AArch64: count call stack when working out if scavenging is needed

[t.p.northover]

Summary

We try to guess whether we'll need extra registers or slots to handle large stack offsets (bigger than a single [sp, #N] can handle). This calculation needs to take into account stack that's in the caller's frame, either our own args or involved in calls we make.

Bit of a messy test, but it needs to both have lots of args to actually use significant stack, and be opaque enough that it also uses many registers. So the usual [8 x i64] tricks to quickly use args don't work.

[t.p.northover]

Ping.

[efriedma]

llvm/lib/Target/AArch64/AArch64FrameLowering.cpp

3084

To be clear, this is looking at arguments passed on the stack by the caller of this function definition? If we use a frame index to access them, and we don't have a frame pointer, they need to count towards the stack limit, sure.

Referring to this as the "callee's stack" is a bit confusing.

[t.p.northover]

llvm/lib/Target/AArch64/AArch64FrameLowering.cpp

3084

It is confusing. It's both callee and caller stack space (i.e. used only for args, but might be incoming or outgoing). Regardless, I did word that badly. I'll try to improve it.

[t.p.northover]

Reword comment to avoid confusion.

[efriedma]

LGTM

[t.p.northover]

Thanks. 2002c82278f6 on main.

llvm/lib/Target/AArch64/AArch64FrameLowering.cpp

void AArch64FrameLowering::determineCalleeSaves(MachineFunction &MF,
   int64_t SVEStackSize =
       alignTo(SVECSStackSize + estimateSVEStackObjectOffsets(MFI), 16);
   bool CanEliminateFrame = (SavedRegs.count() == 0) && !SVEStackSize;
 
   // The CSR spill slots have not been allocated yet, so estimateStackSize
   // won't include them.
   unsigned EstimatedStackSizeLimit = estimateRSStackSizeLimit(MF);
 
+  // We may address some of the callee's stack, either for our own arguments or
+  // during a call. Include that in calculating whether we have complicated
+  // addressing concerns.
+  int64_t CalleeStackUsed = 0;
+  for (int I = MFI.getObjectIndexBegin(); I != 0; ++I) {
+    int64_t FixedOff = MFI.getObjectOffset(I);
+    if (FixedOff > CalleeStackUsed) CalleeStackUsed = FixedOff;
+  }
+
   // Conservatively always assume BigStack when there are SVE spills.
-  bool BigStack = SVEStackSize ||
-                  (EstimatedStackSize + CSStackSize) > EstimatedStackSizeLimit;
+  bool BigStack = SVEStackSize || (EstimatedStackSize + CSStackSize +
+                                   CalleeStackUsed) > EstimatedStackSizeLimit;
   if (BigStack || !CanEliminateFrame || RegInfo->cannotEliminateFrame(MF))
     AFI->setHasStackFrame(true);
 
   // Estimate if we might need to scavenge a register at some point in order
   // to materialize a stack offset. If so, either spill one additional
   // callee-saved register or reserve a special spill slot to facilitate
   // register scavenging. If we already spilled an extra callee-saved register
   // above to keep the number of spills even, we don't need to do anything else

llvm/test/CodeGen/AArch64/arm64-stackmap.ll

 ; CHECK-NEXT:   .quad 1
 ; CHECK-NEXT:   .quad _jsIntCall
 ; CHECK-NEXT:   .quad 16
 ; CHECK-NEXT:   .quad 1
 ; CHECK-NEXT:   .quad _spilledValue
 ; CHECK-NEXT:   .quad 160
 ; CHECK-NEXT:   .quad 1
 ; CHECK-NEXT:   .quad _spilledStackMapValue
-; CHECK-NEXT:   .quad 128
+; CHECK-NEXT:   .quad 144
 ; CHECK-NEXT:   .quad 1
 ; CHECK-NEXT:   .quad _liveConstant
 ; CHECK-NEXT:   .quad 16
 ; CHECK-NEXT:   .quad 1
 ; CHECK-NEXT:   .quad _clobberLR
 ; CHECK-NEXT:   .quad 112
 ; CHECK-NEXT:   .quad 1
 ; CHECK-NEXT:   .quad _floats

llvm/test/CodeGen/AArch64/arm64-subvector-extend.ll

 ; CHECK-NEXT:    ret
   %res = sext <32 x i1> %arg to <32 x i8>
   ret <32 x i8> %res
 }
 
 define <64 x i8> @zext_v64i1(<64 x i1> %arg) {
 ; CHECK-LABEL: zext_v64i1:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ldr w8, [sp, #320]
+; CHECK-NEXT:    str x29, [sp, #-16]! // 8-byte Folded Spill
+; CHECK-NEXT:    .cfi_def_cfa_offset 16
+; CHECK-NEXT:    .cfi_offset w29, -16
+; CHECK-NEXT:    ldr w8, [sp, #336]
 ; CHECK-NEXT:    fmov s0, w0
-; CHECK-NEXT:    ldr w9, [sp, #64]
-; CHECK-NEXT:    ldr w10, [sp, #192]
+; CHECK-NEXT:    ldr w9, [sp, #80]
+; CHECK-NEXT:    ldr w10, [sp, #208]
 ; CHECK-NEXT:    fmov s3, w8
-; CHECK-NEXT:    ldr w8, [sp, #328]
+; CHECK-NEXT:    ldr w8, [sp, #344]
 ; CHECK-NEXT:    fmov s1, w9
-; CHECK-NEXT:    ldr w9, [sp, #200]
+; CHECK-NEXT:    ldr w9, [sp, #216]
 ; CHECK-NEXT:    fmov s2, w10
-; CHECK-NEXT:    ldr w10, [sp, #336]
+; CHECK-NEXT:    ldr w10, [sp, #352]
 ; CHECK-NEXT:    mov.b v3[1], w8
-; CHECK-NEXT:    ldr w8, [sp, #72]
+; CHECK-NEXT:    ldr w8, [sp, #88]
 ; CHECK-NEXT:    mov.b v0[1], w1
-; CHECK-NEXT:    ldr w11, [sp, #352]
+; CHECK-NEXT:    ldr w11, [sp, #368]
 ; CHECK-NEXT:    mov.b v2[1], w9
-; CHECK-NEXT:    ldr w9, [sp, #80]
+; CHECK-NEXT:    ldr w9, [sp, #96]
 ; CHECK-NEXT:    mov.b v1[1], w8
-; CHECK-NEXT:    ldr w8, [sp, #344]
+; CHECK-NEXT:    ldr w8, [sp, #360]
 ; CHECK-NEXT:    mov.b v3[2], w10
-; CHECK-NEXT:    ldr w10, [sp, #208]
+; CHECK-NEXT:    ldr w10, [sp, #224]
 ; CHECK-NEXT:    mov.b v0[2], w2
-; CHECK-NEXT:    ldr w12, [sp, #368]
-; CHECK-NEXT:    ldr w13, [sp, #384]
+; CHECK-NEXT:    ldr w12, [sp, #384]
+; CHECK-NEXT:    ldr w13, [sp, #400]
 ; CHECK-NEXT:    mov.b v1[2], w9
-; CHECK-NEXT:    ldr w9, [sp, #360]
+; CHECK-NEXT:    ldr w9, [sp, #376]
 ; CHECK-NEXT:    mov.b v2[2], w10
-; CHECK-NEXT:    ldr w10, [sp, #88]
+; CHECK-NEXT:    ldr w10, [sp, #104]
 ; CHECK-NEXT:    mov.b v3[3], w8
-; CHECK-NEXT:    ldr w8, [sp, #216]
+; CHECK-NEXT:    ldr w8, [sp, #232]
 ; CHECK-NEXT:    mov.b v0[3], w3
-; CHECK-NEXT:    ldr w14, [sp, #400]
+; CHECK-NEXT:    ldr w14, [sp, #416]
 ; CHECK-NEXT:    mov.b v1[3], w10
-; CHECK-NEXT:    ldr w10, [sp, #376]
+; CHECK-NEXT:    ldr w10, [sp, #392]
 ; CHECK-NEXT:    mov.b v2[3], w8
-; CHECK-NEXT:    ldr w8, [sp, #96]
+; CHECK-NEXT:    ldr w8, [sp, #112]
 ; CHECK-NEXT:    mov.b v3[4], w11
-; CHECK-NEXT:    ldr w11, [sp, #224]
+; CHECK-NEXT:    ldr w11, [sp, #240]
 ; CHECK-NEXT:    mov.b v0[4], w4
-; CHECK-NEXT:    ldr w15, [sp, #416]
+; CHECK-NEXT:    ldr w15, [sp, #432]
 ; CHECK-NEXT:    mov.b v1[4], w8
-; CHECK-NEXT:    ldr w8, [sp, #392]
+; CHECK-NEXT:    ldr w8, [sp, #408]
 ; CHECK-NEXT:    mov.b v2[4], w11
-; CHECK-NEXT:    ldr w11, [sp, #104]
+; CHECK-NEXT:    ldr w11, [sp, #120]
 ; CHECK-NEXT:    mov.b v3[5], w9
-; CHECK-NEXT:    ldr w9, [sp, #232]
+; CHECK-NEXT:    ldr w9, [sp, #248]
 ; CHECK-NEXT:    mov.b v0[5], w5
-; CHECK-NEXT:    ldr w16, [sp, #432]
+; CHECK-NEXT:    ldr w16, [sp, #448]
 ; CHECK-NEXT:    mov.b v1[5], w11
-; CHECK-NEXT:    ldr w11, [sp, #408]
+; CHECK-NEXT:    ldr w11, [sp, #424]
 ; CHECK-NEXT:    mov.b v2[5], w9
-; CHECK-NEXT:    ldr w9, [sp, #112]
+; CHECK-NEXT:    ldr w9, [sp, #128]
 ; CHECK-NEXT:    mov.b v3[6], w12
-; CHECK-NEXT:    ldr w12, [sp, #240]
+; CHECK-NEXT:    ldr w12, [sp, #256]
 ; CHECK-NEXT:    mov.b v0[6], w6
 ; CHECK-NEXT:    mov.b v1[6], w9
-; CHECK-NEXT:    ldr w9, [sp, #424]
+; CHECK-NEXT:    ldr w9, [sp, #440]
 ; CHECK-NEXT:    mov.b v2[6], w12
-; CHECK-NEXT:    ldr w12, [sp, #120]
+; CHECK-NEXT:    ldr w12, [sp, #136]
 ; CHECK-NEXT:    mov.b v3[7], w10
-; CHECK-NEXT:    ldr w10, [sp, #248]
+; CHECK-NEXT:    ldr w10, [sp, #264]
 ; CHECK-NEXT:    mov.b v0[7], w7
 ; CHECK-NEXT:    mov.b v1[7], w12
-; CHECK-NEXT:    ldr w12, [sp]
+; CHECK-NEXT:    ldr w12, [sp, #16]
 ; CHECK-NEXT:    mov.b v2[7], w10
-; CHECK-NEXT:    ldr w10, [sp, #128]
+; CHECK-NEXT:    ldr w10, [sp, #144]
 ; CHECK-NEXT:    mov.b v3[8], w13
-; CHECK-NEXT:    ldr w13, [sp, #256]
+; CHECK-NEXT:    ldr w13, [sp, #272]
 ; CHECK-NEXT:    mov.b v0[8], w12
-; CHECK-NEXT:    ldr w12, [sp, #440]
+; CHECK-NEXT:    ldr w12, [sp, #456]
 ; CHECK-NEXT:    mov.b v1[8], w10
-; CHECK-NEXT:    ldr w10, [sp, #8]
+; CHECK-NEXT:    ldr w10, [sp, #24]
 ; CHECK-NEXT:    mov.b v2[8], w13
-; CHECK-NEXT:    ldr w13, [sp, #136]
+; CHECK-NEXT:    ldr w13, [sp, #152]
 ; CHECK-NEXT:    mov.b v3[9], w8
-; CHECK-NEXT:    ldr w8, [sp, #264]
+; CHECK-NEXT:    ldr w8, [sp, #280]
 ; CHECK-NEXT:    mov.b v0[9], w10
-; CHECK-NEXT:    ldr w10, [sp, #272]
+; CHECK-NEXT:    ldr w10, [sp, #288]
 ; CHECK-NEXT:    mov.b v1[9], w13
-; CHECK-NEXT:    ldr w13, [sp, #16]
+; CHECK-NEXT:    ldr w13, [sp, #32]
 ; CHECK-NEXT:    mov.b v2[9], w8
-; CHECK-NEXT:    ldr w8, [sp, #144]
+; CHECK-NEXT:    ldr w8, [sp, #160]
 ; CHECK-NEXT:    mov.b v3[10], w14
-; CHECK-NEXT:    ldr w14, [sp, #280]
+; CHECK-NEXT:    ldr w14, [sp, #296]
 ; CHECK-NEXT:    mov.b v0[10], w13
-; CHECK-NEXT:    ldr w13, [sp, #296]
+; CHECK-NEXT:    ldr w13, [sp, #312]
 ; CHECK-NEXT:    mov.b v1[10], w8
-; CHECK-NEXT:    ldr w8, [sp, #24]
+; CHECK-NEXT:    ldr w8, [sp, #40]
 ; CHECK-NEXT:    mov.b v2[10], w10
-; CHECK-NEXT:    ldr w10, [sp, #152]
+; CHECK-NEXT:    ldr w10, [sp, #168]
 ; CHECK-NEXT:    mov.b v3[11], w11
-; CHECK-NEXT:    ldr w11, [sp, #288]
+; CHECK-NEXT:    ldr w11, [sp, #304]
 ; CHECK-NEXT:    mov.b v0[11], w8
-; CHECK-NEXT:    ldr w8, [sp, #32]
+; CHECK-NEXT:    ldr w8, [sp, #48]
 ; CHECK-NEXT:    mov.b v1[11], w10
-; CHECK-NEXT:    ldr w10, [sp, #160]
+; CHECK-NEXT:    ldr w10, [sp, #176]
 ; CHECK-NEXT:    mov.b v2[11], w14
 ; CHECK-NEXT:    mov.b v3[12], w15
 ; CHECK-NEXT:    mov.b v0[12], w8
-; CHECK-NEXT:    ldr w8, [sp, #40]
+; CHECK-NEXT:    ldr w8, [sp, #56]
 ; CHECK-NEXT:    mov.b v1[12], w10
-; CHECK-NEXT:    ldr w10, [sp, #168]
+; CHECK-NEXT:    ldr w10, [sp, #184]
 ; CHECK-NEXT:    mov.b v2[12], w11
-; CHECK-NEXT:    ldr w11, [sp, #312]
+; CHECK-NEXT:    ldr w11, [sp, #328]
 ; CHECK-NEXT:    mov.b v3[13], w9
-; CHECK-NEXT:    ldr w9, [sp, #304]
+; CHECK-NEXT:    ldr w9, [sp, #320]
 ; CHECK-NEXT:    mov.b v0[13], w8
-; CHECK-NEXT:    ldr w8, [sp, #48]
+; CHECK-NEXT:    ldr w8, [sp, #64]
 ; CHECK-NEXT:    mov.b v1[13], w10
-; CHECK-NEXT:    ldr w10, [sp, #176]
+; CHECK-NEXT:    ldr w10, [sp, #192]
 ; CHECK-NEXT:    mov.b v2[13], w13
 ; CHECK-NEXT:    mov.b v3[14], w16
 ; CHECK-NEXT:    mov.b v0[14], w8
-; CHECK-NEXT:    ldr w8, [sp, #56]
+; CHECK-NEXT:    ldr w8, [sp, #72]
 ; CHECK-NEXT:    mov.b v1[14], w10
 ; CHECK-NEXT:    mov.b v2[14], w9
-; CHECK-NEXT:    ldr w9, [sp, #184]
+; CHECK-NEXT:    ldr w9, [sp, #200]
 ; CHECK-NEXT:    movi.16b v4, #1
 ; CHECK-NEXT:    mov.b v0[15], w8
 ; CHECK-NEXT:    mov.b v1[15], w9
 ; CHECK-NEXT:    mov.b v2[15], w11
 ; CHECK-NEXT:    mov.b v3[15], w12
 ; CHECK-NEXT:    and.16b v0, v0, v4
 ; CHECK-NEXT:    and.16b v1, v1, v4
 ; CHECK-NEXT:    and.16b v2, v2, v4
 ; CHECK-NEXT:    and.16b v3, v3, v4
+; CHECK-NEXT:    ldr x29, [sp], #16 // 8-byte Folded Reload
 ; CHECK-NEXT:    ret
   %res = zext <64 x i1> %arg to <64 x i8>
   ret <64 x i8> %res
 }
 
