diff --git a/llvm/lib/Target/ARM/ARMISelLowering.cpp b/llvm/lib/Target/ARM/ARMISelLowering.cpp
--- a/llvm/lib/Target/ARM/ARMISelLowering.cpp
+++ b/llvm/lib/Target/ARM/ARMISelLowering.cpp
@@ -4422,10 +4422,8 @@
 bool ARMTargetLowering::splitValueIntoRegisterParts(
     SelectionDAG &DAG, const SDLoc &DL, SDValue Val, SDValue *Parts,
     unsigned NumParts, MVT PartVT, std::optional<CallingConv::ID> CC) const {
-  bool IsABIRegCopy = CC.has_value();
   EVT ValueVT = Val.getValueType();
-  if (IsABIRegCopy && (ValueVT == MVT::f16 || ValueVT == MVT::bf16) &&
-      PartVT == MVT::f32) {
+  if ((ValueVT == MVT::f16 || ValueVT == MVT::bf16) && PartVT == MVT::f32) {
     unsigned ValueBits = ValueVT.getSizeInBits();
     unsigned PartBits = PartVT.getSizeInBits();
     Val = DAG.getNode(ISD::BITCAST, DL, MVT::getIntegerVT(ValueBits), Val);
@@ -4440,9 +4438,7 @@
 SDValue ARMTargetLowering::joinRegisterPartsIntoValue(
     SelectionDAG &DAG, const SDLoc &DL, const SDValue *Parts, unsigned NumParts,
     MVT PartVT, EVT ValueVT, std::optional<CallingConv::ID> CC) const {
-  bool IsABIRegCopy = CC.has_value();
-  if (IsABIRegCopy && (ValueVT == MVT::f16 || ValueVT == MVT::bf16) &&
-      PartVT == MVT::f32) {
+  if ((ValueVT == MVT::f16 || ValueVT == MVT::bf16) && PartVT == MVT::f32) {
     unsigned ValueBits = ValueVT.getSizeInBits();
     unsigned PartBits = PartVT.getSizeInBits();
     SDValue Val = Parts[0];
@@ -20241,8 +20237,12 @@
     case 'w':
       if (VT == MVT::Other)
         break;
-      if (VT == MVT::f16 || VT == MVT::bf16)
-        return RCPair(0U, &ARM::HPRRegClass);
+      if (VT == MVT::f16)
+        return RCPair(0U, Subtarget->hasFullFP16() ? &ARM::HPRRegClass
+                                                   : &ARM::SPRRegClass);
+      if (VT == MVT::bf16)
+        return RCPair(0U, Subtarget->hasBF16() ? &ARM::HPRRegClass
+                                               : &ARM::SPRRegClass);
       if (VT == MVT::f32)
         return RCPair(0U, &ARM::SPRRegClass);
       if (VT.getSizeInBits() == 64)
@@ -20263,8 +20263,12 @@
     case 't':
       if (VT == MVT::Other)
         break;
-      if (VT == MVT::f16 || VT == MVT::bf16)
-        return RCPair(0U, &ARM::HPRRegClass);
+      if (VT == MVT::f16)
+        return RCPair(0U, Subtarget->hasFullFP16() ? &ARM::HPRRegClass
+                                                   : &ARM::SPRRegClass);
+      if (VT == MVT::bf16)
+        return RCPair(0U, Subtarget->hasBF16() ? &ARM::HPRRegClass
+                                               : &ARM::SPRRegClass);
       if (VT == MVT::f32 || VT == MVT::i32)
         return RCPair(0U, &ARM::SPRRegClass);
       if (VT.getSizeInBits() == 64)
diff --git a/llvm/test/CodeGen/ARM/aes-erratum-fix.ll b/llvm/test/CodeGen/ARM/aes-erratum-fix.ll
--- a/llvm/test/CodeGen/ARM/aes-erratum-fix.ll
+++ b/llvm/test/CodeGen/ARM/aes-erratum-fix.ll
@@ -1353,335 +1353,210 @@
 define arm_aapcs_vfpcc void @aese_setf16_cond_via_ptr(i1 zeroext %0, half* %1, <16 x i8> %2, <16 x i8>* %3) nounwind {
 ; CHECK-FIX-NOSCHED-LABEL: aese_setf16_cond_via_ptr:
 ; CHECK-FIX-NOSCHED:       @ %bb.0:
-; CHECK-FIX-NOSCHED-NEXT:    .save {r4, r5, r6, r7, r8, lr}
-; CHECK-FIX-NOSCHED-NEXT:    push {r4, r5, r6, r7, r8, lr}
+; CHECK-FIX-NOSCHED-NEXT:    .save {r4, r5, r6, r7, r8, r9, r10, r11, lr}
+; CHECK-FIX-NOSCHED-NEXT:    push {r4, r5, r6, r7, r8, r9, r10, r11, lr}
+; CHECK-FIX-NOSCHED-NEXT:    .pad #24
+; CHECK-FIX-NOSCHED-NEXT:    sub sp, sp, #24
 ; CHECK-FIX-NOSCHED-NEXT:    cmp r0, #0
 ; CHECK-FIX-NOSCHED-NEXT:    beq .LBB36_3
 ; CHECK-FIX-NOSCHED-NEXT:  @ %bb.1:
 ; CHECK-FIX-NOSCHED-NEXT:    vld1.64 {d16, d17}, [r2]
-; CHECK-FIX-NOSCHED-NEXT:    vorr q9, q8, q8
-; CHECK-FIX-NOSCHED-NEXT:    vmov lr, r12, d17
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r3, d16[1]
-; CHECK-FIX-NOSCHED-NEXT:    vld1.16 {d18[0]}, [r1:16]
-; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r4, d18[0]
-; CHECK-FIX-NOSCHED-NEXT:    vmov s6, lr
-; CHECK-FIX-NOSCHED-NEXT:    lsr r5, r3, #16
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s8, s6
-; CHECK-FIX-NOSCHED-NEXT:    vmov s6, r3
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s12, s6
-; CHECK-FIX-NOSCHED-NEXT:    lsr r3, r12, #16
-; CHECK-FIX-NOSCHED-NEXT:    lsr lr, lr, #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov s4, r12
-; CHECK-FIX-NOSCHED-NEXT:    vmov s10, lr
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s4, s4
-; CHECK-FIX-NOSCHED-NEXT:    vmov s14, r5
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s10, s10
-; CHECK-FIX-NOSCHED-NEXT:    vmov s6, r4
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s14, s14
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s5, s6
-; CHECK-FIX-NOSCHED-NEXT:    vmov s6, r3
-; CHECK-FIX-NOSCHED-NEXT:    lsr r3, r4, #16
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s6, s6
-; CHECK-FIX-NOSCHED-NEXT:    vmov s7, r3
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s7, s7
+; CHECK-FIX-NOSCHED-NEXT:    vmov r7, r6, d17
+; CHECK-FIX-NOSCHED-NEXT:    vld1.16 {d16[0]}, [r1:16]
+; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r5, d16[0]
+; CHECK-FIX-NOSCHED-NEXT:    uxth r4, r3
+; CHECK-FIX-NOSCHED-NEXT:    lsr r3, r3, #16
+; CHECK-FIX-NOSCHED-NEXT:    str r4, [sp, #8] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    uxth r4, r7
+; CHECK-FIX-NOSCHED-NEXT:    str r4, [sp, #12] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    uxth r4, r6
+; CHECK-FIX-NOSCHED-NEXT:    lsr r6, r6, #16
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp, #4] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    str r4, [sp, #20] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    lsr r7, r7, #16
+; CHECK-FIX-NOSCHED-NEXT:    lsr r3, r5, #16
+; CHECK-FIX-NOSCHED-NEXT:    str r6, [sp, #16] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    uxth r10, r5
 ; CHECK-FIX-NOSCHED-NEXT:    cmp r0, #0
 ; CHECK-FIX-NOSCHED-NEXT:    bne .LBB36_4
 ; CHECK-FIX-NOSCHED-NEXT:  .LBB36_2:
-; CHECK-FIX-NOSCHED-NEXT:    vmov r0, r1, d0
-; CHECK-FIX-NOSCHED-NEXT:    vmov r3, r7, d1
-; CHECK-FIX-NOSCHED-NEXT:    vmov s1, r1
-; CHECK-FIX-NOSCHED-NEXT:    lsr r1, r1, #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov s0, r7
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s9, s1
-; CHECK-FIX-NOSCHED-NEXT:    vmov s2, r3
-; CHECK-FIX-NOSCHED-NEXT:    lsr r3, r3, #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov s1, r0
-; CHECK-FIX-NOSCHED-NEXT:    lsr r7, r7, #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov s3, r3
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s11, s1
-; CHECK-FIX-NOSCHED-NEXT:    vmov s1, r7
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s0, s0
-; CHECK-FIX-NOSCHED-NEXT:    vmov s13, r1
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s2, s2
-; CHECK-FIX-NOSCHED-NEXT:    lsr r0, r0, #16
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s1, s1
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s3, s3
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s13, s13
+; CHECK-FIX-NOSCHED-NEXT:    vmov r4, r6, d1
+; CHECK-FIX-NOSCHED-NEXT:    vmov r0, r3, d0
+; CHECK-FIX-NOSCHED-NEXT:    lsr r5, r4, #16
+; CHECK-FIX-NOSCHED-NEXT:    lsr r1, r6, #16
+; CHECK-FIX-NOSCHED-NEXT:    uxth r11, r6
+; CHECK-FIX-NOSCHED-NEXT:    lsr lr, r0, #16
+; CHECK-FIX-NOSCHED-NEXT:    lsr r12, r3, #16
+; CHECK-FIX-NOSCHED-NEXT:    uxth r9, r4
+; CHECK-FIX-NOSCHED-NEXT:    uxth r6, r3
 ; CHECK-FIX-NOSCHED-NEXT:    b .LBB36_5
 ; CHECK-FIX-NOSCHED-NEXT:  .LBB36_3:
-; CHECK-FIX-NOSCHED-NEXT:    ldrh r5, [r2, #10]
-; CHECK-FIX-NOSCHED-NEXT:    ldrh r12, [r2, #6]
-; CHECK-FIX-NOSCHED-NEXT:    ldrh r6, [r2, #2]
-; CHECK-FIX-NOSCHED-NEXT:    ldrh r7, [r2, #14]
-; CHECK-FIX-NOSCHED-NEXT:    vmov s8, r5
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r3, [r2, #14]
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp, #16] @ 4-byte Spill
 ; CHECK-FIX-NOSCHED-NEXT:    ldrh r3, [r2, #12]
-; CHECK-FIX-NOSCHED-NEXT:    vmov s12, r12
-; CHECK-FIX-NOSCHED-NEXT:    ldrh r4, [r2, #8]
-; CHECK-FIX-NOSCHED-NEXT:    vmov s5, r6
-; CHECK-FIX-NOSCHED-NEXT:    ldrh lr, [r2, #4]
-; CHECK-FIX-NOSCHED-NEXT:    vmov s4, r7
-; CHECK-FIX-NOSCHED-NEXT:    ldrh r8, [r2]
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s6, s4
-; CHECK-FIX-NOSCHED-NEXT:    vmov s4, r3
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s10, s8
-; CHECK-FIX-NOSCHED-NEXT:    vmov s8, r4
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s14, s12
-; CHECK-FIX-NOSCHED-NEXT:    vmov s12, lr
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s7, s5
-; CHECK-FIX-NOSCHED-NEXT:    vmov s5, r8
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s4, s4
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s8, s8
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s12, s12
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s5, s5
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp, #20] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r3, [r2, #8]
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp, #12] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r3, [r2, #6]
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r7, [r2, #10]
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r10, [r2]
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp, #4] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r3, [r2, #4]
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp, #8] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r3, [r2, #2]
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp] @ 4-byte Spill
 ; CHECK-FIX-NOSCHED-NEXT:    cmp r0, #0
 ; CHECK-FIX-NOSCHED-NEXT:    beq .LBB36_2
 ; CHECK-FIX-NOSCHED-NEXT:  .LBB36_4:
-; CHECK-FIX-NOSCHED-NEXT:    vorr q8, q0, q0
-; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r3, d0[1]
-; CHECK-FIX-NOSCHED-NEXT:    vld1.16 {d16[0]}, [r1:16]
-; CHECK-FIX-NOSCHED-NEXT:    vmov r0, r1, d1
-; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r7, d16[0]
-; CHECK-FIX-NOSCHED-NEXT:    vmov s1, r3
-; CHECK-FIX-NOSCHED-NEXT:    lsr r3, r3, #16
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s9, s1
-; CHECK-FIX-NOSCHED-NEXT:    vmov s13, r3
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s13, s13
-; CHECK-FIX-NOSCHED-NEXT:    vmov s0, r1
-; CHECK-FIX-NOSCHED-NEXT:    lsr r1, r1, #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov s2, r0
-; CHECK-FIX-NOSCHED-NEXT:    lsr r0, r0, #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov s3, r0
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s0, s0
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s2, s2
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s3, s3
-; CHECK-FIX-NOSCHED-NEXT:    vmov s1, r7
-; CHECK-FIX-NOSCHED-NEXT:    lsr r0, r7, #16
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s11, s1
-; CHECK-FIX-NOSCHED-NEXT:    vmov s1, r1
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s1, s1
+; CHECK-FIX-NOSCHED-NEXT:    vmov r5, r3, d1
+; CHECK-FIX-NOSCHED-NEXT:    mov r4, r7
+; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r7, d0[1]
+; CHECK-FIX-NOSCHED-NEXT:    vld1.16 {d0[0]}, [r1:16]
+; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r0, d0[0]
+; CHECK-FIX-NOSCHED-NEXT:    uxth r9, r5
+; CHECK-FIX-NOSCHED-NEXT:    uxth r11, r3
+; CHECK-FIX-NOSCHED-NEXT:    uxth r6, r7
+; CHECK-FIX-NOSCHED-NEXT:    lsr r12, r7, #16
+; CHECK-FIX-NOSCHED-NEXT:    lsr r1, r3, #16
+; CHECK-FIX-NOSCHED-NEXT:    lsr r5, r5, #16
+; CHECK-FIX-NOSCHED-NEXT:    mov r7, r4
+; CHECK-FIX-NOSCHED-NEXT:    lsr lr, r0, #16
 ; CHECK-FIX-NOSCHED-NEXT:  .LBB36_5:
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s5, s5
-; CHECK-FIX-NOSCHED-NEXT:    vmov s15, r0
-; CHECK-FIX-NOSCHED-NEXT:    vmov r0, s5
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s5, s7
-; CHECK-FIX-NOSCHED-NEXT:    vmov r1, s5
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s15, s15
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s5, s15
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s11, s11
-; CHECK-FIX-NOSCHED-NEXT:    vmov r3, s5
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s12, s12
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s5, s9
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s8, s8
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s2, s2
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s0, s0
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r0, r1, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov r1, s11
+; CHECK-FIX-NOSCHED-NEXT:    uxth r8, r0
+; CHECK-FIX-NOSCHED-NEXT:    ldr r0, [sp] @ 4-byte Reload
+; CHECK-FIX-NOSCHED-NEXT:    ldr r3, [sp, #4] @ 4-byte Reload
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r10, r0, lsl #16
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d16[0], r0
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r1, r3, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov r1, s12
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s12, s14
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r8, lr, lsl #16
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d18[0], r0
-; CHECK-FIX-NOSCHED-NEXT:    vmov r0, s12
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s12, s13
-; CHECK-FIX-NOSCHED-NEXT:    vmov r3, s12
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r1, r0, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov r1, s5
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r6, r12, lsl #16
+; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d18[1], r0
+; CHECK-FIX-NOSCHED-NEXT:    ldr r0, [sp, #8] @ 4-byte Reload
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r0, r3, lsl #16
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d16[1], r0
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r1, r1, r3, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d18[1], r1
-; CHECK-FIX-NOSCHED-NEXT:    vmov r1, s8
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s8, s10
-; CHECK-FIX-NOSCHED-NEXT:    vmov r3, s8
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r1, r3, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov r1, s2
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s2, s3
-; CHECK-FIX-NOSCHED-NEXT:    vmov r3, s2
-; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d17[0], r0
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s2, s4
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r1, r3, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov r1, s0
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s0, s1
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r9, r5, lsl #16
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d19[0], r0
-; CHECK-FIX-NOSCHED-NEXT:    vmov r0, s0
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s0, s6
-; CHECK-FIX-NOSCHED-NEXT:    vmov r3, s0
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r1, r0, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov r1, s2
+; CHECK-FIX-NOSCHED-NEXT:    ldr r0, [sp, #12] @ 4-byte Reload
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r0, r7, lsl #16
+; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d17[0], r0
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r11, r1, lsl #16
+; CHECK-FIX-NOSCHED-NEXT:    ldr r1, [sp, #16] @ 4-byte Reload
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d19[1], r0
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r1, r3, lsl #16
+; CHECK-FIX-NOSCHED-NEXT:    ldr r0, [sp, #20] @ 4-byte Reload
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r0, r1, lsl #16
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d17[1], r0
 ; CHECK-FIX-NOSCHED-NEXT:    aese.8 q8, q9
 ; CHECK-FIX-NOSCHED-NEXT:    aesmc.8 q8, q8
 ; CHECK-FIX-NOSCHED-NEXT:    vst1.64 {d16, d17}, [r2]
-; CHECK-FIX-NOSCHED-NEXT:    pop {r4, r5, r6, r7, r8, pc}
+; CHECK-FIX-NOSCHED-NEXT:    add sp, sp, #24
+; CHECK-FIX-NOSCHED-NEXT:    pop {r4, r5, r6, r7, r8, r9, r10, r11, pc}
 ;
 ; CHECK-CORTEX-FIX-LABEL: aese_setf16_cond_via_ptr:
 ; CHECK-CORTEX-FIX:       @ %bb.0:
-; CHECK-CORTEX-FIX-NEXT:    .save {r4, r5, r6, r7, r8, lr}
-; CHECK-CORTEX-FIX-NEXT:    push {r4, r5, r6, r7, r8, lr}
-; CHECK-CORTEX-FIX-NEXT:    .vsave {d8, d9}
-; CHECK-CORTEX-FIX-NEXT:    vpush {d8, d9}
+; CHECK-CORTEX-FIX-NEXT:    .save {r4, r5, r6, r7, r8, r9, r10, r11, lr}
+; CHECK-CORTEX-FIX-NEXT:    push {r4, r5, r6, r7, r8, r9, r10, r11, lr}
+; CHECK-CORTEX-FIX-NEXT:    .pad #24
