[X86] Scalar mask and scalar move optimizations

[tkrupa]

Summary

1. Introduction of mask scalar TableGen patterns.
2. Introduction of new scalar move TableGen patterns and refactoring of existing ones.
3. Folding of pattern created by introducing scalar masking in Clang header files.

[rja]

looks fine

[craig.topper]

lib/Target/X86/X86ISelLowering.cpp

32068 ↗

(On Diff #147298)

use cast not dyn_cast

[echristo]

Removing rja from the reviewers line here.

[craig.topper]

Is it possible to add tests for any patterns that are new? It's not good to add untested code.

[tkrupa]

Yes we could, but only for X86ISelLowering folding and FMA patterns - the other ones are multiclasses without any definitions, so they never alter anything at this point. They are intended to be a basis for several upcoming patches which will add defms for these patterns with the tests alongside them. I'm OOO tomorrow so I'll be able to add the tests for first two cases on Monday.

[tkrupa]

Added tests for ISelLowering combining and FMA patterns.

[RKSimon]

Add these tests with the current codegen so the patch shows the diff improvement?

[tkrupa]

In D47012#1106702, @RKSimon wrote:

Add these tests with the current codegen so the patch shows the diff improvement?

As in a separate patch?

[craig.topper]

Just commit the test files to trunk with the current codegen. No need for a phabricator review. Then rebase this patch to show the diff with the new codegen

[tkrupa]

Test diffs are now visible.
Note: there is no test for div in combine-select.ll - div pattern gets split into 3 basic block with a branch condition, so this optimization doesn't apply to it.

[craig.topper]

lib/Target/X86/X86ISelLowering.cpp

32216 ↗

(On Diff #147844)

I don't know if you can use ISD::SELECT here if you don't have the ANDing with 1. ISD::SELECT definition is that the condition value is either 0 or 1 regardless of how many bits it is. If you pass in the raw X you violate this rule.

[tkrupa]

lib/Target/X86/X86ISelLowering.cpp

32216 ↗

(On Diff #147844)

I think I can - X86TargetLowering::LowerSELECT takes care of that when operands are scalar and subtarget has AVX512 which is always the case here. I'm talking about this code:

// AVX512 fallback is to lower selects of scalar floats to masked moves.
if ((VT == MVT::f64 || VT == MVT::f32) && Subtarget.hasAVX512()) {
  SDValue Cmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v1i1, Cond);
  return DAG.getNode(X86ISD::SELECTS, DL, VT, Cmp, Op1, Op2);
}

[craig.topper]

It's not really a question of whether it will be lowered correctly. Target independent DAG combines call getBooleanContents and expect that the condition for a target independent ISD:;SELECT node obey that.

[craig.topper]

lib/Target/X86/X86ISelLowering.cpp

32208 ↗

(On Diff #147844)

This isScalarInteger check is unnecessary, ConstantSDNodes can only be scalar, and the condition for an ISD::SELECT can only be scalar which means the AND would have to be scalar.

32213 ↗

(On Diff #147844)

The tests all pass if I change this line to

SDValue Mask = AndNode;

We have a later combine that removes the mask once we've converted to v1i1.

[tkrupa]

Fixed fma mask3 pattern.

[craig.topper]

lib/Target/X86/X86ISelLowering.cpp

32207 ↗

(On Diff #148575)

You might still need a truncate on the AND if it isn't an i8

lib/Target/X86/X86InstrFMA.td

369 ↗

(On Diff #148575)

This should be [HasFMA, NoAVX512]. And the AVX512InstrInfo.td needs an equivalent set of patterns mapped to the unmasked EVEX instructions.

[craig.topper]

LGTM

llvm/trunk/lib/Target/X86/X86ISelLowering.cpp

// Check for x CC y ? y : x -- a min/max with reversed arms.
         break;
       }
     }
 
     if (Opcode)
       return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
   }
 
+  // Some mask scalar intrinsics rely on checking if only one bit is set
+  // and implement it in C code like this:
+  // A[0] = (U & 1) ? A[0] : W[0];
+  // This creates some redundant instructions that break pattern matching.
+  // fold (select (setcc (and (X, 1), 0, seteq), Y, Z)) -> select(and(X, 1),Z,Y)
+  if (Subtarget.hasAVX512() && N->getOpcode() == ISD::SELECT &&
+      Cond.getOpcode() == ISD::SETCC && (VT == MVT::f32 || VT == MVT::f64)) {
+    ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
+    SDValue AndNode = Cond.getOperand(0);
+    if (AndNode.getOpcode() == ISD::AND && CC == ISD::SETEQ &&
+        isNullConstant(Cond.getOperand(1)) &&
+        isa<ConstantSDNode>(AndNode.getOperand(1)) &&
+        cast<ConstantSDNode>(AndNode.getOperand(1))->getAPIntValue() == 1) {
+      // LHS and RHS swapped due to
+      // setcc outputting 1 when AND resulted in 0 and vice versa.
+      if (AndNode.getValueType() != MVT::i8)
+        AndNode = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, AndNode);
+      return DAG.getNode(ISD::SELECT, DL, VT, AndNode, RHS, LHS);
+    }
+  }
+
   // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
   // lowering on KNL. In this case we convert it to
   // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
   // The same situation all vectors of i8 and i16 without BWI.
   // Make sure we extend these even before type legalization gets a chance to
   // split wide vectors.
   // Since SKX these selects have a proper lowering.
   if (Subtarget.hasAVX512() && !Subtarget.hasBWI() && CondVT.isVector() &&

llvm/trunk/lib/Target/X86/X86InstrAVX512.td

defm VFMADD  : avx512_fma3s<0xA9, 0xB9, 0x99, "vfmadd", X86Fmadd, X86Fmadds1,
                             X86FmaddRnds1, X86Fmadds3, X86FmaddRnds3>;
 defm VFMSUB  : avx512_fma3s<0xAB, 0xBB, 0x9B, "vfmsub", X86Fmsub, X86Fmsubs1,
                             X86FmsubRnds1, X86Fmsubs3, X86FmsubRnds3>;
 defm VFNMADD : avx512_fma3s<0xAD, 0xBD, 0x9D, "vfnmadd", X86Fnmadd, X86Fnmadds1,
                             X86FnmaddRnds1, X86Fnmadds3, X86FnmaddRnds3>;
 defm VFNMSUB : avx512_fma3s<0xAF, 0xBF, 0x9F, "vfnmsub", X86Fnmsub, X86Fnmsubs1,
                             X86FnmsubRnds1, X86Fnmsubs3, X86FnmsubRnds3>;
 
+multiclass avx512_scalar_fma_patterns<SDNode Op, string Prefix, string Suffix, SDNode Move,
+                                      ValueType VT, ValueType EltVT, PatLeaf ZeroFP> {
+  let Predicates = [HasFMA, HasAVX512] in {
+    def : Pat<(VT (Move (VT VR128:$src2), (VT (scalar_to_vector
+                (Op (EltVT (extractelt (VT VR128:$src1), (iPTR 0))),
+                    (EltVT (extractelt (VT VR128:$src2), (iPTR 0))),
+                    (EltVT (extractelt (VT VR128:$src3), (iPTR 0)))))))),
+              (!cast<I>(Prefix#"213"#Suffix#"Zr_Int")
+               VR128:$src2, VR128:$src1, VR128:$src3)>;
+
+    def : Pat<(VT (Move (VT VR128:$src2), (VT (scalar_to_vector
+               (X86selects VK1WM:$mask,
+                (Op (EltVT (extractelt (VT VR128:$src1), (iPTR 0))),
+                    (EltVT (extractelt (VT VR128:$src2), (iPTR 0))),
+                    (EltVT (extractelt (VT VR128:$src3), (iPTR 0)))),
+                (EltVT (extractelt (VT VR128:$src2), (iPTR 0)))))))),
+              (!cast<I>(Prefix#"213"#Suffix#"Zr_Intk")
+               VR128:$src2, VK1WM:$mask, VR128:$src1, VR128:$src3)>;
+
+    def : Pat<(VT (Move (VT VR128:$src3), (VT (scalar_to_vector
+               (X86selects VK1WM:$mask,
+                (Op (EltVT (extractelt (VT VR128:$src1), (iPTR 0))),
+                    (EltVT (extractelt (VT VR128:$src2), (iPTR 0))),
+                    (EltVT (extractelt (VT VR128:$src3), (iPTR 0)))),
+                (EltVT (extractelt (VT VR128:$src3), (iPTR 0)))))))),
+              (!cast<I>(Prefix#"231"#Suffix#"Zr_Intk")
+               VR128:$src3, VK1WM:$mask, VR128:$src2, VR128:$src1)>;
+
+    def : Pat<(VT (Move (VT VR128:$src2), (VT (scalar_to_vector
+               (X86selects VK1WM:$mask,
+                (Op (EltVT (extractelt (VT VR128:$src1), (iPTR 0))),
+                    (EltVT (extractelt (VT VR128:$src2), (iPTR 0))),
+                    (EltVT (extractelt (VT VR128:$src3), (iPTR 0)))),
+                (EltVT ZeroFP)))))),
+              (!cast<I>(Prefix#"213"#Suffix#"Zr_Intkz")
+               VR128:$src2, VK1WM:$mask, VR128:$src1, VR128:$src3)>;
+  }
+}
+
+defm : avx512_scalar_fma_patterns<X86Fmadd, "VFMADD", "SS", X86Movss,
+                                  v4f32, f32, fp32imm0>;
+defm : avx512_scalar_fma_patterns<X86Fmsub, "VFMSUB", "SS", X86Movss,
+                                  v4f32, f32, fp32imm0>;
+defm : avx512_scalar_fma_patterns<X86Fnmadd, "VFNMADD", "SS", X86Movss,
+                                  v4f32, f32, fp32imm0>;
+defm : avx512_scalar_fma_patterns<X86Fnmsub, "VFNMSUB", "SS", X86Movss,
+                                  v4f32, f32, fp32imm0>;
+
+defm : avx512_scalar_fma_patterns<X86Fmadd, "VFMADD", "SD", X86Movsd,
+                                  v2f64, f64, fp64imm0>;
+defm : avx512_scalar_fma_patterns<X86Fmsub, "VFMSUB", "SD", X86Movsd,
+                                  v2f64, f64, fp64imm0>;
+defm : avx512_scalar_fma_patterns<X86Fnmadd, "VFNMADD", "SD", X86Movsd,
+                                  v2f64, f64, fp64imm0>;
+defm : avx512_scalar_fma_patterns<X86Fnmsub, "VFNMSUB", "SD", X86Movsd,
+                                  v2f64, f64, fp64imm0>;
+
 //===----------------------------------------------------------------------===//
 // AVX-512  Packed Multiply of Unsigned 52-bit Integers and Add the Low 52-bit IFMA
 //===----------------------------------------------------------------------===//
 let Constraints = "$src1 = $dst" in {
 multiclass avx512_pmadd52_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
                              X86FoldableSchedWrite sched, X86VectorVTInfo _> {
   // NOTE: The SDNode have the multiply operands first with the add last.
   // This enables commuted load patterns to be autogenerated by tablegen.
defm VRNDSCALESS : avx512_rndscale_scalar<0x0A, "vrndscaless",
                                           AVX512AIi8Base, EVEX_4V,
                                           EVEX_CD8<32, CD8VT1>;
 
 defm VRNDSCALESD : avx512_rndscale_scalar<0x0B, "vrndscalesd",
                                           SchedWriteFRnd.Scl, f64x_info>,
                                           VEX_W, AVX512AIi8Base, EVEX_4V,
                                           EVEX_CD8<64, CD8VT1>;
 
