Index: lib/CodeGen/TwoAddressInstructionPass.cpp
===================================================================
--- lib/CodeGen/TwoAddressInstructionPass.cpp
+++ lib/CodeGen/TwoAddressInstructionPass.cpp
@@ -1205,6 +1205,7 @@
   if (!MI->isCommutable())
     return false;
 
+  bool MadeChange = false;
   unsigned DstOpReg = MI->getOperand(DstOpIdx).getReg();
   unsigned BaseOpReg = MI->getOperand(BaseOpIdx).getReg();
   unsigned OpsNum = MI->getDesc().getNumOperands();
@@ -1223,8 +1224,8 @@
 
     // If OtherOp dies but BaseOp does not, swap the OtherOp and BaseOp
     // operands. This makes the live ranges of DstOp and OtherOp joinable.
-    bool DoCommute =
-        !BaseOpKilled && isKilled(*MI, OtherOpReg, MRI, TII, LIS, false);
+    bool OtherOpKilled = isKilled(*MI, OtherOpReg, MRI, TII, LIS, false);
+    bool DoCommute = !BaseOpKilled && OtherOpKilled;
 
     if (!DoCommute &&
         isProfitableToCommute(DstOpReg, BaseOpReg, OtherOpReg, MI, Dist)) {
@@ -1238,10 +1239,15 @@
       ++NumCommuted;
       if (AggressiveCommute)
         ++NumAggrCommuted;
-      return true;
+      // There might be more than two commutable operands, update BaseOp and
+      // contine scanning.
+      BaseOpReg = OtherOpReg;
+      BaseOpKilled = OtherOpKilled;
+      MadeChange = true;
+      continue;
     }
   }
-  return false;
+  return MadeChange;
 }
 
 /// For the case where an instruction has a single pair of tied register
Index: test/CodeGen/X86/avx512-vpternlog-commute.ll
===================================================================
--- test/CodeGen/X86/avx512-vpternlog-commute.ll
+++ test/CodeGen/X86/avx512-vpternlog-commute.ll
@@ -27,8 +27,7 @@
 define <16 x i32> @vpternlog_v16i32_210(<16 x i32> %x0, <16 x i32> %x1, <16 x i32> %x2) {
 ; CHECK-LABEL: vpternlog_v16i32_210:
 ; CHECK:       ## %bb.0:
-; CHECK-NEXT:    vpternlogd $78, %zmm0, %zmm2, %zmm1
-; CHECK-NEXT:    vmovdqa64 %zmm1, %zmm0
+; CHECK-NEXT:    vpternlogd $92, %zmm1, %zmm2, %zmm0
 ; CHECK-NEXT:    retq
   %res = call <16 x i32> @llvm.x86.avx512.mask.pternlog.d.512(<16 x i32> %x2, <16 x i32> %x1, <16 x i32> %x0, i32 114, i16 -1)
   ret <16 x i32> %res
@@ -434,8 +433,7 @@
 ; CHECK-LABEL: vpternlog_v16i32_210_maskz:
 ; CHECK:       ## %bb.0:
 ; CHECK-NEXT:    kmovd %edi, %k1
-; CHECK-NEXT:    vpternlogd $78, %zmm0, %zmm2, %zmm1 {%k1} {z}
-; CHECK-NEXT:    vmovdqa64 %zmm1, %zmm0