+; CHECK-CORTEX-FIX-NEXT:    sub sp, sp, #24
 ; CHECK-CORTEX-FIX-NEXT:    cmp r0, #0
 ; CHECK-CORTEX-FIX-NEXT:    beq .LBB36_3
 ; CHECK-CORTEX-FIX-NEXT:  @ %bb.1:
 ; CHECK-CORTEX-FIX-NEXT:    vld1.64 {d16, d17}, [r2]
 ; CHECK-CORTEX-FIX-NEXT:    vorr q9, q8, q8
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 r3, d16[1]
-; CHECK-CORTEX-FIX-NEXT:    vmov r5, r6, d17
 ; CHECK-CORTEX-FIX-NEXT:    vld1.16 {d18[0]}, [r1:16]
-; CHECK-CORTEX-FIX-NEXT:    lsr r7, r5, #16
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 r3, d18[0]
+; CHECK-CORTEX-FIX-NEXT:    uxth r7, r3
+; CHECK-CORTEX-FIX-NEXT:    lsr r3, r3, #16
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #16] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 r3, d16[1]
+; CHECK-CORTEX-FIX-NEXT:    str r7, [sp, #20] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    uxth r7, r3
+; CHECK-CORTEX-FIX-NEXT:    lsr r3, r3, #16
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #4] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    vmov r3, r6, d17
+; CHECK-CORTEX-FIX-NEXT:    str r7, [sp, #8] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    uxth r7, r3
+; CHECK-CORTEX-FIX-NEXT:    lsr r3, r3, #16
+; CHECK-CORTEX-FIX-NEXT:    uxth r11, r6
 ; CHECK-CORTEX-FIX-NEXT:    lsr r4, r6, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s4, r6
-; CHECK-CORTEX-FIX-NEXT:    vmov s6, r5
-; CHECK-CORTEX-FIX-NEXT:    lsr r8, r3, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s8, r3
-; CHECK-CORTEX-FIX-NEXT:    vmov s12, r4
-; CHECK-CORTEX-FIX-NEXT:    vmov s5, r7
-; CHECK-CORTEX-FIX-NEXT:    vmov s9, r8
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s10, s4
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s4, s6
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 lr, d18[0]
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s6, s5
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s7, s8
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s14, s12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s5, s9
-; CHECK-CORTEX-FIX-NEXT:    lsr r12, lr, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s11, lr
-; CHECK-CORTEX-FIX-NEXT:    vmov s13, r12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s12, s11
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s8, s13
+; CHECK-CORTEX-FIX-NEXT:    str r7, [sp, #12] @ 4-byte Spill
 ; CHECK-CORTEX-FIX-NEXT:    cmp r0, #0
 ; CHECK-CORTEX-FIX-NEXT:    bne .LBB36_4
 ; CHECK-CORTEX-FIX-NEXT:  .LBB36_2:
-; CHECK-CORTEX-FIX-NEXT:    vmov r6, r5, d1
-; CHECK-CORTEX-FIX-NEXT:    vmov r0, r1, d0
-; CHECK-CORTEX-FIX-NEXT:    lsr r7, r1, #16
-; CHECK-CORTEX-FIX-NEXT:    lsr r4, r6, #16
-; CHECK-CORTEX-FIX-NEXT:    lsr r3, r5, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s2, r6
-; CHECK-CORTEX-FIX-NEXT:    vmov s0, r5
-; CHECK-CORTEX-FIX-NEXT:    vmov s3, r1
-; CHECK-CORTEX-FIX-NEXT:    vmov s9, r0
-; CHECK-CORTEX-FIX-NEXT:    lsr r12, r0, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s13, r3
-; CHECK-CORTEX-FIX-NEXT:    vmov s15, r4
-; CHECK-CORTEX-FIX-NEXT:    vmov s16, r7
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s1, s0
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s11, s3
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s3, s9
-; CHECK-CORTEX-FIX-NEXT:    vmov s0, r12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s2, s2
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s9, s15
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s13, s13
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s15, s16
+; CHECK-CORTEX-FIX-NEXT:    vmov r1, r7, d0
+; CHECK-CORTEX-FIX-NEXT:    uxth r0, r1
+; CHECK-CORTEX-FIX-NEXT:    uxth r6, r7
+; CHECK-CORTEX-FIX-NEXT:    lsr r12, r7, #16
+; CHECK-CORTEX-FIX-NEXT:    lsr r9, r1, #16
+; CHECK-CORTEX-FIX-NEXT:    str r0, [sp] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    mov r0, r3
+; CHECK-CORTEX-FIX-NEXT:    vmov r7, r3, d1
+; CHECK-CORTEX-FIX-NEXT:    uxth r10, r7
+; CHECK-CORTEX-FIX-NEXT:    lsr r5, r7, #16
+; CHECK-CORTEX-FIX-NEXT:    uxth lr, r3
+; CHECK-CORTEX-FIX-NEXT:    lsr r8, r3, #16
+; CHECK-CORTEX-FIX-NEXT:    mov r3, r0
 ; CHECK-CORTEX-FIX-NEXT:    b .LBB36_5
 ; CHECK-CORTEX-FIX-NEXT:  .LBB36_3:
-; CHECK-CORTEX-FIX-NEXT:    ldrh r12, [r2]
-; CHECK-CORTEX-FIX-NEXT:    ldrh lr, [r2, #2]
-; CHECK-CORTEX-FIX-NEXT:    ldrh r8, [r2, #4]
-; CHECK-CORTEX-FIX-NEXT:    ldrh r5, [r2, #6]
-; CHECK-CORTEX-FIX-NEXT:    ldrh r4, [r2, #8]
+; CHECK-CORTEX-FIX-NEXT:    ldrh r3, [r2]
+; CHECK-CORTEX-FIX-NEXT:    ldrh r11, [r2, #12]
+; CHECK-CORTEX-FIX-NEXT:    ldrh r4, [r2, #14]
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #20] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    ldrh r3, [r2, #2]
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #16] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    ldrh r3, [r2, #4]
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #8] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    ldrh r3, [r2, #6]
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #4] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    ldrh r3, [r2, #8]
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #12] @ 4-byte Spill
 ; CHECK-CORTEX-FIX-NEXT:    ldrh r3, [r2, #10]
-; CHECK-CORTEX-FIX-NEXT:    ldrh r7, [r2, #12]
-; CHECK-CORTEX-FIX-NEXT:    ldrh r6, [r2, #14]
-; CHECK-CORTEX-FIX-NEXT:    vmov s5, r5
-; CHECK-CORTEX-FIX-NEXT:    vmov s7, r8
-; CHECK-CORTEX-FIX-NEXT:    vmov s4, r6
-; CHECK-CORTEX-FIX-NEXT:    vmov s6, r7
-; CHECK-CORTEX-FIX-NEXT:    vmov s8, r3
-; CHECK-CORTEX-FIX-NEXT:    vmov s12, r4
-; CHECK-CORTEX-FIX-NEXT:    vmov s9, lr
-; CHECK-CORTEX-FIX-NEXT:    vmov s11, r12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s14, s4
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s10, s6
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s6, s8
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s4, s12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s5, s5
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s7, s7
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s8, s9
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s12, s11
 ; CHECK-CORTEX-FIX-NEXT:    cmp r0, #0
 ; CHECK-CORTEX-FIX-NEXT:    beq .LBB36_2
 ; CHECK-CORTEX-FIX-NEXT:  .LBB36_4:
 ; CHECK-CORTEX-FIX-NEXT:    vorr q8, q0, q0
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 r3, d0[1]
-; CHECK-CORTEX-FIX-NEXT:    vmov r6, r5, d1
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 r5, d0[1]
 ; CHECK-CORTEX-FIX-NEXT:    vld1.16 {d16[0]}, [r1:16]
-; CHECK-CORTEX-FIX-NEXT:    lsr r4, r6, #16
-; CHECK-CORTEX-FIX-NEXT:    lsr r1, r5, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s2, r6
-; CHECK-CORTEX-FIX-NEXT:    vmov s0, r5
-; CHECK-CORTEX-FIX-NEXT:    lsr r7, r3, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s3, r3
-; CHECK-CORTEX-FIX-NEXT:    vmov s9, r1
-; CHECK-CORTEX-FIX-NEXT:    vmov s15, r4
-; CHECK-CORTEX-FIX-NEXT:    vmov s16, r7
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s1, s0
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s2, s2
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 r0, d16[0]
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s13, s9
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s9, s15
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s11, s3
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s15, s16
-; CHECK-CORTEX-FIX-NEXT:    vmov s18, r0
-; CHECK-CORTEX-FIX-NEXT:    lsr r12, r0, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s0, r12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s3, s18
+; CHECK-CORTEX-FIX-NEXT:    uxth r6, r5
+; CHECK-CORTEX-FIX-NEXT:    lsr r12, r5, #16
+; CHECK-CORTEX-FIX-NEXT:    vmov r5, r7, d1
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 r1, d16[0]
+; CHECK-CORTEX-FIX-NEXT:    uxth r10, r5
+; CHECK-CORTEX-FIX-NEXT:    lsr r5, r5, #16
+; CHECK-CORTEX-FIX-NEXT:    uxth lr, r7
+; CHECK-CORTEX-FIX-NEXT:    lsr r8, r7, #16
+; CHECK-CORTEX-FIX-NEXT:    uxth r0, r1
+; CHECK-CORTEX-FIX-NEXT:    lsr r9, r1, #16
+; CHECK-CORTEX-FIX-NEXT:    str r0, [sp] @ 4-byte Spill
 ; CHECK-CORTEX-FIX-NEXT:  .LBB36_5:
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s10, s10
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s14, s14
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s7, s7
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s6, s6
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s4, s4
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s8, s8
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s12, s12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s0, s0
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s2, s2
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s9, s9
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s3, s3
-; CHECK-CORTEX-FIX-NEXT:    vmov r0, s10
-; CHECK-CORTEX-FIX-NEXT:    vmov r1, s14
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s10, s5
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s14, s1
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s1, s13
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s5, s11
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s11, s15
-; CHECK-CORTEX-FIX-NEXT:    vmov r5, s6
-; CHECK-CORTEX-FIX-NEXT:    vmov r4, s8
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s0, s0
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r12, r0, r1, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r1, s7
-; CHECK-CORTEX-FIX-NEXT:    vmov r3, s10
-; CHECK-CORTEX-FIX-NEXT:    vmov r7, s1
-; CHECK-CORTEX-FIX-NEXT:    vmov r6, s11
-; CHECK-CORTEX-FIX-NEXT:    vmov r0, s9
-; CHECK-CORTEX-FIX-NEXT:    pkhbt lr, r1, r3, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r3, s14
-; CHECK-CORTEX-FIX-NEXT:    vmov r1, s0
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r3, r3, r7, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r7, s5
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r7, r7, r6, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r6, s4
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r6, r6, r5, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r5, s12
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r5, r5, r4, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r4, s2
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d18[0], r5
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d19[0], r6
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d18[1], lr
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d19[1], r12
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r0, r4, r0, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r4, s3
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r1, r4, r1, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d16[0], r1
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d17[0], r0
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d16[1], r7
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d17[1], r3
+; CHECK-CORTEX-FIX-NEXT:    ldr r0, [sp, #8] @ 4-byte Reload
+; CHECK-CORTEX-FIX-NEXT:    ldr r1, [sp, #4] @ 4-byte Reload
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r11, r11, r4, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    ldr r4, [sp, #16] @ 4-byte Reload
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r6, r6, r12, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r5, r10, r5, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r7, r0, r1, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    ldr r1, [sp, #12] @ 4-byte Reload
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r0, lr, r8, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r1, r1, r3, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    ldr r3, [sp, #20] @ 4-byte Reload
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r4, r3, r4, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    ldr r3, [sp] @ 4-byte Reload
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d18[0], r4
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d19[0], r1
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d18[1], r7
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d19[1], r11
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r3, r3, r9, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d16[0], r3
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d17[0], r5
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d16[1], r6
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d17[1], r0
 ; CHECK-CORTEX-FIX-NEXT:    aese.8 q9, q8
 ; CHECK-CORTEX-FIX-NEXT:    aesmc.8 q8, q9
 ; CHECK-CORTEX-FIX-NEXT:    vst1.64 {d16, d17}, [r2]
-; CHECK-CORTEX-FIX-NEXT:    vpop {d8, d9}
-; CHECK-CORTEX-FIX-NEXT:    pop {r4, r5, r6, r7, r8, pc}
+; CHECK-CORTEX-FIX-NEXT:    add sp, sp, #24
+; CHECK-CORTEX-FIX-NEXT:    pop {r4, r5, r6, r7, r8, r9, r10, r11, pc}
   br i1 %0, label %5, label %12
 
 5:
@@ -1727,341 +1602,213 @@
 define arm_aapcs_vfpcc void @aese_setf16_cond_via_val(i1 zeroext %0, half %1, <16 x i8> %2, <16 x i8>* %3) nounwind {
 ; CHECK-FIX-NOSCHED-LABEL: aese_setf16_cond_via_val:
 ; CHECK-FIX-NOSCHED:       @ %bb.0:
-; CHECK-FIX-NOSCHED-NEXT:    .save {r4, r5, r6, r7, r11, lr}
-; CHECK-FIX-NOSCHED-NEXT:    push {r4, r5, r6, r7, r11, lr}
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s9, s0
+; CHECK-FIX-NOSCHED-NEXT:    .save {r4, r5, r6, r7, r8, r9, r10, r11, lr}
+; CHECK-FIX-NOSCHED-NEXT:    push {r4, r5, r6, r7, r8, r9, r10, r11, lr}
+; CHECK-FIX-NOSCHED-NEXT:    .pad #24
+; CHECK-FIX-NOSCHED-NEXT:    sub sp, sp, #24
+; CHECK-FIX-NOSCHED-NEXT:    vmov r12, s0
 ; CHECK-FIX-NOSCHED-NEXT:    cmp r0, #0
-; CHECK-FIX-NOSCHED-NEXT:    beq .LBB37_3
+; CHECK-FIX-NOSCHED-NEXT:    beq .LBB37_2
 ; CHECK-FIX-NOSCHED-NEXT:  @ %bb.1:
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s0, s9
 ; CHECK-FIX-NOSCHED-NEXT:    vld1.64 {d16, d17}, [r1]
-; CHECK-FIX-NOSCHED-NEXT:    vmov r2, s0
-; CHECK-FIX-NOSCHED-NEXT:    vmov lr, r12, d17
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r3, d16[1]
-; CHECK-FIX-NOSCHED-NEXT:    vmov.16 d16[0], r2
-; CHECK-FIX-NOSCHED-NEXT:    vmov s2, lr
-; CHECK-FIX-NOSCHED-NEXT:    lsr lr, lr, #16
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s8, s2
-; CHECK-FIX-NOSCHED-NEXT:    vmov s2, r3
-; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r2, d16[0]
-; CHECK-FIX-NOSCHED-NEXT:    lsr r4, r3, #16
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s12, s2
-; CHECK-FIX-NOSCHED-NEXT:    lsr r3, r12, #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov s0, r12
-; CHECK-FIX-NOSCHED-NEXT:    vmov s10, lr
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s0, s0
-; CHECK-FIX-NOSCHED-NEXT:    vmov s14, r4
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s10, s10
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s14, s14
-; CHECK-FIX-NOSCHED-NEXT:    vmov s2, r2
-; CHECK-FIX-NOSCHED-NEXT:    lsr r2, r2, #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov s3, r2
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s1, s2
-; CHECK-FIX-NOSCHED-NEXT:    vmov s2, r3
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s3, s3
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s2, s2
-; CHECK-FIX-NOSCHED-NEXT:    cmp r0, #0
-; CHECK-FIX-NOSCHED-NEXT:    bne .LBB37_4
-; CHECK-FIX-NOSCHED-NEXT:  .LBB37_2:
-; CHECK-FIX-NOSCHED-NEXT:    vmov r0, r2, d2
-; CHECK-FIX-NOSCHED-NEXT:    vmov r3, r7, d3
-; CHECK-FIX-NOSCHED-NEXT:    vmov s5, r2
-; CHECK-FIX-NOSCHED-NEXT:    lsr r2, r2, #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov s4, r7
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s9, s5
-; CHECK-FIX-NOSCHED-NEXT:    vmov s6, r3
+; CHECK-FIX-NOSCHED-NEXT:    vmov r7, r6, d17
+; CHECK-FIX-NOSCHED-NEXT:    vmov.16 d16[0], r12
+; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r5, d16[0]
+; CHECK-FIX-NOSCHED-NEXT:    uxth r4, r7
+; CHECK-FIX-NOSCHED-NEXT:    uxth r2, r3
 ; CHECK-FIX-NOSCHED-NEXT:    lsr r3, r3, #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov s5, r0
+; CHECK-FIX-NOSCHED-NEXT:    str r4, [sp, #12] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    uxth r4, r6
+; CHECK-FIX-NOSCHED-NEXT:    lsr r6, r6, #16
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp, #8] @ 4-byte Spill
 ; CHECK-FIX-NOSCHED-NEXT:    lsr r7, r7, #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov s7, r3
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s11, s5