+multiclass avx512_masked_scalar<SDNode OpNode, string OpcPrefix, SDNode Move,
+                                dag Mask, X86VectorVTInfo _, PatLeaf ZeroFP,
+                                dag OutMask, Predicate BasePredicate> {
+  let Predicates = [BasePredicate] in {
+    def : Pat<(Move _.VT:$src1, (scalar_to_vector (X86selects Mask,
+               (OpNode (extractelt _.VT:$src2, (iPTR 0))),
+               (extractelt _.VT:$dst, (iPTR 0))))),
+              (!cast<Instruction>("V"#OpcPrefix#r_Intk)
+               _.VT:$dst, OutMask, _.VT:$src2, _.VT:$src1)>;
+
+    def : Pat<(Move _.VT:$src1, (scalar_to_vector (X86selects Mask,
+               (OpNode (extractelt _.VT:$src2, (iPTR 0))),
+               ZeroFP))),
+              (!cast<Instruction>("V"#OpcPrefix#r_Intkz)
+               OutMask, _.VT:$src2, _.VT:$src1)>;
+  }
+}
+
+multiclass avx512_masked_scalar_imm<SDNode OpNode, string OpcPrefix, SDNode Move,
+                                    dag Mask, X86VectorVTInfo _, PatLeaf ZeroFP,
+                                    bits<8> ImmV, dag OutMask,
+                                    Predicate BasePredicate> {
+  let Predicates = [BasePredicate] in {
+    def : Pat<(Move _.VT:$src1, (scalar_to_vector (X86selects Mask,
+               (OpNode (extractelt _.VT:$src2, (iPTR 0))),
+               (extractelt _.VT:$dst, (iPTR 0))))),
+              (!cast<Instruction>("V"#OpcPrefix#r_Intk)
+               _.VT:$dst, OutMask, _.VT:$src1, _.VT:$src2, (i32 ImmV))>;
+
+    def : Pat<(Move _.VT:$src1, (scalar_to_vector (X86selects Mask,
+               (OpNode (extractelt _.VT:$src2, (iPTR 0))), ZeroFP))),
+              (!cast<Instruction>("V"#OpcPrefix#r_Intkz)
+               OutMask, _.VT:$src1, _.VT:$src2, (i32 ImmV))>;
+  }
+}
+
 //-------------------------------------------------
 // Integer truncate and extend operations
 //-------------------------------------------------
 
 multiclass avx512_trunc_common<bits<8> opc, string OpcodeStr, SDNode OpNode,
                               X86FoldableSchedWrite sched, X86VectorVTInfo SrcInfo,
                               X86VectorVTInfo DestInfo, X86MemOperand x86memop> {
   let ExeDomain = DestInfo.ExeDomain in

 //   addps %xmm0, %xmm1
 //   movss %xmm1, %xmm0
 //
 // We now generate:
 //   addss %xmm1, %xmm0
 
 // TODO: Some canonicalization in lowering would simplify the number of
 // patterns we have to try to match.
-multiclass AVX512_scalar_math_f32_patterns<SDNode Op, string OpcPrefix> {
+multiclass AVX512_scalar_math_fp_patterns<SDNode Op, string OpcPrefix, SDNode MoveNode,
+                                           X86VectorVTInfo _, PatLeaf ZeroFP> {
   let Predicates = [HasAVX512] in {
     // extracted scalar math op with insert via movss
-    def : Pat<(v4f32 (X86Movss (v4f32 VR128X:$dst), (v4f32 (scalar_to_vector
-          (Op (f32 (extractelt (v4f32 VR128X:$dst), (iPTR 0))),
-          FR32X:$src))))),
-      (!cast<I>("V"#OpcPrefix#SSZrr_Int) v4f32:$dst,
-          (COPY_TO_REGCLASS FR32X:$src, VR128X))>;
+    def : Pat<(_.VT (MoveNode (_.VT VR128X:$dst), (_.VT (scalar_to_vector
+          (Op (_.EltVT (extractelt (_.VT VR128X:$dst), (iPTR 0))),
+          _.FRC:$src))))),
+      (!cast<I>("V"#OpcPrefix#Zrr_Int) _.VT:$dst,
+          (COPY_TO_REGCLASS _.FRC:$src, VR128X))>;
 
     // vector math op with insert via movss
-    def : Pat<(v4f32 (X86Movss (v4f32 VR128X:$dst),
-          (Op (v4f32 VR128X:$dst), (v4f32 VR128X:$src)))),
-      (!cast<I>("V"#OpcPrefix#SSZrr_Int) v4f32:$dst, v4f32:$src)>;
+    def : Pat<(_.VT (MoveNode (_.VT VR128X:$dst),
+          (Op (_.VT VR128X:$dst), (_.VT VR128X:$src)))),
+      (!cast<I>("V"#OpcPrefix#Zrr_Int) _.VT:$dst, _.VT:$src)>;
 
     // extracted masked scalar math op with insert via movss
-    def : Pat<(X86Movss (v4f32 VR128X:$src1),
+    def : Pat<(MoveNode (_.VT VR128X:$src1),
                (scalar_to_vector
                 (X86selects VK1WM:$mask,
-                            (Op (f32 (extractelt (v4f32 VR128X:$src1), (iPTR 0))),
-                                FR32X:$src2),
-                            FR32X:$src0))),
-      (!cast<I>("V"#OpcPrefix#SSZrr_Intk) (COPY_TO_REGCLASS FR32X:$src0, VR128X),
-          VK1WM:$mask, v4f32:$src1,
-          (COPY_TO_REGCLASS FR32X:$src2, VR128X))>;
-  }
-}
-
-defm : AVX512_scalar_math_f32_patterns<fadd, "ADD">;
-defm : AVX512_scalar_math_f32_patterns<fsub, "SUB">;
-defm : AVX512_scalar_math_f32_patterns<fmul, "MUL">;
-defm : AVX512_scalar_math_f32_patterns<fdiv, "DIV">;
-
-multiclass AVX512_scalar_math_f64_patterns<SDNode Op, string OpcPrefix> {
-  let Predicates = [HasAVX512] in {
-    // extracted scalar math op with insert via movsd
-    def : Pat<(v2f64 (X86Movsd (v2f64 VR128X:$dst), (v2f64 (scalar_to_vector
-          (Op (f64 (extractelt (v2f64 VR128X:$dst), (iPTR 0))),
-          FR64X:$src))))),
-      (!cast<I>("V"#OpcPrefix#SDZrr_Int) v2f64:$dst,
-          (COPY_TO_REGCLASS FR64X:$src, VR128X))>;
-
-    // vector math op with insert via movsd
-    def : Pat<(v2f64 (X86Movsd (v2f64 VR128X:$dst),
-          (Op (v2f64 VR128X:$dst), (v2f64 VR128X:$src)))),
-      (!cast<I>("V"#OpcPrefix#SDZrr_Int) v2f64:$dst, v2f64:$src)>;
+                            (Op (_.EltVT (extractelt (_.VT VR128X:$src1), (iPTR 0))),
+                                _.FRC:$src2),
+                            _.FRC:$src0))),
+      (!cast<I>("V"#OpcPrefix#Zrr_Intk) (COPY_TO_REGCLASS _.FRC:$src0, VR128X),
+          VK1WM:$mask, _.VT:$src1,
+          (COPY_TO_REGCLASS _.FRC:$src2, VR128X))>;
     
     // extracted masked scalar math op with insert via movss
-    def : Pat<(X86Movsd (v2f64 VR128X:$src1),
+    def : Pat<(MoveNode (_.VT VR128X:$src1),
                (scalar_to_vector
                 (X86selects VK1WM:$mask,
-                            (Op (f64 (extractelt (v2f64 VR128X:$src1), (iPTR 0))),
-                                FR64X:$src2),
-                            FR64X:$src0))),
-      (!cast<I>("V"#OpcPrefix#SDZrr_Intk) (COPY_TO_REGCLASS FR64X:$src0, VR128X),
-          VK1WM:$mask, v2f64:$src1,
-          (COPY_TO_REGCLASS FR64X:$src2, VR128X))>;
+                            (Op (_.EltVT (extractelt (_.VT VR128X:$src1), (iPTR 0))),
+                                _.FRC:$src2), (_.EltVT ZeroFP)))),
+      (!cast<I>("V"#OpcPrefix#Zrr_Intkz) 
+          VK1WM:$mask, _.VT:$src1,
+          (COPY_TO_REGCLASS _.FRC:$src2, VR128X))>;
   }
 }
 
-defm : AVX512_scalar_math_f64_patterns<fadd, "ADD">;
-defm : AVX512_scalar_math_f64_patterns<fsub, "SUB">;
-defm : AVX512_scalar_math_f64_patterns<fmul, "MUL">;
-defm : AVX512_scalar_math_f64_patterns<fdiv, "DIV">;
+defm : AVX512_scalar_math_fp_patterns<fadd, "ADDSS", X86Movss, v4f32x_info, fp32imm0>;
+defm : AVX512_scalar_math_fp_patterns<fsub, "SUBSS", X86Movss, v4f32x_info, fp32imm0>;
+defm : AVX512_scalar_math_fp_patterns<fmul, "MULSS", X86Movss, v4f32x_info, fp32imm0>;
+defm : AVX512_scalar_math_fp_patterns<fdiv, "DIVSS", X86Movss, v4f32x_info, fp32imm0>;
+
+defm : AVX512_scalar_math_fp_patterns<fadd, "ADDSD", X86Movsd, v2f64x_info, fp64imm0>;
+defm : AVX512_scalar_math_fp_patterns<fsub, "SUBSD", X86Movsd, v2f64x_info, fp64imm0>;
+defm : AVX512_scalar_math_fp_patterns<fmul, "MULSD", X86Movsd, v2f64x_info, fp64imm0>;
+defm : AVX512_scalar_math_fp_patterns<fdiv, "DIVSD", X86Movsd, v2f64x_info, fp64imm0>;
+
 