 define <64 x i8> @sext_v64i1(<64 x i1> %arg) {
 ; CHECK-LABEL: sext_v64i1:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ldr w8, [sp, #320]
+; CHECK-NEXT:    str x29, [sp, #-16]! // 8-byte Folded Spill
+; CHECK-NEXT:    .cfi_def_cfa_offset 16
+; CHECK-NEXT:    .cfi_offset w29, -16
+; CHECK-NEXT:    ldr w8, [sp, #336]
 ; CHECK-NEXT:    fmov s3, w0
-; CHECK-NEXT:    ldr w9, [sp, #64]
-; CHECK-NEXT:    ldr w10, [sp, #192]
+; CHECK-NEXT:    ldr w9, [sp, #80]
+; CHECK-NEXT:    ldr w10, [sp, #208]
 ; CHECK-NEXT:    fmov s0, w8
-; CHECK-NEXT:    ldr w8, [sp, #328]
+; CHECK-NEXT:    ldr w8, [sp, #344]
 ; CHECK-NEXT:    fmov s1, w9
-; CHECK-NEXT:    ldr w9, [sp, #72]
+; CHECK-NEXT:    ldr w9, [sp, #88]
 ; CHECK-NEXT:    fmov s2, w10
-; CHECK-NEXT:    ldr w10, [sp, #80]
+; CHECK-NEXT:    ldr w10, [sp, #96]
 ; CHECK-NEXT:    mov.b v0[1], w8
-; CHECK-NEXT:    ldr w8, [sp, #200]
+; CHECK-NEXT:    ldr w8, [sp, #216]
 ; CHECK-NEXT:    mov.b v1[1], w9
-; CHECK-NEXT:    ldr w9, [sp, #336]
+; CHECK-NEXT:    ldr w9, [sp, #352]
 ; CHECK-NEXT:    mov.b v3[1], w1
-; CHECK-NEXT:    ldr w11, [sp, #88]
+; CHECK-NEXT:    ldr w11, [sp, #104]
 ; CHECK-NEXT:    mov.b v2[1], w8
-; CHECK-NEXT:    ldr w8, [sp, #344]
+; CHECK-NEXT:    ldr w8, [sp, #360]
 ; CHECK-NEXT:    mov.b v0[2], w9
-; CHECK-NEXT:    ldr w9, [sp, #208]
+; CHECK-NEXT:    ldr w9, [sp, #224]
 ; CHECK-NEXT:    mov.b v1[2], w10
-; CHECK-NEXT:    ldr w10, [sp, #352]
+; CHECK-NEXT:    ldr w10, [sp, #368]
 ; CHECK-NEXT:    mov.b v3[2], w2
-; CHECK-NEXT:    ldr w12, [sp, #96]
+; CHECK-NEXT:    ldr w12, [sp, #112]
 ; CHECK-NEXT:    mov.b v2[2], w9
-; CHECK-NEXT:    ldr w9, [sp, #360]
+; CHECK-NEXT:    ldr w9, [sp, #376]
 ; CHECK-NEXT:    mov.b v0[3], w8
-; CHECK-NEXT:    ldr w8, [sp, #216]
+; CHECK-NEXT:    ldr w8, [sp, #232]
 ; CHECK-NEXT:    mov.b v1[3], w11
-; CHECK-NEXT:    ldr w13, [sp, #104]
+; CHECK-NEXT:    ldr w13, [sp, #120]
 ; CHECK-NEXT:    mov.b v3[3], w3
-; CHECK-NEXT:    ldr w11, [sp, #368]
+; CHECK-NEXT:    ldr w11, [sp, #384]
 ; CHECK-NEXT:    mov.b v2[3], w8
-; CHECK-NEXT:    ldr w14, [sp, #112]
+; CHECK-NEXT:    ldr w14, [sp, #128]
 ; CHECK-NEXT:    mov.b v0[4], w10
-; CHECK-NEXT:    ldr w10, [sp, #224]
+; CHECK-NEXT:    ldr w10, [sp, #240]
 ; CHECK-NEXT:    mov.b v1[4], w12
-; CHECK-NEXT:    ldr w8, [sp, #376]
+; CHECK-NEXT:    ldr w8, [sp, #392]
 ; CHECK-NEXT:    mov.b v3[4], w4
-; CHECK-NEXT:    ldr w15, [sp, #120]
+; CHECK-NEXT:    ldr w15, [sp, #136]
 ; CHECK-NEXT:    mov.b v2[4], w10
-; CHECK-NEXT:    ldr w12, [sp, #384]
+; CHECK-NEXT:    ldr w12, [sp, #400]
 ; CHECK-NEXT:    mov.b v0[5], w9
-; CHECK-NEXT:    ldr w9, [sp, #232]
+; CHECK-NEXT:    ldr w9, [sp, #248]
 ; CHECK-NEXT:    mov.b v1[5], w13
-; CHECK-NEXT:    ldr w16, [sp, #128]
+; CHECK-NEXT:    ldr w16, [sp, #144]
 ; CHECK-NEXT:    mov.b v3[5], w5
-; CHECK-NEXT:    ldr w10, [sp, #392]
+; CHECK-NEXT:    ldr w10, [sp, #408]
 ; CHECK-NEXT:    mov.b v2[5], w9
-; CHECK-NEXT:    ldr w13, [sp, #400]
+; CHECK-NEXT:    ldr w13, [sp, #416]
 ; CHECK-NEXT:    mov.b v0[6], w11
-; CHECK-NEXT:    ldr w11, [sp, #240]
+; CHECK-NEXT:    ldr w11, [sp, #256]
 ; CHECK-NEXT:    mov.b v1[6], w14
-; CHECK-NEXT:    ldr w9, [sp, #408]
+; CHECK-NEXT:    ldr w9, [sp, #424]
 ; CHECK-NEXT:    mov.b v3[6], w6
-; CHECK-NEXT:    ldr w14, [sp, #416]
+; CHECK-NEXT:    ldr w14, [sp, #432]
 ; CHECK-NEXT:    mov.b v2[6], w11
-; CHECK-NEXT:    ldr w11, [sp, #424]
+; CHECK-NEXT:    ldr w11, [sp, #440]
 ; CHECK-NEXT:    mov.b v0[7], w8
-; CHECK-NEXT:    ldr w8, [sp, #248]
+; CHECK-NEXT:    ldr w8, [sp, #264]
 ; CHECK-NEXT:    mov.b v1[7], w15
-; CHECK-NEXT:    ldr w15, [sp, #432]
+; CHECK-NEXT:    ldr w15, [sp, #448]
 ; CHECK-NEXT:    mov.b v3[7], w7
 ; CHECK-NEXT:    mov.b v2[7], w8
-; CHECK-NEXT:    ldr w8, [sp]
+; CHECK-NEXT:    ldr w8, [sp, #16]
 ; CHECK-NEXT:    mov.b v0[8], w12
-; CHECK-NEXT:    ldr w12, [sp, #256]
+; CHECK-NEXT:    ldr w12, [sp, #272]
 ; CHECK-NEXT:    mov.b v1[8], w16
-; CHECK-NEXT:    ldr w16, [sp, #440]
+; CHECK-NEXT:    ldr w16, [sp, #456]
 ; CHECK-NEXT:    mov.b v3[8], w8
-; CHECK-NEXT:    ldr w8, [sp, #136]
+; CHECK-NEXT:    ldr w8, [sp, #152]
 ; CHECK-NEXT:    mov.b v2[8], w12
-; CHECK-NEXT:    ldr w12, [sp, #8]
+; CHECK-NEXT:    ldr w12, [sp, #24]
 ; CHECK-NEXT:    mov.b v0[9], w10
-; CHECK-NEXT:    ldr w10, [sp, #264]
+; CHECK-NEXT:    ldr w10, [sp, #280]
 ; CHECK-NEXT:    mov.b v1[9], w8
-; CHECK-NEXT:    ldr w8, [sp, #272]
+; CHECK-NEXT:    ldr w8, [sp, #288]
 ; CHECK-NEXT:    mov.b v3[9], w12
-; CHECK-NEXT:    ldr w12, [sp, #144]
+; CHECK-NEXT:    ldr w12, [sp, #160]
 ; CHECK-NEXT:    mov.b v2[9], w10
-; CHECK-NEXT:    ldr w10, [sp, #16]
+; CHECK-NEXT:    ldr w10, [sp, #32]
 ; CHECK-NEXT:    mov.b v0[10], w13
-; CHECK-NEXT:    ldr w13, [sp, #280]
+; CHECK-NEXT:    ldr w13, [sp, #296]
 ; CHECK-NEXT:    mov.b v1[10], w12
-; CHECK-NEXT:    ldr w12, [sp, #152]
+; CHECK-NEXT:    ldr w12, [sp, #168]
 ; CHECK-NEXT:    mov.b v3[10], w10
-; CHECK-NEXT:    ldr w10, [sp, #160]
+; CHECK-NEXT:    ldr w10, [sp, #176]
 ; CHECK-NEXT:    mov.b v2[10], w8
-; CHECK-NEXT:    ldr w8, [sp, #24]
+; CHECK-NEXT:    ldr w8, [sp, #40]
 ; CHECK-NEXT:    mov.b v0[11], w9
-; CHECK-NEXT:    ldr w9, [sp, #288]
+; CHECK-NEXT:    ldr w9, [sp, #304]
 ; CHECK-NEXT:    mov.b v1[11], w12
-; CHECK-NEXT:    ldr w12, [sp, #296]
+; CHECK-NEXT:    ldr w12, [sp, #312]
 ; CHECK-NEXT:    mov.b v3[11], w8
-; CHECK-NEXT:    ldr w8, [sp, #32]
+; CHECK-NEXT:    ldr w8, [sp, #48]
 ; CHECK-NEXT:    mov.b v2[11], w13
 ; CHECK-NEXT:    mov.b v0[12], w14
 ; CHECK-NEXT:    mov.b v1[12], w10
-; CHECK-NEXT:    ldr w10, [sp, #168]
+; CHECK-NEXT:    ldr w10, [sp, #184]
 ; CHECK-NEXT:    mov.b v3[12], w8
-; CHECK-NEXT:    ldr w8, [sp, #40]
+; CHECK-NEXT:    ldr w8, [sp, #56]
 ; CHECK-NEXT:    mov.b v2[12], w9
-; CHECK-NEXT:    ldr w9, [sp, #304]
+; CHECK-NEXT:    ldr w9, [sp, #320]
 ; CHECK-NEXT:    mov.b v0[13], w11
-; CHECK-NEXT:    ldr w11, [sp, #312]
+; CHECK-NEXT:    ldr w11, [sp, #328]
 ; CHECK-NEXT:    mov.b v1[13], w10
-; CHECK-NEXT:    ldr w10, [sp, #176]
+; CHECK-NEXT:    ldr w10, [sp, #192]
 ; CHECK-NEXT:    mov.b v3[13], w8
-; CHECK-NEXT:    ldr w8, [sp, #48]
+; CHECK-NEXT:    ldr w8, [sp, #64]
 ; CHECK-NEXT:    mov.b v2[13], w12
 ; CHECK-NEXT:    mov.b v0[14], w15
 ; CHECK-NEXT:    mov.b v1[14], w10
-; CHECK-NEXT:    ldr w10, [sp, #184]
+; CHECK-NEXT:    ldr w10, [sp, #200]
 ; CHECK-NEXT:    mov.b v3[14], w8
-; CHECK-NEXT:    ldr w8, [sp, #56]
+; CHECK-NEXT:    ldr w8, [sp, #72]
 ; CHECK-NEXT:    mov.b v2[14], w9
 ; CHECK-NEXT:    mov.b v0[15], w16
 ; CHECK-NEXT:    mov.b v1[15], w10
 ; CHECK-NEXT:    mov.b v3[15], w8
 ; CHECK-NEXT:    mov.b v2[15], w11
 ; CHECK-NEXT:    shl.16b v4, v0, #7
 ; CHECK-NEXT:    shl.16b v1, v1, #7
 ; CHECK-NEXT:    shl.16b v3, v3, #7
 ; CHECK-NEXT:    shl.16b v2, v2, #7
 ; CHECK-NEXT:    cmlt.16b v0, v3, #0
 ; CHECK-NEXT:    cmlt.16b v1, v1, #0
 ; CHECK-NEXT:    cmlt.16b v2, v2, #0
 ; CHECK-NEXT:    cmlt.16b v3, v4, #0
+; CHECK-NEXT:    ldr x29, [sp], #16 // 8-byte Folded Reload
 ; CHECK-NEXT:    ret
   %res = sext <64 x i1> %arg to <64 x i8>
   ret <64 x i8> %res
 }
 
 define <1 x i128> @sext_v1x64(<1 x i64> %arg) {
   ; X0 & X1 are the real return registers, SDAG messes with v0 too for unknown reasons.
 ; CHECKDAG-LABEL: sext_v1x64:

llvm/test/CodeGen/AArch64/neon-dotreduce.ll

entry:
   %5 = call i32 @llvm.vector.reduce.add.v33i32(<33 x i32> %4)
   %op.extra = add nsw i32 %5, %sum
   ret i32 %op.extra
 }
 