-; CHECK-FIX-NOSCHED-NEXT:    vmov s5, r7
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s4, s4
-; CHECK-FIX-NOSCHED-NEXT:    vmov s13, r2
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s6, s6
-; CHECK-FIX-NOSCHED-NEXT:    lsr r0, r0, #16
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s5, s5
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s7, s7
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s13, s13
-; CHECK-FIX-NOSCHED-NEXT:    b .LBB37_5
+; CHECK-FIX-NOSCHED-NEXT:    lsr r3, r5, #16
+; CHECK-FIX-NOSCHED-NEXT:    str r4, [sp, #20] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp, #4] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    uxth r3, r5
+; CHECK-FIX-NOSCHED-NEXT:    str r6, [sp, #16] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    b .LBB37_3
+; CHECK-FIX-NOSCHED-NEXT:  .LBB37_2:
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r3, [r1, #14]
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp, #16] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r3, [r1, #12]
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp, #20] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r3, [r1, #8]
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp, #12] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r2, [r1, #6]
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r3, [r1, #2]
+; CHECK-FIX-NOSCHED-NEXT:    str r2, [sp, #8] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp, #4] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r7, [r1, #10]
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r2, [r1, #4]
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r3, [r1]
 ; CHECK-FIX-NOSCHED-NEXT:  .LBB37_3:
-; CHECK-FIX-NOSCHED-NEXT:    ldrh r3, [r1, #10]
-; CHECK-FIX-NOSCHED-NEXT:    ldrh r12, [r1, #6]
-; CHECK-FIX-NOSCHED-NEXT:    ldrh r5, [r1, #2]
-; CHECK-FIX-NOSCHED-NEXT:    ldrh r7, [r1, #14]
-; CHECK-FIX-NOSCHED-NEXT:    vmov s8, r3
-; CHECK-FIX-NOSCHED-NEXT:    ldrh r2, [r1, #12]
-; CHECK-FIX-NOSCHED-NEXT:    vmov s12, r12
-; CHECK-FIX-NOSCHED-NEXT:    ldrh r4, [r1, #8]
-; CHECK-FIX-NOSCHED-NEXT:    vmov s1, r5
-; CHECK-FIX-NOSCHED-NEXT:    ldrh lr, [r1, #4]
-; CHECK-FIX-NOSCHED-NEXT:    vmov s0, r7
-; CHECK-FIX-NOSCHED-NEXT:    ldrh r6, [r1]
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s2, s0
-; CHECK-FIX-NOSCHED-NEXT:    vmov s0, r2
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s10, s8
-; CHECK-FIX-NOSCHED-NEXT:    vmov s8, r4
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s14, s12
-; CHECK-FIX-NOSCHED-NEXT:    vmov s12, lr
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s3, s1
-; CHECK-FIX-NOSCHED-NEXT:    vmov s1, r6
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s0, s0
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s8, s8
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s12, s12
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s1, s1
 ; CHECK-FIX-NOSCHED-NEXT:    cmp r0, #0
-; CHECK-FIX-NOSCHED-NEXT:    beq .LBB37_2
-; CHECK-FIX-NOSCHED-NEXT:  .LBB37_4:
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s9, s9
-; CHECK-FIX-NOSCHED-NEXT:    vmov r0, r2, d3
-; CHECK-FIX-NOSCHED-NEXT:    vmov r7, s9
-; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r3, d2[1]
-; CHECK-FIX-NOSCHED-NEXT:    vmov.16 d2[0], r7
-; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r7, d2[0]
-; CHECK-FIX-NOSCHED-NEXT:    vmov s5, r3
-; CHECK-FIX-NOSCHED-NEXT:    vmov s4, r2
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s9, s5
-; CHECK-FIX-NOSCHED-NEXT:    vmov s6, r0
-; CHECK-FIX-NOSCHED-NEXT:    lsr r0, r0, #16
-; CHECK-FIX-NOSCHED-NEXT:    lsr r3, r3, #16
-; CHECK-FIX-NOSCHED-NEXT:    lsr r2, r2, #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov s7, r0
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s4, s4
-; CHECK-FIX-NOSCHED-NEXT:    vmov s13, r3
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s6, s6
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s7, s7
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s13, s13
-; CHECK-FIX-NOSCHED-NEXT:    vmov s5, r7
-; CHECK-FIX-NOSCHED-NEXT:    lsr r0, r7, #16
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s11, s5
-; CHECK-FIX-NOSCHED-NEXT:    vmov s5, r2
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s5, s5
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    beq .LBB37_5
+; CHECK-FIX-NOSCHED-NEXT:  @ %bb.4:
+; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r6, d2[1]
+; CHECK-FIX-NOSCHED-NEXT:    mov r3, r2
+; CHECK-FIX-NOSCHED-NEXT:    mov r2, r7
+; CHECK-FIX-NOSCHED-NEXT:    vmov r4, r7, d3
+; CHECK-FIX-NOSCHED-NEXT:    vmov.16 d2[0], r12
+; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r0, d2[0]
+; CHECK-FIX-NOSCHED-NEXT:    uxth r5, r6
+; CHECK-FIX-NOSCHED-NEXT:    lsr r12, r6, #16
+; CHECK-FIX-NOSCHED-NEXT:    uxth r10, r4
+; CHECK-FIX-NOSCHED-NEXT:    uxth r11, r7
+; CHECK-FIX-NOSCHED-NEXT:    lsr r9, r7, #16
+; CHECK-FIX-NOSCHED-NEXT:    mov r7, r2
+; CHECK-FIX-NOSCHED-NEXT:    mov r2, r3
+; CHECK-FIX-NOSCHED-NEXT:    lsr r4, r4, #16
+; CHECK-FIX-NOSCHED-NEXT:    lsr lr, r0, #16
+; CHECK-FIX-NOSCHED-NEXT:    b .LBB37_6
 ; CHECK-FIX-NOSCHED-NEXT:  .LBB37_5:
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s1, s1
-; CHECK-FIX-NOSCHED-NEXT:    vmov s15, r0
-; CHECK-FIX-NOSCHED-NEXT:    vmov r0, s1
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s1, s3
-; CHECK-FIX-NOSCHED-NEXT:    vmov r2, s1
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s15, s15
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s1, s15
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s11, s11
-; CHECK-FIX-NOSCHED-NEXT:    vmov r3, s1
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s12, s12
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s1, s9
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s8, s8
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s6, s6
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s4, s4
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s0, s0
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r0, r2, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov r2, s11
+; CHECK-FIX-NOSCHED-NEXT:    vmov r3, r6, d3
+; CHECK-FIX-NOSCHED-NEXT:    vmov r0, r5, d2
+; CHECK-FIX-NOSCHED-NEXT:    lsr r4, r3, #16
+; CHECK-FIX-NOSCHED-NEXT:    lsr r9, r6, #16
+; CHECK-FIX-NOSCHED-NEXT:    lsr r12, r5, #16
+; CHECK-FIX-NOSCHED-NEXT:    lsr lr, r0, #16
+; CHECK-FIX-NOSCHED-NEXT:    uxth r11, r6
+; CHECK-FIX-NOSCHED-NEXT:    uxth r10, r3
+; CHECK-FIX-NOSCHED-NEXT:    uxth r5, r5
+; CHECK-FIX-NOSCHED-NEXT:  .LBB37_6:
+; CHECK-FIX-NOSCHED-NEXT:    uxth r8, r0
+; CHECK-FIX-NOSCHED-NEXT:    ldr r0, [sp, #4] @ 4-byte Reload
+; CHECK-FIX-NOSCHED-NEXT:    ldr r3, [sp] @ 4-byte Reload
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r3, r0, lsl #16
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d16[0], r0
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r2, r3, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov r2, s12
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s12, s14
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r8, lr, lsl #16
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d18[0], r0
-; CHECK-FIX-NOSCHED-NEXT:    vmov r0, s12
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s12, s13
-; CHECK-FIX-NOSCHED-NEXT:    vmov r3, s12
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r5, r12, lsl #16
+; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d18[1], r0
+; CHECK-FIX-NOSCHED-NEXT:    ldr r0, [sp, #8] @ 4-byte Reload
 ; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r2, r0, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov r2, s1
+; CHECK-FIX-NOSCHED-NEXT:    ldr r2, [sp, #16] @ 4-byte Reload
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d16[1], r0
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r2, r2, r3, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d18[1], r2
-; CHECK-FIX-NOSCHED-NEXT:    vmov r2, s8
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s8, s10
-; CHECK-FIX-NOSCHED-NEXT:    vmov r3, s8
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r2, r3, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov r2, s6
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s6, s7
-; CHECK-FIX-NOSCHED-NEXT:    vmov r3, s6
-; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d17[0], r0
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r2, r3, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov r2, s4
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s4, s5
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r10, r4, lsl #16
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d19[0], r0
-; CHECK-FIX-NOSCHED-NEXT:    vmov r0, s4
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r2, r0, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov r2, s0
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s0, s2
-; CHECK-FIX-NOSCHED-NEXT:    vmov r3, s0
+; CHECK-FIX-NOSCHED-NEXT:    ldr r0, [sp, #12] @ 4-byte Reload
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r0, r7, lsl #16
+; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d17[0], r0
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r11, r9, lsl #16
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d19[1], r0
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r2, r3, lsl #16
+; CHECK-FIX-NOSCHED-NEXT:    ldr r0, [sp, #20] @ 4-byte Reload
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r0, r2, lsl #16
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d17[1], r0
 ; CHECK-FIX-NOSCHED-NEXT:    aese.8 q8, q9
 ; CHECK-FIX-NOSCHED-NEXT:    aesmc.8 q8, q8
 ; CHECK-FIX-NOSCHED-NEXT:    vst1.64 {d16, d17}, [r1]
-; CHECK-FIX-NOSCHED-NEXT:    pop {r4, r5, r6, r7, r11, pc}
+; CHECK-FIX-NOSCHED-NEXT:    add sp, sp, #24
+; CHECK-FIX-NOSCHED-NEXT:    pop {r4, r5, r6, r7, r8, r9, r10, r11, pc}
 ;
 ; CHECK-CORTEX-FIX-LABEL: aese_setf16_cond_via_val:
 ; CHECK-CORTEX-FIX:       @ %bb.0:
-; CHECK-CORTEX-FIX-NEXT:    .save {r4, r5, r6, r7, r11, lr}
-; CHECK-CORTEX-FIX-NEXT:    push {r4, r5, r6, r7, r11, lr}
-; CHECK-CORTEX-FIX-NEXT:    .vsave {d8, d9}
-; CHECK-CORTEX-FIX-NEXT:    vpush {d8, d9}
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s9, s0
+; CHECK-CORTEX-FIX-NEXT:    .save {r4, r5, r6, r7, r8, r9, r10, r11, lr}
+; CHECK-CORTEX-FIX-NEXT:    push {r4, r5, r6, r7, r8, r9, r10, r11, lr}
+; CHECK-CORTEX-FIX-NEXT:    .pad #28
+; CHECK-CORTEX-FIX-NEXT:    sub sp, sp, #28
+; CHECK-CORTEX-FIX-NEXT:    vmov r2, s0
 ; CHECK-CORTEX-FIX-NEXT:    cmp r0, #0
-; CHECK-CORTEX-FIX-NEXT:    beq .LBB37_3
+; CHECK-CORTEX-FIX-NEXT:    beq .LBB37_2
 ; CHECK-CORTEX-FIX-NEXT:  @ %bb.1:
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s0, s9
 ; CHECK-CORTEX-FIX-NEXT:    vld1.64 {d16, d17}, [r1]
-; CHECK-CORTEX-FIX-NEXT:    vmov r2, s0
 ; CHECK-CORTEX-FIX-NEXT:    vmov.32 r3, d16[1]
 ; CHECK-CORTEX-FIX-NEXT:    vmov.16 d16[0], r2
-; CHECK-CORTEX-FIX-NEXT:    vmov r4, r5, d17
-; CHECK-CORTEX-FIX-NEXT:    lsr lr, r3, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s8, r3
-; CHECK-CORTEX-FIX-NEXT:    vmov s11, lr
-; CHECK-CORTEX-FIX-NEXT:    lsr r6, r4, #16
-; CHECK-CORTEX-FIX-NEXT:    lsr r7, r5, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s0, r5
-; CHECK-CORTEX-FIX-NEXT:    vmov s2, r4
-; CHECK-CORTEX-FIX-NEXT:    vmov s12, r7
-; CHECK-CORTEX-FIX-NEXT:    vmov s1, r6
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s3, s8
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 r2, d16[0]
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s10, s0
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s0, s2
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s2, s1
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s14, s12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s1, s11
-; CHECK-CORTEX-FIX-NEXT:    lsr r12, r2, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s13, r2
-; CHECK-CORTEX-FIX-NEXT:    vmov s15, r12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s12, s13
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s8, s15
-; CHECK-CORTEX-FIX-NEXT:    cmp r0, #0
-; CHECK-CORTEX-FIX-NEXT:    bne .LBB37_4
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 r7, d16[0]
+; CHECK-CORTEX-FIX-NEXT:    uxth r6, r7
+; CHECK-CORTEX-FIX-NEXT:    lsr r7, r7, #16
+; CHECK-CORTEX-FIX-NEXT:    str r7, [sp, #20] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    uxth r7, r3
+; CHECK-CORTEX-FIX-NEXT:    lsr r3, r3, #16
+; CHECK-CORTEX-FIX-NEXT:    str r6, [sp, #24] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    str r7, [sp, #8] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #4] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    vmov r3, r7, d17
+; CHECK-CORTEX-FIX-NEXT:    uxth r6, r3
+; CHECK-CORTEX-FIX-NEXT:    lsr r3, r3, #16
+; CHECK-CORTEX-FIX-NEXT:    uxth r11, r7
+; CHECK-CORTEX-FIX-NEXT:    lsr r7, r7, #16
+; CHECK-CORTEX-FIX-NEXT:    str r6, [sp, #16] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    b .LBB37_3
 ; CHECK-CORTEX-FIX-NEXT:  .LBB37_2:
-; CHECK-CORTEX-FIX-NEXT:    vmov r6, r5, d3
-; CHECK-CORTEX-FIX-NEXT:    vmov r0, r2, d2
-; CHECK-CORTEX-FIX-NEXT:    lsr r7, r2, #16
-; CHECK-CORTEX-FIX-NEXT:    lsr r4, r6, #16
-; CHECK-CORTEX-FIX-NEXT:    lsr r3, r5, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s6, r6
-; CHECK-CORTEX-FIX-NEXT:    vmov s4, r5
-; CHECK-CORTEX-FIX-NEXT:    vmov s7, r2
-; CHECK-CORTEX-FIX-NEXT:    vmov s9, r0
-; CHECK-CORTEX-FIX-NEXT:    lsr r12, r0, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s13, r3
-; CHECK-CORTEX-FIX-NEXT:    vmov s15, r4
-; CHECK-CORTEX-FIX-NEXT:    vmov s16, r7
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s5, s4
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s11, s7
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s7, s9
-; CHECK-CORTEX-FIX-NEXT:    vmov s4, r12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s6, s6
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s9, s15
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s13, s13
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s15, s16
-; CHECK-CORTEX-FIX-NEXT:    b .LBB37_5
+; CHECK-CORTEX-FIX-NEXT:    ldrh r3, [r1]
+; CHECK-CORTEX-FIX-NEXT:    ldrh r11, [r1, #12]
+; CHECK-CORTEX-FIX-NEXT:    ldrh r7, [r1, #14]
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #24] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    ldrh r3, [r1, #2]
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #20] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    ldrh r3, [r1, #4]
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #8] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    ldrh r3, [r1, #6]
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #4] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    ldrh r3, [r1, #8]
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #16] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    ldrh r3, [r1, #10]
 ; CHECK-CORTEX-FIX-NEXT:  .LBB37_3:
-; CHECK-CORTEX-FIX-NEXT:    ldrh r12, [r1]
-; CHECK-CORTEX-FIX-NEXT:    ldrh lr, [r1, #2]
-; CHECK-CORTEX-FIX-NEXT:    ldrh r7, [r1, #4]
-; CHECK-CORTEX-FIX-NEXT:    ldrh r6, [r1, #6]
-; CHECK-CORTEX-FIX-NEXT:    ldrh r5, [r1, #8]
-; CHECK-CORTEX-FIX-NEXT:    ldrh r4, [r1, #10]
-; CHECK-CORTEX-FIX-NEXT:    ldrh r2, [r1, #12]
-; CHECK-CORTEX-FIX-NEXT:    ldrh r3, [r1, #14]
-; CHECK-CORTEX-FIX-NEXT:    vmov s1, r6
-; CHECK-CORTEX-FIX-NEXT:    vmov s3, r7
-; CHECK-CORTEX-FIX-NEXT:    vmov s0, r3
-; CHECK-CORTEX-FIX-NEXT:    vmov s2, r2
-; CHECK-CORTEX-FIX-NEXT:    vmov s8, r4
-; CHECK-CORTEX-FIX-NEXT:    vmov s12, r5
-; CHECK-CORTEX-FIX-NEXT:    vmov s11, lr
-; CHECK-CORTEX-FIX-NEXT:    vmov s13, r12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s14, s0
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s10, s2
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s2, s8
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s0, s12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s1, s1
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s3, s3
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s8, s11
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s12, s13
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #12] @ 4-byte Spill
 ; CHECK-CORTEX-FIX-NEXT:    cmp r0, #0
-; CHECK-CORTEX-FIX-NEXT:    beq .LBB37_2
-; CHECK-CORTEX-FIX-NEXT:  .LBB37_4:
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s9, s9