 //===----------------------------------------------------------------------===//
 // AES instructions
 //===----------------------------------------------------------------------===//
 
 multiclass avx512_vaes<bits<8> Op, string OpStr, string IntPrefix> {
   let Predicates = [HasVLX, HasVAES] in {
     defm Z128 : AESI_binop_rm_int<Op, OpStr,

llvm/trunk/lib/Target/X86/X86InstrFMA.td

 defm VFMSUB : fma3s<0x9B, 0xAB, 0xBB, "vfmsub", X86Fmsubs1, X86Fmsub,
                     SchedWriteFMA.Scl>, VEX_LIG;
 
 defm VFNMADD : fma3s<0x9D, 0xAD, 0xBD, "vfnmadd", X86Fnmadds1, X86Fnmadd,
                      SchedWriteFMA.Scl>, VEX_LIG;
 defm VFNMSUB : fma3s<0x9F, 0xAF, 0xBF, "vfnmsub", X86Fnmsubs1, X86Fnmsub,
                      SchedWriteFMA.Scl>, VEX_LIG;
 
+multiclass scalar_fma_patterns<SDNode Op, string Prefix, string Suffix, SDNode Move,
+                               ValueType VT, ValueType EltVT> {
+  let Predicates = [HasFMA, NoAVX512] in {
+    def : Pat<(VT (Move (VT VR128:$src2), (VT (scalar_to_vector
+                (Op (EltVT (extractelt (VT VR128:$src1), (iPTR 0))),
+                    (EltVT (extractelt (VT VR128:$src2), (iPTR 0))),
+                    (EltVT (extractelt (VT VR128:$src3), (iPTR 0)))))))),
+              (!cast<I>(Prefix#"213"#Suffix#"r_Int")
+               VR128:$src2, VR128:$src1, VR128:$src3)>;
+  }
+}
+
+defm : scalar_fma_patterns<X86Fmadd, "VFMADD", "SS", X86Movss, v4f32, f32>;
+defm : scalar_fma_patterns<X86Fmsub, "VFMSUB", "SS", X86Movss, v4f32, f32>;
+defm : scalar_fma_patterns<X86Fnmadd, "VFNMADD", "SS", X86Movss, v4f32, f32>;
+defm : scalar_fma_patterns<X86Fnmsub, "VFNMSUB", "SS", X86Movss, v4f32, f32>;
+
+defm : scalar_fma_patterns<X86Fmadd, "VFMADD", "SD", X86Movsd, v2f64, f64>;
+defm : scalar_fma_patterns<X86Fmsub, "VFMSUB", "SD", X86Movsd, v2f64, f64>;
+defm : scalar_fma_patterns<X86Fnmadd, "VFNMADD", "SD", X86Movsd, v2f64, f64>;
+defm : scalar_fma_patterns<X86Fnmsub, "VFNMSUB", "SD", X86Movsd, v2f64, f64>;
+
 //===----------------------------------------------------------------------===//
 // FMA4 - AMD 4 operand Fused Multiply-Add instructions
 //===----------------------------------------------------------------------===//
 
 multiclass fma4s<bits<8> opc, string OpcodeStr, RegisterClass RC,
                  X86MemOperand x86memop, ValueType OpVT, SDNode OpNode,
                  PatFrag mem_frag, X86FoldableSchedWrite sched> {
   let isCommutable = 1 in

llvm/trunk/lib/Target/X86/X86InstrSSE.td

 //   addps %xmm0, %xmm1
 //   movss %xmm1, %xmm0
 //
 // We now generate:
 //   addss %xmm1, %xmm0
 
 // TODO: Some canonicalization in lowering would simplify the number of
 // patterns we have to try to match.
-multiclass scalar_math_f32_patterns<SDNode Op, string OpcPrefix> {
-  let Predicates = [UseSSE1] in {
-    // extracted scalar math op with insert via movss
-    def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
-          (Op (f32 (extractelt (v4f32 VR128:$dst), (iPTR 0))),
-          FR32:$src))))),
-      (!cast<I>(OpcPrefix#SSrr_Int) v4f32:$dst,
-          (COPY_TO_REGCLASS FR32:$src, VR128))>;
-
-    // vector math op with insert via movss
-    def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
-          (Op (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
-      (!cast<I>(OpcPrefix#SSrr_Int) v4f32:$dst, v4f32:$src)>;
-  }
-
-  // Repeat everything for AVX.
-  let Predicates = [UseAVX] in {
-    // extracted scalar math op with insert via movss
-    def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst), (v4f32 (scalar_to_vector
-          (Op (f32 (extractelt (v4f32 VR128:$dst), (iPTR 0))),
-          FR32:$src))))),
-      (!cast<I>("V"#OpcPrefix#SSrr_Int) v4f32:$dst,
-          (COPY_TO_REGCLASS FR32:$src, VR128))>;
-
-    // vector math op with insert via movss
-    def : Pat<(v4f32 (X86Movss (v4f32 VR128:$dst),
-          (Op (v4f32 VR128:$dst), (v4f32 VR128:$src)))),
-      (!cast<I>("V"#OpcPrefix#SSrr_Int) v4f32:$dst, v4f32:$src)>;
-  }
-}
-
-defm : scalar_math_f32_patterns<fadd, "ADD">;
-defm : scalar_math_f32_patterns<fsub, "SUB">;
-defm : scalar_math_f32_patterns<fmul, "MUL">;
-defm : scalar_math_f32_patterns<fdiv, "DIV">;
-
-multiclass scalar_math_f64_patterns<SDNode Op, string OpcPrefix> {
-  let Predicates = [UseSSE2] in {
-    // extracted scalar math op with insert via movsd
-    def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst), (v2f64 (scalar_to_vector
-          (Op (f64 (extractelt (v2f64 VR128:$dst), (iPTR 0))),
-          FR64:$src))))),
-      (!cast<I>(OpcPrefix#SDrr_Int) v2f64:$dst,
-          (COPY_TO_REGCLASS FR64:$src, VR128))>;
-
-    // vector math op with insert via movsd
-    def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
-          (Op (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
-      (!cast<I>(OpcPrefix#SDrr_Int) v2f64:$dst, v2f64:$src)>;
+multiclass scalar_math_patterns<SDNode Op, string OpcPrefix, SDNode Move,
+                                    ValueType VT, ValueType EltTy,
+                                    RegisterClass RC, Predicate BasePredicate> {
+   let Predicates = [BasePredicate] in {
+    // extracted scalar math op with insert via movss/movsd
+    def : Pat<(VT (Move (VT VR128:$dst), (VT (scalar_to_vector
+          (Op (EltTy (extractelt (VT VR128:$dst), (iPTR 0))),
+          RC:$src))))),
+      (!cast<I>(OpcPrefix#rr_Int) VT:$dst,
+          (COPY_TO_REGCLASS RC:$src, VR128))>;
+
+    // vector math op with insert via movss/movsd
+    def : Pat<(VT (Move (VT VR128:$dst),
+          (Op (VT VR128:$dst), (VT VR128:$src)))),
+      (!cast<I>(OpcPrefix#rr_Int) VT:$dst, VT:$src)>;
    }
 
-  // Repeat everything for AVX.
+   // Repeat for AVX versions of the instructions.
    let Predicates = [UseAVX] in {
-    // extracted scalar math op with insert via movsd
-    def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst), (v2f64 (scalar_to_vector
-          (Op (f64 (extractelt (v2f64 VR128:$dst), (iPTR 0))),
-          FR64:$src))))),
-      (!cast<I>("V"#OpcPrefix#SDrr_Int) v2f64:$dst,
-          (COPY_TO_REGCLASS FR64:$src, VR128))>;
-
-    // vector math op with insert via movsd
-    def : Pat<(v2f64 (X86Movsd (v2f64 VR128:$dst),
-          (Op (v2f64 VR128:$dst), (v2f64 VR128:$src)))),
-      (!cast<I>("V"#OpcPrefix#SDrr_Int) v2f64:$dst, v2f64:$src)>;
+    // extracted scalar math op with insert via movss/movsd
+    def : Pat<(VT (Move (VT VR128:$dst), (VT (scalar_to_vector
+          (Op (EltTy (extractelt (VT VR128:$dst), (iPTR 0))),
+          RC:$src))))),
+      (!cast<I>("V"#OpcPrefix#rr_Int) VT:$dst,
+          (COPY_TO_REGCLASS RC:$src, VR128))>;
+
+    // vector math op with insert via movss/movsd
+    def : Pat<(VT (Move (VT VR128:$dst),
+          (Op (VT VR128:$dst), (VT VR128:$src)))),
+      (!cast<I>("V"#OpcPrefix#rr_Int) VT:$dst, VT:$src)>;
    }
 }
 
-defm : scalar_math_f64_patterns<fadd, "ADD">;
-defm : scalar_math_f64_patterns<fsub, "SUB">;
-defm : scalar_math_f64_patterns<fmul, "MUL">;
-defm : scalar_math_f64_patterns<fdiv, "DIV">;
+defm : scalar_math_patterns<fadd, "ADDSS", X86Movss, v4f32, f32, FR32, UseSSE1>;
+defm : scalar_math_patterns<fsub, "SUBSS", X86Movss, v4f32, f32, FR32, UseSSE1>;
+defm : scalar_math_patterns<fmul, "MULSS", X86Movss, v4f32, f32, FR32, UseSSE1>;
+defm : scalar_math_patterns<fdiv, "DIVSS", X86Movss, v4f32, f32, FR32, UseSSE1>;
+
+defm : scalar_math_patterns<fadd, "ADDSD", X86Movsd, v2f64, f64, FR64, UseSSE2>;
+defm : scalar_math_patterns<fsub, "SUBSD", X86Movsd, v2f64, f64, FR64, UseSSE2>;
+defm : scalar_math_patterns<fmul, "MULSD", X86Movsd, v2f64, f64, FR64, UseSSE2>;
+defm : scalar_math_patterns<fdiv, "DIVSD", X86Movsd, v2f64, f64, FR64, UseSSE2>;
  
 /// Unop Arithmetic
 /// In addition, we also have a special variant of the scalar form here to
 /// represent the associated intrinsic operation.  This form is unlike the
 /// plain scalar form, in that it takes an entire vector (instead of a
 /// scalar) and leaves the top elements undefined.
 ///
 /// And, we have a special variant form for a full-vector intrinsic form.
 