+; CHECK-NEXT:    vpternlogd $92, %zmm1, %zmm2, %zmm0 {%k1} {z}
 ; CHECK-NEXT:    retq
   %res = call <16 x i32> @llvm.x86.avx512.maskz.pternlog.d.512(<16 x i32> %x2, <16 x i32> %x1, <16 x i32> %x0, i32 114, i16 %mask)
   ret <16 x i32> %res
Index: test/CodeGen/X86/recip-fastmath.ll
===================================================================
--- test/CodeGen/X86/recip-fastmath.ll
+++ test/CodeGen/X86/recip-fastmath.ll
@@ -390,10 +390,11 @@
 ;
 ; HASWELL-LABEL: v4f32_one_step:
 ; HASWELL:       # %bb.0:
-; HASWELL-NEXT:    vrcpps %xmm0, %xmm1 # sched: [5:1.00]
-; HASWELL-NEXT:    vbroadcastss {{.*#+}} xmm2 = [1,1,1,1] sched: [6:0.50]
-; HASWELL-NEXT:    vfnmadd213ps %xmm2, %xmm1, %xmm0 # sched: [5:0.50]
-; HASWELL-NEXT:    vfmadd132ps %xmm1, %xmm1, %xmm0 # sched: [5:0.50]
+; HASWELL-NEXT:    vrcpps %xmm0, %xmm2 # sched: [5:1.00]
+; HASWELL-NEXT:    vbroadcastss {{.*#+}} xmm1 = [1,1,1,1] sched: [6:0.50]
+; HASWELL-NEXT:    vfnmadd231ps %xmm0, %xmm2, %xmm1 # sched: [5:0.50]
+; HASWELL-NEXT:    vfmadd132ps %xmm2, %xmm2, %xmm1 # sched: [5:0.50]
+; HASWELL-NEXT:    vmovaps %xmm1, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [7:1.00]
 ;
 ; HASWELL-NO-FMA-LABEL: v4f32_one_step:
@@ -408,10 +409,11 @@
 ;
 ; KNL-LABEL: v4f32_one_step:
 ; KNL:       # %bb.0:
-; KNL-NEXT:    vrcpps %xmm0, %xmm1 # sched: [5:1.00]
-; KNL-NEXT:    vbroadcastss {{.*#+}} xmm2 = [1,1,1,1] sched: [6:0.50]
-; KNL-NEXT:    vfnmadd213ps %xmm2, %xmm1, %xmm0 # sched: [5:0.50]
-; KNL-NEXT:    vfmadd132ps %xmm1, %xmm1, %xmm0 # sched: [5:0.50]
+; KNL-NEXT:    vrcpps %xmm0, %xmm2 # sched: [5:1.00]
+; KNL-NEXT:    vbroadcastss {{.*#+}} xmm1 = [1,1,1,1] sched: [6:0.50]
+; KNL-NEXT:    vfnmadd231ps %xmm0, %xmm2, %xmm1 # sched: [5:0.50]
+; KNL-NEXT:    vfmadd132ps %xmm2, %xmm2, %xmm1 # sched: [5:0.50]
+; KNL-NEXT:    vmovaps %xmm1, %xmm0 # sched: [1:1.00]
 ; KNL-NEXT:    retq # sched: [7:1.00]
 ;
 ; SKX-LABEL: v4f32_one_step:
@@ -665,10 +667,11 @@
 ;
 ; HASWELL-LABEL: v8f32_one_step:
 ; HASWELL:       # %bb.0:
-; HASWELL-NEXT:    vrcpps %ymm0, %ymm1 # sched: [11:2.00]
-; HASWELL-NEXT:    vbroadcastss {{.*#+}} ymm2 = [1,1,1,1,1,1,1,1] sched: [7:0.50]
-; HASWELL-NEXT:    vfnmadd213ps %ymm2, %ymm1, %ymm0 # sched: [5:0.50]
-; HASWELL-NEXT:    vfmadd132ps %ymm1, %ymm1, %ymm0 # sched: [5:0.50]
+; HASWELL-NEXT:    vrcpps %ymm0, %ymm2 # sched: [11:2.00]
+; HASWELL-NEXT:    vbroadcastss {{.*#+}} ymm1 = [1,1,1,1,1,1,1,1] sched: [7:0.50]