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 r2, d2[1]
-; CHECK-CORTEX-FIX-NEXT:    vmov r0, s9
-; CHECK-CORTEX-FIX-NEXT:    lsr r7, r2, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s16, r7
-; CHECK-CORTEX-FIX-NEXT:    vmov.16 d2[0], r0
-; CHECK-CORTEX-FIX-NEXT:    vmov r6, r5, d3
-; CHECK-CORTEX-FIX-NEXT:    vmov s7, r2
-; CHECK-CORTEX-FIX-NEXT:    lsr r4, r6, #16
-; CHECK-CORTEX-FIX-NEXT:    lsr r3, r5, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s6, r6
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s11, s7
-; CHECK-CORTEX-FIX-NEXT:    vmov s9, r3
-; CHECK-CORTEX-FIX-NEXT:    vmov s15, r4
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 r0, d2[0]
-; CHECK-CORTEX-FIX-NEXT:    vmov s4, r5
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s6, s6
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s13, s9
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s9, s15
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s15, s16
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s5, s4
-; CHECK-CORTEX-FIX-NEXT:    vmov s18, r0
-; CHECK-CORTEX-FIX-NEXT:    lsr r12, r0, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s4, r12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s7, s18
+; CHECK-CORTEX-FIX-NEXT:    beq .LBB37_5
+; CHECK-CORTEX-FIX-NEXT:  @ %bb.4:
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 r3, d2[1]
+; CHECK-CORTEX-FIX-NEXT:    vmov.16 d2[0], r2
+; CHECK-CORTEX-FIX-NEXT:    vmov r4, r6, d3
+; CHECK-CORTEX-FIX-NEXT:    uxth r10, r4
+; CHECK-CORTEX-FIX-NEXT:    lsr r4, r4, #16
+; CHECK-CORTEX-FIX-NEXT:    uxth lr, r6
+; CHECK-CORTEX-FIX-NEXT:    lsr r8, r6, #16
+; CHECK-CORTEX-FIX-NEXT:    uxth r5, r3
+; CHECK-CORTEX-FIX-NEXT:    lsr r12, r3, #16
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 r2, d2[0]
+; CHECK-CORTEX-FIX-NEXT:    uxth r0, r2
+; CHECK-CORTEX-FIX-NEXT:    lsr r9, r2, #16
+; CHECK-CORTEX-FIX-NEXT:    str r0, [sp] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    b .LBB37_6
 ; CHECK-CORTEX-FIX-NEXT:  .LBB37_5:
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s10, s10
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s14, s14
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s3, s3
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s2, s2
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s0, s0
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s8, s8
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s12, s12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s4, s4
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s6, s6
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s9, s9
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s7, s7
-; CHECK-CORTEX-FIX-NEXT:    vmov r0, s10
-; CHECK-CORTEX-FIX-NEXT:    vmov r2, s14
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s10, s1
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s14, s5
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s1, s13
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s5, s11
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s11, s15
-; CHECK-CORTEX-FIX-NEXT:    vmov r5, s2
-; CHECK-CORTEX-FIX-NEXT:    vmov r4, s8
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r12, r0, r2, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r2, s3
-; CHECK-CORTEX-FIX-NEXT:    vmov r3, s10
-; CHECK-CORTEX-FIX-NEXT:    vmov r7, s1
-; CHECK-CORTEX-FIX-NEXT:    vmov r6, s11
-; CHECK-CORTEX-FIX-NEXT:    vmov r0, s9
-; CHECK-CORTEX-FIX-NEXT:    pkhbt lr, r2, r3, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r3, s14
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r3, r3, r7, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r7, s5
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r7, r7, r6, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r6, s0
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s0, s4
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r6, r6, r5, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r5, s12
-; CHECK-CORTEX-FIX-NEXT:    vmov r2, s0
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r5, r5, r4, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r4, s6
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d18[0], r5
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d19[0], r6
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d18[1], lr
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d19[1], r12
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r0, r4, r0, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r4, s7
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r2, r4, r2, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d16[0], r2
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d17[0], r0
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d16[1], r7
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d17[1], r3
+; CHECK-CORTEX-FIX-NEXT:    vmov r2, r3, d2
+; CHECK-CORTEX-FIX-NEXT:    uxth r0, r2
+; CHECK-CORTEX-FIX-NEXT:    lsr r9, r2, #16
+; CHECK-CORTEX-FIX-NEXT:    uxth r5, r3
+; CHECK-CORTEX-FIX-NEXT:    lsr r12, r3, #16
+; CHECK-CORTEX-FIX-NEXT:    str r0, [sp] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    mov r0, r7
+; CHECK-CORTEX-FIX-NEXT:    vmov r6, r7, d3
+; CHECK-CORTEX-FIX-NEXT:    uxth r10, r6
+; CHECK-CORTEX-FIX-NEXT:    lsr r4, r6, #16
+; CHECK-CORTEX-FIX-NEXT:    uxth lr, r7
+; CHECK-CORTEX-FIX-NEXT:    lsr r8, r7, #16
+; CHECK-CORTEX-FIX-NEXT:    mov r7, r0
+; CHECK-CORTEX-FIX-NEXT:  .LBB37_6:
+; CHECK-CORTEX-FIX-NEXT:    ldr r0, [sp, #8] @ 4-byte Reload
+; CHECK-CORTEX-FIX-NEXT:    ldr r2, [sp, #4] @ 4-byte Reload
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r11, r11, r7, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    ldr r3, [sp, #12] @ 4-byte Reload
+; CHECK-CORTEX-FIX-NEXT:    ldr r6, [sp, #20] @ 4-byte Reload
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r5, r5, r12, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r4, r10, r4, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r7, r0, r2, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    ldr r2, [sp, #16] @ 4-byte Reload
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r0, lr, r8, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r2, r2, r3, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    ldr r3, [sp, #24] @ 4-byte Reload
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r3, r3, r6, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    ldr r6, [sp] @ 4-byte Reload
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d18[0], r3
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d19[0], r2
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d18[1], r7
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d19[1], r11
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r6, r6, r9, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d16[0], r6
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d17[0], r4
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d16[1], r5
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d17[1], r0
 ; CHECK-CORTEX-FIX-NEXT:    aese.8 q9, q8
 ; CHECK-CORTEX-FIX-NEXT:    aesmc.8 q8, q9
 ; CHECK-CORTEX-FIX-NEXT:    vst1.64 {d16, d17}, [r1]
-; CHECK-CORTEX-FIX-NEXT:    vpop {d8, d9}
-; CHECK-CORTEX-FIX-NEXT:    pop {r4, r5, r6, r7, r11, pc}
+; CHECK-CORTEX-FIX-NEXT:    add sp, sp, #28
+; CHECK-CORTEX-FIX-NEXT:    pop {r4, r5, r6, r7, r8, r9, r10, r11, pc}
   br i1 %0, label %5, label %11
 
 5:
@@ -2159,10 +1906,7 @@
 ; CHECK-FIX-NEXT:    cmp r0, #0
 ; CHECK-FIX-NEXT:    bxeq lr
 ; CHECK-FIX-NEXT:  .LBB39_1:
-; CHECK-FIX-NEXT:    vcvtb.f32.f16 s0, s0
-; CHECK-FIX-NEXT:    vcvtb.f16.f32 s0, s0
 ; CHECK-FIX-NEXT:    vmov r2, s0
-; CHECK-FIX-NEXT:    uxth r2, r2
 ; CHECK-FIX-NEXT:    vmov.16 d2[0], r2
 ; CHECK-FIX-NEXT:  .LBB39_2: @ =>This Inner Loop Header: Depth=1
 ; CHECK-FIX-NEXT:    vld1.64 {d16, d17}, [r1]
@@ -3821,335 +3565,210 @@
 define arm_aapcs_vfpcc void @aesd_setf16_cond_via_ptr(i1 zeroext %0, half* %1, <16 x i8> %2, <16 x i8>* %3) nounwind {
 ; CHECK-FIX-NOSCHED-LABEL: aesd_setf16_cond_via_ptr:
 ; CHECK-FIX-NOSCHED:       @ %bb.0:
-; CHECK-FIX-NOSCHED-NEXT:    .save {r4, r5, r6, r7, r8, lr}
-; CHECK-FIX-NOSCHED-NEXT:    push {r4, r5, r6, r7, r8, lr}
+; CHECK-FIX-NOSCHED-NEXT:    .save {r4, r5, r6, r7, r8, r9, r10, r11, lr}
+; CHECK-FIX-NOSCHED-NEXT:    push {r4, r5, r6, r7, r8, r9, r10, r11, lr}
+; CHECK-FIX-NOSCHED-NEXT:    .pad #24
+; CHECK-FIX-NOSCHED-NEXT:    sub sp, sp, #24
 ; CHECK-FIX-NOSCHED-NEXT:    cmp r0, #0
 ; CHECK-FIX-NOSCHED-NEXT:    beq .LBB82_3
 ; CHECK-FIX-NOSCHED-NEXT:  @ %bb.1:
 ; CHECK-FIX-NOSCHED-NEXT:    vld1.64 {d16, d17}, [r2]
-; CHECK-FIX-NOSCHED-NEXT:    vorr q9, q8, q8
-; CHECK-FIX-NOSCHED-NEXT:    vmov lr, r12, d17
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r3, d16[1]
-; CHECK-FIX-NOSCHED-NEXT:    vld1.16 {d18[0]}, [r1:16]
-; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r4, d18[0]
-; CHECK-FIX-NOSCHED-NEXT:    vmov s6, lr
-; CHECK-FIX-NOSCHED-NEXT:    lsr r5, r3, #16
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s8, s6
-; CHECK-FIX-NOSCHED-NEXT:    vmov s6, r3
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s12, s6
-; CHECK-FIX-NOSCHED-NEXT:    lsr r3, r12, #16
-; CHECK-FIX-NOSCHED-NEXT:    lsr lr, lr, #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov s4, r12
-; CHECK-FIX-NOSCHED-NEXT:    vmov s10, lr
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s4, s4
-; CHECK-FIX-NOSCHED-NEXT:    vmov s14, r5
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s10, s10
-; CHECK-FIX-NOSCHED-NEXT:    vmov s6, r4
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s14, s14
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s5, s6
-; CHECK-FIX-NOSCHED-NEXT:    vmov s6, r3
-; CHECK-FIX-NOSCHED-NEXT:    lsr r3, r4, #16
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s6, s6
-; CHECK-FIX-NOSCHED-NEXT:    vmov s7, r3
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s7, s7
+; CHECK-FIX-NOSCHED-NEXT:    vmov r7, r6, d17
+; CHECK-FIX-NOSCHED-NEXT:    vld1.16 {d16[0]}, [r1:16]
+; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r5, d16[0]
+; CHECK-FIX-NOSCHED-NEXT:    uxth r4, r3
+; CHECK-FIX-NOSCHED-NEXT:    lsr r3, r3, #16
+; CHECK-FIX-NOSCHED-NEXT:    str r4, [sp, #8] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    uxth r4, r7
+; CHECK-FIX-NOSCHED-NEXT:    str r4, [sp, #12] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    uxth r4, r6
+; CHECK-FIX-NOSCHED-NEXT:    lsr r6, r6, #16
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp, #4] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    str r4, [sp, #20] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    lsr r7, r7, #16
+; CHECK-FIX-NOSCHED-NEXT:    lsr r3, r5, #16
+; CHECK-FIX-NOSCHED-NEXT:    str r6, [sp, #16] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    uxth r10, r5
 ; CHECK-FIX-NOSCHED-NEXT:    cmp r0, #0
 ; CHECK-FIX-NOSCHED-NEXT:    bne .LBB82_4
 ; CHECK-FIX-NOSCHED-NEXT:  .LBB82_2:
-; CHECK-FIX-NOSCHED-NEXT:    vmov r0, r1, d0
-; CHECK-FIX-NOSCHED-NEXT:    vmov r3, r7, d1
-; CHECK-FIX-NOSCHED-NEXT:    vmov s1, r1
-; CHECK-FIX-NOSCHED-NEXT:    lsr r1, r1, #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov s0, r7
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s9, s1
-; CHECK-FIX-NOSCHED-NEXT:    vmov s2, r3
-; CHECK-FIX-NOSCHED-NEXT:    lsr r3, r3, #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov s1, r0
-; CHECK-FIX-NOSCHED-NEXT:    lsr r7, r7, #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov s3, r3
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s11, s1
-; CHECK-FIX-NOSCHED-NEXT:    vmov s1, r7
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s0, s0
-; CHECK-FIX-NOSCHED-NEXT:    vmov s13, r1
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s2, s2
-; CHECK-FIX-NOSCHED-NEXT:    lsr r0, r0, #16
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s1, s1
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s3, s3
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s13, s13
+; CHECK-FIX-NOSCHED-NEXT:    vmov r4, r6, d1
+; CHECK-FIX-NOSCHED-NEXT:    vmov r0, r3, d0
+; CHECK-FIX-NOSCHED-NEXT:    lsr r5, r4, #16
+; CHECK-FIX-NOSCHED-NEXT:    lsr r1, r6, #16
+; CHECK-FIX-NOSCHED-NEXT:    uxth r11, r6
+; CHECK-FIX-NOSCHED-NEXT:    lsr lr, r0, #16
+; CHECK-FIX-NOSCHED-NEXT:    lsr r12, r3, #16
+; CHECK-FIX-NOSCHED-NEXT:    uxth r9, r4
+; CHECK-FIX-NOSCHED-NEXT:    uxth r6, r3
 ; CHECK-FIX-NOSCHED-NEXT:    b .LBB82_5
 ; CHECK-FIX-NOSCHED-NEXT:  .LBB82_3:
-; CHECK-FIX-NOSCHED-NEXT:    ldrh r5, [r2, #10]
-; CHECK-FIX-NOSCHED-NEXT:    ldrh r12, [r2, #6]
-; CHECK-FIX-NOSCHED-NEXT:    ldrh r6, [r2, #2]
-; CHECK-FIX-NOSCHED-NEXT:    ldrh r7, [r2, #14]
-; CHECK-FIX-NOSCHED-NEXT:    vmov s8, r5
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r3, [r2, #14]
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp, #16] @ 4-byte Spill
 ; CHECK-FIX-NOSCHED-NEXT:    ldrh r3, [r2, #12]
-; CHECK-FIX-NOSCHED-NEXT:    vmov s12, r12
-; CHECK-FIX-NOSCHED-NEXT:    ldrh r4, [r2, #8]
-; CHECK-FIX-NOSCHED-NEXT:    vmov s5, r6
-; CHECK-FIX-NOSCHED-NEXT:    ldrh lr, [r2, #4]
-; CHECK-FIX-NOSCHED-NEXT:    vmov s4, r7
-; CHECK-FIX-NOSCHED-NEXT:    ldrh r8, [r2]
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s6, s4
-; CHECK-FIX-NOSCHED-NEXT:    vmov s4, r3
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s10, s8
-; CHECK-FIX-NOSCHED-NEXT:    vmov s8, r4
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s14, s12
-; CHECK-FIX-NOSCHED-NEXT:    vmov s12, lr
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s7, s5
-; CHECK-FIX-NOSCHED-NEXT:    vmov s5, r8
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s4, s4
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s8, s8
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s12, s12
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s5, s5
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp, #20] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r3, [r2, #8]
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp, #12] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r3, [r2, #6]
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r7, [r2, #10]
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r10, [r2]
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp, #4] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r3, [r2, #4]
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp, #8] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r3, [r2, #2]
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp] @ 4-byte Spill
 ; CHECK-FIX-NOSCHED-NEXT:    cmp r0, #0
 ; CHECK-FIX-NOSCHED-NEXT:    beq .LBB82_2
 ; CHECK-FIX-NOSCHED-NEXT:  .LBB82_4:
-; CHECK-FIX-NOSCHED-NEXT:    vorr q8, q0, q0
-; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r3, d0[1]
-; CHECK-FIX-NOSCHED-NEXT:    vld1.16 {d16[0]}, [r1:16]
-; CHECK-FIX-NOSCHED-NEXT:    vmov r0, r1, d1
-; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r7, d16[0]
-; CHECK-FIX-NOSCHED-NEXT:    vmov s1, r3
-; CHECK-FIX-NOSCHED-NEXT:    lsr r3, r3, #16
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s9, s1
-; CHECK-FIX-NOSCHED-NEXT:    vmov s13, r3
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s13, s13
-; CHECK-FIX-NOSCHED-NEXT:    vmov s0, r1
-; CHECK-FIX-NOSCHED-NEXT:    lsr r1, r1, #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov s2, r0
-; CHECK-FIX-NOSCHED-NEXT:    lsr r0, r0, #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov s3, r0
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s0, s0
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s2, s2
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s3, s3
-; CHECK-FIX-NOSCHED-NEXT:    vmov s1, r7
-; CHECK-FIX-NOSCHED-NEXT:    lsr r0, r7, #16
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s11, s1
-; CHECK-FIX-NOSCHED-NEXT:    vmov s1, r1
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s1, s1
+; CHECK-FIX-NOSCHED-NEXT:    vmov r5, r3, d1
+; CHECK-FIX-NOSCHED-NEXT:    mov r4, r7
+; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r7, d0[1]
+; CHECK-FIX-NOSCHED-NEXT:    vld1.16 {d0[0]}, [r1:16]
+; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r0, d0[0]
+; CHECK-FIX-NOSCHED-NEXT:    uxth r9, r5
+; CHECK-FIX-NOSCHED-NEXT:    uxth r11, r3
+; CHECK-FIX-NOSCHED-NEXT:    uxth r6, r7
+; CHECK-FIX-NOSCHED-NEXT:    lsr r12, r7, #16