 // in order to obtain suitable precision.
 defm RSQRT : sse1_fp_unop_s<0x52, "rsqrt", X86frsqrt, SchedWriteFRsqrt, HasAVX>,
              sse1_fp_unop_p<0x52, "rsqrt", X86frsqrt, SchedWriteFRsqrt, [HasAVX]>;
 defm RCP   : sse1_fp_unop_s<0x53, "rcp", X86frcp, SchedWriteFRcp, HasAVX>,
              sse1_fp_unop_p<0x53, "rcp", X86frcp, SchedWriteFRcp, [HasAVX]>;
 
 // There is no f64 version of the reciprocal approximation instructions.
 
-// TODO: We should add *scalar* op patterns for these just like we have for
-// the binops above. If the binop and unop patterns could all be unified
-// that would be even better.
+multiclass scalar_unary_math_patterns<SDNode OpNode, string OpcPrefix, SDNode Move,
+                                      ValueType VT, Predicate BasePredicate> {
+  let Predicates = [BasePredicate] in {
+    def : Pat<(VT (Move VT:$dst, (scalar_to_vector
+                                  (OpNode (extractelt VT:$src, 0))))),
+              (!cast<I>(OpcPrefix#r_Int) VT:$dst, VT:$src)>;
+  }
+
+  // Repeat for AVX versions of the instructions.
+  let Predicates = [HasAVX] in {
+    def : Pat<(VT (Move VT:$dst, (scalar_to_vector
+                                  (OpNode (extractelt VT:$src, 0))))),
+              (!cast<I>("V"#OpcPrefix#r_Int) VT:$dst, VT:$src)>;
+  }
+}
+
+multiclass scalar_unary_math_imm_patterns<SDNode OpNode, string OpcPrefix, SDNode Move,
+                                          ValueType VT, bits<8> ImmV,
+                                          Predicate BasePredicate> {
+  let Predicates = [BasePredicate] in {
+    def : Pat<(VT (Move VT:$dst, (scalar_to_vector
+                                  (OpNode (extractelt VT:$src, 0))))),
+              (!cast<Ii8>(OpcPrefix#r_Int) VT:$dst, VT:$src, (i32 ImmV))>;
+  }
+
+  // Repeat for AVX versions of the instructions.
+  let Predicates = [HasAVX] in {
+    def : Pat<(VT (Move VT:$dst, (scalar_to_vector
+                                  (OpNode (extractelt VT:$src, 0))))),
+              (!cast<Ii8>("V"#OpcPrefix#r_Int) VT:$dst, VT:$src, (i32 ImmV))>;
+  }
+}
 
-multiclass scalar_unary_math_patterns<Intrinsic Intr, string OpcPrefix,
+multiclass scalar_unary_math_intr_patterns<Intrinsic Intr, string OpcPrefix,
                                            SDNode Move, ValueType VT,
                                            Predicate BasePredicate> {
   let Predicates = [BasePredicate] in {
     def : Pat<(VT (Move VT:$dst, (Intr VT:$src))),
               (!cast<I>(OpcPrefix#r_Int) VT:$dst, VT:$src)>;
   }
 
   // Repeat for AVX versions of the instructions.
   let Predicates = [HasAVX] in {
     def : Pat<(VT (Move VT:$dst, (Intr VT:$src))),
               (!cast<I>("V"#OpcPrefix#r_Int) VT:$dst, VT:$src)>;
   }
 }
 
-defm : scalar_unary_math_patterns<int_x86_sse_rcp_ss, "RCPSS", X86Movss,
+defm : scalar_unary_math_intr_patterns<int_x86_sse_rcp_ss, "RCPSS", X86Movss,
                                        v4f32, UseSSE1>;
-defm : scalar_unary_math_patterns<int_x86_sse_rsqrt_ss, "RSQRTSS", X86Movss,
+defm : scalar_unary_math_intr_patterns<int_x86_sse_rsqrt_ss, "RSQRTSS", X86Movss,
                                        v4f32, UseSSE1>;
-defm : scalar_unary_math_patterns<int_x86_sse_sqrt_ss, "SQRTSS", X86Movss,
+defm : scalar_unary_math_intr_patterns<int_x86_sse_sqrt_ss, "SQRTSS", X86Movss,
                                        v4f32, UseSSE1>;
-defm : scalar_unary_math_patterns<int_x86_sse2_sqrt_sd, "SQRTSD", X86Movsd,
+defm : scalar_unary_math_intr_patterns<int_x86_sse2_sqrt_sd, "SQRTSD", X86Movsd,
                                        v2f64, UseSSE2>;
 
 
 //===----------------------------------------------------------------------===//
 // SSE 1 & 2 - Non-temporal stores
 //===----------------------------------------------------------------------===//
 
 let AddedComplexity = 400 in { // Prefer non-temporal versions
 let Predicates = [HasAVX, NoVLX] in {

llvm/trunk/test/CodeGen/X86/combine-select.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=knl | FileCheck %s
 
 define <4 x float> @select_mask_add_ss(<4 x float> %w, i8 zeroext %u, <4 x float> %a, <4 x float> %b) {
 ; CHECK-LABEL: select_mask_add_ss:
 ; CHECK:       ## %bb.0: ## %entry
-; CHECK-NEXT:    vaddss %xmm2, %xmm1, %xmm2
-; CHECK-NEXT:    testb $1, %dil
-; CHECK-NEXT:    sete %al
-; CHECK-NEXT:    kmovw %eax, %k1
-; CHECK-NEXT:    vmovss %xmm0, %xmm1, %xmm2 {%k1}
-; CHECK-NEXT:    vmovaps %xmm2, %xmm0
+; CHECK-NEXT:    kmovw %edi, %k1
+; CHECK-NEXT:    vaddss %xmm2, %xmm1, %xmm0 {%k1}
 ; CHECK-NEXT:    retq
 entry:
   %0 = extractelement <4 x float> %b, i32 0
   %1 = extractelement <4 x float> %a, i32 0
   %2 = fadd float %1, %0
   %3 = and i8 %u, 1
   %4 = icmp eq i8 %3, 0
   %5 = extractelement <4 x float> %w, i32 0
   %6 = select i1 %4, float %5, float %2
   %7 = insertelement <4 x float> %a, float %6, i32 0
   ret <4 x float> %7
 }
 
 define <4 x float> @select_maskz_add_ss(i8 zeroext %u, <4 x float> %a, <4 x float> %b) {
 ; CHECK-LABEL: select_maskz_add_ss:
 ; CHECK:       ## %bb.0: ## %entry
-; CHECK-NEXT:    vaddss %xmm1, %xmm0, %xmm1
-; CHECK-NEXT:    testb $1, %dil
-; CHECK-NEXT:    sete %al
-; CHECK-NEXT:    kmovw %eax, %k1
-; CHECK-NEXT:    vxorps %xmm2, %xmm2, %xmm2
-; CHECK-NEXT:    vmovss %xmm2, %xmm0, %xmm1 {%k1}
-; CHECK-NEXT:    vmovaps %xmm1, %xmm0
+; CHECK-NEXT:    kmovw %edi, %k1
+; CHECK-NEXT:    vaddss %xmm1, %xmm0, %xmm0 {%k1} {z}
 ; CHECK-NEXT:    retq
 entry:
   %0 = extractelement <4 x float> %b, i32 0
   %1 = extractelement <4 x float> %a, i32 0
   %2 = fadd float %1, %0
   %3 = and i8 %u, 1
   %4 = icmp eq i8 %3, 0
   %5 = select i1 %4, float 0.000000e+00, float %2
   %6 = insertelement <4 x float> %a, float %5, i32 0
   ret <4 x float> %6
 }
 
 define <4 x float> @select_mask_sub_ss(<4 x float> %w, i8 zeroext %u, <4 x float> %a, <4 x float> %b) {
 ; CHECK-LABEL: select_mask_sub_ss:
 ; CHECK:       ## %bb.0: ## %entry
-; CHECK-NEXT:    vsubss %xmm2, %xmm1, %xmm2
-; CHECK-NEXT:    testb $1, %dil
-; CHECK-NEXT:    sete %al
-; CHECK-NEXT:    kmovw %eax, %k1
-; CHECK-NEXT:    vmovss %xmm0, %xmm1, %xmm2 {%k1}
-; CHECK-NEXT:    vmovaps %xmm2, %xmm0
+; CHECK-NEXT:    kmovw %edi, %k1
+; CHECK-NEXT:    vsubss %xmm2, %xmm1, %xmm0 {%k1}
 ; CHECK-NEXT:    retq
 entry:
   %0 = extractelement <4 x float> %b, i32 0
   %1 = extractelement <4 x float> %a, i32 0
   %2 = fsub float %1, %0
   %3 = and i8 %u, 1
   %4 = icmp eq i8 %3, 0
   %5 = extractelement <4 x float> %w, i32 0
   %6 = select i1 %4, float %5, float %2
   %7 = insertelement <4 x float> %a, float %6, i32 0
   ret <4 x float> %7
 }
 
 define <4 x float> @select_maskz_sub_ss(i8 zeroext %u, <4 x float> %a, <4 x float> %b) {
 ; CHECK-LABEL: select_maskz_sub_ss:
 ; CHECK:       ## %bb.0: ## %entry
-; CHECK-NEXT:    vsubss %xmm1, %xmm0, %xmm1
-; CHECK-NEXT:    testb $1, %dil
-; CHECK-NEXT:    sete %al
-; CHECK-NEXT:    kmovw %eax, %k1
-; CHECK-NEXT:    vxorps %xmm2, %xmm2, %xmm2
-; CHECK-NEXT:    vmovss %xmm2, %xmm0, %xmm1 {%k1}
-; CHECK-NEXT:    vmovaps %xmm1, %xmm0
+; CHECK-NEXT:    kmovw %edi, %k1
+; CHECK-NEXT:    vsubss %xmm1, %xmm0, %xmm0 {%k1} {z}
 ; CHECK-NEXT:    retq
 entry:
   %0 = extractelement <4 x float> %b, i32 0
   %1 = extractelement <4 x float> %a, i32 0
   %2 = fsub float %1, %0
   %3 = and i8 %u, 1
   %4 = icmp eq i8 %3, 0
   %5 = select i1 %4, float 0.000000e+00, float %2
   %6 = insertelement <4 x float> %a, float %5, i32 0
   ret <4 x float> %6
 }
 