+; HASWELL-NEXT:    vfnmadd231ps %ymm0, %ymm2, %ymm1 # sched: [5:0.50]
+; HASWELL-NEXT:    vfmadd132ps %ymm2, %ymm2, %ymm1 # sched: [5:0.50]
+; HASWELL-NEXT:    vmovaps %ymm1, %ymm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [7:1.00]
 ;
 ; HASWELL-NO-FMA-LABEL: v8f32_one_step:
@@ -683,10 +686,11 @@
 ;
 ; KNL-LABEL: v8f32_one_step:
 ; KNL:       # %bb.0:
-; KNL-NEXT:    vrcpps %ymm0, %ymm1 # sched: [11:2.00]
-; KNL-NEXT:    vbroadcastss {{.*#+}} ymm2 = [1,1,1,1,1,1,1,1] sched: [7:0.50]
-; KNL-NEXT:    vfnmadd213ps %ymm2, %ymm1, %ymm0 # sched: [5:0.50]
-; KNL-NEXT:    vfmadd132ps %ymm1, %ymm1, %ymm0 # sched: [5:0.50]
+; KNL-NEXT:    vrcpps %ymm0, %ymm2 # sched: [11:2.00]
+; KNL-NEXT:    vbroadcastss {{.*#+}} ymm1 = [1,1,1,1,1,1,1,1] sched: [7:0.50]
+; KNL-NEXT:    vfnmadd231ps %ymm0, %ymm2, %ymm1 # sched: [5:0.50]
+; KNL-NEXT:    vfmadd132ps %ymm2, %ymm2, %ymm1 # sched: [5:0.50]
+; KNL-NEXT:    vmovaps %ymm1, %ymm0 # sched: [1:1.00]
 ; KNL-NEXT:    retq # sched: [7:1.00]
 ;
 ; SKX-LABEL: v8f32_one_step:
Index: test/CodeGen/X86/recip-fastmath2.ll
===================================================================
--- test/CodeGen/X86/recip-fastmath2.ll
+++ test/CodeGen/X86/recip-fastmath2.ll
@@ -451,9 +451,9 @@
 ; HASWELL:       # %bb.0:
 ; HASWELL-NEXT:    vrcpps %xmm0, %xmm1 # sched: [5:1.00]
 ; HASWELL-NEXT:    vbroadcastss {{.*#+}} xmm2 = [1,1,1,1] sched: [6:0.50]
-; HASWELL-NEXT:    vfnmadd213ps %xmm2, %xmm1, %xmm0 # sched: [5:0.50]
-; HASWELL-NEXT:    vfmadd132ps %xmm1, %xmm1, %xmm0 # sched: [5:0.50]
-; HASWELL-NEXT:    vmulps {{.*}}(%rip), %xmm0, %xmm0 # sched: [11:0.50]
+; HASWELL-NEXT:    vfnmadd231ps %xmm0, %xmm1, %xmm2 # sched: [5:0.50]
+; HASWELL-NEXT:    vfmadd132ps %xmm1, %xmm1, %xmm2 # sched: [5:0.50]
+; HASWELL-NEXT:    vmulps {{.*}}(%rip), %xmm2, %xmm0 # sched: [11:0.50]
 ; HASWELL-NEXT:    retq # sched: [7:1.00]
 ;
 ; HASWELL-NO-FMA-LABEL: v4f32_one_step2:
@@ -471,9 +471,9 @@
 ; KNL:       # %bb.0:
 ; KNL-NEXT:    vrcpps %xmm0, %xmm1 # sched: [5:1.00]
 ; KNL-NEXT:    vbroadcastss {{.*#+}} xmm2 = [1,1,1,1] sched: [6:0.50]
-; KNL-NEXT:    vfnmadd213ps %xmm2, %xmm1, %xmm0 # sched: [5:0.50]
-; KNL-NEXT:    vfmadd132ps %xmm1, %xmm1, %xmm0 # sched: [5:0.50]
-; KNL-NEXT:    vmulps {{.*}}(%rip), %xmm0, %xmm0 # sched: [11:0.50]
+; KNL-NEXT:    vfnmadd231ps %xmm0, %xmm1, %xmm2 # sched: [5:0.50]