+; CHECK-FIX-NOSCHED-NEXT:    lsr r1, r3, #16
+; CHECK-FIX-NOSCHED-NEXT:    lsr r5, r5, #16
+; CHECK-FIX-NOSCHED-NEXT:    mov r7, r4
+; CHECK-FIX-NOSCHED-NEXT:    lsr lr, r0, #16
 ; CHECK-FIX-NOSCHED-NEXT:  .LBB82_5:
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s5, s5
-; CHECK-FIX-NOSCHED-NEXT:    vmov s15, r0
-; CHECK-FIX-NOSCHED-NEXT:    vmov r0, s5
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s5, s7
-; CHECK-FIX-NOSCHED-NEXT:    vmov r1, s5
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s15, s15
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s5, s15
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s11, s11
-; CHECK-FIX-NOSCHED-NEXT:    vmov r3, s5
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s12, s12
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s5, s9
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s8, s8
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s2, s2
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s0, s0
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r0, r1, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov r1, s11
+; CHECK-FIX-NOSCHED-NEXT:    uxth r8, r0
+; CHECK-FIX-NOSCHED-NEXT:    ldr r0, [sp] @ 4-byte Reload
+; CHECK-FIX-NOSCHED-NEXT:    ldr r3, [sp, #4] @ 4-byte Reload
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r10, r0, lsl #16
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d16[0], r0
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r1, r3, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov r1, s12
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s12, s14
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r8, lr, lsl #16
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d18[0], r0
-; CHECK-FIX-NOSCHED-NEXT:    vmov r0, s12
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s12, s13
-; CHECK-FIX-NOSCHED-NEXT:    vmov r3, s12
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r1, r0, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov r1, s5
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r6, r12, lsl #16
+; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d18[1], r0
+; CHECK-FIX-NOSCHED-NEXT:    ldr r0, [sp, #8] @ 4-byte Reload
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r0, r3, lsl #16
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d16[1], r0
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r1, r1, r3, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d18[1], r1
-; CHECK-FIX-NOSCHED-NEXT:    vmov r1, s8
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s8, s10
-; CHECK-FIX-NOSCHED-NEXT:    vmov r3, s8
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r1, r3, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov r1, s2
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s2, s3
-; CHECK-FIX-NOSCHED-NEXT:    vmov r3, s2
-; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d17[0], r0
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s2, s4
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r1, r3, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov r1, s0
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s0, s1
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r9, r5, lsl #16
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d19[0], r0
-; CHECK-FIX-NOSCHED-NEXT:    vmov r0, s0
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s0, s6
-; CHECK-FIX-NOSCHED-NEXT:    vmov r3, s0
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r1, r0, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov r1, s2
+; CHECK-FIX-NOSCHED-NEXT:    ldr r0, [sp, #12] @ 4-byte Reload
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r0, r7, lsl #16
+; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d17[0], r0
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r11, r1, lsl #16
+; CHECK-FIX-NOSCHED-NEXT:    ldr r1, [sp, #16] @ 4-byte Reload
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d19[1], r0
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r1, r3, lsl #16
+; CHECK-FIX-NOSCHED-NEXT:    ldr r0, [sp, #20] @ 4-byte Reload
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r0, r1, lsl #16
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d17[1], r0
 ; CHECK-FIX-NOSCHED-NEXT:    aesd.8 q8, q9
 ; CHECK-FIX-NOSCHED-NEXT:    aesimc.8 q8, q8
 ; CHECK-FIX-NOSCHED-NEXT:    vst1.64 {d16, d17}, [r2]
-; CHECK-FIX-NOSCHED-NEXT:    pop {r4, r5, r6, r7, r8, pc}
+; CHECK-FIX-NOSCHED-NEXT:    add sp, sp, #24
+; CHECK-FIX-NOSCHED-NEXT:    pop {r4, r5, r6, r7, r8, r9, r10, r11, pc}
 ;
 ; CHECK-CORTEX-FIX-LABEL: aesd_setf16_cond_via_ptr:
 ; CHECK-CORTEX-FIX:       @ %bb.0:
-; CHECK-CORTEX-FIX-NEXT:    .save {r4, r5, r6, r7, r8, lr}
-; CHECK-CORTEX-FIX-NEXT:    push {r4, r5, r6, r7, r8, lr}
-; CHECK-CORTEX-FIX-NEXT:    .vsave {d8, d9}
-; CHECK-CORTEX-FIX-NEXT:    vpush {d8, d9}
+; CHECK-CORTEX-FIX-NEXT:    .save {r4, r5, r6, r7, r8, r9, r10, r11, lr}
+; CHECK-CORTEX-FIX-NEXT:    push {r4, r5, r6, r7, r8, r9, r10, r11, lr}
+; CHECK-CORTEX-FIX-NEXT:    .pad #24
+; CHECK-CORTEX-FIX-NEXT:    sub sp, sp, #24
 ; CHECK-CORTEX-FIX-NEXT:    cmp r0, #0
 ; CHECK-CORTEX-FIX-NEXT:    beq .LBB82_3
 ; CHECK-CORTEX-FIX-NEXT:  @ %bb.1:
 ; CHECK-CORTEX-FIX-NEXT:    vld1.64 {d16, d17}, [r2]
 ; CHECK-CORTEX-FIX-NEXT:    vorr q9, q8, q8
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 r3, d16[1]
-; CHECK-CORTEX-FIX-NEXT:    vmov r5, r6, d17
 ; CHECK-CORTEX-FIX-NEXT:    vld1.16 {d18[0]}, [r1:16]
-; CHECK-CORTEX-FIX-NEXT:    lsr r7, r5, #16
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 r3, d18[0]
+; CHECK-CORTEX-FIX-NEXT:    uxth r7, r3
+; CHECK-CORTEX-FIX-NEXT:    lsr r3, r3, #16
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #16] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 r3, d16[1]
+; CHECK-CORTEX-FIX-NEXT:    str r7, [sp, #20] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    uxth r7, r3
+; CHECK-CORTEX-FIX-NEXT:    lsr r3, r3, #16
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #4] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    vmov r3, r6, d17
+; CHECK-CORTEX-FIX-NEXT:    str r7, [sp, #8] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    uxth r7, r3
+; CHECK-CORTEX-FIX-NEXT:    lsr r3, r3, #16
+; CHECK-CORTEX-FIX-NEXT:    uxth r11, r6
 ; CHECK-CORTEX-FIX-NEXT:    lsr r4, r6, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s4, r6
-; CHECK-CORTEX-FIX-NEXT:    vmov s6, r5
-; CHECK-CORTEX-FIX-NEXT:    lsr r8, r3, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s8, r3
-; CHECK-CORTEX-FIX-NEXT:    vmov s12, r4
-; CHECK-CORTEX-FIX-NEXT:    vmov s5, r7
-; CHECK-CORTEX-FIX-NEXT:    vmov s9, r8
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s10, s4
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s4, s6
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 lr, d18[0]
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s6, s5
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s7, s8
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s14, s12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s5, s9
-; CHECK-CORTEX-FIX-NEXT:    lsr r12, lr, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s11, lr
-; CHECK-CORTEX-FIX-NEXT:    vmov s13, r12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s12, s11
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s8, s13
+; CHECK-CORTEX-FIX-NEXT:    str r7, [sp, #12] @ 4-byte Spill
 ; CHECK-CORTEX-FIX-NEXT:    cmp r0, #0
 ; CHECK-CORTEX-FIX-NEXT:    bne .LBB82_4
 ; CHECK-CORTEX-FIX-NEXT:  .LBB82_2:
-; CHECK-CORTEX-FIX-NEXT:    vmov r6, r5, d1
-; CHECK-CORTEX-FIX-NEXT:    vmov r0, r1, d0
-; CHECK-CORTEX-FIX-NEXT:    lsr r7, r1, #16
-; CHECK-CORTEX-FIX-NEXT:    lsr r4, r6, #16
-; CHECK-CORTEX-FIX-NEXT:    lsr r3, r5, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s2, r6
-; CHECK-CORTEX-FIX-NEXT:    vmov s0, r5
-; CHECK-CORTEX-FIX-NEXT:    vmov s3, r1
-; CHECK-CORTEX-FIX-NEXT:    vmov s9, r0
-; CHECK-CORTEX-FIX-NEXT:    lsr r12, r0, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s13, r3
-; CHECK-CORTEX-FIX-NEXT:    vmov s15, r4
-; CHECK-CORTEX-FIX-NEXT:    vmov s16, r7
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s1, s0
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s11, s3
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s3, s9
-; CHECK-CORTEX-FIX-NEXT:    vmov s0, r12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s2, s2
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s9, s15
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s13, s13
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s15, s16
+; CHECK-CORTEX-FIX-NEXT:    vmov r1, r7, d0
+; CHECK-CORTEX-FIX-NEXT:    uxth r0, r1
+; CHECK-CORTEX-FIX-NEXT:    uxth r6, r7
+; CHECK-CORTEX-FIX-NEXT:    lsr r12, r7, #16
+; CHECK-CORTEX-FIX-NEXT:    lsr r9, r1, #16
+; CHECK-CORTEX-FIX-NEXT:    str r0, [sp] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    mov r0, r3
+; CHECK-CORTEX-FIX-NEXT:    vmov r7, r3, d1
+; CHECK-CORTEX-FIX-NEXT:    uxth r10, r7
+; CHECK-CORTEX-FIX-NEXT:    lsr r5, r7, #16
+; CHECK-CORTEX-FIX-NEXT:    uxth lr, r3
+; CHECK-CORTEX-FIX-NEXT:    lsr r8, r3, #16
+; CHECK-CORTEX-FIX-NEXT:    mov r3, r0
 ; CHECK-CORTEX-FIX-NEXT:    b .LBB82_5
 ; CHECK-CORTEX-FIX-NEXT:  .LBB82_3:
-; CHECK-CORTEX-FIX-NEXT:    ldrh r12, [r2]
-; CHECK-CORTEX-FIX-NEXT:    ldrh lr, [r2, #2]
-; CHECK-CORTEX-FIX-NEXT:    ldrh r8, [r2, #4]
-; CHECK-CORTEX-FIX-NEXT:    ldrh r5, [r2, #6]
-; CHECK-CORTEX-FIX-NEXT:    ldrh r4, [r2, #8]
+; CHECK-CORTEX-FIX-NEXT:    ldrh r3, [r2]
+; CHECK-CORTEX-FIX-NEXT:    ldrh r11, [r2, #12]
+; CHECK-CORTEX-FIX-NEXT:    ldrh r4, [r2, #14]
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #20] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    ldrh r3, [r2, #2]
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #16] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    ldrh r3, [r2, #4]
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #8] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    ldrh r3, [r2, #6]
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #4] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    ldrh r3, [r2, #8]
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #12] @ 4-byte Spill
 ; CHECK-CORTEX-FIX-NEXT:    ldrh r3, [r2, #10]
-; CHECK-CORTEX-FIX-NEXT:    ldrh r7, [r2, #12]
-; CHECK-CORTEX-FIX-NEXT:    ldrh r6, [r2, #14]
-; CHECK-CORTEX-FIX-NEXT:    vmov s5, r5
-; CHECK-CORTEX-FIX-NEXT:    vmov s7, r8
-; CHECK-CORTEX-FIX-NEXT:    vmov s4, r6
-; CHECK-CORTEX-FIX-NEXT:    vmov s6, r7
-; CHECK-CORTEX-FIX-NEXT:    vmov s8, r3
-; CHECK-CORTEX-FIX-NEXT:    vmov s12, r4
-; CHECK-CORTEX-FIX-NEXT:    vmov s9, lr
-; CHECK-CORTEX-FIX-NEXT:    vmov s11, r12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s14, s4
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s10, s6
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s6, s8
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s4, s12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s5, s5
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s7, s7
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s8, s9
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s12, s11
 ; CHECK-CORTEX-FIX-NEXT:    cmp r0, #0
 ; CHECK-CORTEX-FIX-NEXT:    beq .LBB82_2
 ; CHECK-CORTEX-FIX-NEXT:  .LBB82_4:
 ; CHECK-CORTEX-FIX-NEXT:    vorr q8, q0, q0
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 r3, d0[1]
-; CHECK-CORTEX-FIX-NEXT:    vmov r6, r5, d1
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 r5, d0[1]
 ; CHECK-CORTEX-FIX-NEXT:    vld1.16 {d16[0]}, [r1:16]
-; CHECK-CORTEX-FIX-NEXT:    lsr r4, r6, #16
-; CHECK-CORTEX-FIX-NEXT:    lsr r1, r5, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s2, r6
-; CHECK-CORTEX-FIX-NEXT:    vmov s0, r5
-; CHECK-CORTEX-FIX-NEXT:    lsr r7, r3, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s3, r3
-; CHECK-CORTEX-FIX-NEXT:    vmov s9, r1
-; CHECK-CORTEX-FIX-NEXT:    vmov s15, r4
-; CHECK-CORTEX-FIX-NEXT:    vmov s16, r7
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s1, s0
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s2, s2
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 r0, d16[0]
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s13, s9
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s9, s15
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s11, s3
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s15, s16
-; CHECK-CORTEX-FIX-NEXT:    vmov s18, r0
-; CHECK-CORTEX-FIX-NEXT:    lsr r12, r0, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s0, r12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s3, s18
+; CHECK-CORTEX-FIX-NEXT:    uxth r6, r5
+; CHECK-CORTEX-FIX-NEXT:    lsr r12, r5, #16
+; CHECK-CORTEX-FIX-NEXT:    vmov r5, r7, d1
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 r1, d16[0]
+; CHECK-CORTEX-FIX-NEXT:    uxth r10, r5
+; CHECK-CORTEX-FIX-NEXT:    lsr r5, r5, #16
+; CHECK-CORTEX-FIX-NEXT:    uxth lr, r7
+; CHECK-CORTEX-FIX-NEXT:    lsr r8, r7, #16
+; CHECK-CORTEX-FIX-NEXT:    uxth r0, r1
+; CHECK-CORTEX-FIX-NEXT:    lsr r9, r1, #16
+; CHECK-CORTEX-FIX-NEXT:    str r0, [sp] @ 4-byte Spill
 ; CHECK-CORTEX-FIX-NEXT:  .LBB82_5:
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s10, s10
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s14, s14
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s7, s7
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s6, s6
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s4, s4
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s8, s8
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s12, s12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s0, s0
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s2, s2
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s9, s9
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s3, s3
-; CHECK-CORTEX-FIX-NEXT:    vmov r0, s10
-; CHECK-CORTEX-FIX-NEXT:    vmov r1, s14
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s10, s5
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s14, s1
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s1, s13
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s5, s11
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s11, s15
-; CHECK-CORTEX-FIX-NEXT:    vmov r5, s6
-; CHECK-CORTEX-FIX-NEXT:    vmov r4, s8
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s0, s0
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r12, r0, r1, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r1, s7
-; CHECK-CORTEX-FIX-NEXT:    vmov r3, s10
-; CHECK-CORTEX-FIX-NEXT:    vmov r7, s1
-; CHECK-CORTEX-FIX-NEXT:    vmov r6, s11
-; CHECK-CORTEX-FIX-NEXT:    vmov r0, s9
-; CHECK-CORTEX-FIX-NEXT:    pkhbt lr, r1, r3, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r3, s14
-; CHECK-CORTEX-FIX-NEXT:    vmov r1, s0
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r3, r3, r7, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r7, s5
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r7, r7, r6, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r6, s4
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r6, r6, r5, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r5, s12
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r5, r5, r4, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r4, s2
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d18[0], r5
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d19[0], r6
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d18[1], lr
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d19[1], r12
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r0, r4, r0, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r4, s3
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r1, r4, r1, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d16[0], r1
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d17[0], r0
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d16[1], r7
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d17[1], r3
+; CHECK-CORTEX-FIX-NEXT:    ldr r0, [sp, #8] @ 4-byte Reload
+; CHECK-CORTEX-FIX-NEXT:    ldr r1, [sp, #4] @ 4-byte Reload
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r11, r11, r4, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    ldr r4, [sp, #16] @ 4-byte Reload