 define <4 x float> @select_mask_mul_ss(<4 x float> %w, i8 zeroext %u, <4 x float> %a, <4 x float> %b) {
 ; CHECK-LABEL: select_mask_mul_ss:
 ; CHECK:       ## %bb.0: ## %entry
-; CHECK-NEXT:    vmulss %xmm2, %xmm1, %xmm2
-; CHECK-NEXT:    testb $1, %dil
-; CHECK-NEXT:    sete %al
-; CHECK-NEXT:    kmovw %eax, %k1
-; CHECK-NEXT:    vmovss %xmm0, %xmm1, %xmm2 {%k1}
-; CHECK-NEXT:    vmovaps %xmm2, %xmm0
+; CHECK-NEXT:    kmovw %edi, %k1
+; CHECK-NEXT:    vmulss %xmm2, %xmm1, %xmm0 {%k1}
 ; CHECK-NEXT:    retq
 entry:
   %0 = extractelement <4 x float> %b, i32 0
   %1 = extractelement <4 x float> %a, i32 0
   %2 = fmul float %1, %0
   %3 = and i8 %u, 1
   %4 = icmp eq i8 %3, 0
   %5 = extractelement <4 x float> %w, i32 0
   %6 = select i1 %4, float %5, float %2
   %7 = insertelement <4 x float> %a, float %6, i32 0
   ret <4 x float> %7
 }
 
 define <4 x float> @select_maskz_mul_ss(i8 zeroext %u, <4 x float> %a, <4 x float> %b) {
 ; CHECK-LABEL: select_maskz_mul_ss:
 ; CHECK:       ## %bb.0: ## %entry
-; CHECK-NEXT:    vmulss %xmm1, %xmm0, %xmm1
-; CHECK-NEXT:    testb $1, %dil
-; CHECK-NEXT:    sete %al
-; CHECK-NEXT:    kmovw %eax, %k1
-; CHECK-NEXT:    vxorps %xmm2, %xmm2, %xmm2
-; CHECK-NEXT:    vmovss %xmm2, %xmm0, %xmm1 {%k1}
-; CHECK-NEXT:    vmovaps %xmm1, %xmm0
+; CHECK-NEXT:    kmovw %edi, %k1
+; CHECK-NEXT:    vmulss %xmm1, %xmm0, %xmm0 {%k1} {z}
 ; CHECK-NEXT:    retq
 entry:
   %0 = extractelement <4 x float> %b, i32 0
   %1 = extractelement <4 x float> %a, i32 0
   %2 = fmul float %1, %0
   %3 = and i8 %u, 1
   %4 = icmp eq i8 %3, 0
   %5 = select i1 %4, float 0.000000e+00, float %2
   %6 = insertelement <4 x float> %a, float %5, i32 0
   ret <4 x float> %6
 }

llvm/trunk/test/CodeGen/X86/fma-scalar-combine.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mattr=+avx512f -mattr=+fma -show-mc-encoding | FileCheck %s --check-prefix=CHECK --check-prefix=SKX
 
 define <2 x double> @combine_scalar_mask_fmadd_f32(<2 x double> %a, i8 zeroext %k, <2 x double> %b, <2 x double> %c) {
 ; CHECK-LABEL: combine_scalar_mask_fmadd_f32:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vfmadd213ss %xmm2, %xmm0, %xmm1 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0xa9,0xca]
-; CHECK-NEXT:    # xmm1 = (xmm0 * xmm1) + xmm2
 ; CHECK-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; CHECK-NEXT:    vmovss %xmm1, %xmm0, %xmm0 {%k1} # encoding: [0x62,0xf1,0x7e,0x09,0x10,0xc1]
+; CHECK-NEXT:    vfmadd213ss %xmm2, %xmm1, %xmm0 {%k1} # encoding: [0x62,0xf2,0x75,0x09,0xa9,0xc2]
+; CHECK-NEXT:    # xmm0 = (xmm1 * xmm0) + xmm2
 ; CHECK-NEXT:    retq # encoding: [0xc3]
 entry:
   %0 = bitcast <2 x double> %a to <4 x float>
   %1 = bitcast <2 x double> %b to <4 x float>
   %2 = bitcast <2 x double> %c to <4 x float>
   %3 = extractelement <4 x float> %0, i64 0
   %4 = extractelement <4 x float> %1, i64 0
   %5 = extractelement <4 x float> %2, i64 0
   %6 = fmul fast float %4, %3
   %7 = fadd fast float %6, %5
   %8 = bitcast i8 %k to <8 x i1>
   %9 = extractelement <8 x i1> %8, i64 0
   %10 = select i1 %9, float %7, float %3
   %11 = insertelement <4 x float> %0, float %10, i64 0
   %12 = bitcast <4 x float> %11 to <2 x double>
   ret <2 x double> %12
 }
 
 define <2 x double> @combine_scalar_mask_fmadd_f64(<2 x double> %a, i8 zeroext %k, <2 x double> %b, <2 x double> %c) {
 ; CHECK-LABEL: combine_scalar_mask_fmadd_f64:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vfmadd213sd %xmm2, %xmm0, %xmm1 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf9,0xa9,0xca]
-; CHECK-NEXT:    # xmm1 = (xmm0 * xmm1) + xmm2
 ; CHECK-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; CHECK-NEXT:    vmovsd %xmm1, %xmm0, %xmm0 {%k1} # encoding: [0x62,0xf1,0xff,0x09,0x10,0xc1]
+; CHECK-NEXT:    vfmadd213sd %xmm2, %xmm1, %xmm0 {%k1} # encoding: [0x62,0xf2,0xf5,0x09,0xa9,0xc2]
+; CHECK-NEXT:    # xmm0 = (xmm1 * xmm0) + xmm2
 ; CHECK-NEXT:    retq # encoding: [0xc3]
 entry:
   %0 = extractelement <2 x double> %a, i64 0
   %1 = extractelement <2 x double> %b, i64 0
   %2 = extractelement <2 x double> %c, i64 0
   %3 = fmul fast double %1, %0
   %4 = fadd fast double %3, %2
   %5 = bitcast i8 %k to <8 x i1>
   %6 = extractelement <8 x i1> %5, i64 0
   %7 = select i1 %6, double %4, double %0
   %8 = insertelement <2 x double> %a, double %7, i64 0
   ret <2 x double> %8
 }
 
 define <2 x double> @combine_scalar_maskz_fmadd_32(i8 zeroext %k, <2 x double> %a, <2 x double> %b, <2 x double> %c) {
 ; CHECK-LABEL: combine_scalar_maskz_fmadd_32:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vfmadd213ss %xmm2, %xmm0, %xmm1 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0xa9,0xca]
-; CHECK-NEXT:    # xmm1 = (xmm0 * xmm1) + xmm2
 ; CHECK-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; CHECK-NEXT:    vmovss %xmm1, %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7e,0x89,0x10,0xc1]
+; CHECK-NEXT:    vfmadd213ss %xmm2, %xmm1, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x75,0x89,0xa9,0xc2]
+; CHECK-NEXT:    # xmm0 = (xmm1 * xmm0) + xmm2
 ; CHECK-NEXT:    retq # encoding: [0xc3]
 entry:
   %0 = bitcast <2 x double> %a to <4 x float>
   %1 = bitcast <2 x double> %b to <4 x float>
   %2 = bitcast <2 x double> %c to <4 x float>
   %3 = extractelement <4 x float> %0, i64 0
   %4 = extractelement <4 x float> %1, i64 0
   %5 = extractelement <4 x float> %2, i64 0
   %6 = fmul fast float %4, %3
   %7 = fadd fast float %6, %5
   %8 = bitcast i8 %k to <8 x i1>
   %9 = extractelement <8 x i1> %8, i64 0
   %10 = select i1 %9, float %7, float 0.000000e+00
   %11 = insertelement <4 x float> %0, float %10, i64 0
   %12 = bitcast <4 x float> %11 to <2 x double>
   ret <2 x double> %12
 }
 
 define <2 x double> @combine_scalar_maskz_fmadd_64(i8 zeroext %k, <2 x double> %a, <2 x double> %b, <2 x double> %c) {
 ; CHECK-LABEL: combine_scalar_maskz_fmadd_64:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vfmadd213sd %xmm2, %xmm0, %xmm1 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf9,0xa9,0xca]
-; CHECK-NEXT:    # xmm1 = (xmm0 * xmm1) + xmm2
 ; CHECK-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; CHECK-NEXT:    vmovsd %xmm1, %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf1,0xff,0x89,0x10,0xc1]
+; CHECK-NEXT:    vfmadd213sd %xmm2, %xmm1, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0xf5,0x89,0xa9,0xc2]
+; CHECK-NEXT:    # xmm0 = (xmm1 * xmm0) + xmm2
 ; CHECK-NEXT:    retq # encoding: [0xc3]
 entry:
   %0 = extractelement <2 x double> %a, i64 0
   %1 = extractelement <2 x double> %b, i64 0
   %2 = extractelement <2 x double> %c, i64 0
   %3 = fmul fast double %1, %0
   %4 = fadd fast double %3, %2
   %5 = bitcast i8 %k to <8 x i1>
   %6 = extractelement <8 x i1> %5, i64 0
   %7 = select i1 %6, double %4, double 0.000000e+00
   %8 = insertelement <2 x double> %a, double %7, i64 0
   ret <2 x double> %8
 }
 
 define <2 x double> @combine_scalar_mask3_fmadd_32(<2 x double> %a, <2 x double> %b, <2 x double> %c, i8 zeroext %k) {
 ; CHECK-LABEL: combine_scalar_mask3_fmadd_32:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vfmadd213ss %xmm2, %xmm0, %xmm1 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0xa9,0xca]
-; CHECK-NEXT:    # xmm1 = (xmm0 * xmm1) + xmm2
 ; CHECK-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; CHECK-NEXT:    vmovss %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf1,0x7e,0x09,0x10,0xd1]
+; CHECK-NEXT:    vfmadd231ss %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0xb9,0xd1]
+; CHECK-NEXT:    # xmm2 = (xmm0 * xmm1) + xmm2
 ; CHECK-NEXT:    vmovaps %xmm2, %xmm0 # encoding: [0xc5,0xf8,0x28,0xc2]
 ; CHECK-NEXT:    retq # encoding: [0xc3]
 entry:
   %0 = bitcast <2 x double> %a to <4 x float>
   %1 = bitcast <2 x double> %b to <4 x float>
   %2 = bitcast <2 x double> %c to <4 x float>
   %3 = extractelement <4 x float> %0, i64 0
   %4 = extractelement <4 x float> %1, i64 0
   %5 = extractelement <4 x float> %2, i64 0
   %6 = fmul fast float %4, %3
   %7 = fadd fast float %6, %5
   %8 = bitcast i8 %k to <8 x i1>
   %9 = extractelement <8 x i1> %8, i64 0
   %10 = select i1 %9, float %7, float %5
-  %11 = insertelement <4 x float> %0, float %10, i64 0
+  %11 = insertelement <4 x float> %2, float %10, i64 0
   %12 = bitcast <4 x float> %11 to <2 x double>
   ret <2 x double> %12
 }
 