+; KNL-NEXT:    vfmadd132ps %xmm1, %xmm1, %xmm2 # sched: [5:0.50]
+; KNL-NEXT:    vmulps {{.*}}(%rip), %xmm2, %xmm0 # sched: [11:0.50]
 ; KNL-NEXT:    retq # sched: [7:1.00]
 ;
 ; SKX-LABEL: v4f32_one_step2:
@@ -550,10 +550,10 @@
 ; HASWELL:       # %bb.0:
 ; HASWELL-NEXT:    vrcpps %xmm0, %xmm1 # sched: [5:1.00]
 ; HASWELL-NEXT:    vbroadcastss {{.*#+}} xmm2 = [1,1,1,1] sched: [6:0.50]
-; HASWELL-NEXT:    vfnmadd213ps %xmm2, %xmm1, %xmm0 # sched: [5:0.50]
-; HASWELL-NEXT:    vfmadd132ps %xmm1, %xmm1, %xmm0 # sched: [5:0.50]
-; HASWELL-NEXT:    vmulps {{.*}}(%rip), %xmm0, %xmm1 # sched: [11:0.50]
-; HASWELL-NEXT:    vmulps %xmm0, %xmm1, %xmm0 # sched: [5:0.50]
+; HASWELL-NEXT:    vfnmadd231ps %xmm0, %xmm1, %xmm2 # sched: [5:0.50]
+; HASWELL-NEXT:    vfmadd132ps %xmm1, %xmm1, %xmm2 # sched: [5:0.50]
+; HASWELL-NEXT:    vmulps {{.*}}(%rip), %xmm2, %xmm0 # sched: [11:0.50]
+; HASWELL-NEXT:    vmulps %xmm2, %xmm0, %xmm0 # sched: [5:0.50]
 ; HASWELL-NEXT:    retq # sched: [7:1.00]
 ;
 ; HASWELL-NO-FMA-LABEL: v4f32_one_step_2_divs:
@@ -572,10 +572,10 @@
 ; KNL:       # %bb.0:
 ; KNL-NEXT:    vrcpps %xmm0, %xmm1 # sched: [5:1.00]
 ; KNL-NEXT:    vbroadcastss {{.*#+}} xmm2 = [1,1,1,1] sched: [6:0.50]
-; KNL-NEXT:    vfnmadd213ps %xmm2, %xmm1, %xmm0 # sched: [5:0.50]
-; KNL-NEXT:    vfmadd132ps %xmm1, %xmm1, %xmm0 # sched: [5:0.50]
-; KNL-NEXT:    vmulps {{.*}}(%rip), %xmm0, %xmm1 # sched: [11:0.50]
-; KNL-NEXT:    vmulps %xmm0, %xmm1, %xmm0 # sched: [5:0.50]
+; KNL-NEXT:    vfnmadd231ps %xmm0, %xmm1, %xmm2 # sched: [5:0.50]
+; KNL-NEXT:    vfmadd132ps %xmm1, %xmm1, %xmm2 # sched: [5:0.50]
+; KNL-NEXT:    vmulps {{.*}}(%rip), %xmm2, %xmm0 # sched: [11:0.50]
+; KNL-NEXT:    vmulps %xmm2, %xmm0, %xmm0 # sched: [5:0.50]
 ; KNL-NEXT:    retq # sched: [7:1.00]
 ;
 ; SKX-LABEL: v4f32_one_step_2_divs:
@@ -787,9 +787,9 @@
 ; HASWELL:       # %bb.0:
 ; HASWELL-NEXT:    vrcpps %ymm0, %ymm1 # sched: [11:2.00]
 ; HASWELL-NEXT:    vbroadcastss {{.*#+}} ymm2 = [1,1,1,1,1,1,1,1] sched: [7:0.50]
-; HASWELL-NEXT:    vfnmadd213ps %ymm2, %ymm1, %ymm0 # sched: [5:0.50]
-; HASWELL-NEXT:    vfmadd132ps %ymm1, %ymm1, %ymm0 # sched: [5:0.50]
-; HASWELL-NEXT:    vmulps {{.*}}(%rip), %ymm0, %ymm0 # sched: [12:0.50]
+; HASWELL-NEXT:    vfnmadd231ps %ymm0, %ymm1, %ymm2 # sched: [5:0.50]
+; HASWELL-NEXT:    vfmadd132ps %ymm1, %ymm1, %ymm2 # sched: [5:0.50]
+; HASWELL-NEXT:    vmulps {{.*}}(%rip), %ymm2, %ymm0 # sched: [12:0.50]