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r6, r6, r12, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r5, r10, r5, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r7, r0, r1, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    ldr r1, [sp, #12] @ 4-byte Reload
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r0, lr, r8, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r1, r1, r3, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    ldr r3, [sp, #20] @ 4-byte Reload
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r4, r3, r4, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    ldr r3, [sp] @ 4-byte Reload
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d18[0], r4
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d19[0], r1
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d18[1], r7
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d19[1], r11
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r3, r3, r9, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d16[0], r3
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d17[0], r5
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d16[1], r6
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d17[1], r0
 ; CHECK-CORTEX-FIX-NEXT:    aesd.8 q9, q8
 ; CHECK-CORTEX-FIX-NEXT:    aesimc.8 q8, q9
 ; CHECK-CORTEX-FIX-NEXT:    vst1.64 {d16, d17}, [r2]
-; CHECK-CORTEX-FIX-NEXT:    vpop {d8, d9}
-; CHECK-CORTEX-FIX-NEXT:    pop {r4, r5, r6, r7, r8, pc}
+; CHECK-CORTEX-FIX-NEXT:    add sp, sp, #24
+; CHECK-CORTEX-FIX-NEXT:    pop {r4, r5, r6, r7, r8, r9, r10, r11, pc}
   br i1 %0, label %5, label %12
 
 5:
@@ -4195,341 +3814,213 @@
 define arm_aapcs_vfpcc void @aesd_setf16_cond_via_val(i1 zeroext %0, half %1, <16 x i8> %2, <16 x i8>* %3) nounwind {
 ; CHECK-FIX-NOSCHED-LABEL: aesd_setf16_cond_via_val:
 ; CHECK-FIX-NOSCHED:       @ %bb.0:
-; CHECK-FIX-NOSCHED-NEXT:    .save {r4, r5, r6, r7, r11, lr}
-; CHECK-FIX-NOSCHED-NEXT:    push {r4, r5, r6, r7, r11, lr}
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s9, s0
+; CHECK-FIX-NOSCHED-NEXT:    .save {r4, r5, r6, r7, r8, r9, r10, r11, lr}
+; CHECK-FIX-NOSCHED-NEXT:    push {r4, r5, r6, r7, r8, r9, r10, r11, lr}
+; CHECK-FIX-NOSCHED-NEXT:    .pad #24
+; CHECK-FIX-NOSCHED-NEXT:    sub sp, sp, #24
+; CHECK-FIX-NOSCHED-NEXT:    vmov r12, s0
 ; CHECK-FIX-NOSCHED-NEXT:    cmp r0, #0
-; CHECK-FIX-NOSCHED-NEXT:    beq .LBB83_3
+; CHECK-FIX-NOSCHED-NEXT:    beq .LBB83_2
 ; CHECK-FIX-NOSCHED-NEXT:  @ %bb.1:
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s0, s9
 ; CHECK-FIX-NOSCHED-NEXT:    vld1.64 {d16, d17}, [r1]
-; CHECK-FIX-NOSCHED-NEXT:    vmov r2, s0
-; CHECK-FIX-NOSCHED-NEXT:    vmov lr, r12, d17
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r3, d16[1]
-; CHECK-FIX-NOSCHED-NEXT:    vmov.16 d16[0], r2
-; CHECK-FIX-NOSCHED-NEXT:    vmov s2, lr
-; CHECK-FIX-NOSCHED-NEXT:    lsr lr, lr, #16
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s8, s2
-; CHECK-FIX-NOSCHED-NEXT:    vmov s2, r3
-; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r2, d16[0]
-; CHECK-FIX-NOSCHED-NEXT:    lsr r4, r3, #16
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s12, s2
-; CHECK-FIX-NOSCHED-NEXT:    lsr r3, r12, #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov s0, r12
-; CHECK-FIX-NOSCHED-NEXT:    vmov s10, lr
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s0, s0
-; CHECK-FIX-NOSCHED-NEXT:    vmov s14, r4
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s10, s10
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s14, s14
-; CHECK-FIX-NOSCHED-NEXT:    vmov s2, r2
-; CHECK-FIX-NOSCHED-NEXT:    lsr r2, r2, #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov s3, r2
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s1, s2
-; CHECK-FIX-NOSCHED-NEXT:    vmov s2, r3
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s3, s3
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s2, s2
-; CHECK-FIX-NOSCHED-NEXT:    cmp r0, #0
-; CHECK-FIX-NOSCHED-NEXT:    bne .LBB83_4
-; CHECK-FIX-NOSCHED-NEXT:  .LBB83_2:
-; CHECK-FIX-NOSCHED-NEXT:    vmov r0, r2, d2
-; CHECK-FIX-NOSCHED-NEXT:    vmov r3, r7, d3
-; CHECK-FIX-NOSCHED-NEXT:    vmov s5, r2
-; CHECK-FIX-NOSCHED-NEXT:    lsr r2, r2, #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov s4, r7
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s9, s5
-; CHECK-FIX-NOSCHED-NEXT:    vmov s6, r3
+; CHECK-FIX-NOSCHED-NEXT:    vmov r7, r6, d17
+; CHECK-FIX-NOSCHED-NEXT:    vmov.16 d16[0], r12
+; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r5, d16[0]
+; CHECK-FIX-NOSCHED-NEXT:    uxth r4, r7
+; CHECK-FIX-NOSCHED-NEXT:    uxth r2, r3
 ; CHECK-FIX-NOSCHED-NEXT:    lsr r3, r3, #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov s5, r0
+; CHECK-FIX-NOSCHED-NEXT:    str r4, [sp, #12] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    uxth r4, r6
+; CHECK-FIX-NOSCHED-NEXT:    lsr r6, r6, #16
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp, #8] @ 4-byte Spill
 ; CHECK-FIX-NOSCHED-NEXT:    lsr r7, r7, #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov s7, r3
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s11, s5
-; CHECK-FIX-NOSCHED-NEXT:    vmov s5, r7
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s4, s4
-; CHECK-FIX-NOSCHED-NEXT:    vmov s13, r2
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s6, s6
-; CHECK-FIX-NOSCHED-NEXT:    lsr r0, r0, #16
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s5, s5
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s7, s7
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s13, s13
-; CHECK-FIX-NOSCHED-NEXT:    b .LBB83_5
+; CHECK-FIX-NOSCHED-NEXT:    lsr r3, r5, #16
+; CHECK-FIX-NOSCHED-NEXT:    str r4, [sp, #20] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp, #4] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    uxth r3, r5
+; CHECK-FIX-NOSCHED-NEXT:    str r6, [sp, #16] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    b .LBB83_3
+; CHECK-FIX-NOSCHED-NEXT:  .LBB83_2:
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r3, [r1, #14]
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp, #16] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r3, [r1, #12]
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp, #20] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r3, [r1, #8]
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp, #12] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r2, [r1, #6]
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r3, [r1, #2]
+; CHECK-FIX-NOSCHED-NEXT:    str r2, [sp, #8] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp, #4] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r7, [r1, #10]
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r2, [r1, #4]
+; CHECK-FIX-NOSCHED-NEXT:    ldrh r3, [r1]
 ; CHECK-FIX-NOSCHED-NEXT:  .LBB83_3:
-; CHECK-FIX-NOSCHED-NEXT:    ldrh r3, [r1, #10]
-; CHECK-FIX-NOSCHED-NEXT:    ldrh r12, [r1, #6]
-; CHECK-FIX-NOSCHED-NEXT:    ldrh r5, [r1, #2]
-; CHECK-FIX-NOSCHED-NEXT:    ldrh r7, [r1, #14]
-; CHECK-FIX-NOSCHED-NEXT:    vmov s8, r3
-; CHECK-FIX-NOSCHED-NEXT:    ldrh r2, [r1, #12]
-; CHECK-FIX-NOSCHED-NEXT:    vmov s12, r12
-; CHECK-FIX-NOSCHED-NEXT:    ldrh r4, [r1, #8]
-; CHECK-FIX-NOSCHED-NEXT:    vmov s1, r5
-; CHECK-FIX-NOSCHED-NEXT:    ldrh lr, [r1, #4]
-; CHECK-FIX-NOSCHED-NEXT:    vmov s0, r7
-; CHECK-FIX-NOSCHED-NEXT:    ldrh r6, [r1]
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s2, s0
-; CHECK-FIX-NOSCHED-NEXT:    vmov s0, r2
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s10, s8
-; CHECK-FIX-NOSCHED-NEXT:    vmov s8, r4
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s14, s12
-; CHECK-FIX-NOSCHED-NEXT:    vmov s12, lr
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s3, s1
-; CHECK-FIX-NOSCHED-NEXT:    vmov s1, r6
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s0, s0
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s8, s8
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s12, s12
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s1, s1
 ; CHECK-FIX-NOSCHED-NEXT:    cmp r0, #0
-; CHECK-FIX-NOSCHED-NEXT:    beq .LBB83_2
-; CHECK-FIX-NOSCHED-NEXT:  .LBB83_4:
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s9, s9
-; CHECK-FIX-NOSCHED-NEXT:    vmov r0, r2, d3
-; CHECK-FIX-NOSCHED-NEXT:    vmov r7, s9
-; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r3, d2[1]
-; CHECK-FIX-NOSCHED-NEXT:    vmov.16 d2[0], r7
-; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r7, d2[0]
-; CHECK-FIX-NOSCHED-NEXT:    vmov s5, r3
-; CHECK-FIX-NOSCHED-NEXT:    vmov s4, r2
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s9, s5
-; CHECK-FIX-NOSCHED-NEXT:    vmov s6, r0
-; CHECK-FIX-NOSCHED-NEXT:    lsr r0, r0, #16
-; CHECK-FIX-NOSCHED-NEXT:    lsr r3, r3, #16
-; CHECK-FIX-NOSCHED-NEXT:    lsr r2, r2, #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov s7, r0
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s4, s4
-; CHECK-FIX-NOSCHED-NEXT:    vmov s13, r3
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s6, s6
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s7, s7
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s13, s13
-; CHECK-FIX-NOSCHED-NEXT:    vmov s5, r7
-; CHECK-FIX-NOSCHED-NEXT:    lsr r0, r7, #16
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s11, s5
-; CHECK-FIX-NOSCHED-NEXT:    vmov s5, r2
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s5, s5
+; CHECK-FIX-NOSCHED-NEXT:    str r3, [sp] @ 4-byte Spill
+; CHECK-FIX-NOSCHED-NEXT:    beq .LBB83_5
+; CHECK-FIX-NOSCHED-NEXT:  @ %bb.4:
+; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r6, d2[1]
+; CHECK-FIX-NOSCHED-NEXT:    mov r3, r2
+; CHECK-FIX-NOSCHED-NEXT:    mov r2, r7
+; CHECK-FIX-NOSCHED-NEXT:    vmov r4, r7, d3
+; CHECK-FIX-NOSCHED-NEXT:    vmov.16 d2[0], r12
+; CHECK-FIX-NOSCHED-NEXT:    vmov.32 r0, d2[0]
+; CHECK-FIX-NOSCHED-NEXT:    uxth r5, r6
+; CHECK-FIX-NOSCHED-NEXT:    lsr r12, r6, #16
+; CHECK-FIX-NOSCHED-NEXT:    uxth r10, r4
+; CHECK-FIX-NOSCHED-NEXT:    uxth r11, r7
+; CHECK-FIX-NOSCHED-NEXT:    lsr r9, r7, #16
+; CHECK-FIX-NOSCHED-NEXT:    mov r7, r2
+; CHECK-FIX-NOSCHED-NEXT:    mov r2, r3
+; CHECK-FIX-NOSCHED-NEXT:    lsr r4, r4, #16
+; CHECK-FIX-NOSCHED-NEXT:    lsr lr, r0, #16
+; CHECK-FIX-NOSCHED-NEXT:    b .LBB83_6
 ; CHECK-FIX-NOSCHED-NEXT:  .LBB83_5:
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s1, s1
-; CHECK-FIX-NOSCHED-NEXT:    vmov s15, r0
-; CHECK-FIX-NOSCHED-NEXT:    vmov r0, s1
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s1, s3
-; CHECK-FIX-NOSCHED-NEXT:    vmov r2, s1
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f32.f16 s15, s15
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s1, s15
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s11, s11
-; CHECK-FIX-NOSCHED-NEXT:    vmov r3, s1
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s12, s12
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s1, s9
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s8, s8
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s6, s6
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s4, s4
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s0, s0
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r0, r2, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov r2, s11
+; CHECK-FIX-NOSCHED-NEXT:    vmov r3, r6, d3
+; CHECK-FIX-NOSCHED-NEXT:    vmov r0, r5, d2
+; CHECK-FIX-NOSCHED-NEXT:    lsr r4, r3, #16
+; CHECK-FIX-NOSCHED-NEXT:    lsr r9, r6, #16
+; CHECK-FIX-NOSCHED-NEXT:    lsr r12, r5, #16
+; CHECK-FIX-NOSCHED-NEXT:    lsr lr, r0, #16
+; CHECK-FIX-NOSCHED-NEXT:    uxth r11, r6
+; CHECK-FIX-NOSCHED-NEXT:    uxth r10, r3
+; CHECK-FIX-NOSCHED-NEXT:    uxth r5, r5
+; CHECK-FIX-NOSCHED-NEXT:  .LBB83_6:
+; CHECK-FIX-NOSCHED-NEXT:    uxth r8, r0
+; CHECK-FIX-NOSCHED-NEXT:    ldr r0, [sp, #4] @ 4-byte Reload
+; CHECK-FIX-NOSCHED-NEXT:    ldr r3, [sp] @ 4-byte Reload
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r3, r0, lsl #16
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d16[0], r0
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r2, r3, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov r2, s12
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s12, s14
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r8, lr, lsl #16
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d18[0], r0
-; CHECK-FIX-NOSCHED-NEXT:    vmov r0, s12
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s12, s13
-; CHECK-FIX-NOSCHED-NEXT:    vmov r3, s12
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r5, r12, lsl #16
+; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d18[1], r0
+; CHECK-FIX-NOSCHED-NEXT:    ldr r0, [sp, #8] @ 4-byte Reload
 ; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r2, r0, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov r2, s1
+; CHECK-FIX-NOSCHED-NEXT:    ldr r2, [sp, #16] @ 4-byte Reload
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d16[1], r0
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r2, r2, r3, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d18[1], r2
-; CHECK-FIX-NOSCHED-NEXT:    vmov r2, s8
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s8, s10
-; CHECK-FIX-NOSCHED-NEXT:    vmov r3, s8
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r2, r3, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov r2, s6
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s6, s7
-; CHECK-FIX-NOSCHED-NEXT:    vmov r3, s6
-; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d17[0], r0
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r2, r3, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov r2, s4
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s4, s5
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r10, r4, lsl #16
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d19[0], r0
-; CHECK-FIX-NOSCHED-NEXT:    vmov r0, s4
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r2, r0, lsl #16
-; CHECK-FIX-NOSCHED-NEXT:    vmov r2, s0
-; CHECK-FIX-NOSCHED-NEXT:    vcvtb.f16.f32 s0, s2
-; CHECK-FIX-NOSCHED-NEXT:    vmov r3, s0
+; CHECK-FIX-NOSCHED-NEXT:    ldr r0, [sp, #12] @ 4-byte Reload
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r0, r7, lsl #16
+; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d17[0], r0
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r11, r9, lsl #16
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d19[1], r0
-; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r2, r3, lsl #16
+; CHECK-FIX-NOSCHED-NEXT:    ldr r0, [sp, #20] @ 4-byte Reload
+; CHECK-FIX-NOSCHED-NEXT:    pkhbt r0, r0, r2, lsl #16
 ; CHECK-FIX-NOSCHED-NEXT:    vmov.32 d17[1], r0
 ; CHECK-FIX-NOSCHED-NEXT:    aesd.8 q8, q9
 ; CHECK-FIX-NOSCHED-NEXT:    aesimc.8 q8, q8
 ; CHECK-FIX-NOSCHED-NEXT:    vst1.64 {d16, d17}, [r1]
-; CHECK-FIX-NOSCHED-NEXT:    pop {r4, r5, r6, r7, r11, pc}
+; CHECK-FIX-NOSCHED-NEXT:    add sp, sp, #24
+; CHECK-FIX-NOSCHED-NEXT:    pop {r4, r5, r6, r7, r8, r9, r10, r11, pc}
 ;
 ; CHECK-CORTEX-FIX-LABEL: aesd_setf16_cond_via_val:
 ; CHECK-CORTEX-FIX:       @ %bb.0:
-; CHECK-CORTEX-FIX-NEXT:    .save {r4, r5, r6, r7, r11, lr}
-; CHECK-CORTEX-FIX-NEXT:    push {r4, r5, r6, r7, r11, lr}
-; CHECK-CORTEX-FIX-NEXT:    .vsave {d8, d9}
-; CHECK-CORTEX-FIX-NEXT:    vpush {d8, d9}
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s9, s0
+; CHECK-CORTEX-FIX-NEXT:    .save {r4, r5, r6, r7, r8, r9, r10, r11, lr}
+; CHECK-CORTEX-FIX-NEXT:    push {r4, r5, r6, r7, r8, r9, r10, r11, lr}
+; CHECK-CORTEX-FIX-NEXT:    .pad #28
+; CHECK-CORTEX-FIX-NEXT:    sub sp, sp, #28
+; CHECK-CORTEX-FIX-NEXT:    vmov r2, s0
 ; CHECK-CORTEX-FIX-NEXT:    cmp r0, #0
-; CHECK-CORTEX-FIX-NEXT:    beq .LBB83_3
+; CHECK-CORTEX-FIX-NEXT:    beq .LBB83_2
 ; CHECK-CORTEX-FIX-NEXT:  @ %bb.1:
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s0, s9
 ; CHECK-CORTEX-FIX-NEXT:    vld1.64 {d16, d17}, [r1]
-; CHECK-CORTEX-FIX-NEXT:    vmov r2, s0
 ; CHECK-CORTEX-FIX-NEXT:    vmov.32 r3, d16[1]
 ; CHECK-CORTEX-FIX-NEXT:    vmov.16 d16[0], r2
-; CHECK-CORTEX-FIX-NEXT:    vmov r4, r5, d17
-; CHECK-CORTEX-FIX-NEXT:    lsr lr, r3, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s8, r3
-; CHECK-CORTEX-FIX-NEXT:    vmov s11, lr
-; CHECK-CORTEX-FIX-NEXT:    lsr r6, r4, #16
-; CHECK-CORTEX-FIX-NEXT:    lsr r7, r5, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s0, r5
-; CHECK-CORTEX-FIX-NEXT:    vmov s2, r4
-; CHECK-CORTEX-FIX-NEXT:    vmov s12, r7
-; CHECK-CORTEX-FIX-NEXT:    vmov s1, r6
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s3, s8