 define <2 x double> @combine_scalar_mask3_fmadd_64(<2 x double> %a, <2 x double> %b, <2 x double> %c, i8 zeroext %k) {
 ; CHECK-LABEL: combine_scalar_mask3_fmadd_64:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vfmadd213sd %xmm2, %xmm0, %xmm1 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf9,0xa9,0xca]
-; CHECK-NEXT:    # xmm1 = (xmm0 * xmm1) + xmm2
 ; CHECK-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; CHECK-NEXT:    vmovsd %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf1,0xff,0x09,0x10,0xd1]
+; CHECK-NEXT:    vfmadd231sd %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0xb9,0xd1]
+; CHECK-NEXT:    # xmm2 = (xmm0 * xmm1) + xmm2
 ; CHECK-NEXT:    vmovapd %xmm2, %xmm0 # encoding: [0xc5,0xf9,0x28,0xc2]
 ; CHECK-NEXT:    retq # encoding: [0xc3]
 entry:
   %0 = extractelement <2 x double> %a, i64 0
   %1 = extractelement <2 x double> %b, i64 0
   %2 = extractelement <2 x double> %c, i64 0
   %3 = fmul fast double %1, %0
   %4 = fadd fast double %3, %2
   %5 = bitcast i8 %k to <8 x i1>
   %6 = extractelement <8 x i1> %5, i64 0
   %7 = select i1 %6, double %4, double %2
-  %8 = insertelement <2 x double> %a, double %7, i64 0
+  %8 = insertelement <2 x double> %c, double %7, i64 0
   ret <2 x double> %8
 }
 
 define <2 x double> @combine_scalar_mask_fmsub_f32(<2 x double> %a, i8 zeroext %k, <2 x double> %b, <2 x double> %c) {
 ; CHECK-LABEL: combine_scalar_mask_fmsub_f32:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vfmsub213ss %xmm2, %xmm0, %xmm1 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0xab,0xca]
-; CHECK-NEXT:    # xmm1 = (xmm0 * xmm1) - xmm2
 ; CHECK-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; CHECK-NEXT:    vmovss %xmm1, %xmm0, %xmm0 {%k1} # encoding: [0x62,0xf1,0x7e,0x09,0x10,0xc1]
+; CHECK-NEXT:    vfmsub213ss %xmm2, %xmm1, %xmm0 {%k1} # encoding: [0x62,0xf2,0x75,0x09,0xab,0xc2]
+; CHECK-NEXT:    # xmm0 = (xmm1 * xmm0) - xmm2
 ; CHECK-NEXT:    retq # encoding: [0xc3]
 entry:
   %0 = bitcast <2 x double> %a to <4 x float>
   %1 = bitcast <2 x double> %b to <4 x float>
   %2 = bitcast <2 x double> %c to <4 x float>
   %3 = extractelement <4 x float> %0, i64 0
   %4 = extractelement <4 x float> %1, i64 0
   %5 = extractelement <4 x float> %2, i64 0
   %6 = fmul fast float %4, %3
   %7 = fsub fast float %6, %5
   %8 = bitcast i8 %k to <8 x i1>
   %9 = extractelement <8 x i1> %8, i64 0
   %10 = select i1 %9, float %7, float %3
   %11 = insertelement <4 x float> %0, float %10, i64 0
   %12 = bitcast <4 x float> %11 to <2 x double>
   ret <2 x double> %12
 }
 
 define <2 x double> @combine_scalar_mask_fmsub_f64(<2 x double> %a, i8 zeroext %k, <2 x double> %b, <2 x double> %c) {
 ; CHECK-LABEL: combine_scalar_mask_fmsub_f64:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vfmsub213sd %xmm2, %xmm0, %xmm1 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf9,0xab,0xca]
-; CHECK-NEXT:    # xmm1 = (xmm0 * xmm1) - xmm2
 ; CHECK-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; CHECK-NEXT:    vmovsd %xmm1, %xmm0, %xmm0 {%k1} # encoding: [0x62,0xf1,0xff,0x09,0x10,0xc1]
+; CHECK-NEXT:    vfmsub213sd %xmm2, %xmm1, %xmm0 {%k1} # encoding: [0x62,0xf2,0xf5,0x09,0xab,0xc2]
+; CHECK-NEXT:    # xmm0 = (xmm1 * xmm0) - xmm2
 ; CHECK-NEXT:    retq # encoding: [0xc3]
 entry:
   %0 = extractelement <2 x double> %a, i64 0
   %1 = extractelement <2 x double> %b, i64 0
   %2 = extractelement <2 x double> %c, i64 0
   %3 = fmul fast double %1, %0
   %4 = fsub fast double %3, %2
   %5 = bitcast i8 %k to <8 x i1>
   %6 = extractelement <8 x i1> %5, i64 0
   %7 = select i1 %6, double %4, double %0
   %8 = insertelement <2 x double> %a, double %7, i64 0
   ret <2 x double> %8
 }
 
 define <2 x double> @combine_scalar_maskz_fmsub_32(i8 zeroext %k, <2 x double> %a, <2 x double> %b, <2 x double> %c) {
 ; CHECK-LABEL: combine_scalar_maskz_fmsub_32:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vfmsub213ss %xmm2, %xmm0, %xmm1 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0xab,0xca]
-; CHECK-NEXT:    # xmm1 = (xmm0 * xmm1) - xmm2
 ; CHECK-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; CHECK-NEXT:    vmovss %xmm1, %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7e,0x89,0x10,0xc1]
+; CHECK-NEXT:    vfmsub213ss %xmm2, %xmm1, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x75,0x89,0xab,0xc2]
+; CHECK-NEXT:    # xmm0 = (xmm1 * xmm0) - xmm2
 ; CHECK-NEXT:    retq # encoding: [0xc3]
 entry:
   %0 = bitcast <2 x double> %a to <4 x float>
   %1 = bitcast <2 x double> %b to <4 x float>
   %2 = bitcast <2 x double> %c to <4 x float>
   %3 = extractelement <4 x float> %0, i64 0
   %4 = extractelement <4 x float> %1, i64 0
   %5 = extractelement <4 x float> %2, i64 0
   %6 = fmul fast float %4, %3
   %7 = fsub fast float %6, %5
   %8 = bitcast i8 %k to <8 x i1>
   %9 = extractelement <8 x i1> %8, i64 0
   %10 = select i1 %9, float %7, float 0.000000e+00
   %11 = insertelement <4 x float> %0, float %10, i64 0
   %12 = bitcast <4 x float> %11 to <2 x double>
   ret <2 x double> %12
 }
 
 define <2 x double> @combine_scalar_maskz_fmsub_64(i8 zeroext %k, <2 x double> %a, <2 x double> %b, <2 x double> %c) {
 ; CHECK-LABEL: combine_scalar_maskz_fmsub_64:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vfmsub213sd %xmm2, %xmm0, %xmm1 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf9,0xab,0xca]
-; CHECK-NEXT:    # xmm1 = (xmm0 * xmm1) - xmm2
 ; CHECK-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; CHECK-NEXT:    vmovsd %xmm1, %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf1,0xff,0x89,0x10,0xc1]
+; CHECK-NEXT:    vfmsub213sd %xmm2, %xmm1, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0xf5,0x89,0xab,0xc2]
+; CHECK-NEXT:    # xmm0 = (xmm1 * xmm0) - xmm2
 ; CHECK-NEXT:    retq # encoding: [0xc3]
 entry:
   %0 = extractelement <2 x double> %a, i64 0
   %1 = extractelement <2 x double> %b, i64 0
   %2 = extractelement <2 x double> %c, i64 0
   %3 = fmul fast double %1, %0
   %4 = fsub fast double %3, %2
   %5 = bitcast i8 %k to <8 x i1>
   %6 = extractelement <8 x i1> %5, i64 0
   %7 = select i1 %6, double %4, double 0.000000e+00
   %8 = insertelement <2 x double> %a, double %7, i64 0
   ret <2 x double> %8
 }
 
 define <2 x double> @combine_scalar_mask3_fmsub_32(<2 x double> %a, <2 x double> %b, <2 x double> %c, i8 zeroext %k) {
 ; CHECK-LABEL: combine_scalar_mask3_fmsub_32:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vfmsub213ss %xmm2, %xmm0, %xmm1 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0xab,0xca]
-; CHECK-NEXT:    # xmm1 = (xmm0 * xmm1) - xmm2
 ; CHECK-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; CHECK-NEXT:    vmovss %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf1,0x7e,0x09,0x10,0xd1]
+; CHECK-NEXT:    vfmsub231ss %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0xbb,0xd1]
+; CHECK-NEXT:    # xmm2 = (xmm0 * xmm1) - xmm2
 ; CHECK-NEXT:    vmovaps %xmm2, %xmm0 # encoding: [0xc5,0xf8,0x28,0xc2]
 ; CHECK-NEXT:    retq # encoding: [0xc3]
 entry:
   %0 = bitcast <2 x double> %a to <4 x float>
   %1 = bitcast <2 x double> %b to <4 x float>
   %2 = bitcast <2 x double> %c to <4 x float>
   %3 = extractelement <4 x float> %0, i64 0
   %4 = extractelement <4 x float> %1, i64 0
   %5 = extractelement <4 x float> %2, i64 0
   %6 = fmul fast float %4, %3
   %7 = fsub fast float %6, %5
   %8 = bitcast i8 %k to <8 x i1>
   %9 = extractelement <8 x i1> %8, i64 0
   %10 = select i1 %9, float %7, float %5
-  %11 = insertelement <4 x float> %0, float %10, i64 0
+  %11 = insertelement <4 x float> %2, float %10, i64 0
   %12 = bitcast <4 x float> %11 to <2 x double>
   ret <2 x double> %12
 }
 