 define i32 @test_sdot_v33i8_double(<33 x i8> %a, <33 x i8> %b, <33 x i8> %c, <33 x i8> %d) {
 ; CHECK-LABEL: test_sdot_v33i8_double:
 ; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    ldr b0, [sp, #64]
-; CHECK-NEXT:    add x8, sp, #72
-; CHECK-NEXT:    ldr b3, [sp]
-; CHECK-NEXT:    add x9, sp, #136
-; CHECK-NEXT:    ldr b2, [sp, #128]
-; CHECK-NEXT:    add x10, sp, #88
+; CHECK-NEXT:    str x29, [sp, #-16]! // 8-byte Folded Spill
+; CHECK-NEXT:    .cfi_def_cfa_offset 16
+; CHECK-NEXT:    .cfi_offset w29, -16
+; CHECK-NEXT:    ldr b0, [sp, #80]
+; CHECK-NEXT:    add x8, sp, #88
+; CHECK-NEXT:    ldr b3, [sp, #16]
+; CHECK-NEXT:    add x9, sp, #152
+; CHECK-NEXT:    ldr b2, [sp, #144]
+; CHECK-NEXT:    add x10, sp, #104
 ; CHECK-NEXT:    ld1 { v0.b }[1], [x8]
-; CHECK-NEXT:    add x8, sp, #80
-; CHECK-NEXT:    ldr b4, [sp, #328]
-; CHECK-NEXT:    add x11, sp, #352
+; CHECK-NEXT:    add x8, sp, #96
+; CHECK-NEXT:    ldr b4, [sp, #344]
+; CHECK-NEXT:    add x11, sp, #368
 ; CHECK-NEXT:    ld1 { v2.b }[1], [x9]
-; CHECK-NEXT:    add x9, sp, #144
-; CHECK-NEXT:    ldr b7, [sp, #200]
+; CHECK-NEXT:    add x9, sp, #160
+; CHECK-NEXT:    ldr b7, [sp, #216]
 ; CHECK-NEXT:    fmov s1, w0
 ; CHECK-NEXT:    ld1 { v0.b }[2], [x8]
-; CHECK-NEXT:    add x8, sp, #8
-; CHECK-NEXT:    ldr b17, [sp, #264]
+; CHECK-NEXT:    add x8, sp, #24
+; CHECK-NEXT:    ldr b17, [sp, #280]
 ; CHECK-NEXT:    ld1 { v2.b }[2], [x9]
-; CHECK-NEXT:    add x9, sp, #152
+; CHECK-NEXT:    add x9, sp, #168
 ; CHECK-NEXT:    ld1 { v3.b }[1], [x8]
-; CHECK-NEXT:    add x8, sp, #16
+; CHECK-NEXT:    add x8, sp, #32
 ; CHECK-NEXT:    ld1 { v0.b }[3], [x10]
-; CHECK-NEXT:    add x10, sp, #96
-; CHECK-NEXT:    ldr b16, [sp, #392]
+; CHECK-NEXT:    add x10, sp, #112
+; CHECK-NEXT:    ldr b16, [sp, #408]
 ; CHECK-NEXT:    ld1 { v2.b }[3], [x9]
-; CHECK-NEXT:    add x9, sp, #160
+; CHECK-NEXT:    add x9, sp, #176
 ; CHECK-NEXT:    ld1 { v3.b }[2], [x8]
-; CHECK-NEXT:    add x8, sp, #24
+; CHECK-NEXT:    add x8, sp, #40
 ; CHECK-NEXT:    ld1 { v0.b }[4], [x10]
-; CHECK-NEXT:    add x10, sp, #104
-; CHECK-NEXT:    ldr b5, [sp, #192]
+; CHECK-NEXT:    add x10, sp, #120
+; CHECK-NEXT:    ldr b5, [sp, #208]
 ; CHECK-NEXT:    ld1 { v2.b }[4], [x9]
-; CHECK-NEXT:    add x9, sp, #168
+; CHECK-NEXT:    add x9, sp, #184
 ; CHECK-NEXT:    ld1 { v3.b }[3], [x8]
-; CHECK-NEXT:    add x8, sp, #32
+; CHECK-NEXT:    add x8, sp, #48
 ; CHECK-NEXT:    ld1 { v0.b }[5], [x10]
-; CHECK-NEXT:    add x10, sp, #112
+; CHECK-NEXT:    add x10, sp, #128
 ; CHECK-NEXT:    sshll v5.8h, v5.8b, #0
-; CHECK-NEXT:    ldr b20, [sp, #856]
+; CHECK-NEXT:    ldr b20, [sp, #872]
 ; CHECK-NEXT:    ld1 { v2.b }[5], [x9]
-; CHECK-NEXT:    add x9, sp, #176
+; CHECK-NEXT:    add x9, sp, #192
 ; CHECK-NEXT:    ld1 { v3.b }[4], [x8]
-; CHECK-NEXT:    add x8, sp, #40
+; CHECK-NEXT:    add x8, sp, #56
 ; CHECK-NEXT:    ld1 { v0.b }[6], [x10]
-; CHECK-NEXT:    add x10, sp, #48
+; CHECK-NEXT:    add x10, sp, #64
 ; CHECK-NEXT:    mov v1.b[1], w1
-; CHECK-NEXT:    ldr b21, [sp, #728]
+; CHECK-NEXT:    ldr b21, [sp, #744]
 ; CHECK-NEXT:    ld1 { v2.b }[6], [x9]
-; CHECK-NEXT:    add x9, sp, #184
+; CHECK-NEXT:    add x9, sp, #200
 ; CHECK-NEXT:    ld1 { v3.b }[5], [x8]
-; CHECK-NEXT:    add x8, sp, #120
+; CHECK-NEXT:    add x8, sp, #136
 ; CHECK-NEXT:    mov v1.b[2], w2
 ; CHECK-NEXT:    ld1 { v0.b }[7], [x8]
-; CHECK-NEXT:    add x8, sp, #56
+; CHECK-NEXT:    add x8, sp, #72
 ; CHECK-NEXT:    ld1 { v3.b }[6], [x10]
-; CHECK-NEXT:    add x10, sp, #272
+; CHECK-NEXT:    add x10, sp, #288
 ; CHECK-NEXT:    ld1 { v2.b }[7], [x9]
-; CHECK-NEXT:    add x9, sp, #208
+; CHECK-NEXT:    add x9, sp, #224
 ; CHECK-NEXT:    mov v1.b[3], w3
 ; CHECK-NEXT:    ld1 { v17.b }[1], [x10]
-; CHECK-NEXT:    add x10, sp, #224
+; CHECK-NEXT:    add x10, sp, #240
 ; CHECK-NEXT:    ld1 { v3.b }[7], [x8]
-; CHECK-NEXT:    add x8, sp, #336
+; CHECK-NEXT:    add x8, sp, #352
 ; CHECK-NEXT:    ld1 { v7.b }[1], [x9]
-; CHECK-NEXT:    add x9, sp, #400
+; CHECK-NEXT:    add x9, sp, #416
 ; CHECK-NEXT:    sshll v19.8h, v2.8b, #0
-; CHECK-NEXT:    ldr b2, [sp, #456]
+; CHECK-NEXT:    ldr b2, [sp, #472]
 ; CHECK-NEXT:    ld1 { v4.b }[1], [x8]
-; CHECK-NEXT:    add x8, sp, #344
+; CHECK-NEXT:    add x8, sp, #360
 ; CHECK-NEXT:    ld1 { v16.b }[1], [x9]
-; CHECK-NEXT:    add x9, sp, #408
+; CHECK-NEXT:    add x9, sp, #424
 ; CHECK-NEXT:    sshll v3.8h, v3.8b, #0
 ; CHECK-NEXT:    mov v1.b[4], w4
 ; CHECK-NEXT:    ld1 { v4.b }[2], [x8]
-; CHECK-NEXT:    add x8, sp, #216
+; CHECK-NEXT:    add x8, sp, #232
 ; CHECK-NEXT:    ld1 { v16.b }[2], [x9]
-; CHECK-NEXT:    add x9, sp, #416
+; CHECK-NEXT:    add x9, sp, #432
 ; CHECK-NEXT:    sshll v0.8h, v0.8b, #0
 ; CHECK-NEXT:    ld1 { v7.b }[2], [x8]
-; CHECK-NEXT:    add x8, sp, #280
+; CHECK-NEXT:    add x8, sp, #296
 ; CHECK-NEXT:    ld1 { v4.b }[3], [x11]
-; CHECK-NEXT:    add x11, sp, #360
+; CHECK-NEXT:    add x11, sp, #376
 ; CHECK-NEXT:    ld1 { v16.b }[3], [x9]
-; CHECK-NEXT:    add x9, sp, #424
+; CHECK-NEXT:    add x9, sp, #440
 ; CHECK-NEXT:    ld1 { v17.b }[2], [x8]
-; CHECK-NEXT:    add x8, sp, #288
+; CHECK-NEXT:    add x8, sp, #304
 ; CHECK-NEXT:    ld1 { v7.b }[3], [x10]
-; CHECK-NEXT:    add x10, sp, #232
+; CHECK-NEXT:    add x10, sp, #248
 ; CHECK-NEXT:    ld1 { v4.b }[4], [x11]
-; CHECK-NEXT:    add x11, sp, #368
+; CHECK-NEXT:    add x11, sp, #384
 ; CHECK-NEXT:    ld1 { v16.b }[4], [x9]
-; CHECK-NEXT:    add x9, sp, #432
+; CHECK-NEXT:    add x9, sp, #448
 ; CHECK-NEXT:    ld1 { v17.b }[3], [x8]
-; CHECK-NEXT:    add x8, sp, #296
+; CHECK-NEXT:    add x8, sp, #312
 ; CHECK-NEXT:    ld1 { v7.b }[4], [x10]
-; CHECK-NEXT:    add x10, sp, #240
+; CHECK-NEXT:    add x10, sp, #256
 ; CHECK-NEXT:    ld1 { v4.b }[5], [x11]
-; CHECK-NEXT:    add x11, sp, #376
+; CHECK-NEXT:    add x11, sp, #392
 ; CHECK-NEXT:    ld1 { v16.b }[5], [x9]
-; CHECK-NEXT:    add x9, sp, #440
+; CHECK-NEXT:    add x9, sp, #456
 ; CHECK-NEXT:    ld1 { v17.b }[4], [x8]
-; CHECK-NEXT:    add x8, sp, #304
+; CHECK-NEXT:    add x8, sp, #320
 ; CHECK-NEXT:    ld1 { v7.b }[5], [x10]
-; CHECK-NEXT:    add x10, sp, #248
+; CHECK-NEXT:    add x10, sp, #264
 ; CHECK-NEXT:    ld1 { v4.b }[6], [x11]
-; CHECK-NEXT:    add x11, sp, #384
+; CHECK-NEXT:    add x11, sp, #400
 ; CHECK-NEXT:    ld1 { v16.b }[6], [x9]
-; CHECK-NEXT:    add x9, sp, #448
+; CHECK-NEXT:    add x9, sp, #464
 ; CHECK-NEXT:    ld1 { v17.b }[5], [x8]
-; CHECK-NEXT:    add x8, sp, #312
+; CHECK-NEXT:    add x8, sp, #328
 ; CHECK-NEXT:    ld1 { v7.b }[6], [x10]
-; CHECK-NEXT:    add x10, sp, #256
+; CHECK-NEXT:    add x10, sp, #272
 ; CHECK-NEXT:    ld1 { v4.b }[7], [x11]
-; CHECK-NEXT:    add x11, sp, #480
+; CHECK-NEXT:    add x11, sp, #496
 ; CHECK-NEXT:    ld1 { v16.b }[7], [x9]
-; CHECK-NEXT:    add x9, sp, #472
+; CHECK-NEXT:    add x9, sp, #488
 ; CHECK-NEXT:    ld1 { v17.b }[6], [x8]
-; CHECK-NEXT:    add x8, sp, #320
+; CHECK-NEXT:    add x8, sp, #336
 ; CHECK-NEXT:    ld1 { v7.b }[7], [x10]
-; CHECK-NEXT:    add x10, sp, #624
+; CHECK-NEXT:    add x10, sp, #640
 ; CHECK-NEXT:    sshll v6.8h, v4.8b, #0
 ; CHECK-NEXT:    sshll v16.8h, v16.8b, #0
 ; CHECK-NEXT:    ld1 { v17.b }[7], [x8]
-; CHECK-NEXT:    add x8, sp, #600
+; CHECK-NEXT:    add x8, sp, #616
 ; CHECK-NEXT:    sshll v18.8h, v7.8b, #0
-; CHECK-NEXT:    ldr b7, [sp, #592]
+; CHECK-NEXT:    ldr b7, [sp, #608]
 ; CHECK-NEXT:    mov v1.b[5], w5
 ; CHECK-NEXT:    ld1 { v7.b }[1], [x8]
-; CHECK-NEXT:    add x8, sp, #608
+; CHECK-NEXT:    add x8, sp, #624
 ; CHECK-NEXT:    sshll v4.8h, v17.8b, #0
 ; CHECK-NEXT:    sshll v17.8h, v2.8b, #0
 ; CHECK-NEXT:    smull2 v2.4s, v3.8h, v4.8h
 ; CHECK-NEXT:    smull v3.4s, v3.4h, v4.4h
 ; CHECK-NEXT:    ld1 { v7.b }[2], [x8]
-; CHECK-NEXT:    add x8, sp, #616
+; CHECK-NEXT:    add x8, sp, #632
 ; CHECK-NEXT:    smull v4.4s, v5.4h, v17.4h
-; CHECK-NEXT:    ldr b17, [sp, #528]
+; CHECK-NEXT:    ldr b17, [sp, #544]
 ; CHECK-NEXT:    smlal2 v2.4s, v19.8h, v16.8h
 ; CHECK-NEXT:    smlal v3.4s, v19.4h, v16.4h
-; CHECK-NEXT:    ldr b16, [sp, #464]
+; CHECK-NEXT:    ldr b16, [sp, #480]
 ; CHECK-NEXT:    ld1 { v7.b }[3], [x8]
-; CHECK-NEXT:    add x8, sp, #536
-; CHECK-NEXT:    ldr b19, [sp, #656]
+; CHECK-NEXT:    add x8, sp, #552
+; CHECK-NEXT:    ldr b19, [sp, #672]
 ; CHECK-NEXT:    ld1 { v16.b }[1], [x9]
-; CHECK-NEXT:    add x9, sp, #664
+; CHECK-NEXT:    add x9, sp, #680
 ; CHECK-NEXT:    ld1 { v17.b }[1], [x8]
-; CHECK-NEXT:    add x8, sp, #544
+; CHECK-NEXT:    add x8, sp, #560
 ; CHECK-NEXT:    ld1 { v7.b }[4], [x10]
-; CHECK-NEXT:    add x10, sp, #632
+; CHECK-NEXT:    add x10, sp, #648
 ; CHECK-NEXT:    ld1 { v19.b }[1], [x9]
-; CHECK-NEXT:    add x9, sp, #672
+; CHECK-NEXT:    add x9, sp, #688
 ; CHECK-NEXT:    ld1 { v16.b }[2], [x11]
-; CHECK-NEXT:    add x11, sp, #488
+; CHECK-NEXT:    add x11, sp, #504
 ; CHECK-NEXT:    ld1 { v17.b }[2], [x8]
-; CHECK-NEXT:    add x8, sp, #552
+; CHECK-NEXT:    add x8, sp, #568
 ; CHECK-NEXT:    ld1 { v7.b }[5], [x10]
-; CHECK-NEXT:    add x10, sp, #640
+; CHECK-NEXT:    add x10, sp, #656
 ; CHECK-NEXT:    ld1 { v19.b }[2], [x9]
-; CHECK-NEXT:    add x9, sp, #680
+; CHECK-NEXT:    add x9, sp, #696
 ; CHECK-NEXT:    ld1 { v16.b }[3], [x11]
-; CHECK-NEXT:    add x11, sp, #496
+; CHECK-NEXT:    add x11, sp, #512
 ; CHECK-NEXT:    ld1 { v17.b }[3], [x8]
-; CHECK-NEXT:    add x8, sp, #560
+; CHECK-NEXT:    add x8, sp, #576
 ; CHECK-NEXT:    mov v4.s[1], wzr
 ; CHECK-NEXT:    ld1 { v7.b }[6], [x10]
 ; CHECK-NEXT:    ld1 { v19.b }[3], [x9]
-; CHECK-NEXT:    add x9, sp, #688
+; CHECK-NEXT:    add x9, sp, #704
 ; CHECK-NEXT:    ld1 { v16.b }[4], [x11]
-; CHECK-NEXT:    add x11, sp, #504
+; CHECK-NEXT:    add x11, sp, #520
 ; CHECK-NEXT:    ld1 { v17.b }[4], [x8]
-; CHECK-NEXT:    add x8, sp, #568
+; CHECK-NEXT:    add x8, sp, #584
 ; CHECK-NEXT:    mov v1.b[6], w6
-; CHECK-NEXT:    add x10, sp, #648
+; CHECK-NEXT:    add x10, sp, #664
 ; CHECK-NEXT:    ld1 { v19.b }[4], [x9]
-; CHECK-NEXT:    add x9, sp, #696
+; CHECK-NEXT:    add x9, sp, #712
 ; CHECK-NEXT:    mov v4.s[2], wzr
 ; CHECK-NEXT:    ld1 { v16.b }[5], [x11]
 ; CHECK-NEXT:    ld1 { v17.b }[5], [x8]
-; CHECK-NEXT:    add x8, sp, #864
+; CHECK-NEXT:    add x8, sp, #880
 ; CHECK-NEXT:    mov v1.b[7], w7
 ; CHECK-NEXT:    ld1 { v7.b }[7], [x10]
 ; CHECK-NEXT:    ld1 { v19.b }[5], [x9]
-; CHECK-NEXT:    add x9, sp, #704
+; CHECK-NEXT:    add x9, sp, #720
 ; CHECK-NEXT:    ld1 { v20.b }[1], [x8]
-; CHECK-NEXT:    add x10, sp, #512
-; CHECK-NEXT:    add x8, sp, #872
-; CHECK-NEXT:    add x11, sp, #576
+; CHECK-NEXT:    add x10, sp, #528
+; CHECK-NEXT:    add x8, sp, #888
+; CHECK-NEXT:    add x11, sp, #592
 ; CHECK-NEXT:    mov v4.s[3], wzr
 ; CHECK-NEXT:    ld1 { v19.b }[6], [x9]
-; CHECK-NEXT:    add x9, sp, #712
+; CHECK-NEXT:    add x9, sp, #728
 ; CHECK-NEXT:    ld1 { v16.b }[6], [x10]
-; CHECK-NEXT:    add x10, sp, #520
+; CHECK-NEXT:    add x10, sp, #536
 ; CHECK-NEXT:    ld1 { v20.b }[2], [x8]
-; CHECK-NEXT:    add x8, sp, #736
+; CHECK-NEXT:    add x8, sp, #752
 ; CHECK-NEXT:    sshll v1.8h, v1.8b, #0
 ; CHECK-NEXT:    ld1 { v17.b }[6], [x11]
 ; CHECK-NEXT:    ld1 { v19.b }[7], [x9]
-; CHECK-NEXT:    add x9, sp, #928
+; CHECK-NEXT:    add x9, sp, #944
 ; CHECK-NEXT:    smull2 v5.4s, v1.8h, v18.8h
 ; CHECK-NEXT:    ld1 { v21.b }[1], [x8]
 ; CHECK-NEXT:    smlal v4.4s, v1.4h, v18.4h
-; CHECK-NEXT:    ldr b1, [sp, #920]
-; CHECK-NEXT:    add x11, sp, #584
+; CHECK-NEXT:    ldr b1, [sp, #936]
+; CHECK-NEXT:    add x11, sp, #600
 ; CHECK-NEXT:    ld1 { v16.b }[7], [x10]
-; CHECK-NEXT:    add x10, sp, #880
-; CHECK-NEXT:    add x8, sp, #744
+; CHECK-NEXT:    add x10, sp, #896
+; CHECK-NEXT:    add x8, sp, #760
 ; CHECK-NEXT:    ld1 { v1.b }[1], [x9]
-; CHECK-NEXT:    add x9, sp, #936
+; CHECK-NEXT:    add x9, sp, #952
 ; CHECK-NEXT:    ld1 { v17.b }[7], [x11]
-; CHECK-NEXT:    add x11, sp, #800
+; CHECK-NEXT:    add x11, sp, #816
 ; CHECK-NEXT:    smlal2 v5.4s, v0.8h, v6.8h
 ; CHECK-NEXT:    ld1 { v20.b }[3], [x10]
 ; CHECK-NEXT:    smlal v4.4s, v0.4h, v6.4h
 ; CHECK-NEXT:    ld1 { v21.b }[2], [x8]
-; CHECK-NEXT:    ldr b0, [sp, #792]
-; CHECK-NEXT:    add x10, sp, #888
-; CHECK-NEXT:    add x8, sp, #752
+; CHECK-NEXT:    ldr b0, [sp, #808]
+; CHECK-NEXT:    add x10, sp, #904
+; CHECK-NEXT:    add x8, sp, #768
 ; CHECK-NEXT:    ld1 { v1.b }[2], [x9]
-; CHECK-NEXT:    add x9, sp, #944
-; CHECK-NEXT:    ldr b18, [sp, #720]
+; CHECK-NEXT:    add x9, sp, #960
+; CHECK-NEXT:    ldr b18, [sp, #736]
 ; CHECK-NEXT:    ld1 { v0.b }[1], [x11]
-; CHECK-NEXT:    add x11, sp, #808
+; CHECK-NEXT:    add x11, sp, #824
 ; CHECK-NEXT:    ld1 { v20.b }[4], [x10]
-; CHECK-NEXT:    add x10, sp, #896
+; CHECK-NEXT:    add x10, sp, #912
 ; CHECK-NEXT:    ld1 { v21.b }[3], [x8]
-; CHECK-NEXT:    add x8, sp, #760
+; CHECK-NEXT:    add x8, sp, #776
 ; CHECK-NEXT:    ld1 { v1.b }[3], [x9]
-; CHECK-NEXT:    add x9, sp, #952
+; CHECK-NEXT:    add x9, sp, #968
 ; CHECK-NEXT:    ld1 { v0.b }[2], [x11]
-; CHECK-NEXT:    add x11, sp, #816
+; CHECK-NEXT:    add x11, sp, #832
 ; CHECK-NEXT:    ld1 { v20.b }[5], [x10]
-; CHECK-NEXT:    add x10, sp, #904
+; CHECK-NEXT:    add x10, sp, #920
 ; CHECK-NEXT:    ld1 { v21.b }[4], [x8]
-; CHECK-NEXT:    add x8, sp, #768
+; CHECK-NEXT:    add x8, sp, #784
 ; CHECK-NEXT:    sshll v6.8h, v7.8b, #0
 ; CHECK-NEXT:    ld1 { v1.b }[4], [x9]
 ; CHECK-NEXT:    sshll v7.8h, v16.8b, #0
 ; CHECK-NEXT:    ld1 { v0.b }[3], [x11]
 ; CHECK-NEXT:    sshll v16.8h, v18.8b, #0
-; CHECK-NEXT:    ldr b18, [sp, #984]
+; CHECK-NEXT:    ldr b18, [sp, #1000]
 ; CHECK-NEXT:    ld1 { v20.b }[6], [x10]
-; CHECK-NEXT:    add x9, sp, #960
+; CHECK-NEXT:    add x9, sp, #976
 ; CHECK-NEXT:    ld1 { v21.b }[5], [x8]
-; CHECK-NEXT:    add x8, sp, #824
+; CHECK-NEXT:    add x8, sp, #840
 ; CHECK-NEXT:    sshll v18.8h, v18.8b, #0
-; CHECK-NEXT:    add x10, sp, #912
+; CHECK-NEXT:    add x10, sp, #928
 ; CHECK-NEXT:    ld1 { v1.b }[5], [x9]
-; CHECK-NEXT:    add x9, sp, #776
+; CHECK-NEXT:    add x9, sp, #792
 ; CHECK-NEXT:    smull v16.4s, v16.4h, v18.4h
 ; CHECK-NEXT:    ld1 { v0.b }[4], [x8]
 ; CHECK-NEXT:    ld1 { v20.b }[7], [x10]
-; CHECK-NEXT:    add x10, sp, #968
-; CHECK-NEXT:    add x8, sp, #832
+; CHECK-NEXT:    add x10, sp, #984
+; CHECK-NEXT:    add x8, sp, #848
 ; CHECK-NEXT:    ld1 { v21.b }[6], [x9]
-; CHECK-NEXT:    add x9, sp, #784
+; CHECK-NEXT:    add x9, sp, #800
 ; CHECK-NEXT:    mov v16.s[1], wzr
 ; CHECK-NEXT:    ld1 { v1.b }[6], [x10]
 ; CHECK-NEXT:    ld1 { v0.b }[5], [x8]
-; CHECK-NEXT:    add x10, sp, #976
-; CHECK-NEXT:    add x8, sp, #840
+; CHECK-NEXT:    add x10, sp, #992
+; CHECK-NEXT:    add x8, sp, #856
 ; CHECK-NEXT:    ld1 { v21.b }[7], [x9]
 ; CHECK-NEXT:    sshll v18.8h, v19.8b, #0
 ; CHECK-NEXT:    mov v16.s[2], wzr
 ; CHECK-NEXT:    ld1 { v1.b }[7], [x10]
 ; CHECK-NEXT:    ld1 { v0.b }[6], [x8]
-; CHECK-NEXT:    add x8, sp, #848
+; CHECK-NEXT:    add x8, sp, #864
 ; CHECK-NEXT:    sshll v19.8h, v20.8b, #0
 ; CHECK-NEXT:    sshll v21.8h, v21.8b, #0
 ; CHECK-NEXT:    mov v16.s[3], wzr
 ; CHECK-NEXT:    sshll v1.8h, v1.8b, #0
 ; CHECK-NEXT:    ld1 { v0.b }[7], [x8]
 ; CHECK-NEXT:    sshll v17.8h, v17.8b, #0
 ; CHECK-NEXT:    smull v20.4s, v18.4h, v1.4h
 ; CHECK-NEXT:    smull2 v1.4s, v18.8h, v1.8h
 ; CHECK-NEXT:    smull2 v18.4s, v6.8h, v19.8h
 ; CHECK-NEXT:    smlal v16.4s, v7.4h, v21.4h
 ; CHECK-NEXT:    sshll v0.8h, v0.8b, #0
 ; CHECK-NEXT:    smlal2 v18.4s, v7.8h, v21.8h
 ; CHECK-NEXT:    smlal2 v1.4s, v17.8h, v0.8h
 ; CHECK-NEXT:    smlal v20.4s, v17.4h, v0.4h
 ; CHECK-NEXT:    smlal v16.4s, v6.4h, v19.4h
 ; CHECK-NEXT:    add v0.4s, v5.4s, v2.4s
 ; CHECK-NEXT:    add v2.4s, v4.4s, v3.4s
 ; CHECK-NEXT:    add v1.4s, v18.4s, v1.4s
 ; CHECK-NEXT:    add v3.4s, v16.4s, v20.4s
 ; CHECK-NEXT:    add v0.4s, v2.4s, v0.4s
 ; CHECK-NEXT:    add v1.4s, v3.4s, v1.4s
 ; CHECK-NEXT:    add v0.4s, v0.4s, v1.4s
 ; CHECK-NEXT:    addv s0, v0.4s
 ; CHECK-NEXT:    fmov w0, s0
+; CHECK-NEXT:    ldr x29, [sp], #16 // 8-byte Folded Reload
 ; CHECK-NEXT:    ret
 entry:
   %az = sext <33 x i8> %a to <33 x i32>
   %bz = sext <33 x i8> %b to <33 x i32>
   %m1 = mul nuw nsw <33 x i32> %az, %bz
   %r1 = call i32 @llvm.vector.reduce.add.v33i32(<33 x i32> %m1)
   %cz = sext <33 x i8> %c to <33 x i32>
   %dz = sext <33 x i8> %d to <33 x i32>
   %m2 = mul nuw nsw <33 x i32> %cz, %dz
   %r2 = call i32 @llvm.vector.reduce.add.v33i32(<33 x i32> %m2)
   %x = add i32 %r1, %r2
   ret i32 %x
 }
 