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 r2, d16[0]
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s10, s0
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s0, s2
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s2, s1
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s14, s12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s1, s11
-; CHECK-CORTEX-FIX-NEXT:    lsr r12, r2, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s13, r2
-; CHECK-CORTEX-FIX-NEXT:    vmov s15, r12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s12, s13
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s8, s15
-; CHECK-CORTEX-FIX-NEXT:    cmp r0, #0
-; CHECK-CORTEX-FIX-NEXT:    bne .LBB83_4
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 r7, d16[0]
+; CHECK-CORTEX-FIX-NEXT:    uxth r6, r7
+; CHECK-CORTEX-FIX-NEXT:    lsr r7, r7, #16
+; CHECK-CORTEX-FIX-NEXT:    str r7, [sp, #20] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    uxth r7, r3
+; CHECK-CORTEX-FIX-NEXT:    lsr r3, r3, #16
+; CHECK-CORTEX-FIX-NEXT:    str r6, [sp, #24] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    str r7, [sp, #8] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #4] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    vmov r3, r7, d17
+; CHECK-CORTEX-FIX-NEXT:    uxth r6, r3
+; CHECK-CORTEX-FIX-NEXT:    lsr r3, r3, #16
+; CHECK-CORTEX-FIX-NEXT:    uxth r11, r7
+; CHECK-CORTEX-FIX-NEXT:    lsr r7, r7, #16
+; CHECK-CORTEX-FIX-NEXT:    str r6, [sp, #16] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    b .LBB83_3
 ; CHECK-CORTEX-FIX-NEXT:  .LBB83_2:
-; CHECK-CORTEX-FIX-NEXT:    vmov r6, r5, d3
-; CHECK-CORTEX-FIX-NEXT:    vmov r0, r2, d2
-; CHECK-CORTEX-FIX-NEXT:    lsr r7, r2, #16
-; CHECK-CORTEX-FIX-NEXT:    lsr r4, r6, #16
-; CHECK-CORTEX-FIX-NEXT:    lsr r3, r5, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s6, r6
-; CHECK-CORTEX-FIX-NEXT:    vmov s4, r5
-; CHECK-CORTEX-FIX-NEXT:    vmov s7, r2
-; CHECK-CORTEX-FIX-NEXT:    vmov s9, r0
-; CHECK-CORTEX-FIX-NEXT:    lsr r12, r0, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s13, r3
-; CHECK-CORTEX-FIX-NEXT:    vmov s15, r4
-; CHECK-CORTEX-FIX-NEXT:    vmov s16, r7
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s5, s4
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s11, s7
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s7, s9
-; CHECK-CORTEX-FIX-NEXT:    vmov s4, r12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s6, s6
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s9, s15
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s13, s13
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s15, s16
-; CHECK-CORTEX-FIX-NEXT:    b .LBB83_5
+; CHECK-CORTEX-FIX-NEXT:    ldrh r3, [r1]
+; CHECK-CORTEX-FIX-NEXT:    ldrh r11, [r1, #12]
+; CHECK-CORTEX-FIX-NEXT:    ldrh r7, [r1, #14]
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #24] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    ldrh r3, [r1, #2]
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #20] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    ldrh r3, [r1, #4]
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #8] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    ldrh r3, [r1, #6]
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #4] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    ldrh r3, [r1, #8]
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #16] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    ldrh r3, [r1, #10]
 ; CHECK-CORTEX-FIX-NEXT:  .LBB83_3:
-; CHECK-CORTEX-FIX-NEXT:    ldrh r12, [r1]
-; CHECK-CORTEX-FIX-NEXT:    ldrh lr, [r1, #2]
-; CHECK-CORTEX-FIX-NEXT:    ldrh r7, [r1, #4]
-; CHECK-CORTEX-FIX-NEXT:    ldrh r6, [r1, #6]
-; CHECK-CORTEX-FIX-NEXT:    ldrh r5, [r1, #8]
-; CHECK-CORTEX-FIX-NEXT:    ldrh r4, [r1, #10]
-; CHECK-CORTEX-FIX-NEXT:    ldrh r2, [r1, #12]
-; CHECK-CORTEX-FIX-NEXT:    ldrh r3, [r1, #14]
-; CHECK-CORTEX-FIX-NEXT:    vmov s1, r6
-; CHECK-CORTEX-FIX-NEXT:    vmov s3, r7
-; CHECK-CORTEX-FIX-NEXT:    vmov s0, r3
-; CHECK-CORTEX-FIX-NEXT:    vmov s2, r2
-; CHECK-CORTEX-FIX-NEXT:    vmov s8, r4
-; CHECK-CORTEX-FIX-NEXT:    vmov s12, r5
-; CHECK-CORTEX-FIX-NEXT:    vmov s11, lr
-; CHECK-CORTEX-FIX-NEXT:    vmov s13, r12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s14, s0
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s10, s2
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s2, s8
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s0, s12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s1, s1
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s3, s3
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s8, s11
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s12, s13
+; CHECK-CORTEX-FIX-NEXT:    str r3, [sp, #12] @ 4-byte Spill
 ; CHECK-CORTEX-FIX-NEXT:    cmp r0, #0
-; CHECK-CORTEX-FIX-NEXT:    beq .LBB83_2
-; CHECK-CORTEX-FIX-NEXT:  .LBB83_4:
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s9, s9
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 r2, d2[1]
-; CHECK-CORTEX-FIX-NEXT:    vmov r0, s9
-; CHECK-CORTEX-FIX-NEXT:    lsr r7, r2, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s16, r7
-; CHECK-CORTEX-FIX-NEXT:    vmov.16 d2[0], r0
-; CHECK-CORTEX-FIX-NEXT:    vmov r6, r5, d3
-; CHECK-CORTEX-FIX-NEXT:    vmov s7, r2
-; CHECK-CORTEX-FIX-NEXT:    lsr r4, r6, #16
-; CHECK-CORTEX-FIX-NEXT:    lsr r3, r5, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s6, r6
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s11, s7
-; CHECK-CORTEX-FIX-NEXT:    vmov s9, r3
-; CHECK-CORTEX-FIX-NEXT:    vmov s15, r4
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 r0, d2[0]
-; CHECK-CORTEX-FIX-NEXT:    vmov s4, r5
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s6, s6
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s13, s9
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s9, s15
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s15, s16
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s5, s4
-; CHECK-CORTEX-FIX-NEXT:    vmov s18, r0
-; CHECK-CORTEX-FIX-NEXT:    lsr r12, r0, #16
-; CHECK-CORTEX-FIX-NEXT:    vmov s4, r12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s7, s18
+; CHECK-CORTEX-FIX-NEXT:    beq .LBB83_5
+; CHECK-CORTEX-FIX-NEXT:  @ %bb.4:
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 r3, d2[1]
+; CHECK-CORTEX-FIX-NEXT:    vmov.16 d2[0], r2
+; CHECK-CORTEX-FIX-NEXT:    vmov r4, r6, d3
+; CHECK-CORTEX-FIX-NEXT:    uxth r10, r4
+; CHECK-CORTEX-FIX-NEXT:    lsr r4, r4, #16
+; CHECK-CORTEX-FIX-NEXT:    uxth lr, r6
+; CHECK-CORTEX-FIX-NEXT:    lsr r8, r6, #16
+; CHECK-CORTEX-FIX-NEXT:    uxth r5, r3
+; CHECK-CORTEX-FIX-NEXT:    lsr r12, r3, #16
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 r2, d2[0]
+; CHECK-CORTEX-FIX-NEXT:    uxth r0, r2
+; CHECK-CORTEX-FIX-NEXT:    lsr r9, r2, #16
+; CHECK-CORTEX-FIX-NEXT:    str r0, [sp] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    b .LBB83_6
 ; CHECK-CORTEX-FIX-NEXT:  .LBB83_5:
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s10, s10
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s14, s14
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s3, s3
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s2, s2
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s0, s0
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s8, s8
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s12, s12
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f32.f16 s4, s4
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s6, s6
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s9, s9
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s7, s7
-; CHECK-CORTEX-FIX-NEXT:    vmov r0, s10
-; CHECK-CORTEX-FIX-NEXT:    vmov r2, s14
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s10, s1
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s14, s5
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s1, s13
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s5, s11
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s11, s15
-; CHECK-CORTEX-FIX-NEXT:    vmov r5, s2
-; CHECK-CORTEX-FIX-NEXT:    vmov r4, s8
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r12, r0, r2, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r2, s3
-; CHECK-CORTEX-FIX-NEXT:    vmov r3, s10
-; CHECK-CORTEX-FIX-NEXT:    vmov r7, s1
-; CHECK-CORTEX-FIX-NEXT:    vmov r6, s11
-; CHECK-CORTEX-FIX-NEXT:    vmov r0, s9
-; CHECK-CORTEX-FIX-NEXT:    pkhbt lr, r2, r3, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r3, s14
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r3, r3, r7, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r7, s5
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r7, r7, r6, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r6, s0
-; CHECK-CORTEX-FIX-NEXT:    vcvtb.f16.f32 s0, s4
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r6, r6, r5, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r5, s12
-; CHECK-CORTEX-FIX-NEXT:    vmov r2, s0
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r5, r5, r4, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r4, s6
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d18[0], r5
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d19[0], r6
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d18[1], lr
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d19[1], r12
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r0, r4, r0, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov r4, s7
-; CHECK-CORTEX-FIX-NEXT:    pkhbt r2, r4, r2, lsl #16
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d16[0], r2
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d17[0], r0
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d16[1], r7
-; CHECK-CORTEX-FIX-NEXT:    vmov.32 d17[1], r3
+; CHECK-CORTEX-FIX-NEXT:    vmov r2, r3, d2
+; CHECK-CORTEX-FIX-NEXT:    uxth r0, r2
+; CHECK-CORTEX-FIX-NEXT:    lsr r9, r2, #16
+; CHECK-CORTEX-FIX-NEXT:    uxth r5, r3
+; CHECK-CORTEX-FIX-NEXT:    lsr r12, r3, #16
+; CHECK-CORTEX-FIX-NEXT:    str r0, [sp] @ 4-byte Spill
+; CHECK-CORTEX-FIX-NEXT:    mov r0, r7
+; CHECK-CORTEX-FIX-NEXT:    vmov r6, r7, d3
+; CHECK-CORTEX-FIX-NEXT:    uxth r10, r6
+; CHECK-CORTEX-FIX-NEXT:    lsr r4, r6, #16
+; CHECK-CORTEX-FIX-NEXT:    uxth lr, r7
+; CHECK-CORTEX-FIX-NEXT:    lsr r8, r7, #16
+; CHECK-CORTEX-FIX-NEXT:    mov r7, r0
+; CHECK-CORTEX-FIX-NEXT:  .LBB83_6:
+; CHECK-CORTEX-FIX-NEXT:    ldr r0, [sp, #8] @ 4-byte Reload
+; CHECK-CORTEX-FIX-NEXT:    ldr r2, [sp, #4] @ 4-byte Reload
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r11, r11, r7, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    ldr r3, [sp, #12] @ 4-byte Reload
+; CHECK-CORTEX-FIX-NEXT:    ldr r6, [sp, #20] @ 4-byte Reload
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r5, r5, r12, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r4, r10, r4, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r7, r0, r2, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    ldr r2, [sp, #16] @ 4-byte Reload
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r0, lr, r8, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r2, r2, r3, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    ldr r3, [sp, #24] @ 4-byte Reload
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r3, r3, r6, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    ldr r6, [sp] @ 4-byte Reload
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d18[0], r3
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d19[0], r2
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d18[1], r7
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d19[1], r11
+; CHECK-CORTEX-FIX-NEXT:    pkhbt r6, r6, r9, lsl #16
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d16[0], r6
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d17[0], r4
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d16[1], r5
+; CHECK-CORTEX-FIX-NEXT:    vmov.32 d17[1], r0
 ; CHECK-CORTEX-FIX-NEXT:    aesd.8 q9, q8
 ; CHECK-CORTEX-FIX-NEXT:    aesimc.8 q8, q9
 ; CHECK-CORTEX-FIX-NEXT:    vst1.64 {d16, d17}, [r1]
-; CHECK-CORTEX-FIX-NEXT:    vpop {d8, d9}
-; CHECK-CORTEX-FIX-NEXT:    pop {r4, r5, r6, r7, r11, pc}
+; CHECK-CORTEX-FIX-NEXT:    add sp, sp, #28
+; CHECK-CORTEX-FIX-NEXT:    pop {r4, r5, r6, r7, r8, r9, r10, r11, pc}
   br i1 %0, label %5, label %11
 
 5:
@@ -4627,10 +4118,7 @@
 ; CHECK-FIX-NEXT:    cmp r0, #0
 ; CHECK-FIX-NEXT:    bxeq lr
 ; CHECK-FIX-NEXT:  .LBB85_1:
-; CHECK-FIX-NEXT:    vcvtb.f32.f16 s0, s0
-; CHECK-FIX-NEXT:    vcvtb.f16.f32 s0, s0
 ; CHECK-FIX-NEXT:    vmov r2, s0
-; CHECK-FIX-NEXT:    uxth r2, r2
 ; CHECK-FIX-NEXT:    vmov.16 d2[0], r2
 ; CHECK-FIX-NEXT:  .LBB85_2: @ =>This Inner Loop Header: Depth=1
 ; CHECK-FIX-NEXT:    vld1.64 {d16, d17}, [r1]
diff --git a/llvm/test/CodeGen/ARM/fp16-promote.ll b/llvm/test/CodeGen/ARM/fp16-promote.ll
--- a/llvm/test/CodeGen/ARM/fp16-promote.ll
+++ b/llvm/test/CodeGen/ARM/fp16-promote.ll
@@ -223,18 +223,12 @@
 
 declare i1 @test_dummy(ptr %p) #0
 ; CHECK-ALL-LABEL: test_phi:
-; CHECK-FP16: vcvtb.f32.f16
-; CHECK-FP16: [[LOOP:.LBB[1-9_]+]]:
-; CHECK-FP16: vcvtb.f32.f16
+; CHECK-FP16: [[LOOP:.LBB[0-9_]+]]:
 ; CHECK-FP16: bl      test_dummy
 ; CHECK-FP16: bne     [[LOOP]]
-; CHECK-FP16: vcvtb.f16.f32
-; CHECK-LIBCALL-VFP: bl __aeabi_h2f
-; CHECK-LIBCALL: [[LOOP:.LBB[1-9_]+]]:
-; CHECK-LIBCALL-VFP: bl __aeabi_h2f
+; CHECK-LIBCALL: [[LOOP:.LBB[0-9_]+]]:
 ; CHECK-LIBCALL: bl test_dummy
 ; CHECK-LIBCALL: bne     [[LOOP]]
-; CHECK-LIBCALL-VFP: bl __aeabi_f2h
 define void @test_phi(ptr %p) #0 {
 entry:
   %a = load half, ptr %p
diff --git a/llvm/test/CodeGen/ARM/inlineasm-fp-half.ll b/llvm/test/CodeGen/ARM/inlineasm-fp-half.ll
--- a/llvm/test/CodeGen/ARM/inlineasm-fp-half.ll
+++ b/llvm/test/CodeGen/ARM/inlineasm-fp-half.ll
@@ -1,60 +1,173 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc -mtriple=arm -mattr=+armv8.2-a,+fp-armv8,+fullfp16,+bf16,-neon %s -o - | FileCheck %s
-; RUN: llc -mtriple=thumb -mattr=+armv8.2-a,+fp-armv8,+fullfp16,+bf16,-neon %s -o - | FileCheck %s
+; No FP16/BF16
+; RUN: llc -mtriple=arm-none-eabi -mattr=+armv8.2-a,+fp-armv8,-fullfp16,-bf16,-neon %s -o - | FileCheck %s --check-prefixes=NO-FP16-SOFTFP
+; RUN: llc -mtriple=thumb-none-eabi -mattr=+armv8.2-a,+fp-armv8,-fullfp16,-bf16,-neon %s -o - | FileCheck %s --check-prefixes=NO-FP16-SOFTFP
+; RUN: llc -mtriple=arm-none-eabihf -mattr=+armv8.2-a,+fp-armv8,-fullfp16,-bf16,-neon %s -o - | FileCheck %s --check-prefixes=NO-FP16-HARD
+; RUN: llc -mtriple=thumb-none-eabihf -mattr=+armv8.2-a,+fp-armv8,-fullfp16,-bf16,-neon %s -o - | FileCheck %s --check-prefixes=NO-FP16-HARD
 
+; With FP16, Without BF16
+; RUN: llc -mtriple=arm-none-eabi -mattr=+armv8.2-a,+fp-armv8,+fullfp16,-bf16,-neon %s -o - | FileCheck %s --check-prefixes=FP16-SOFTFP
+; RUN: llc -mtriple=thumb-none-eabi -mattr=+armv8.2-a,+fp-armv8,+fullfp16,-bf16,-neon %s -o - | FileCheck %s --check-prefixes=FP16-SOFTFP
+; RUN: llc -mtriple=arm-none-eabihf -mattr=+armv8.2-a,+fp-armv8,+fullfp16,-bf16,-neon %s -o - | FileCheck %s --check-prefixes=FP16-HARD
+; RUN: llc -mtriple=thumb-none-eabihf -mattr=+armv8.2-a,+fp-armv8,+fullfp16,-bf16,-neon %s -o - | FileCheck %s --check-prefixes=FP16-HARD
 
-define arm_aapcscc half @f_t(half %x) nounwind {
-; CHECK-LABEL: f_t:
-; CHECK:       @ %bb.0: @ %entry
-; CHECK-NEXT:    vmov.f16 s0, r0
-; CHECK-NEXT:    @APP
-; CHECK-NEXT:    vsqrt.f16 s0, s0
-; CHECK-NEXT:    @NO_APP
-; CHECK-NEXT:    vmov r0, s0
-; CHECK-NEXT:    bx lr
+; With FP16/BF16
+; RUN: llc -mtriple=arm-none-eabi -mattr=+armv8.2-a,+fp-armv8,+fullfp16,+bf16,-neon %s -o - | FileCheck %s --check-prefixes=FP16-SOFTFP,BF16-SOFTFP
+; RUN: llc -mtriple=thumb-none-eabi -mattr=+armv8.2-a,+fp-armv8,+fullfp16,+bf16,-neon %s -o - | FileCheck %s --check-prefixes=FP16-SOFTFP,BF16-SOFTFP
+; RUN: llc -mtriple=arm-none-eabihf -mattr=+armv8.2-a,+fp-armv8,+fullfp16,+bf16,-neon %s -o - | FileCheck %s --check-prefixes=FP16-HARD,BF16-HARD
+; RUN: llc -mtriple=thumb-none-eabihf -mattr=+armv8.2-a,+fp-armv8,+fullfp16,+bf16,-neon %s -o - | FileCheck %s --check-prefixes=FP16-HARD,BF16-HARD
+
+; This test ensures that we can use `w` and `t` constraints to allocate
+; S-registers for 16-bit FP inputs and outputs for inline assembly, with either
+; the softfp or hard float ABIs. (With the soft abi, no S-regs are available).