 define <2 x double> @combine_scalar_mask3_fmsub_64(<2 x double> %a, <2 x double> %b, <2 x double> %c, i8 zeroext %k) {
 ; CHECK-LABEL: combine_scalar_mask3_fmsub_64:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vfmsub213sd %xmm2, %xmm0, %xmm1 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf9,0xab,0xca]
-; CHECK-NEXT:    # xmm1 = (xmm0 * xmm1) - xmm2
 ; CHECK-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; CHECK-NEXT:    vmovsd %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf1,0xff,0x09,0x10,0xd1]
+; CHECK-NEXT:    vfmsub231sd %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0xbb,0xd1]
+; CHECK-NEXT:    # xmm2 = (xmm0 * xmm1) - xmm2
 ; CHECK-NEXT:    vmovapd %xmm2, %xmm0 # encoding: [0xc5,0xf9,0x28,0xc2]
 ; CHECK-NEXT:    retq # encoding: [0xc3]
 entry:
   %0 = extractelement <2 x double> %a, i64 0
   %1 = extractelement <2 x double> %b, i64 0
   %2 = extractelement <2 x double> %c, i64 0
   %3 = fmul fast double %1, %0
   %4 = fsub fast double %3, %2
   %5 = bitcast i8 %k to <8 x i1>
   %6 = extractelement <8 x i1> %5, i64 0
   %7 = select i1 %6, double %4, double %2
-  %8 = insertelement <2 x double> %a, double %7, i64 0
+  %8 = insertelement <2 x double> %c, double %7, i64 0
   ret <2 x double> %8
 }
 
 define <2 x double> @combine_scalar_mask_fnmadd_f32(<2 x double> %a, i8 zeroext %k, <2 x double> %b, <2 x double> %c) {
 ; CHECK-LABEL: combine_scalar_mask_fnmadd_f32:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vfnmadd213ss %xmm2, %xmm0, %xmm1 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0xad,0xca]
-; CHECK-NEXT:    # xmm1 = -(xmm0 * xmm1) + xmm2
 ; CHECK-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; CHECK-NEXT:    vmovss %xmm1, %xmm0, %xmm0 {%k1} # encoding: [0x62,0xf1,0x7e,0x09,0x10,0xc1]
+; CHECK-NEXT:    vfnmadd213ss %xmm2, %xmm1, %xmm0 {%k1} # encoding: [0x62,0xf2,0x75,0x09,0xad,0xc2]
+; CHECK-NEXT:    # xmm0 = -(xmm1 * xmm0) + xmm2
 ; CHECK-NEXT:    retq # encoding: [0xc3]
 entry:
   %0 = bitcast <2 x double> %a to <4 x float>
   %1 = bitcast <2 x double> %b to <4 x float>
   %2 = bitcast <2 x double> %c to <4 x float>
   %3 = extractelement <4 x float> %0, i64 0
   %4 = extractelement <4 x float> %1, i64 0
   %5 = extractelement <4 x float> %2, i64 0
   %6 = fmul fast float %4, %3
   %7 = fsub fast float %5, %6
   %8 = bitcast i8 %k to <8 x i1>
   %9 = extractelement <8 x i1> %8, i64 0
   %10 = select i1 %9, float %7, float %3
   %11 = insertelement <4 x float> %0, float %10, i64 0
   %12 = bitcast <4 x float> %11 to <2 x double>
   ret <2 x double> %12
 }
 
 define <2 x double> @combine_scalar_mask_fnmadd_f64(<2 x double> %a, i8 zeroext %k, <2 x double> %b, <2 x double> %c) {
 ; CHECK-LABEL: combine_scalar_mask_fnmadd_f64:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vfnmadd213sd %xmm2, %xmm0, %xmm1 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf9,0xad,0xca]
-; CHECK-NEXT:    # xmm1 = -(xmm0 * xmm1) + xmm2
 ; CHECK-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; CHECK-NEXT:    vmovsd %xmm1, %xmm0, %xmm0 {%k1} # encoding: [0x62,0xf1,0xff,0x09,0x10,0xc1]
+; CHECK-NEXT:    vfnmadd213sd %xmm2, %xmm1, %xmm0 {%k1} # encoding: [0x62,0xf2,0xf5,0x09,0xad,0xc2]
+; CHECK-NEXT:    # xmm0 = -(xmm1 * xmm0) + xmm2
 ; CHECK-NEXT:    retq # encoding: [0xc3]
 entry:
   %0 = extractelement <2 x double> %a, i64 0
   %1 = extractelement <2 x double> %b, i64 0
   %2 = extractelement <2 x double> %c, i64 0
   %3 = fmul fast double %1, %0
   %4 = fsub fast double %2, %3
   %5 = bitcast i8 %k to <8 x i1>
   %6 = extractelement <8 x i1> %5, i64 0
   %7 = select i1 %6, double %4, double %0
   %8 = insertelement <2 x double> %a, double %7, i64 0
   ret <2 x double> %8
 }
 
 define <2 x double> @combine_scalar_maskz_fnmadd_32(i8 zeroext %k, <2 x double> %a, <2 x double> %b, <2 x double> %c) {
 ; CHECK-LABEL: combine_scalar_maskz_fnmadd_32:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vfnmadd213ss %xmm2, %xmm0, %xmm1 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0xad,0xca]
-; CHECK-NEXT:    # xmm1 = -(xmm0 * xmm1) + xmm2
 ; CHECK-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; CHECK-NEXT:    vmovss %xmm1, %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7e,0x89,0x10,0xc1]
+; CHECK-NEXT:    vfnmadd213ss %xmm2, %xmm1, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x75,0x89,0xad,0xc2]
+; CHECK-NEXT:    # xmm0 = -(xmm1 * xmm0) + xmm2
 ; CHECK-NEXT:    retq # encoding: [0xc3]
 entry:
   %0 = bitcast <2 x double> %a to <4 x float>
   %1 = bitcast <2 x double> %b to <4 x float>
   %2 = bitcast <2 x double> %c to <4 x float>
   %3 = extractelement <4 x float> %0, i64 0
   %4 = extractelement <4 x float> %1, i64 0
   %5 = extractelement <4 x float> %2, i64 0
   %6 = fmul fast float %4, %3
   %7 = fsub fast float %5, %6
   %8 = bitcast i8 %k to <8 x i1>
   %9 = extractelement <8 x i1> %8, i64 0
   %10 = select i1 %9, float %7, float 0.000000e+00
   %11 = insertelement <4 x float> %0, float %10, i64 0
   %12 = bitcast <4 x float> %11 to <2 x double>
   ret <2 x double> %12
 }
 
 define <2 x double> @combine_scalar_maskz_fnmadd_64(i8 zeroext %k, <2 x double> %a, <2 x double> %b, <2 x double> %c) {
 ; CHECK-LABEL: combine_scalar_maskz_fnmadd_64:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vfnmadd213sd %xmm2, %xmm0, %xmm1 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf9,0xad,0xca]
-; CHECK-NEXT:    # xmm1 = -(xmm0 * xmm1) + xmm2
 ; CHECK-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; CHECK-NEXT:    vmovsd %xmm1, %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf1,0xff,0x89,0x10,0xc1]
+; CHECK-NEXT:    vfnmadd213sd %xmm2, %xmm1, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0xf5,0x89,0xad,0xc2]
+; CHECK-NEXT:    # xmm0 = -(xmm1 * xmm0) + xmm2
 ; CHECK-NEXT:    retq # encoding: [0xc3]
 entry:
   %0 = extractelement <2 x double> %a, i64 0
   %1 = extractelement <2 x double> %b, i64 0
   %2 = extractelement <2 x double> %c, i64 0
   %3 = fmul fast double %1, %0
   %4 = fsub fast double %2, %3
   %5 = bitcast i8 %k to <8 x i1>
   %6 = extractelement <8 x i1> %5, i64 0
   %7 = select i1 %6, double %4, double 0.000000e+00
   %8 = insertelement <2 x double> %a, double %7, i64 0
   ret <2 x double> %8
 }
 
 define <2 x double> @combine_scalar_mask3_fnmadd_32(<2 x double> %a, <2 x double> %b, <2 x double> %c, i8 zeroext %k) {
 ; CHECK-LABEL: combine_scalar_mask3_fnmadd_32:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vfnmadd213ss %xmm2, %xmm0, %xmm1 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0xad,0xca]
-; CHECK-NEXT:    # xmm1 = -(xmm0 * xmm1) + xmm2
 ; CHECK-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; CHECK-NEXT:    vmovss %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf1,0x7e,0x09,0x10,0xd1]
+; CHECK-NEXT:    vfnmadd231ss %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0xbd,0xd1]
+; CHECK-NEXT:    # xmm2 = -(xmm0 * xmm1) + xmm2
 ; CHECK-NEXT:    vmovaps %xmm2, %xmm0 # encoding: [0xc5,0xf8,0x28,0xc2]
 ; CHECK-NEXT:    retq # encoding: [0xc3]
 entry:
   %0 = bitcast <2 x double> %a to <4 x float>
   %1 = bitcast <2 x double> %b to <4 x float>
   %2 = bitcast <2 x double> %c to <4 x float>
   %3 = extractelement <4 x float> %0, i64 0
   %4 = extractelement <4 x float> %1, i64 0
   %5 = extractelement <4 x float> %2, i64 0
   %6 = fmul fast float %4, %3
   %7 = fsub fast float %5, %6
   %8 = bitcast i8 %k to <8 x i1>
   %9 = extractelement <8 x i1> %8, i64 0
   %10 = select i1 %9, float %7, float %5
-  %11 = insertelement <4 x float> %0, float %10, i64 0
+  %11 = insertelement <4 x float> %2, float %10, i64 0
   %12 = bitcast <4 x float> %11 to <2 x double>
   ret <2 x double> %12
 }
 
 define <2 x double> @combine_scalar_mask3_fnmadd_64(<2 x double> %a, <2 x double> %b, <2 x double> %c, i8 zeroext %k) {
 ; CHECK-LABEL: combine_scalar_mask3_fnmadd_64:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vfnmadd213sd %xmm2, %xmm0, %xmm1 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf9,0xad,0xca]
-; CHECK-NEXT:    # xmm1 = -(xmm0 * xmm1) + xmm2
 ; CHECK-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; CHECK-NEXT:    vmovsd %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf1,0xff,0x09,0x10,0xd1]
+; CHECK-NEXT:    vfnmadd231sd %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0xbd,0xd1]
+; CHECK-NEXT:    # xmm2 = -(xmm0 * xmm1) + xmm2
 ; CHECK-NEXT:    vmovapd %xmm2, %xmm0 # encoding: [0xc5,0xf9,0x28,0xc2]
 ; CHECK-NEXT:    retq # encoding: [0xc3]
 entry:
   %0 = extractelement <2 x double> %a, i64 0
   %1 = extractelement <2 x double> %b, i64 0
   %2 = extractelement <2 x double> %c, i64 0
   %3 = fmul fast double %1, %0
   %4 = fsub fast double %2, %3
   %5 = bitcast i8 %k to <8 x i1>
   %6 = extractelement <8 x i1> %5, i64 0
   %7 = select i1 %6, double %4, double %2
-  %8 = insertelement <2 x double> %a, double %7, i64 0
+  %8 = insertelement <2 x double> %c, double %7, i64 0
   ret <2 x double> %8
 }
 
 define <2 x double> @combine_scalar_mask_fnmsub_f32(<2 x double> %a, i8 zeroext %k, <2 x double> %b, <2 x double> %c) {
 ; CHECK-LABEL: combine_scalar_mask_fnmsub_f32:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vfnmsub213ss %xmm2, %xmm0, %xmm1 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0xaf,0xca]
-; CHECK-NEXT:    # xmm1 = -(xmm0 * xmm1) - xmm2
 ; CHECK-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; CHECK-NEXT:    vmovss %xmm1, %xmm0, %xmm0 {%k1} # encoding: [0x62,0xf1,0x7e,0x09,0x10,0xc1]
+; CHECK-NEXT:    vfnmsub213ss %xmm2, %xmm1, %xmm0 {%k1} # encoding: [0x62,0xf2,0x75,0x09,0xaf,0xc2]
+; CHECK-NEXT:    # xmm0 = -(xmm1 * xmm0) - xmm2
 ; CHECK-NEXT:    retq # encoding: [0xc3]
 entry:
   %0 = bitcast <2 x double> %a to <4 x float>
   %1 = bitcast <2 x double> %b to <4 x float>
   %2 = bitcast <2 x double> %c to <4 x float>
   %3 = extractelement <4 x float> %0, i64 0
   %4 = extractelement <4 x float> %1, i64 0
   %5 = extractelement <4 x float> %2, i64 0
   %sub = fsub fast float -0.000000e+00, %5
   %6 = fmul fast float %4, %3
   %7 = fsub fast float %sub, %6
   %8 = bitcast i8 %k to <8 x i1>
   %9 = extractelement <8 x i1> %8, i64 0
   %10 = select i1 %9, float %7, float %3
   %11 = insertelement <4 x float> %0, float %10, i64 0
   %12 = bitcast <4 x float> %11 to <2 x double>
   ret <2 x double> %12
 }
 
 define <2 x double> @combine_scalar_mask_fnmsub_f64(<2 x double> %a, i8 zeroext %k, <2 x double> %b, <2 x double> %c) {
 ; CHECK-LABEL: combine_scalar_mask_fnmsub_f64:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vfnmsub213sd %xmm2, %xmm0, %xmm1 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf9,0xaf,0xca]
-; CHECK-NEXT:    # xmm1 = -(xmm0 * xmm1) - xmm2
 ; CHECK-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; CHECK-NEXT:    vmovsd %xmm1, %xmm0, %xmm0 {%k1} # encoding: [0x62,0xf1,0xff,0x09,0x10,0xc1]
+; CHECK-NEXT:    vfnmsub213sd %xmm2, %xmm1, %xmm0 {%k1} # encoding: [0x62,0xf2,0xf5,0x09,0xaf,0xc2]
+; CHECK-NEXT:    # xmm0 = -(xmm1 * xmm0) - xmm2
 ; CHECK-NEXT:    retq # encoding: [0xc3]
 entry:
   %0 = extractelement <2 x double> %a, i64 0
   %1 = extractelement <2 x double> %b, i64 0
   %2 = extractelement <2 x double> %c, i64 0
   %sub = fsub fast double -0.000000e+00, %2
   %3 = fmul fast double %1, %0
   %4 = fsub fast double %sub, %3
   %5 = bitcast i8 %k to <8 x i1>
   %6 = extractelement <8 x i1> %5, i64 0
   %7 = select i1 %6, double %4, double %0
   %8 = insertelement <2 x double> %a, double %7, i64 0
   ret <2 x double> %8
 }
 
 define <2 x double> @combine_scalar_maskz_fnmsub_32(i8 zeroext %k, <2 x double> %a, <2 x double> %b, <2 x double> %c) {
 ; CHECK-LABEL: combine_scalar_maskz_fnmsub_32:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vfnmsub213ss %xmm2, %xmm0, %xmm1 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0xaf,0xca]
-; CHECK-NEXT:    # xmm1 = -(xmm0 * xmm1) - xmm2
 ; CHECK-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; CHECK-NEXT:    vmovss %xmm1, %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf1,0x7e,0x89,0x10,0xc1]
+; CHECK-NEXT:    vfnmsub213ss %xmm2, %xmm1, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0x75,0x89,0xaf,0xc2]
+; CHECK-NEXT:    # xmm0 = -(xmm1 * xmm0) - xmm2
 ; CHECK-NEXT:    retq # encoding: [0xc3]
 entry:
   %0 = bitcast <2 x double> %a to <4 x float>
   %1 = bitcast <2 x double> %b to <4 x float>
   %2 = bitcast <2 x double> %c to <4 x float>
   %3 = extractelement <4 x float> %0, i64 0
   %4 = extractelement <4 x float> %1, i64 0
   %5 = extractelement <4 x float> %2, i64 0
   %sub = fsub fast float -0.000000e+00, %5
   %6 = fmul fast float %4, %3
   %7 = fsub fast float %sub, %6
   %8 = bitcast i8 %k to <8 x i1>
   %9 = extractelement <8 x i1> %8, i64 0
   %10 = select i1 %9, float %7, float 0.000000e+00
   %11 = insertelement <4 x float> %0, float %10, i64 0
   %12 = bitcast <4 x float> %11 to <2 x double>
   ret <2 x double> %12
 }
 
 define <2 x double> @combine_scalar_maskz_fnmsub_64(i8 zeroext %k, <2 x double> %a, <2 x double> %b, <2 x double> %c) {
 ; CHECK-LABEL: combine_scalar_maskz_fnmsub_64:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vfnmsub213sd %xmm2, %xmm0, %xmm1 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf9,0xaf,0xca]
-; CHECK-NEXT:    # xmm1 = -(xmm0 * xmm1) - xmm2
 ; CHECK-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; CHECK-NEXT:    vmovsd %xmm1, %xmm0, %xmm0 {%k1} {z} # encoding: [0x62,0xf1,0xff,0x89,0x10,0xc1]
+; CHECK-NEXT:    vfnmsub213sd %xmm2, %xmm1, %xmm0 {%k1} {z} # encoding: [0x62,0xf2,0xf5,0x89,0xaf,0xc2]
+; CHECK-NEXT:    # xmm0 = -(xmm1 * xmm0) - xmm2
 ; CHECK-NEXT:    retq # encoding: [0xc3]
 entry:
   %0 = extractelement <2 x double> %a, i64 0
   %1 = extractelement <2 x double> %b, i64 0
   %2 = extractelement <2 x double> %c, i64 0
   %sub = fsub fast double -0.000000e+00, %2
   %3 = fmul fast double %1, %0
   %4 = fsub fast double %sub, %3
   %5 = bitcast i8 %k to <8 x i1>
   %6 = extractelement <8 x i1> %5, i64 0
   %7 = select i1 %6, double %4, double 0.000000e+00
   %8 = insertelement <2 x double> %a, double %7, i64 0
   ret <2 x double> %8
 }
 
 define <2 x double> @combine_scalar_mask3_fnmsub_32(<2 x double> %a, <2 x double> %b, <2 x double> %c, i8 zeroext %k) {
 ; CHECK-LABEL: combine_scalar_mask3_fnmsub_32:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vfnmsub213ss %xmm2, %xmm0, %xmm1 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0xaf,0xca]
-; CHECK-NEXT:    # xmm1 = -(xmm0 * xmm1) - xmm2
 ; CHECK-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; CHECK-NEXT:    vmovss %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf1,0x7e,0x09,0x10,0xd1]
+; CHECK-NEXT:    vfnmsub231ss %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0x7d,0x09,0xbf,0xd1]
+; CHECK-NEXT:    # xmm2 = -(xmm0 * xmm1) - xmm2
 ; CHECK-NEXT:    vmovaps %xmm2, %xmm0 # encoding: [0xc5,0xf8,0x28,0xc2]
 ; CHECK-NEXT:    retq # encoding: [0xc3]
 entry:
   %0 = bitcast <2 x double> %a to <4 x float>
   %1 = bitcast <2 x double> %b to <4 x float>
   %2 = bitcast <2 x double> %c to <4 x float>
   %3 = extractelement <4 x float> %0, i64 0
   %4 = extractelement <4 x float> %1, i64 0
   %5 = extractelement <4 x float> %2, i64 0
   %sub = fsub fast float -0.000000e+00, %5
   %6 = fmul fast float %4, %3
   %7 = fsub fast float %sub, %6
   %8 = bitcast i8 %k to <8 x i1>
   %9 = extractelement <8 x i1> %8, i64 0
   %10 = select i1 %9, float %7, float %5
-  %11 = insertelement <4 x float> %0, float %10, i64 0
+  %11 = insertelement <4 x float> %2, float %10, i64 0
   %12 = bitcast <4 x float> %11 to <2 x double>
   ret <2 x double> %12
 }
 
 define <2 x double> @combine_scalar_mask3_fnmsub_64(<2 x double> %a, <2 x double> %b, <2 x double> %c, i8 zeroext %k) {
 ; CHECK-LABEL: combine_scalar_mask3_fnmsub_64:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vfnmsub213sd %xmm2, %xmm0, %xmm1 # EVEX TO VEX Compression encoding: [0xc4,0xe2,0xf9,0xaf,0xca]
-; CHECK-NEXT:    # xmm1 = -(xmm0 * xmm1) - xmm2
 ; CHECK-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; CHECK-NEXT:    vmovsd %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf1,0xff,0x09,0x10,0xd1]
+; CHECK-NEXT:    vfnmsub231sd %xmm1, %xmm0, %xmm2 {%k1} # encoding: [0x62,0xf2,0xfd,0x09,0xbf,0xd1]
+; CHECK-NEXT:    # xmm2 = -(xmm0 * xmm1) - xmm2
 ; CHECK-NEXT:    vmovapd %xmm2, %xmm0 # encoding: [0xc5,0xf9,0x28,0xc2]
 ; CHECK-NEXT:    retq # encoding: [0xc3]
 entry:
   %0 = extractelement <2 x double> %a, i64 0
   %1 = extractelement <2 x double> %b, i64 0
   %2 = extractelement <2 x double> %c, i64 0
   %sub = fsub fast double -0.000000e+00, %2
   %3 = fmul fast double %1, %0
   %4 = fsub fast double %sub, %3
   %5 = bitcast i8 %k to <8 x i1>
   %6 = extractelement <8 x i1> %5, i64 0
   %7 = select i1 %6, double %4, double %2
-  %8 = insertelement <2 x double> %a, double %7, i64 0
+  %8 = insertelement <2 x double> %c, double %7, i64 0
   ret <2 x double> %8
 }