 define i32 @test_sdot_v33i8_double_nomla(<33 x i8> %a, <33 x i8> %b, <33 x i8> %c, <33 x i8> %d) {
 ; CHECK-LABEL: test_sdot_v33i8_double_nomla:
 ; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    ldr b0, [sp, #64]
-; CHECK-NEXT:    add x8, sp, #72
-; CHECK-NEXT:    ldr b3, [sp, #128]
-; CHECK-NEXT:    add x10, sp, #80
-; CHECK-NEXT:    ldr b5, [sp]
-; CHECK-NEXT:    add x9, sp, #8
+; CHECK-NEXT:    str x29, [sp, #-16]! // 8-byte Folded Spill
+; CHECK-NEXT:    .cfi_def_cfa_offset 16
+; CHECK-NEXT:    .cfi_offset w29, -16
+; CHECK-NEXT:    ldr b0, [sp, #80]
+; CHECK-NEXT:    add x8, sp, #88
+; CHECK-NEXT:    ldr b3, [sp, #144]
+; CHECK-NEXT:    add x10, sp, #96
+; CHECK-NEXT:    ldr b5, [sp, #16]
+; CHECK-NEXT:    add x9, sp, #24
 ; CHECK-NEXT:    ld1 { v0.b }[1], [x8]
-; CHECK-NEXT:    add x8, sp, #136
+; CHECK-NEXT:    add x8, sp, #152
 ; CHECK-NEXT:    fmov s2, w0
-; CHECK-NEXT:    ldr b1, [sp, #192]
+; CHECK-NEXT:    ldr b1, [sp, #208]
 ; CHECK-NEXT:    ld1 { v5.b }[1], [x9]
-; CHECK-NEXT:    add x9, sp, #16
+; CHECK-NEXT:    add x9, sp, #32
 ; CHECK-NEXT:    ld1 { v3.b }[1], [x8]
-; CHECK-NEXT:    add x8, sp, #144
+; CHECK-NEXT:    add x8, sp, #160
 ; CHECK-NEXT:    ld1 { v0.b }[2], [x10]
-; CHECK-NEXT:    add x10, sp, #88
+; CHECK-NEXT:    add x10, sp, #104
 ; CHECK-NEXT:    mov v2.b[1], w1
-; CHECK-NEXT:    ldr b17, [sp, #656]
+; CHECK-NEXT:    ldr b17, [sp, #672]
 ; CHECK-NEXT:    ld1 { v5.b }[2], [x9]
-; CHECK-NEXT:    add x9, sp, #24
+; CHECK-NEXT:    add x9, sp, #40
 ; CHECK-NEXT:    ld1 { v3.b }[2], [x8]
-; CHECK-NEXT:    add x8, sp, #152
+; CHECK-NEXT:    add x8, sp, #168
 ; CHECK-NEXT:    ld1 { v0.b }[3], [x10]
-; CHECK-NEXT:    add x10, sp, #96
+; CHECK-NEXT:    add x10, sp, #112
 ; CHECK-NEXT:    mov v2.b[2], w2
-; CHECK-NEXT:    add x11, sp, #632
+; CHECK-NEXT:    add x11, sp, #648
 ; CHECK-NEXT:    ld1 { v5.b }[3], [x9]
-; CHECK-NEXT:    add x9, sp, #32
+; CHECK-NEXT:    add x9, sp, #48
 ; CHECK-NEXT:    ld1 { v3.b }[3], [x8]
-; CHECK-NEXT:    add x8, sp, #160
+; CHECK-NEXT:    add x8, sp, #176
 ; CHECK-NEXT:    ld1 { v0.b }[4], [x10]
-; CHECK-NEXT:    add x10, sp, #104
+; CHECK-NEXT:    add x10, sp, #120
 ; CHECK-NEXT:    mov v2.b[3], w3
 ; CHECK-NEXT:    ld1 { v5.b }[4], [x9]
-; CHECK-NEXT:    add x9, sp, #40
+; CHECK-NEXT:    add x9, sp, #56
 ; CHECK-NEXT:    ld1 { v3.b }[4], [x8]
-; CHECK-NEXT:    add x8, sp, #168
+; CHECK-NEXT:    add x8, sp, #184
 ; CHECK-NEXT:    ld1 { v0.b }[5], [x10]
-; CHECK-NEXT:    add x10, sp, #112
+; CHECK-NEXT:    add x10, sp, #128
 ; CHECK-NEXT:    mov v2.b[4], w4
 ; CHECK-NEXT:    ld1 { v5.b }[5], [x9]
-; CHECK-NEXT:    add x9, sp, #48
+; CHECK-NEXT:    add x9, sp, #64
 ; CHECK-NEXT:    ld1 { v3.b }[5], [x8]
-; CHECK-NEXT:    add x8, sp, #176
+; CHECK-NEXT:    add x8, sp, #192
 ; CHECK-NEXT:    ld1 { v0.b }[6], [x10]
-; CHECK-NEXT:    add x10, sp, #120
+; CHECK-NEXT:    add x10, sp, #136
 ; CHECK-NEXT:    sshll v4.8h, v1.8b, #0
 ; CHECK-NEXT:    mov v2.b[5], w5
 ; CHECK-NEXT:    ld1 { v5.b }[6], [x9]
 ; CHECK-NEXT:    ld1 { v3.b }[6], [x8]
-; CHECK-NEXT:    add x8, sp, #184
+; CHECK-NEXT:    add x8, sp, #200
 ; CHECK-NEXT:    ld1 { v0.b }[7], [x10]
-; CHECK-NEXT:    add x9, sp, #56
-; CHECK-NEXT:    add x10, sp, #664
+; CHECK-NEXT:    add x9, sp, #72
+; CHECK-NEXT:    add x10, sp, #680
 ; CHECK-NEXT:    mov v2.b[6], w6
 ; CHECK-NEXT:    ld1 { v3.b }[7], [x8]
-; CHECK-NEXT:    add x8, sp, #600
+; CHECK-NEXT:    add x8, sp, #616
 ; CHECK-NEXT:    sshll v1.8h, v0.8b, #0
-; CHECK-NEXT:    ldr b0, [sp, #592]
+; CHECK-NEXT:    ldr b0, [sp, #608]
 ; CHECK-NEXT:    ld1 { v5.b }[7], [x9]
-; CHECK-NEXT:    add x9, sp, #472
+; CHECK-NEXT:    add x9, sp, #488
 ; CHECK-NEXT:    mov v2.b[7], w7
 ; CHECK-NEXT:    ld1 { v17.b }[1], [x10]
 ; CHECK-NEXT:    ld1 { v0.b }[1], [x8]
-; CHECK-NEXT:    add x8, sp, #608
+; CHECK-NEXT:    add x8, sp, #624
 ; CHECK-NEXT:    sshll v7.8h, v3.8b, #0
-; CHECK-NEXT:    add x10, sp, #488
+; CHECK-NEXT:    add x10, sp, #504
 ; CHECK-NEXT:    sshll v16.8h, v5.8b, #0
-; CHECK-NEXT:    ldr b5, [sp, #464]
+; CHECK-NEXT:    ldr b5, [sp, #480]
 ; CHECK-NEXT:    sshll v6.8h, v2.8b, #0
 ; CHECK-NEXT:    ld1 { v0.b }[2], [x8]
-; CHECK-NEXT:    add x8, sp, #616
+; CHECK-NEXT:    add x8, sp, #632
 ; CHECK-NEXT:    saddl2 v2.4s, v16.8h, v7.8h
 ; CHECK-NEXT:    ld1 { v5.b }[1], [x9]
 ; CHECK-NEXT:    saddl v7.4s, v16.4h, v7.4h
-; CHECK-NEXT:    ldr b16, [sp, #528]
-; CHECK-NEXT:    add x9, sp, #536
+; CHECK-NEXT:    ldr b16, [sp, #544]
+; CHECK-NEXT:    add x9, sp, #552
 ; CHECK-NEXT:    ld1 { v0.b }[3], [x8]
-; CHECK-NEXT:    add x8, sp, #624
+; CHECK-NEXT:    add x8, sp, #640
 ; CHECK-NEXT:    sshll v4.4s, v4.4h, #0
 ; CHECK-NEXT:    ld1 { v16.b }[1], [x9]
-; CHECK-NEXT:    add x9, sp, #544
+; CHECK-NEXT:    add x9, sp, #560
 ; CHECK-NEXT:    mov v4.s[1], wzr
 ; CHECK-NEXT:    ld1 { v0.b }[4], [x8]
-; CHECK-NEXT:    add x8, sp, #480
+; CHECK-NEXT:    add x8, sp, #496
 ; CHECK-NEXT:    saddl2 v3.4s, v6.8h, v1.8h
 ; CHECK-NEXT:    ld1 { v16.b }[2], [x9]
-; CHECK-NEXT:    add x9, sp, #552
+; CHECK-NEXT:    add x9, sp, #568
 ; CHECK-NEXT:    ld1 { v5.b }[2], [x8]
-; CHECK-NEXT:    add x8, sp, #672
+; CHECK-NEXT:    add x8, sp, #688
 ; CHECK-NEXT:    mov v4.s[2], wzr
 ; CHECK-NEXT:    ld1 { v0.b }[5], [x11]
-; CHECK-NEXT:    add x11, sp, #640
+; CHECK-NEXT:    add x11, sp, #656
 ; CHECK-NEXT:    ld1 { v17.b }[2], [x8]
-; CHECK-NEXT:    add x8, sp, #680
+; CHECK-NEXT:    add x8, sp, #696
 ; CHECK-NEXT:    ld1 { v5.b }[3], [x10]
-; CHECK-NEXT:    add x10, sp, #496
+; CHECK-NEXT:    add x10, sp, #512
 ; CHECK-NEXT:    mov v4.s[3], wzr
 ; CHECK-NEXT:    ld1 { v16.b }[3], [x9]
-; CHECK-NEXT:    add x9, sp, #560
+; CHECK-NEXT:    add x9, sp, #576
 ; CHECK-NEXT:    ld1 { v0.b }[6], [x11]
 ; CHECK-NEXT:    ld1 { v17.b }[3], [x8]
-; CHECK-NEXT:    add x8, sp, #688
+; CHECK-NEXT:    add x8, sp, #704
 ; CHECK-NEXT:    ld1 { v5.b }[4], [x10]
-; CHECK-NEXT:    add x11, sp, #504
+; CHECK-NEXT:    add x11, sp, #520
 ; CHECK-NEXT:    saddw v4.4s, v4.4s, v6.4h
-; CHECK-NEXT:    ldr b6, [sp, #720]
+; CHECK-NEXT:    ldr b6, [sp, #736]
 ; CHECK-NEXT:    ld1 { v16.b }[4], [x9]
-; CHECK-NEXT:    add x9, sp, #568
+; CHECK-NEXT:    add x9, sp, #584
 ; CHECK-NEXT:    ld1 { v17.b }[4], [x8]
-; CHECK-NEXT:    add x8, sp, #696
-; CHECK-NEXT:    add x10, sp, #648
+; CHECK-NEXT:    add x8, sp, #712
+; CHECK-NEXT:    add x10, sp, #664
 ; CHECK-NEXT:    ld1 { v5.b }[5], [x11]
 ; CHECK-NEXT:    saddw v1.4s, v4.4s, v1.4h
 ; CHECK-NEXT:    sshll v4.8h, v6.8b, #0
 ; CHECK-NEXT:    ld1 { v16.b }[5], [x9]
 ; CHECK-NEXT:    ld1 { v17.b }[5], [x8]
-; CHECK-NEXT:    add x9, sp, #576
+; CHECK-NEXT:    add x9, sp, #592
 ; CHECK-NEXT:    ld1 { v0.b }[7], [x10]
-; CHECK-NEXT:    add x8, sp, #512
-; CHECK-NEXT:    add x10, sp, #704
+; CHECK-NEXT:    add x8, sp, #528
+; CHECK-NEXT:    add x10, sp, #720
 ; CHECK-NEXT:    sshll v4.4s, v4.4h, #0
 ; CHECK-NEXT:    ld1 { v16.b }[6], [x9]
 ; CHECK-NEXT:    ld1 { v5.b }[6], [x8]
-; CHECK-NEXT:    add x9, sp, #584
+; CHECK-NEXT:    add x9, sp, #600
 ; CHECK-NEXT:    mov v4.s[1], wzr
 ; CHECK-NEXT:    ld1 { v17.b }[6], [x10]
-; CHECK-NEXT:    add x10, sp, #712
-; CHECK-NEXT:    add x8, sp, #520
+; CHECK-NEXT:    add x10, sp, #728
+; CHECK-NEXT:    add x8, sp, #536
 ; CHECK-NEXT:    ld1 { v16.b }[7], [x9]
 ; CHECK-NEXT:    add v2.4s, v3.4s, v2.4s
 ; CHECK-NEXT:    mov v4.s[2], wzr
 ; CHECK-NEXT:    ld1 { v17.b }[7], [x10]
 ; CHECK-NEXT:    ld1 { v5.b }[7], [x8]
 ; CHECK-NEXT:    sshll v0.8h, v0.8b, #0
 ; CHECK-NEXT:    sshll v3.8h, v16.8b, #0
 ; CHECK-NEXT:    mov v4.s[3], wzr
 ; CHECK-NEXT:    sshll v6.8h, v17.8b, #0
 ; CHECK-NEXT:    sshll v5.8h, v5.8b, #0
 ; CHECK-NEXT:    add v1.4s, v1.4s, v7.4s
 ; CHECK-NEXT:    saddl2 v7.4s, v6.8h, v3.8h
 ; CHECK-NEXT:    saddl2 v16.4s, v0.8h, v5.8h
 ; CHECK-NEXT:    saddw v4.4s, v4.4s, v5.4h
 ; CHECK-NEXT:    saddl v3.4s, v6.4h, v3.4h
 ; CHECK-NEXT:    add v5.4s, v16.4s, v7.4s
 ; CHECK-NEXT:    saddw v0.4s, v4.4s, v0.4h
 ; CHECK-NEXT:    add v3.4s, v3.4s, v5.4s
 ; CHECK-NEXT:    add v1.4s, v1.4s, v2.4s
 ; CHECK-NEXT:    add v0.4s, v0.4s, v3.4s
 ; CHECK-NEXT:    add v0.4s, v1.4s, v0.4s
 ; CHECK-NEXT:    addv s0, v0.4s
 ; CHECK-NEXT:    fmov w0, s0
+; CHECK-NEXT:    ldr x29, [sp], #16 // 8-byte Folded Reload
 ; CHECK-NEXT:    ret
 entry:
   %az = sext <33 x i8> %a to <33 x i32>
   %r1 = call i32 @llvm.vector.reduce.add.v33i32(<33 x i32> %az)
   %cz = sext <33 x i8> %c to <33 x i32>
   %r2 = call i32 @llvm.vector.reduce.add.v33i32(<33 x i32> %cz)
   %x = add i32 %r1, %r2
   ret i32 %x
entry:
   %5 = call i32 @llvm.vector.reduce.add.v48i32(<48 x i32> %4)
   %op.extra = add nsw i32 %5, %sum
   ret i32 %op.extra
 }
 