 ; HASWELL-NEXT:    retq # sched: [7:1.00]
 ;
 ; HASWELL-NO-FMA-LABEL: v8f32_one_step2:
@@ -807,9 +807,9 @@
 ; KNL:       # %bb.0:
 ; KNL-NEXT:    vrcpps %ymm0, %ymm1 # sched: [11:2.00]
 ; KNL-NEXT:    vbroadcastss {{.*#+}} ymm2 = [1,1,1,1,1,1,1,1] sched: [7:0.50]
-; KNL-NEXT:    vfnmadd213ps %ymm2, %ymm1, %ymm0 # sched: [5:0.50]
-; KNL-NEXT:    vfmadd132ps %ymm1, %ymm1, %ymm0 # sched: [5:0.50]
-; KNL-NEXT:    vmulps {{.*}}(%rip), %ymm0, %ymm0 # sched: [12:0.50]
+; KNL-NEXT:    vfnmadd231ps %ymm0, %ymm1, %ymm2 # sched: [5:0.50]
+; KNL-NEXT:    vfmadd132ps %ymm1, %ymm1, %ymm2 # sched: [5:0.50]
+; KNL-NEXT:    vmulps {{.*}}(%rip), %ymm2, %ymm0 # sched: [12:0.50]
 ; KNL-NEXT:    retq # sched: [7:1.00]
 ;
 ; SKX-LABEL: v8f32_one_step2:
@@ -895,10 +895,10 @@
 ; HASWELL:       # %bb.0:
 ; HASWELL-NEXT:    vrcpps %ymm0, %ymm1 # sched: [11:2.00]
 ; HASWELL-NEXT:    vbroadcastss {{.*#+}} ymm2 = [1,1,1,1,1,1,1,1] sched: [7:0.50]
-; HASWELL-NEXT:    vfnmadd213ps %ymm2, %ymm1, %ymm0 # sched: [5:0.50]
-; HASWELL-NEXT:    vfmadd132ps %ymm1, %ymm1, %ymm0 # sched: [5:0.50]
-; HASWELL-NEXT:    vmulps {{.*}}(%rip), %ymm0, %ymm1 # sched: [12:0.50]
-; HASWELL-NEXT:    vmulps %ymm0, %ymm1, %ymm0 # sched: [5:0.50]
+; HASWELL-NEXT:    vfnmadd231ps %ymm0, %ymm1, %ymm2 # sched: [5:0.50]
+; HASWELL-NEXT:    vfmadd132ps %ymm1, %ymm1, %ymm2 # sched: [5:0.50]
+; HASWELL-NEXT:    vmulps {{.*}}(%rip), %ymm2, %ymm0 # sched: [12:0.50]
+; HASWELL-NEXT:    vmulps %ymm2, %ymm0, %ymm0 # sched: [5:0.50]
 ; HASWELL-NEXT:    retq # sched: [7:1.00]
 ;
 ; HASWELL-NO-FMA-LABEL: v8f32_one_step_2_divs:
@@ -917,10 +917,10 @@
 ; KNL:       # %bb.0:
 ; KNL-NEXT:    vrcpps %ymm0, %ymm1 # sched: [11:2.00]
 ; KNL-NEXT:    vbroadcastss {{.*#+}} ymm2 = [1,1,1,1,1,1,1,1] sched: [7:0.50]
-; KNL-NEXT:    vfnmadd213ps %ymm2, %ymm1, %ymm0 # sched: [5:0.50]
-; KNL-NEXT:    vfmadd132ps %ymm1, %ymm1, %ymm0 # sched: [5:0.50]
-; KNL-NEXT:    vmulps {{.*}}(%rip), %ymm0, %ymm1 # sched: [12:0.50]
-; KNL-NEXT:    vmulps %ymm0, %ymm1, %ymm0 # sched: [5:0.50]
+; KNL-NEXT:    vfnmadd231ps %ymm0, %ymm1, %ymm2 # sched: [5:0.50]
+; KNL-NEXT:    vfmadd132ps %ymm1, %ymm1, %ymm2 # sched: [5:0.50]
+; KNL-NEXT:    vmulps {{.*}}(%rip), %ymm2, %ymm0 # sched: [12:0.50]
+; KNL-NEXT:    vmulps %ymm2, %ymm0, %ymm0 # sched: [5:0.50]
 ; KNL-NEXT:    retq # sched: [7:1.00]
 ;
 ; SKX-LABEL: v8f32_one_step_2_divs: