAR-NEXT:    xorl %r12d, %r15d
+; SCALAR-NEXT:    movl %r11d, %r12d
+; SCALAR-NEXT:    imull %edi, %r12d
+; SCALAR-NEXT:    imull %edi, %r14d
+; SCALAR-NEXT:    imull %r8d, %edi
+; SCALAR-NEXT:    imull %eax, %ebp
+; SCALAR-NEXT:    xorl %ebx, %ebp
+; SCALAR-NEXT:    movl %esi, %ebx
+; SCALAR-NEXT:    imull %r9d, %ebx
+; SCALAR-NEXT:    xorl %r12d, %ebx
+; SCALAR-NEXT:    andl $8704, %r15d # imm = 0x2200
+; SCALAR-NEXT:    xorl %ebp, %ebx
+; SCALAR-NEXT:    andl $4352, %ebx # imm = 0x1100
+; SCALAR-NEXT:    orl %r15d, %ebx
+; SCALAR-NEXT:    imull %r9d, %r8d
+; SCALAR-NEXT:    xorl %r8d, %r14d
+; SCALAR-NEXT:    movl %r11d, %r8d
+; SCALAR-NEXT:    imull %eax, %r8d
+; SCALAR-NEXT:    xorl %r8d, %r13d
+; SCALAR-NEXT:    xorl %r14d, %r13d
+; SCALAR-NEXT:    andl $17408, %r13d # imm = 0x4400
+; SCALAR-NEXT:    orl %ebx, %r13d
+; SCALAR-NEXT:    xorl %edi, %r10d
+; SCALAR-NEXT:    imull %r9d, %r11d
+; SCALAR-NEXT:    imull %eax, %esi
+; SCALAR-NEXT:    xorl %r11d, %esi
+; SCALAR-NEXT:    xorl %r10d, %esi
+; SCALAR-NEXT:    andl $2176, %esi # imm = 0x880
+; SCALAR-NEXT:    orl %r13d, %esi
+; SCALAR-NEXT:    shrl $7, %esi
+; SCALAR-NEXT:    movb %sil, (%rdx)
+; SCALAR-NEXT:    movb %sil, (%rcx)
 ; SCALAR-NEXT:    popq %rbx
 ; SCALAR-NEXT:    popq %r12
 ; SCALAR-NEXT:    popq %r13



More information about the llvm-commits mailing list