 define i32 @test_sdot_v48i8_double(<48 x i8> %a, <48 x i8> %b, <48 x i8> %c, <48 x i8> %d) {
 ; CHECK-LABEL: test_sdot_v48i8_double:
 ; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    ldr b0, [sp, #64]
-; CHECK-NEXT:    add x8, sp, #72
-; CHECK-NEXT:    add x9, sp, #80
-; CHECK-NEXT:    ldr b1, [sp, #192]
-; CHECK-NEXT:    ldr b3, [sp, #128]
-; CHECK-NEXT:    add x10, sp, #88
+; CHECK-NEXT:    str x29, [sp, #-16]! // 8-byte Folded Spill
+; CHECK-NEXT:    .cfi_def_cfa_offset 16
+; CHECK-NEXT:    .cfi_offset w29, -16
+; CHECK-NEXT:    ldr b0, [sp, #80]
+; CHECK-NEXT:    add x8, sp, #88
+; CHECK-NEXT:    add x9, sp, #96
+; CHECK-NEXT:    ldr b1, [sp, #208]
+; CHECK-NEXT:    ldr b3, [sp, #144]
+; CHECK-NEXT:    add x10, sp, #104
 ; CHECK-NEXT:    ld1 { v0.b }[1], [x8]
-; CHECK-NEXT:    add x8, sp, #200
-; CHECK-NEXT:    ldr b4, [sp, #256]
+; CHECK-NEXT:    add x8, sp, #216
+; CHECK-NEXT:    ldr b4, [sp, #272]
 ; CHECK-NEXT:    fmov s17, w0
-; CHECK-NEXT:    ldr b18, [sp]
-; CHECK-NEXT:    add x11, sp, #528
+; CHECK-NEXT:    ldr b18, [sp, #16]
+; CHECK-NEXT:    add x11, sp, #544
 ; CHECK-NEXT:    ld1 { v1.b }[1], [x8]
-; CHECK-NEXT:    add x8, sp, #264
+; CHECK-NEXT:    add x8, sp, #280
 ; CHECK-NEXT:    ld1 { v0.b }[2], [x9]
-; CHECK-NEXT:    add x9, sp, #136
-; CHECK-NEXT:    ldr b5, [sp, #576]
+; CHECK-NEXT:    add x9, sp, #152
+; CHECK-NEXT:    ldr b5, [sp, #592]
 ; CHECK-NEXT:    ld1 { v4.b }[1], [x8]
-; CHECK-NEXT:    add x8, sp, #144
+; CHECK-NEXT:    add x8, sp, #160
 ; CHECK-NEXT:    ld1 { v3.b }[1], [x9]
-; CHECK-NEXT:    add x9, sp, #208
+; CHECK-NEXT:    add x9, sp, #224
 ; CHECK-NEXT:    ld1 { v0.b }[3], [x10]
-; CHECK-NEXT:    add x10, sp, #96
-; CHECK-NEXT:    ldr b2, [sp, #448]
+; CHECK-NEXT:    add x10, sp, #112
+; CHECK-NEXT:    ldr b2, [sp, #464]
 ; CHECK-NEXT:    ld1 { v1.b }[2], [x9]
-; CHECK-NEXT:    add x9, sp, #8
+; CHECK-NEXT:    add x9, sp, #24
 ; CHECK-NEXT:    ld1 { v3.b }[2], [x8]
-; CHECK-NEXT:    add x8, sp, #104
+; CHECK-NEXT:    add x8, sp, #120
 ; CHECK-NEXT:    ld1 { v0.b }[4], [x10]
-; CHECK-NEXT:    add x10, sp, #216
+; CHECK-NEXT:    add x10, sp, #232
 ; CHECK-NEXT:    ld1 { v18.b }[1], [x9]
-; CHECK-NEXT:    add x9, sp, #152
-; CHECK-NEXT:    ldr b21, [sp, #320]
+; CHECK-NEXT:    add x9, sp, #168
+; CHECK-NEXT:    ldr b21, [sp, #336]
 ; CHECK-NEXT:    ld1 { v1.b }[3], [x10]
-; CHECK-NEXT:    add x10, sp, #272
+; CHECK-NEXT:    add x10, sp, #288
 ; CHECK-NEXT:    ld1 { v0.b }[5], [x8]
-; CHECK-NEXT:    add x8, sp, #224
+; CHECK-NEXT:    add x8, sp, #240
 ; CHECK-NEXT:    ld1 { v3.b }[3], [x9]
-; CHECK-NEXT:    add x9, sp, #112
+; CHECK-NEXT:    add x9, sp, #128
 ; CHECK-NEXT:    ld1 { v4.b }[2], [x10]
-; CHECK-NEXT:    add x10, sp, #160
+; CHECK-NEXT:    add x10, sp, #176
 ; CHECK-NEXT:    ld1 { v1.b }[4], [x8]
-; CHECK-NEXT:    add x8, sp, #280
+; CHECK-NEXT:    add x8, sp, #296
 ; CHECK-NEXT:    ld1 { v0.b }[6], [x9]
-; CHECK-NEXT:    add x9, sp, #232
+; CHECK-NEXT:    add x9, sp, #248
 ; CHECK-NEXT:    ld1 { v3.b }[4], [x10]
-; CHECK-NEXT:    add x10, sp, #120
+; CHECK-NEXT:    add x10, sp, #136
 ; CHECK-NEXT:    ld1 { v4.b }[3], [x8]
-; CHECK-NEXT:    add x8, sp, #168
+; CHECK-NEXT:    add x8, sp, #184
 ; CHECK-NEXT:    ld1 { v1.b }[5], [x9]
-; CHECK-NEXT:    add x9, sp, #288
+; CHECK-NEXT:    add x9, sp, #304
 ; CHECK-NEXT:    ld1 { v0.b }[7], [x10]
-; CHECK-NEXT:    add x10, sp, #16
+; CHECK-NEXT:    add x10, sp, #32
 ; CHECK-NEXT:    ld1 { v3.b }[5], [x8]
-; CHECK-NEXT:    add x8, sp, #240
+; CHECK-NEXT:    add x8, sp, #256
 ; CHECK-NEXT:    ld1 { v4.b }[4], [x9]
-; CHECK-NEXT:    add x9, sp, #176
+; CHECK-NEXT:    add x9, sp, #192
 ; CHECK-NEXT:    ld1 { v18.b }[2], [x10]
-; CHECK-NEXT:    add x10, sp, #296
+; CHECK-NEXT:    add x10, sp, #312
 ; CHECK-NEXT:    ld1 { v1.b }[6], [x8]
-; CHECK-NEXT:    add x8, sp, #24
+; CHECK-NEXT:    add x8, sp, #40
 ; CHECK-NEXT:    ld1 { v3.b }[6], [x9]
-; CHECK-NEXT:    add x9, sp, #248
+; CHECK-NEXT:    add x9, sp, #264
 ; CHECK-NEXT:    ld1 { v4.b }[5], [x10]
-; CHECK-NEXT:    add x10, sp, #184
+; CHECK-NEXT:    add x10, sp, #200
 ; CHECK-NEXT:    ld1 { v18.b }[3], [x8]
-; CHECK-NEXT:    add x8, sp, #304
+; CHECK-NEXT:    add x8, sp, #320
 ; CHECK-NEXT:    ld1 { v1.b }[7], [x9]
-; CHECK-NEXT:    add x9, sp, #32
+; CHECK-NEXT:    add x9, sp, #48
 ; CHECK-NEXT:    ld1 { v3.b }[7], [x10]
-; CHECK-NEXT:    add x10, sp, #584
+; CHECK-NEXT:    add x10, sp, #600
 ; CHECK-NEXT:    ld1 { v4.b }[6], [x8]
-; CHECK-NEXT:    add x8, sp, #312
+; CHECK-NEXT:    add x8, sp, #328
 ; CHECK-NEXT:    ld1 { v18.b }[4], [x9]
-; CHECK-NEXT:    add x9, sp, #456
+; CHECK-NEXT:    add x9, sp, #472
 ; CHECK-NEXT:    ld1 { v5.b }[1], [x10]
-; CHECK-NEXT:    add x10, sp, #600
+; CHECK-NEXT:    add x10, sp, #616
 ; CHECK-NEXT:    sshll v6.8h, v0.8b, #0
-; CHECK-NEXT:    ldr b0, [sp, #384]
+; CHECK-NEXT:    ldr b0, [sp, #400]
 ; CHECK-NEXT:    ld1 { v4.b }[7], [x8]
-; CHECK-NEXT:    add x8, sp, #40
+; CHECK-NEXT:    add x8, sp, #56
 ; CHECK-NEXT:    ld1 { v2.b }[1], [x9]
-; CHECK-NEXT:    add x9, sp, #464
-; CHECK-NEXT:    ldr b19, [sp, #512]
+; CHECK-NEXT:    add x9, sp, #480
+; CHECK-NEXT:    ldr b19, [sp, #528]
 ; CHECK-NEXT:    ld1 { v18.b }[5], [x8]
-; CHECK-NEXT:    add x8, sp, #592
+; CHECK-NEXT:    add x8, sp, #608
 ; CHECK-NEXT:    sshll v16.8h, v1.8b, #0
-; CHECK-NEXT:    ldr b1, [sp, #640]
+; CHECK-NEXT:    ldr b1, [sp, #656]
 ; CHECK-NEXT:    ld1 { v2.b }[2], [x9]
-; CHECK-NEXT:    add x9, sp, #472
+; CHECK-NEXT:    add x9, sp, #488
 ; CHECK-NEXT:    ld1 { v5.b }[2], [x8]
-; CHECK-NEXT:    add x8, sp, #48
+; CHECK-NEXT:    add x8, sp, #64
 ; CHECK-NEXT:    mov v17.b[1], w1
 ; CHECK-NEXT:    sshll v7.8h, v3.8b, #0
 ; CHECK-NEXT:    ld1 { v18.b }[6], [x8]
-; CHECK-NEXT:    add x8, sp, #608
+; CHECK-NEXT:    add x8, sp, #624
 ; CHECK-NEXT:    ld1 { v5.b }[3], [x10]
-; CHECK-NEXT:    add x10, sp, #616
+; CHECK-NEXT:    add x10, sp, #632
 ; CHECK-NEXT:    ld1 { v2.b }[3], [x9]
-; CHECK-NEXT:    add x9, sp, #480
+; CHECK-NEXT:    add x9, sp, #496
 ; CHECK-NEXT:    mov v17.b[2], w2
 ; CHECK-NEXT:    sshll v4.8h, v4.8b, #0
 ; CHECK-NEXT:    ld1 { v5.b }[4], [x8]
-; CHECK-NEXT:    add x8, sp, #56
+; CHECK-NEXT:    add x8, sp, #72
 ; CHECK-NEXT:    ld1 { v2.b }[4], [x9]
-; CHECK-NEXT:    add x9, sp, #488
+; CHECK-NEXT:    add x9, sp, #504
 ; CHECK-NEXT:    mov v17.b[3], w3
 ; CHECK-NEXT:    ld1 { v18.b }[7], [x8]
-; CHECK-NEXT:    add x8, sp, #624
+; CHECK-NEXT:    add x8, sp, #640
 ; CHECK-NEXT:    ld1 { v5.b }[5], [x10]
-; CHECK-NEXT:    add x10, sp, #648
+; CHECK-NEXT:    add x10, sp, #664
 ; CHECK-NEXT:    ld1 { v2.b }[5], [x9]
-; CHECK-NEXT:    add x9, sp, #496
+; CHECK-NEXT:    add x9, sp, #512
 ; CHECK-NEXT:    mov v17.b[4], w4
 ; CHECK-NEXT:    ld1 { v1.b }[1], [x10]
-; CHECK-NEXT:    add x10, sp, #656
+; CHECK-NEXT:    add x10, sp, #672
 ; CHECK-NEXT:    ld1 { v5.b }[6], [x8]
-; CHECK-NEXT:    add x8, sp, #632
+; CHECK-NEXT:    add x8, sp, #648
 ; CHECK-NEXT:    ld1 { v2.b }[6], [x9]
-; CHECK-NEXT:    add x9, sp, #504
+; CHECK-NEXT:    add x9, sp, #520
 ; CHECK-NEXT:    mov v17.b[5], w5
 ; CHECK-NEXT:    ld1 { v1.b }[2], [x10]
-; CHECK-NEXT:    add x10, sp, #664
+; CHECK-NEXT:    add x10, sp, #680
 ; CHECK-NEXT:    ld1 { v5.b }[7], [x8]
-; CHECK-NEXT:    add x8, sp, #392
+; CHECK-NEXT:    add x8, sp, #408
 ; CHECK-NEXT:    ld1 { v2.b }[7], [x9]
-; CHECK-NEXT:    add x9, sp, #520
+; CHECK-NEXT:    add x9, sp, #536
 ; CHECK-NEXT:    mov v17.b[6], w6
 ; CHECK-NEXT:    ld1 { v0.b }[1], [x8]
-; CHECK-NEXT:    add x8, sp, #328
+; CHECK-NEXT:    add x8, sp, #344
 ; CHECK-NEXT:    ld1 { v19.b }[1], [x9]
-; CHECK-NEXT:    add x9, sp, #400
+; CHECK-NEXT:    add x9, sp, #416
 ; CHECK-NEXT:    ld1 { v1.b }[3], [x10]
-; CHECK-NEXT:    add x10, sp, #672
+; CHECK-NEXT:    add x10, sp, #688
 ; CHECK-NEXT:    ld1 { v21.b }[1], [x8]
-; CHECK-NEXT:    add x8, sp, #336
+; CHECK-NEXT:    add x8, sp, #352
 ; CHECK-NEXT:    ld1 { v0.b }[2], [x9]
-; CHECK-NEXT:    add x9, sp, #408
+; CHECK-NEXT:    add x9, sp, #424
 ; CHECK-NEXT:    ld1 { v19.b }[2], [x11]
-; CHECK-NEXT:    add x11, sp, #536
+; CHECK-NEXT:    add x11, sp, #552
 ; CHECK-NEXT:    ld1 { v1.b }[4], [x10]
-; CHECK-NEXT:    add x10, sp, #680
+; CHECK-NEXT:    add x10, sp, #696
 ; CHECK-NEXT:    ld1 { v21.b }[2], [x8]
-; CHECK-NEXT:    add x8, sp, #344
+; CHECK-NEXT:    add x8, sp, #360
 ; CHECK-NEXT:    ld1 { v0.b }[3], [x9]
-; CHECK-NEXT:    add x9, sp, #416
+; CHECK-NEXT:    add x9, sp, #432
 ; CHECK-NEXT:    sshll v3.8h, v18.8b, #0
 ; CHECK-NEXT:    ld1 { v19.b }[3], [x11]
 ; CHECK-NEXT:    ld1 { v1.b }[5], [x10]
-; CHECK-NEXT:    add x10, sp, #688
+; CHECK-NEXT:    add x10, sp, #704
 ; CHECK-NEXT:    ld1 { v21.b }[3], [x8]
-; CHECK-NEXT:    add x8, sp, #352
+; CHECK-NEXT:    add x8, sp, #368
 ; CHECK-NEXT:    ld1 { v0.b }[4], [x9]
-; CHECK-NEXT:    add x9, sp, #424
+; CHECK-NEXT:    add x9, sp, #440
 ; CHECK-NEXT:    mov v17.b[7], w7
-; CHECK-NEXT:    add x11, sp, #544
+; CHECK-NEXT:    add x11, sp, #560
 ; CHECK-NEXT:    ld1 { v1.b }[6], [x10]
-; CHECK-NEXT:    add x10, sp, #696
+; CHECK-NEXT:    add x10, sp, #712
 ; CHECK-NEXT:    ld1 { v21.b }[4], [x8]
-; CHECK-NEXT:    add x8, sp, #360
+; CHECK-NEXT:    add x8, sp, #376
 ; CHECK-NEXT:    ld1 { v0.b }[5], [x9]
-; CHECK-NEXT:    add x9, sp, #432
+; CHECK-NEXT:    add x9, sp, #448
 ; CHECK-NEXT:    sshll v18.8h, v2.8b, #0
 ; CHECK-NEXT:    ld1 { v19.b }[4], [x11]
 ; CHECK-NEXT:    sshll v20.8h, v5.8b, #0
 ; CHECK-NEXT:    ld1 { v1.b }[7], [x10]
 ; CHECK-NEXT:    ld1 { v21.b }[5], [x8]
-; CHECK-NEXT:    add x8, sp, #368
+; CHECK-NEXT:    add x8, sp, #384
 ; CHECK-NEXT:    ld1 { v0.b }[6], [x9]
-; CHECK-NEXT:    add x9, sp, #440
+; CHECK-NEXT:    add x9, sp, #456
 ; CHECK-NEXT:    sshll v17.8h, v17.8b, #0
-; CHECK-NEXT:    add x10, sp, #1032
+; CHECK-NEXT:    add x10, sp, #1048
 ; CHECK-NEXT:    sshll v22.8h, v1.8b, #0
-; CHECK-NEXT:    add x11, sp, #552
+; CHECK-NEXT:    add x11, sp, #568
 ; CHECK-NEXT:    ld1 { v21.b }[6], [x8]
-; CHECK-NEXT:    add x8, sp, #376
+; CHECK-NEXT:    add x8, sp, #392
 ; CHECK-NEXT:    ld1 { v0.b }[7], [x9]
-; CHECK-NEXT:    add x9, sp, #968
+; CHECK-NEXT:    add x9, sp, #984
 ; CHECK-NEXT:    ld1 { v19.b }[5], [x11]
-; CHECK-NEXT:    add x11, sp, #560
+; CHECK-NEXT:    add x11, sp, #576
 ; CHECK-NEXT:    ld1 { v21.b }[7], [x8]
-; CHECK-NEXT:    add x8, sp, #840
+; CHECK-NEXT:    add x8, sp, #856
 ; CHECK-NEXT:    sshll v5.8h, v0.8b, #0
 ; CHECK-NEXT:    ld1 { v19.b }[6], [x11]
-; CHECK-NEXT:    add x11, sp, #568
+; CHECK-NEXT:    add x11, sp, #584
 ; CHECK-NEXT:    smull2 v0.4s, v3.8h, v5.8h
 ; CHECK-NEXT:    sshll v2.8h, v21.8b, #0
-; CHECK-NEXT:    ldr b21, [sp, #832]
+; CHECK-NEXT:    ldr b21, [sp, #848]
 ; CHECK-NEXT:    smull v3.4s, v3.4h, v5.4h
-; CHECK-NEXT:    ldr b5, [sp, #960]
+; CHECK-NEXT:    ldr b5, [sp, #976]
 ; CHECK-NEXT:    smull2 v1.4s, v17.8h, v2.8h
 ; CHECK-NEXT:    ld1 { v19.b }[7], [x11]
 ; CHECK-NEXT:    ld1 { v21.b }[1], [x8]
-; CHECK-NEXT:    add x8, sp, #848
+; CHECK-NEXT:    add x8, sp, #864
 ; CHECK-NEXT:    smull v2.4s, v17.4h, v2.4h
 ; CHECK-NEXT:    ld1 { v5.b }[1], [x9]
-; CHECK-NEXT:    ldr b17, [sp, #1024]
-; CHECK-NEXT:    add x9, sp, #904
+; CHECK-NEXT:    ldr b17, [sp, #1040]
+; CHECK-NEXT:    add x9, sp, #920
 ; CHECK-NEXT:    smlal2 v0.4s, v4.8h, v22.8h
-; CHECK-NEXT:    add x11, sp, #856
+; CHECK-NEXT:    add x11, sp, #872
 ; CHECK-NEXT:    ld1 { v21.b }[2], [x8]
-; CHECK-NEXT:    add x8, sp, #976
+; CHECK-NEXT:    add x8, sp, #992
 ; CHECK-NEXT:    smlal v3.4s, v4.4h, v22.4h
-; CHECK-NEXT:    ldr b4, [sp, #896]
+; CHECK-NEXT:    ldr b4, [sp, #912]
 ; CHECK-NEXT:    ld1 { v17.b }[1], [x10]
-; CHECK-NEXT:    add x10, sp, #984
+; CHECK-NEXT:    add x10, sp, #1000
 ; CHECK-NEXT:    ld1 { v5.b }[2], [x8]
-; CHECK-NEXT:    add x8, sp, #1040
+; CHECK-NEXT:    add x8, sp, #1056
 ; CHECK-NEXT:    ld1 { v4.b }[1], [x9]
-; CHECK-NEXT:    add x9, sp, #912
+; CHECK-NEXT:    add x9, sp, #928
 ; CHECK-NEXT:    ld1 { v21.b }[3], [x11]
-; CHECK-NEXT:    add x11, sp, #864
+; CHECK-NEXT:    add x11, sp, #880
 ; CHECK-NEXT:    ld1 { v17.b }[2], [x8]
-; CHECK-NEXT:    add x8, sp, #1048
+; CHECK-NEXT:    add x8, sp, #1064
 ; CHECK-NEXT:    ld1 { v5.b }[3], [x10]
-; CHECK-NEXT:    add x10, sp, #992
+; CHECK-NEXT:    add x10, sp, #1008
 ; CHECK-NEXT:    ld1 { v4.b }[2], [x9]
-; CHECK-NEXT:    add x9, sp, #920
+; CHECK-NEXT:    add x9, sp, #936
 ; CHECK-NEXT:    ld1 { v21.b }[4], [x11]
-; CHECK-NEXT:    add x11, sp, #872
+; CHECK-NEXT:    add x11, sp, #888
 ; CHECK-NEXT:    ld1 { v17.b }[3], [x8]
-; CHECK-NEXT:    add x8, sp, #1056
+; CHECK-NEXT:    add x8, sp, #1072
 ; CHECK-NEXT:    ld1 { v5.b }[4], [x10]
-; CHECK-NEXT:    add x10, sp, #1000
+; CHECK-NEXT:    add x10, sp, #1016
 ; CHECK-NEXT:    ld1 { v4.b }[3], [x9]
-; CHECK-NEXT:    add x9, sp, #928
+; CHECK-NEXT:    add x9, sp, #944
 ; CHECK-NEXT:    ld1 { v21.b }[5], [x11]
-; CHECK-NEXT:    add x11, sp, #880
+; CHECK-NEXT:    add x11, sp, #896
 ; CHECK-NEXT:    ld1 { v17.b }[4], [x8]
-; CHECK-NEXT:    add x8, sp, #1064
+; CHECK-NEXT:    add x8, sp, #1080
 ; CHECK-NEXT:    ld1 { v5.b }[5], [x10]
-; CHECK-NEXT:    add x10, sp, #1008
+; CHECK-NEXT:    add x10, sp, #1024
 ; CHECK-NEXT:    ld1 { v4.b }[4], [x9]
-; CHECK-NEXT:    add x9, sp, #936
+; CHECK-NEXT:    add x9, sp, #952
 ; CHECK-NEXT:    sshll v19.8h, v19.8b, #0
 ; CHECK-NEXT:    ld1 { v21.b }[6], [x11]
 ; CHECK-NEXT:    ld1 { v17.b }[5], [x8]
-; CHECK-NEXT:    add x8, sp, #1072
+; CHECK-NEXT:    add x8, sp, #1088
 ; CHECK-NEXT:    ld1 { v5.b }[6], [x10]
-; CHECK-NEXT:    add x10, sp, #1016
+; CHECK-NEXT:    add x10, sp, #1032
 ; CHECK-NEXT:    ld1 { v4.b }[5], [x9]
-; CHECK-NEXT:    add x9, sp, #944
+; CHECK-NEXT:    add x9, sp, #960
 ; CHECK-NEXT:    smlal2 v1.4s, v16.8h, v20.8h
-; CHECK-NEXT:    add x11, sp, #888
+; CHECK-NEXT:    add x11, sp, #904
 ; CHECK-NEXT:    ld1 { v17.b }[6], [x8]
-; CHECK-NEXT:    add x8, sp, #1080
+; CHECK-NEXT:    add x8, sp, #1096
 ; CHECK-NEXT:    smlal2 v0.4s, v7.8h, v19.8h
 ; CHECK-NEXT:    ld1 { v5.b }[7], [x10]
 ; CHECK-NEXT:    ld1 { v4.b }[6], [x9]
-; CHECK-NEXT:    add x9, sp, #952
+; CHECK-NEXT:    add x9, sp, #968
 ; CHECK-NEXT:    smlal v2.4s, v16.4h, v20.4h
-; CHECK-NEXT:    ldr b16, [sp, #768]
-; CHECK-NEXT:    add x10, sp, #776
+; CHECK-NEXT:    ldr b16, [sp, #784]
+; CHECK-NEXT:    add x10, sp, #792
 ; CHECK-NEXT:    ld1 { v17.b }[7], [x8]
 ; CHECK-NEXT:    smlal v3.4s, v7.4h, v19.4h
-; CHECK-NEXT:    ldr b19, [sp, #1152]
-; CHECK-NEXT:    add x8, sp, #1160
+; CHECK-NEXT:    ldr b19, [sp, #1168]
+; CHECK-NEXT:    add x8, sp, #1176
 ; CHECK-NEXT:    ld1 { v4.b }[7], [x9]
-; CHECK-NEXT:    ldr b7, [sp, #704]
-; CHECK-NEXT:    add x9, sp, #712
+; CHECK-NEXT:    ldr b7, [sp, #720]
+; CHECK-NEXT:    add x9, sp, #728
 ; CHECK-NEXT:    ld1 { v16.b }[1], [x10]
-; CHECK-NEXT:    add x10, sp, #784
+; CHECK-NEXT:    add x10, sp, #800
 ; CHECK-NEXT:    ld1 { v19.b }[1], [x8]
-; CHECK-NEXT:    add x8, sp, #1168
+; CHECK-NEXT:    add x8, sp, #1184
 ; CHECK-NEXT:    smlal2 v1.4s, v6.8h, v18.8h
 ; CHECK-NEXT:    ld1 { v21.b }[7], [x11]
 ; CHECK-NEXT:    smlal v2.4s, v6.4h, v18.4h
-; CHECK-NEXT:    ldr b18, [sp, #1408]
-; CHECK-NEXT:    add x11, sp, #1416
+; CHECK-NEXT:    ldr b18, [sp, #1424]
+; CHECK-NEXT:    add x11, sp, #1432
 ; CHECK-NEXT:    ld1 { v7.b }[1], [x9]
-; CHECK-NEXT:    add x9, sp, #720
+; CHECK-NEXT:    add x9, sp, #736
 ; CHECK-NEXT:    ld1 { v16.b }[2], [x10]
 ; CHECK-NEXT:    ld1 { v19.b }[2], [x8]
-; CHECK-NEXT:    add x10, sp, #792
+; CHECK-NEXT:    add x10, sp, #808
 ; CHECK-NEXT:    ld1 { v18.b }[1], [x11]
-; CHECK-NEXT:    add x11, sp, #1424
-; CHECK-NEXT:    add x8, sp, #1176
+; CHECK-NEXT:    add x11, sp, #1440
+; CHECK-NEXT:    add x8, sp, #1192
 ; CHECK-NEXT:    ld1 { v7.b }[2], [x9]
-; CHECK-NEXT:    add x9, sp, #728
+; CHECK-NEXT:    add x9, sp, #744
 ; CHECK-NEXT:    ld1 { v16.b }[3], [x10]
-; CHECK-NEXT:    add x10, sp, #800
+; CHECK-NEXT:    add x10, sp, #816
 ; CHECK-NEXT:    ld1 { v18.b }[2], [x11]
-; CHECK-NEXT:    add x11, sp, #1432
+; CHECK-NEXT:    add x11, sp, #1448
 ; CHECK-NEXT:    ld1 { v19.b }[3], [x8]
-; CHECK-NEXT:    add x8, sp, #1184
+; CHECK-NEXT:    add x8, sp, #1200
 ; CHECK-NEXT:    ld1 { v7.b }[3], [x9]
-; CHECK-NEXT:    add x9, sp, #736
+; CHECK-NEXT:    add x9, sp, #752
 ; CHECK-NEXT:    ld1 { v16.b }[4], [x10]
-; CHECK-NEXT:    add x10, sp, #808
+; CHECK-NEXT:    add x10, sp, #824
 ; CHECK-NEXT:    ld1 { v18.b }[3], [x11]
-; CHECK-NEXT:    add x11, sp, #1440
+; CHECK-NEXT:    add x11, sp, #1456
 ; CHECK-NEXT:    ld1 { v19.b }[4], [x8]
-; CHECK-NEXT:    add x8, sp, #1192
+; CHECK-NEXT:    add x8, sp, #1208
 ; CHECK-NEXT:    ld1 { v7.b }[4], [x9]
-; CHECK-NEXT:    add x9, sp, #744
+; CHECK-NEXT:    add x9, sp, #760
 ; CHECK-NEXT:    ld1 { v16.b }[5], [x10]
-; CHECK-NEXT:    add x10, sp, #816
+; CHECK-NEXT:    add x10, sp, #832
 ; CHECK-NEXT:    ld1 { v18.b }[4], [x11]
-; CHECK-NEXT:    add x11, sp, #1448
+; CHECK-NEXT:    add x11, sp, #1464
 ; CHECK-NEXT:    ld1 { v19.b }[5], [x8]
-; CHECK-NEXT:    add x8, sp, #1200
+; CHECK-NEXT:    add x8, sp, #1216
 ; CHECK-NEXT:    ld1 { v7.b }[5], [x9]
-; CHECK-NEXT:    add x9, sp, #752
+; CHECK-NEXT:    add x9, sp, #768
 ; CHECK-NEXT:    ld1 { v16.b }[6], [x10]
-; CHECK-NEXT:    add x10, sp, #824
+; CHECK-NEXT:    add x10, sp, #840
 ; CHECK-NEXT:    ld1 { v18.b }[5], [x11]
-; CHECK-NEXT:    add x11, sp, #1456
+; CHECK-NEXT:    add x11, sp, #1472
 ; CHECK-NEXT:    ld1 { v19.b }[6], [x8]
-; CHECK-NEXT:    add x8, sp, #1208
+; CHECK-NEXT:    add x8, sp, #1224
 ; CHECK-NEXT:    ld1 { v7.b }[6], [x9]
-; CHECK-NEXT:    add x9, sp, #760
+; CHECK-NEXT:    add x9, sp, #776
 ; CHECK-NEXT:    ld1 { v16.b }[7], [x10]
-; CHECK-NEXT:    add x10, sp, #1464
+; CHECK-NEXT:    add x10, sp, #1480
 ; CHECK-NEXT:    ld1 { v18.b }[6], [x11]
-; CHECK-NEXT:    add x11, sp, #1288
+; CHECK-NEXT:    add x11, sp, #1304
 ; CHECK-NEXT:    ld1 { v19.b }[7], [x8]
-; CHECK-NEXT:    add x8, sp, #1352
+; CHECK-NEXT:    add x8, sp, #1368
 ; CHECK-NEXT:    sshll v20.8h, v17.8b, #0
 ; CHECK-NEXT:    ld1 { v7.b }[7], [x9]
-; CHECK-NEXT:    ldr b17, [sp, #1088]
-; CHECK-NEXT:    add x9, sp, #1096
+; CHECK-NEXT:    ldr b17, [sp, #1104]
+; CHECK-NEXT:    add x9, sp, #1112
 ; CHECK-NEXT:    sshll v6.8h, v21.8b, #0
 ; CHECK-NEXT:    ld1 { v18.b }[7], [x10]
 ; CHECK-NEXT:    sshll v21.8h, v16.8b, #0
-; CHECK-NEXT:    ldr b16, [sp, #1344]
+; CHECK-NEXT:    ldr b16, [sp, #1360]
 ; CHECK-NEXT:    sshll v19.8h, v19.8b, #0
 ; CHECK-NEXT:    ld1 { v17.b }[1], [x9]
-; CHECK-NEXT:    add x9, sp, #1104
-; CHECK-NEXT:    add x10, sp, #1224
+; CHECK-NEXT:    add x9, sp, #1120
+; CHECK-NEXT:    add x10, sp, #1240
 ; CHECK-NEXT:    smull v22.4s, v21.4h, v19.4h
 ; CHECK-NEXT:    ld1 { v16.b }[1], [x8]
 ; CHECK-NEXT:    smull2 v19.4s, v21.8h, v19.8h
-; CHECK-NEXT:    ldr b21, [sp, #1216]
+; CHECK-NEXT:    ldr b21, [sp, #1232]
 ; CHECK-NEXT:    sshll v18.8h, v18.8b, #0
 ; CHECK-NEXT:    ld1 { v17.b }[2], [x9]
-; CHECK-NEXT:    add x8, sp, #1360
-; CHECK-NEXT:    add x9, sp, #1112
+; CHECK-NEXT:    add x8, sp, #1376
+; CHECK-NEXT:    add x9, sp, #1128
 ; CHECK-NEXT:    smlal v22.4s, v20.4h, v18.4h
 ; CHECK-NEXT:    ld1 { v21.b }[1], [x10]
 ; CHECK-NEXT:    smlal2 v19.4s, v20.8h, v18.8h
-; CHECK-NEXT:    ldr b18, [sp, #1280]
+; CHECK-NEXT:    ldr b18, [sp, #1296]
 ; CHECK-NEXT:    ld1 { v16.b }[2], [x8]
-; CHECK-NEXT:    add x10, sp, #1232
+; CHECK-NEXT:    add x10, sp, #1248
 ; CHECK-NEXT:    ld1 { v17.b }[3], [x9]
-; CHECK-NEXT:    add x8, sp, #1368
+; CHECK-NEXT:    add x8, sp, #1384
 ; CHECK-NEXT:    ld1 { v18.b }[1], [x11]
-; CHECK-NEXT:    add x11, sp, #1296
-; CHECK-NEXT:    add x9, sp, #1120
+; CHECK-NEXT:    add x11, sp, #1312
+; CHECK-NEXT:    add x9, sp, #1136
 ; CHECK-NEXT:    ld1 { v21.b }[2], [x10]
 ; CHECK-NEXT:    ld1 { v16.b }[3], [x8]
-; CHECK-NEXT:    add x10, sp, #1240
-; CHECK-NEXT:    add x8, sp, #1376
+; CHECK-NEXT:    add x10, sp, #1256
+; CHECK-NEXT:    add x8, sp, #1392
 ; CHECK-NEXT:    ld1 { v18.b }[2], [x11]
-; CHECK-NEXT:    add x11, sp, #1304
+; CHECK-NEXT:    add x11, sp, #1320
 ; CHECK-NEXT:    ld1 { v17.b }[4], [x9]
-; CHECK-NEXT:    add x9, sp, #1128
+; CHECK-NEXT:    add x9, sp, #1144
 ; CHECK-NEXT:    ld1 { v21.b }[3], [x10]
-; CHECK-NEXT:    add x10, sp, #1248
+; CHECK-NEXT:    add x10, sp, #1264
 ; CHECK-NEXT:    ld1 { v16.b }[4], [x8]
-; CHECK-NEXT:    add x8, sp, #1384
+; CHECK-NEXT:    add x8, sp, #1400
 ; CHECK-NEXT:    ld1 { v18.b }[3], [x11]
-; CHECK-NEXT:    add x11, sp, #1312
+; CHECK-NEXT:    add x11, sp, #1328
 ; CHECK-NEXT:    ld1 { v17.b }[5], [x9]
-; CHECK-NEXT:    add x9, sp, #1136
+; CHECK-NEXT:    add x9, sp, #1152
 ; CHECK-NEXT:    ld1 { v21.b }[4], [x10]
-; CHECK-NEXT:    add x10, sp, #1256
+; CHECK-NEXT:    add x10, sp, #1272
 ; CHECK-NEXT:    ld1 { v16.b }[5], [x8]
-; CHECK-NEXT:    add x8, sp, #1392
+; CHECK-NEXT:    add x8, sp, #1408
 ; CHECK-NEXT:    ld1 { v18.b }[4], [x11]
-; CHECK-NEXT:    add x11, sp, #1320
+; CHECK-NEXT:    add x11, sp, #1336
 ; CHECK-NEXT:    ld1 { v17.b }[6], [x9]
-; CHECK-NEXT:    add x9, sp, #1144
+; CHECK-NEXT:    add x9, sp, #1160
 ; CHECK-NEXT:    ld1 { v21.b }[5], [x10]
-; CHECK-NEXT:    add x10, sp, #1264
+; CHECK-NEXT:    add x10, sp, #1280
 ; CHECK-NEXT:    ld1 { v16.b }[6], [x8]
-; CHECK-NEXT:    add x8, sp, #1400
+; CHECK-NEXT:    add x8, sp, #1416
 ; CHECK-NEXT:    ld1 { v18.b }[5], [x11]
-; CHECK-NEXT:    add x11, sp, #1328
+; CHECK-NEXT:    add x11, sp, #1344
 ; CHECK-NEXT:    ld1 { v17.b }[7], [x9]
-; CHECK-NEXT:    add x9, sp, #1272
+; CHECK-NEXT:    add x9, sp, #1288
 ; CHECK-NEXT:    sshll v7.8h, v7.8b, #0
 ; CHECK-NEXT:    ld1 { v21.b }[6], [x10]
 ; CHECK-NEXT:    ld1 { v16.b }[7], [x8]
-; CHECK-NEXT:    add x8, sp, #1336
+; CHECK-NEXT:    add x8, sp, #1352
 ; CHECK-NEXT:    ld1 { v18.b }[6], [x11]
 ; CHECK-NEXT:    sshll v17.8h, v17.8b, #0
 ; CHECK-NEXT:    sshll v5.8h, v5.8b, #0
 ; CHECK-NEXT:    ld1 { v21.b }[7], [x9]
 ; CHECK-NEXT:    smull2 v20.4s, v7.8h, v17.8h
 ; CHECK-NEXT:    smull v7.4s, v7.4h, v17.4h
 ; CHECK-NEXT:    ld1 { v18.b }[7], [x8]
 ; CHECK-NEXT:    sshll v16.8h, v16.8b, #0