+
+define half @half_t(half %x) nounwind {
+; NO-FP16-SOFTFP-LABEL: half_t:
+; NO-FP16-SOFTFP:       @ %bb.0: @ %entry
+; NO-FP16-SOFTFP-NEXT:    vmov s0, r0
+; NO-FP16-SOFTFP-NEXT:    @APP
+; NO-FP16-SOFTFP-NEXT:    vmov.f32 s0, s0
+; NO-FP16-SOFTFP-NEXT:    @NO_APP
+; NO-FP16-SOFTFP-NEXT:    vmov r0, s0
+; NO-FP16-SOFTFP-NEXT:    bx lr
+;
+; NO-FP16-HARD-LABEL: half_t:
+; NO-FP16-HARD:       @ %bb.0: @ %entry
+; NO-FP16-HARD-NEXT:    @APP
+; NO-FP16-HARD-NEXT:    vmov.f32 s0, s0
+; NO-FP16-HARD-NEXT:    @NO_APP
+; NO-FP16-HARD-NEXT:    bx lr
+;
+; FP16-SOFTFP-LABEL: half_t:
+; FP16-SOFTFP:       @ %bb.0: @ %entry
+; FP16-SOFTFP-NEXT:    vmov.f16 s0, r0
+; FP16-SOFTFP-NEXT:    @APP
+; FP16-SOFTFP-NEXT:    vmov.f32 s0, s0
+; FP16-SOFTFP-NEXT:    @NO_APP
+; FP16-SOFTFP-NEXT:    vmov r0, s0
+; FP16-SOFTFP-NEXT:    bx lr
+;
+; FP16-HARD-LABEL: half_t:
+; FP16-HARD:       @ %bb.0: @ %entry
+; FP16-HARD-NEXT:    @APP
+; FP16-HARD-NEXT:    vmov.f32 s0, s0
+; FP16-HARD-NEXT:    @NO_APP
+; FP16-HARD-NEXT:    bx lr
 entry:
-  %0 = tail call half asm "vsqrt.f16 $0, $1", "=t,t"(half %x)
+  %0 = tail call half asm "vmov $0, $1", "=t,t"(half %x)
   ret half %0
 }
 
-define arm_aapcscc half @f_w(half %x) nounwind {
-; CHECK-LABEL: f_w:
-; CHECK:       @ %bb.0: @ %entry
-; CHECK-NEXT:    vmov.f16 s0, r0
-; CHECK-NEXT:    @APP
-; CHECK-NEXT:    vsqrt.f16 s0, s0
-; CHECK-NEXT:    @NO_APP
-; CHECK-NEXT:    vmov r0, s0
-; CHECK-NEXT:    bx lr
+define half @half_w(half %x) nounwind {
+; NO-FP16-SOFTFP-LABEL: half_w:
+; NO-FP16-SOFTFP:       @ %bb.0: @ %entry
+; NO-FP16-SOFTFP-NEXT:    vmov s0, r0
+; NO-FP16-SOFTFP-NEXT:    @APP
+; NO-FP16-SOFTFP-NEXT:    vmov.f32 s0, s0
+; NO-FP16-SOFTFP-NEXT:    @NO_APP
+; NO-FP16-SOFTFP-NEXT:    vmov r0, s0
+; NO-FP16-SOFTFP-NEXT:    bx lr
+;
+; NO-FP16-HARD-LABEL: half_w:
+; NO-FP16-HARD:       @ %bb.0: @ %entry
+; NO-FP16-HARD-NEXT:    @APP
+; NO-FP16-HARD-NEXT:    vmov.f32 s0, s0
+; NO-FP16-HARD-NEXT:    @NO_APP
+; NO-FP16-HARD-NEXT:    bx lr
+;
+; FP16-SOFTFP-LABEL: half_w:
+; FP16-SOFTFP:       @ %bb.0: @ %entry
+; FP16-SOFTFP-NEXT:    vmov.f16 s0, r0
+; FP16-SOFTFP-NEXT:    @APP
+; FP16-SOFTFP-NEXT:    vmov.f32 s0, s0
+; FP16-SOFTFP-NEXT:    @NO_APP
+; FP16-SOFTFP-NEXT:    vmov r0, s0
+; FP16-SOFTFP-NEXT:    bx lr
+;
+; FP16-HARD-LABEL: half_w:
+; FP16-HARD:       @ %bb.0: @ %entry
+; FP16-HARD-NEXT:    @APP
+; FP16-HARD-NEXT:    vmov.f32 s0, s0
+; FP16-HARD-NEXT:    @NO_APP
+; FP16-HARD-NEXT:    bx lr
 entry:
-  %0 = tail call half asm "vsqrt.f16 $0, $1", "=w,w"(half %x)
+  %0 = tail call half asm "vmov $0, $1", "=w,w"(half %x)
   ret half %0
 }
 
-define arm_aapcscc bfloat @h_t(bfloat %x) nounwind {
-; CHECK-LABEL: h_t:
-; CHECK:       @ %bb.0: @ %entry
-; CHECK-NEXT:    vmov.f16 s0, r0
-; CHECK-NEXT:    @APP
-; CHECK-NEXT:    vmov.f32 s0, s0
-; CHECK-NEXT:    @NO_APP
-; CHECK-NEXT:    vmov.f16 r0, s0
-; CHECK-NEXT:    bx lr
+define bfloat @bf16_t(bfloat %x) nounwind {
+; NO-FP16-SOFTFP-LABEL: bf16_t:
+; NO-FP16-SOFTFP:       @ %bb.0: @ %entry
+; NO-FP16-SOFTFP-NEXT:    vmov s0, r0
+; NO-FP16-SOFTFP-NEXT:    @APP
+; NO-FP16-SOFTFP-NEXT:    vmov.f32 s0, s0
+; NO-FP16-SOFTFP-NEXT:    @NO_APP
+; NO-FP16-SOFTFP-NEXT:    vmov r0, s0
+; NO-FP16-SOFTFP-NEXT:    bx lr
+;
+; NO-FP16-HARD-LABEL: bf16_t:
+; NO-FP16-HARD:       @ %bb.0: @ %entry
+; NO-FP16-HARD-NEXT:    @APP
+; NO-FP16-HARD-NEXT:    vmov.f32 s0, s0
+; NO-FP16-HARD-NEXT:    @NO_APP
+; NO-FP16-HARD-NEXT:    bx lr
+;
+; FP16-SOFTFP-LABEL: bf16_t:
+; FP16-SOFTFP:       @ %bb.0: @ %entry
+; FP16-SOFTFP-NEXT:    vmov s0, r0
+; FP16-SOFTFP-NEXT:    @APP
+; FP16-SOFTFP-NEXT:    vmov.f32 s0, s0
+; FP16-SOFTFP-NEXT:    @NO_APP
+; FP16-SOFTFP-NEXT:    vmov r0, s0
+; FP16-SOFTFP-NEXT:    bx lr
+;
+; FP16-HARD-LABEL: bf16_t:
+; FP16-HARD:       @ %bb.0: @ %entry
+; FP16-HARD-NEXT:    @APP
+; FP16-HARD-NEXT:    vmov.f32 s0, s0
+; FP16-HARD-NEXT:    @NO_APP
+; FP16-HARD-NEXT:    bx lr
 entry:
-  %0 = tail call bfloat asm "vmov.f32 $0, $1", "=t,t"(bfloat %x)
+  %0 = tail call bfloat asm "vmov $0, $1", "=t,t"(bfloat %x)
   ret bfloat %0
 }
 
-define arm_aapcscc bfloat @h_w(bfloat %x) nounwind {
-; CHECK-LABEL: h_w:
-; CHECK:       @ %bb.0: @ %entry
-; CHECK-NEXT:    vmov.f16 s0, r0
-; CHECK-NEXT:    @APP
-; CHECK-NEXT:    vmov.f32 s0, s0
-; CHECK-NEXT:    @NO_APP
-; CHECK-NEXT:    vmov.f16 r0, s0
-; CHECK-NEXT:    bx lr
+define bfloat @bf16_w(bfloat %x) nounwind {
+; NO-FP16-SOFTFP-LABEL: bf16_w:
+; NO-FP16-SOFTFP:       @ %bb.0: @ %entry
+; NO-FP16-SOFTFP-NEXT:    vmov s0, r0
+; NO-FP16-SOFTFP-NEXT:    @APP
+; NO-FP16-SOFTFP-NEXT:    vmov.f32 s0, s0
+; NO-FP16-SOFTFP-NEXT:    @NO_APP
+; NO-FP16-SOFTFP-NEXT:    vmov r0, s0
+; NO-FP16-SOFTFP-NEXT:    bx lr
+;
+; NO-FP16-HARD-LABEL: bf16_w:
+; NO-FP16-HARD:       @ %bb.0: @ %entry
+; NO-FP16-HARD-NEXT:    @APP
+; NO-FP16-HARD-NEXT:    vmov.f32 s0, s0
+; NO-FP16-HARD-NEXT:    @NO_APP
+; NO-FP16-HARD-NEXT:    bx lr
+;
+; FP16-SOFTFP-LABEL: bf16_w:
+; FP16-SOFTFP:       @ %bb.0: @ %entry
+; FP16-SOFTFP-NEXT:    vmov s0, r0
+; FP16-SOFTFP-NEXT:    @APP
+; FP16-SOFTFP-NEXT:    vmov.f32 s0, s0
+; FP16-SOFTFP-NEXT:    @NO_APP
+; FP16-SOFTFP-NEXT:    vmov r0, s0
+; FP16-SOFTFP-NEXT:    bx lr
+;
+; FP16-HARD-LABEL: bf16_w:
+; FP16-HARD:       @ %bb.0: @ %entry
+; FP16-HARD-NEXT:    @APP
+; FP16-HARD-NEXT:    vmov.f32 s0, s0
+; FP16-HARD-NEXT:    @NO_APP
+; FP16-HARD-NEXT:    bx lr
 entry:
-  %0 = tail call bfloat asm "vmov.f32 $0, $1", "=w,w"(bfloat %x)
+  %0 = tail call bfloat asm "vmov $0, $1", "=w,w"(bfloat %x)
   ret bfloat %0
 }
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; BF16-HARD: {{.*}}
+; BF16-SOFTFP: {{.*}}
diff --git a/llvm/test/CodeGen/ARM/pr47454.ll b/llvm/test/CodeGen/ARM/pr47454.ll
--- a/llvm/test/CodeGen/ARM/pr47454.ll
+++ b/llvm/test/CodeGen/ARM/pr47454.ll
@@ -10,29 +10,15 @@
 ; CHECK:       @ %bb.0: @ %Entry
 ; CHECK-NEXT:    push {r11, lr}
 ; CHECK-NEXT:    mov r11, sp
-; CHECK-NEXT:    sub sp, sp, #16
+; CHECK-NEXT:    sub sp, sp, #8
 ; CHECK-NEXT:    mov r0, #31744
 ; CHECK-NEXT:    strh r0, [r11, #-2]
 ; CHECK-NEXT:    ldrh r0, [r11, #-2]
-; CHECK-NEXT:    bl __gnu_h2f_ieee
 ; CHECK-NEXT:    vmov s0, r0
-; CHECK-NEXT:    vstr s0, [sp, #4] @ 4-byte Spill
+; CHECK-NEXT:    vstr s0, [sp] @ 4-byte Spill
 ; CHECK-NEXT:    bl getConstant
-; CHECK-NEXT:    vmov r0, s0
-; CHECK-NEXT:    bl __gnu_h2f_ieee
-; CHECK-NEXT:    vmov s0, r0
-; CHECK-NEXT:    vmov r0, s0
-; CHECK-NEXT:    bl __gnu_f2h_ieee
-; CHECK-NEXT:    vldr s0, [sp, #4] @ 4-byte Reload
-; CHECK-NEXT:    str r0, [sp, #8] @ 4-byte Spill
-; CHECK-NEXT:    vmov r0, s0
-; CHECK-NEXT:    bl __gnu_f2h_ieee
-; CHECK-NEXT:    mov r1, r0
-; CHECK-NEXT:    ldr r0, [sp, #8] @ 4-byte Reload
-; CHECK-NEXT:    uxth r1, r1
-; CHECK-NEXT:    vmov s0, r1
-; CHECK-NEXT:    uxth r0, r0
-; CHECK-NEXT:    vmov s1, r0
+; CHECK-NEXT:    vmov.f32 s1, s0
+; CHECK-NEXT:    vldr s0, [sp] @ 4-byte Reload
 ; CHECK-NEXT:    bl isEqual
 ; CHECK-NEXT:    mov sp, r11
 ; CHECK-NEXT:    pop {r11, pc}