 ; CHECK-NEXT:    add v1.4s, v2.4s, v3.4s
 ; CHECK-NEXT:    add v2.4s, v20.4s, v19.4s
 ; CHECK-NEXT:    add v3.4s, v7.4s, v22.4s
 ; CHECK-NEXT:    add v0.4s, v1.4s, v0.4s
 ; CHECK-NEXT:    add v1.4s, v3.4s, v2.4s
 ; CHECK-NEXT:    add v0.4s, v0.4s, v1.4s
 ; CHECK-NEXT:    addv s0, v0.4s
 ; CHECK-NEXT:    fmov w0, s0
+; CHECK-NEXT:    ldr x29, [sp], #16 // 8-byte Folded Reload
 ; CHECK-NEXT:    ret
 entry:
   %az = sext <48 x i8> %a to <48 x i32>
   %bz = sext <48 x i8> %b to <48 x i32>
   %m1 = mul nuw nsw <48 x i32> %az, %bz
   %r1 = call i32 @llvm.vector.reduce.add.v48i32(<48 x i32> %m1)
   %cz = sext <48 x i8> %c to <48 x i32>
   %dz = sext <48 x i8> %d to <48 x i32>
   %m2 = mul nuw nsw <48 x i32> %cz, %dz
   %r2 = call i32 @llvm.vector.reduce.add.v48i32(<48 x i32> %m2)
   %x = add i32 %r1, %r2
   ret i32 %x
 }
 
 define i32 @test_sdot_v48i8_double_nomla(<48 x i8> %a, <48 x i8> %b, <48 x i8> %c, <48 x i8> %d) {
 ; CHECK-LABEL: test_sdot_v48i8_double_nomla:
 ; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    ldr b0, [sp, #64]
-; CHECK-NEXT:    add x8, sp, #72
-; CHECK-NEXT:    ldr b1, [sp, #192]
-; CHECK-NEXT:    add x9, sp, #80
-; CHECK-NEXT:    ldr b4, [sp, #128]
-; CHECK-NEXT:    add x10, sp, #136
+; CHECK-NEXT:    str x29, [sp, #-16]! // 8-byte Folded Spill
+; CHECK-NEXT:    .cfi_def_cfa_offset 16
+; CHECK-NEXT:    .cfi_offset w29, -16
+; CHECK-NEXT:    ldr b0, [sp, #80]
+; CHECK-NEXT:    add x8, sp, #88
+; CHECK-NEXT:    ldr b1, [sp, #208]
+; CHECK-NEXT:    add x9, sp, #96
+; CHECK-NEXT:    ldr b4, [sp, #144]
+; CHECK-NEXT:    add x10, sp, #152
 ; CHECK-NEXT:    ld1 { v0.b }[1], [x8]
-; CHECK-NEXT:    add x8, sp, #200
-; CHECK-NEXT:    add x11, sp, #152
-; CHECK-NEXT:    ldr b6, [sp, #256]
+; CHECK-NEXT:    add x8, sp, #216
+; CHECK-NEXT:    add x11, sp, #168
+; CHECK-NEXT:    ldr b6, [sp, #272]
 ; CHECK-NEXT:    ld1 { v4.b }[1], [x10]
-; CHECK-NEXT:    add x10, sp, #144
+; CHECK-NEXT:    add x10, sp, #160
 ; CHECK-NEXT:    ld1 { v1.b }[1], [x8]
-; CHECK-NEXT:    add x8, sp, #88
+; CHECK-NEXT:    add x8, sp, #104
 ; CHECK-NEXT:    ld1 { v0.b }[2], [x9]
-; CHECK-NEXT:    add x9, sp, #208
-; CHECK-NEXT:    ldr b7, [sp]
+; CHECK-NEXT:    add x9, sp, #224
+; CHECK-NEXT:    ldr b7, [sp, #16]
 ; CHECK-NEXT:    fmov s5, w0
 ; CHECK-NEXT:    ld1 { v4.b }[2], [x10]
-; CHECK-NEXT:    add x10, sp, #112
+; CHECK-NEXT:    add x10, sp, #128
 ; CHECK-NEXT:    ld1 { v1.b }[2], [x9]
-; CHECK-NEXT:    add x9, sp, #96
+; CHECK-NEXT:    add x9, sp, #112
 ; CHECK-NEXT:    ld1 { v0.b }[3], [x8]
-; CHECK-NEXT:    add x8, sp, #216
+; CHECK-NEXT:    add x8, sp, #232
 ; CHECK-NEXT:    mov v5.b[1], w1
-; CHECK-NEXT:    ldr b3, [sp, #960]
+; CHECK-NEXT:    ldr b3, [sp, #976]
 ; CHECK-NEXT:    ld1 { v4.b }[3], [x11]
-; CHECK-NEXT:    add x11, sp, #264
+; CHECK-NEXT:    add x11, sp, #280
 ; CHECK-NEXT:    ld1 { v1.b }[3], [x8]
-; CHECK-NEXT:    add x8, sp, #104
+; CHECK-NEXT:    add x8, sp, #120
 ; CHECK-NEXT:    ld1 { v0.b }[4], [x9]
-; CHECK-NEXT:    add x9, sp, #224
+; CHECK-NEXT:    add x9, sp, #240
 ; CHECK-NEXT:    ld1 { v6.b }[1], [x11]
-; CHECK-NEXT:    add x11, sp, #304
+; CHECK-NEXT:    add x11, sp, #320
 ; CHECK-NEXT:    mov v5.b[2], w2
-; CHECK-NEXT:    ldr b19, [sp, #768]
+; CHECK-NEXT:    ldr b19, [sp, #784]
 ; CHECK-NEXT:    ld1 { v1.b }[4], [x9]
-; CHECK-NEXT:    add x9, sp, #232
+; CHECK-NEXT:    add x9, sp, #248
 ; CHECK-NEXT:    ld1 { v0.b }[5], [x8]
-; CHECK-NEXT:    add x8, sp, #160
+; CHECK-NEXT:    add x8, sp, #176
 ; CHECK-NEXT:    mov v5.b[3], w3
 ; CHECK-NEXT:    ld1 { v1.b }[5], [x9]
-; CHECK-NEXT:    add x9, sp, #120
+; CHECK-NEXT:    add x9, sp, #136
 ; CHECK-NEXT:    ld1 { v0.b }[6], [x10]
-; CHECK-NEXT:    add x10, sp, #240
+; CHECK-NEXT:    add x10, sp, #256
 ; CHECK-NEXT:    ld1 { v4.b }[4], [x8]
-; CHECK-NEXT:    add x8, sp, #272
+; CHECK-NEXT:    add x8, sp, #288
 ; CHECK-NEXT:    mov v5.b[4], w4
 ; CHECK-NEXT:    ld1 { v1.b }[6], [x10]
-; CHECK-NEXT:    add x10, sp, #8
+; CHECK-NEXT:    add x10, sp, #24
 ; CHECK-NEXT:    ld1 { v0.b }[7], [x9]
-; CHECK-NEXT:    add x9, sp, #248
+; CHECK-NEXT:    add x9, sp, #264
 ; CHECK-NEXT:    ld1 { v6.b }[2], [x8]
-; CHECK-NEXT:    add x8, sp, #16
+; CHECK-NEXT:    add x8, sp, #32
 ; CHECK-NEXT:    ld1 { v7.b }[1], [x10]
-; CHECK-NEXT:    add x10, sp, #168
+; CHECK-NEXT:    add x10, sp, #184
 ; CHECK-NEXT:    ld1 { v1.b }[7], [x9]
-; CHECK-NEXT:    add x9, sp, #280
+; CHECK-NEXT:    add x9, sp, #296
 ; CHECK-NEXT:    mov v5.b[5], w5
 ; CHECK-NEXT:    ld1 { v4.b }[5], [x10]
-; CHECK-NEXT:    add x10, sp, #176
+; CHECK-NEXT:    add x10, sp, #192
 ; CHECK-NEXT:    ld1 { v6.b }[3], [x9]
-; CHECK-NEXT:    add x9, sp, #288
+; CHECK-NEXT:    add x9, sp, #304
 ; CHECK-NEXT:    ld1 { v7.b }[2], [x8]
-; CHECK-NEXT:    add x8, sp, #24
+; CHECK-NEXT:    add x8, sp, #40
 ; CHECK-NEXT:    sshll v2.8h, v1.8b, #0
-; CHECK-NEXT:    ldr b1, [sp, #832]
+; CHECK-NEXT:    ldr b1, [sp, #848]
 ; CHECK-NEXT:    ld1 { v4.b }[6], [x10]
-; CHECK-NEXT:    add x10, sp, #296
+; CHECK-NEXT:    add x10, sp, #312
 ; CHECK-NEXT:    ld1 { v6.b }[4], [x9]
-; CHECK-NEXT:    add x9, sp, #840
+; CHECK-NEXT:    add x9, sp, #856
 ; CHECK-NEXT:    ld1 { v7.b }[3], [x8]
-; CHECK-NEXT:    add x8, sp, #32
+; CHECK-NEXT:    add x8, sp, #48
 ; CHECK-NEXT:    mov v5.b[6], w6
 ; CHECK-NEXT:    ld1 { v1.b }[1], [x9]
-; CHECK-NEXT:    add x9, sp, #848
+; CHECK-NEXT:    add x9, sp, #864
 ; CHECK-NEXT:    ld1 { v6.b }[5], [x10]
-; CHECK-NEXT:    add x10, sp, #968
+; CHECK-NEXT:    add x10, sp, #984
 ; CHECK-NEXT:    ld1 { v7.b }[4], [x8]
-; CHECK-NEXT:    add x8, sp, #40
+; CHECK-NEXT:    add x8, sp, #56
 ; CHECK-NEXT:    mov v5.b[7], w7
 ; CHECK-NEXT:    ld1 { v3.b }[1], [x10]
-; CHECK-NEXT:    add x10, sp, #976
+; CHECK-NEXT:    add x10, sp, #992
 ; CHECK-NEXT:    ld1 { v1.b }[2], [x9]
-; CHECK-NEXT:    add x9, sp, #856
+; CHECK-NEXT:    add x9, sp, #872
 ; CHECK-NEXT:    ld1 { v7.b }[5], [x8]
-; CHECK-NEXT:    add x8, sp, #48
+; CHECK-NEXT:    add x8, sp, #64
 ; CHECK-NEXT:    ld1 { v6.b }[6], [x11]
-; CHECK-NEXT:    add x11, sp, #312
+; CHECK-NEXT:    add x11, sp, #328
 ; CHECK-NEXT:    ld1 { v3.b }[2], [x10]
-; CHECK-NEXT:    add x10, sp, #984
+; CHECK-NEXT:    add x10, sp, #1000
 ; CHECK-NEXT:    ld1 { v1.b }[3], [x9]
-; CHECK-NEXT:    add x9, sp, #864
+; CHECK-NEXT:    add x9, sp, #880
 ; CHECK-NEXT:    ld1 { v7.b }[6], [x8]
-; CHECK-NEXT:    add x8, sp, #56
+; CHECK-NEXT:    add x8, sp, #72
 ; CHECK-NEXT:    ld1 { v6.b }[7], [x11]
-; CHECK-NEXT:    add x11, sp, #184
+; CHECK-NEXT:    add x11, sp, #200
 ; CHECK-NEXT:    ld1 { v3.b }[3], [x10]
-; CHECK-NEXT:    add x10, sp, #712
+; CHECK-NEXT:    add x10, sp, #728
 ; CHECK-NEXT:    ld1 { v1.b }[4], [x9]
-; CHECK-NEXT:    add x9, sp, #992
+; CHECK-NEXT:    add x9, sp, #1008
 ; CHECK-NEXT:    ld1 { v7.b }[7], [x8]
-; CHECK-NEXT:    add x8, sp, #872
+; CHECK-NEXT:    add x8, sp, #888
 ; CHECK-NEXT:    sshll v16.8h, v6.8b, #0
-; CHECK-NEXT:    ldr b6, [sp, #704]
+; CHECK-NEXT:    ldr b6, [sp, #720]
 ; CHECK-NEXT:    ld1 { v3.b }[4], [x9]
-; CHECK-NEXT:    add x9, sp, #1000
+; CHECK-NEXT:    add x9, sp, #1016
 ; CHECK-NEXT:    ld1 { v1.b }[5], [x8]
-; CHECK-NEXT:    add x8, sp, #880
+; CHECK-NEXT:    add x8, sp, #896
 ; CHECK-NEXT:    sshll v18.8h, v7.8b, #0
 ; CHECK-NEXT:    ld1 { v6.b }[1], [x10]
 ; CHECK-NEXT:    sshll v17.8h, v5.8b, #0
-; CHECK-NEXT:    add x10, sp, #776
+; CHECK-NEXT:    add x10, sp, #792
 ; CHECK-NEXT:    ld1 { v3.b }[5], [x9]
-; CHECK-NEXT:    add x9, sp, #1008
+; CHECK-NEXT:    add x9, sp, #1024
 ; CHECK-NEXT:    ld1 { v1.b }[6], [x8]
-; CHECK-NEXT:    add x8, sp, #888
+; CHECK-NEXT:    add x8, sp, #904
 ; CHECK-NEXT:    saddl2 v5.4s, v18.8h, v16.8h
 ; CHECK-NEXT:    ld1 { v19.b }[1], [x10]
 ; CHECK-NEXT:    saddl v16.4s, v18.4h, v16.4h
-; CHECK-NEXT:    ldr b18, [sp, #1024]
+; CHECK-NEXT:    ldr b18, [sp, #1040]
 ; CHECK-NEXT:    ld1 { v3.b }[6], [x9]
-; CHECK-NEXT:    add x9, sp, #1032
+; CHECK-NEXT:    add x9, sp, #1048
 ; CHECK-NEXT:    ld1 { v1.b }[7], [x8]
-; CHECK-NEXT:    add x8, sp, #1016
+; CHECK-NEXT:    add x8, sp, #1032
 ; CHECK-NEXT:    saddl2 v7.4s, v17.8h, v2.8h
-; CHECK-NEXT:    add x10, sp, #904
+; CHECK-NEXT:    add x10, sp, #920
 ; CHECK-NEXT:    saddl v2.4s, v17.4h, v2.4h
 ; CHECK-NEXT:    ld1 { v18.b }[1], [x9]
 ; CHECK-NEXT:    ld1 { v3.b }[7], [x8]
-; CHECK-NEXT:    add x8, sp, #720
-; CHECK-NEXT:    ldr b17, [sp, #896]
-; CHECK-NEXT:    add x9, sp, #1040
+; CHECK-NEXT:    add x8, sp, #736
+; CHECK-NEXT:    ldr b17, [sp, #912]
+; CHECK-NEXT:    add x9, sp, #1056
 ; CHECK-NEXT:    ld1 { v4.b }[7], [x11]
-; CHECK-NEXT:    add x11, sp, #728
+; CHECK-NEXT:    add x11, sp, #744
 ; CHECK-NEXT:    ld1 { v6.b }[2], [x8]
-; CHECK-NEXT:    add x8, sp, #784
+; CHECK-NEXT:    add x8, sp, #800
 ; CHECK-NEXT:    ld1 { v17.b }[1], [x10]
-; CHECK-NEXT:    add x10, sp, #912
+; CHECK-NEXT:    add x10, sp, #928
 ; CHECK-NEXT:    ld1 { v18.b }[2], [x9]
-; CHECK-NEXT:    add x9, sp, #1048
+; CHECK-NEXT:    add x9, sp, #1064
 ; CHECK-NEXT:    ld1 { v19.b }[2], [x8]
-; CHECK-NEXT:    add x8, sp, #792
+; CHECK-NEXT:    add x8, sp, #808
 ; CHECK-NEXT:    ld1 { v6.b }[3], [x11]
-; CHECK-NEXT:    add x11, sp, #736
+; CHECK-NEXT:    add x11, sp, #752
 ; CHECK-NEXT:    ld1 { v17.b }[2], [x10]
-; CHECK-NEXT:    add x10, sp, #920
+; CHECK-NEXT:    add x10, sp, #936
 ; CHECK-NEXT:    ld1 { v18.b }[3], [x9]
-; CHECK-NEXT:    add x9, sp, #1056
+; CHECK-NEXT:    add x9, sp, #1072
 ; CHECK-NEXT:    ld1 { v19.b }[3], [x8]
-; CHECK-NEXT:    add x8, sp, #800
+; CHECK-NEXT:    add x8, sp, #816
 ; CHECK-NEXT:    ld1 { v6.b }[4], [x11]
-; CHECK-NEXT:    add x11, sp, #744
+; CHECK-NEXT:    add x11, sp, #760
 ; CHECK-NEXT:    ld1 { v17.b }[3], [x10]
-; CHECK-NEXT:    add x10, sp, #928
+; CHECK-NEXT:    add x10, sp, #944
 ; CHECK-NEXT:    ld1 { v18.b }[4], [x9]
-; CHECK-NEXT:    add x9, sp, #1064
+; CHECK-NEXT:    add x9, sp, #1080
 ; CHECK-NEXT:    ld1 { v19.b }[4], [x8]
-; CHECK-NEXT:    add x8, sp, #808
+; CHECK-NEXT:    add x8, sp, #824
 ; CHECK-NEXT:    ld1 { v6.b }[5], [x11]
-; CHECK-NEXT:    add x11, sp, #752
+; CHECK-NEXT:    add x11, sp, #768
 ; CHECK-NEXT:    ld1 { v17.b }[4], [x10]
-; CHECK-NEXT:    add x10, sp, #936
+; CHECK-NEXT:    add x10, sp, #952
 ; CHECK-NEXT:    ld1 { v18.b }[5], [x9]
-; CHECK-NEXT:    add x9, sp, #1072
+; CHECK-NEXT:    add x9, sp, #1088
 ; CHECK-NEXT:    ld1 { v19.b }[5], [x8]
-; CHECK-NEXT:    add x8, sp, #816
+; CHECK-NEXT:    add x8, sp, #832
 ; CHECK-NEXT:    ld1 { v6.b }[6], [x11]
-; CHECK-NEXT:    add x11, sp, #760
+; CHECK-NEXT:    add x11, sp, #776
 ; CHECK-NEXT:    ld1 { v17.b }[5], [x10]
-; CHECK-NEXT:    add x10, sp, #944
+; CHECK-NEXT:    add x10, sp, #960
 ; CHECK-NEXT:    ld1 { v18.b }[6], [x9]
-; CHECK-NEXT:    add x9, sp, #1080
+; CHECK-NEXT:    add x9, sp, #1096
 ; CHECK-NEXT:    ld1 { v19.b }[6], [x8]
-; CHECK-NEXT:    add x8, sp, #824
+; CHECK-NEXT:    add x8, sp, #840
 ; CHECK-NEXT:    sshll v0.8h, v0.8b, #0
 ; CHECK-NEXT:    ld1 { v6.b }[7], [x11]
 ; CHECK-NEXT:    sshll v4.8h, v4.8b, #0
 ; CHECK-NEXT:    ld1 { v17.b }[6], [x10]
 ; CHECK-NEXT:    ld1 { v18.b }[7], [x9]
 ; CHECK-NEXT:    ld1 { v19.b }[7], [x8]
-; CHECK-NEXT:    add x8, sp, #952
+; CHECK-NEXT:    add x8, sp, #968
 ; CHECK-NEXT:    saddw2 v5.4s, v5.4s, v4.8h
 ; CHECK-NEXT:    saddw2 v7.4s, v7.4s, v0.8h
 ; CHECK-NEXT:    ld1 { v17.b }[7], [x8]
 ; CHECK-NEXT:    saddw v0.4s, v2.4s, v0.4h
 ; CHECK-NEXT:    add v5.4s, v7.4s, v5.4s
 ; CHECK-NEXT:    sshll v2.8h, v3.8b, #0
 ; CHECK-NEXT:    sshll v3.8h, v18.8b, #0
 ; CHECK-NEXT:    sshll v7.8h, v19.8b, #0

 ; CHECK-NEXT:    add v0.4s, v0.4s, v4.4s
 ; CHECK-NEXT:    add v2.4s, v7.4s, v16.4s
 ; CHECK-NEXT:    add v1.4s, v1.4s, v3.4s
 ; CHECK-NEXT:    add v0.4s, v0.4s, v5.4s
 ; CHECK-NEXT:    add v1.4s, v1.4s, v2.4s
 ; CHECK-NEXT:    add v0.4s, v0.4s, v1.4s
 ; CHECK-NEXT:    addv s0, v0.4s
 ; CHECK-NEXT:    fmov w0, s0
+; CHECK-NEXT:    ldr x29, [sp], #16 // 8-byte Folded Reload
 ; CHECK-NEXT:    ret
 entry:
   %az = sext <48 x i8> %a to <48 x i32>
   %r1 = call i32 @llvm.vector.reduce.add.v48i32(<48 x i32> %az)
   %cz = sext <48 x i8> %c to <48 x i32>
   %r2 = call i32 @llvm.vector.reduce.add.v48i32(<48 x i32> %cz)
   %x = add i32 %r1, %r2
   ret i32 %x

 ; CHECK-NEXT:    ret
 entry:
   %az = sext <64 x i8> %a to <64 x i32>
   %r1 = call i32 @llvm.vector.reduce.add.v64i32(<64 x i32> %az)
   %cz = sext <64 x i8> %c to <64 x i32>
   %r2 = call i32 @llvm.vector.reduce.add.v64i32(<64 x i32> %cz)
   %x = add i32 %r1, %r2
   ret i32 %x
-}
+}
 

llvm/test/CodeGen/AArch64/scavenge-large-call.ll

+; RUN: llc -mtriple=arm64_32-apple-watchos %s -o - | FileCheck %s
+
+; CHECK-LABEL: caller:
+; CHECK: add {{x[0-9]+}}, sp,
+
+define void @caller(ptr %0, i16 %1, i16 %2, i8 %3, double %4, i16 %5, i8 %6, ptr %7, double %8, i32 %9, ptr %10, double %11, double %12, [2 x i64] %13, [2 x i64] %14, [2 x i64] %15, double %16, double %17, [2 x i64] %18, [2 x i64] %19, i16 %20, i32 %21, double %22, i8 %23, [2 x i64] %24, [2 x i64] %25, [2 x i64] %26, i8 %27, i16 %28, i16 %29, i16 %30, i32 %31, [2 x i64] %32, [2 x i64] %33, [2 x i64] %34, [2 x i64] %35, [2 x i64] %36, i32 %37, i32 %38) {
+  tail call void @callee(ptr %0, i16 %1, i16 %2, i8 %3, double 0.000000e+00, i16 %5, i8 %6, ptr %7, double 0.000000e+00, i32 %9, ptr %10, double 0.000000e+00, double 0.000000e+00, [2 x i64] %13, [2 x i64] %14, [2 x i64] %15, double 0.000000e+00, double 0.000000e+00, [2 x i64] %18, [2 x i64] %19, i16 %20, i32 %21, double 0.000000e+00, i8 %23, [2 x i64] %24, [2 x i64] %25, [2 x i64] zeroinitializer, i8 %27, i16 0, i16 0, i16 %28, i32 0, [2 x i64] zeroinitializer, [2 x i64] zeroinitializer, [2 x i64] zeroinitializer, [2 x i64] %35, [2 x i64] %36, i32 0, i32 0)
+  ret void
+}
+
+declare void @callee(ptr, i16, i16, i8, double, i16, i8, ptr, double, i32, ptr, double, double, [2 x i64], [2 x i64], [2 x i64], double, double, [2 x i64], [2 x i64], i16, i32, double, i8, [2 x i64], [2 x i64], [2 x i64], i8, i16, i16, i16, i32, [2 x i64], [2 x i64], [2 x i64], [2 x i64], [2 x i64], i32, i32)

llvm/test/CodeGen/AArch64/stackmap.ll

 ; CHECK-NEXT:   .xword 1
 ; CHECK-NEXT:   .xword jsVoidCall
 ; CHECK-NEXT:   .xword 16
 ; CHECK-NEXT:   .xword 1
 ; CHECK-NEXT:   .xword jsIntCall
 ; CHECK-NEXT:   .xword 16
 ; CHECK-NEXT:   .xword 1
 ; CHECK-NEXT:   .xword spilledValue
-; CHECK-NEXT:   .xword 144
+; CHECK-NEXT:   .xword 160
 ; CHECK-NEXT:   .xword 1
 ; CHECK-NEXT:   .xword spilledStackMapValue
 ; CHECK-NEXT:   .xword 128
 ; CHECK-NEXT:   .xword 1
 ; CHECK-NEXT:   .xword liveConstant
 ; CHECK-NEXT:   .xword 16
 ; CHECK-NEXT:   .xword 1
 ; CHECK-NEXT:   .xword directFrameIdx

llvm/test/CodeGen/AArch64/sve-fixed-length-frame-offests-crash.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s | FileCheck %s
 
 target triple = "aarch64-unknown-linux-gnu"
 
 ; Ensure we don't crash by trying to fold fixed length frame indexes into
 ; loads/stores that don't support an appropriate addressing mode, hence creating
 ; too many extra vregs during frame lowering, when we don't have an emergency
 ; spill slot.
 
 define dso_local void @func1(ptr %v1, ptr %v2, ptr %v3, ptr %v4, ptr %v5, ptr %v6, ptr %v7, ptr %v8,
 ; CHECK-LABEL: func1:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    str x25, [sp, #-64]! // 8-byte Folded Spill
+; CHECK-NEXT:    stp x29, x25, [sp, #-64]! // 16-byte Folded Spill
 ; CHECK-NEXT:    .cfi_def_cfa_offset 64
 ; CHECK-NEXT:    stp x24, x23, [sp, #16] // 16-byte Folded Spill
 ; CHECK-NEXT:    stp x22, x21, [sp, #32] // 16-byte Folded Spill
 ; CHECK-NEXT:    stp x20, x19, [sp, #48] // 16-byte Folded Spill
 ; CHECK-NEXT:    .cfi_offset w19, -8
 ; CHECK-NEXT:    .cfi_offset w20, -16
 ; CHECK-NEXT:    .cfi_offset w21, -24
 ; CHECK-NEXT:    .cfi_offset w22, -32
 ; CHECK-NEXT:    .cfi_offset w23, -40
 ; CHECK-NEXT:    .cfi_offset w24, -48
-; CHECK-NEXT:    .cfi_offset w25, -64
+; CHECK-NEXT:    .cfi_offset w25, -56
+; CHECK-NEXT:    .cfi_offset w29, -64
 ; CHECK-NEXT:    add x8, sp, #64
 ; CHECK-NEXT:    add x9, sp, #128
 ; CHECK-NEXT:    add x10, sp, #160
 ; CHECK-NEXT:    add x11, sp, #192
 ; CHECK-NEXT:    ptrue p0.d
 ; CHECK-NEXT:    add x20, sp, #192
 ; CHECK-NEXT:    ld1d { z0.d }, p0/z, [x8]
 ; CHECK-NEXT:    ld1d { z1.d }, p0/z, [x9]

 ; CHECK-NEXT:    ldp x22, x21, [sp, #32] // 16-byte Folded Reload
 ; CHECK-NEXT:    stur q4, [sp, #104]
 ; CHECK-NEXT:    ldp x24, x23, [sp, #16] // 16-byte Folded Reload
 ; CHECK-NEXT:    str x15, [sp, #120]
 ; CHECK-NEXT:    stp x14, x13, [sp, #304]
 ; CHECK-NEXT:    stp x12, x11, [sp, #320]
 ; CHECK-NEXT:    stp x10, x9, [sp, #336]
 ; CHECK-NEXT:    str x8, [sp, #352]
-; CHECK-NEXT:    ldr x25, [sp], #64 // 8-byte Folded Reload
+; CHECK-NEXT:    ldp x29, x25, [sp], #64 // 16-byte Folded Reload
 ; CHECK-NEXT:    b func2
                              ptr %v9, ptr %v10, ptr %v11, ptr %v12, ptr %v13, ptr %v14,  ptr %v15, ptr %v16,
                              ptr %v17, ptr %v18, ptr %v19, ptr %v20, ptr %v21, ptr %v22, ptr %v23, ptr %v24,
                              ptr %v25, ptr %v26, ptr %v27, ptr %v28, ptr %v29, ptr %v30, ptr %v31, ptr %v32,
                              ptr %v33, ptr %v34, ptr %v35, ptr %v36, ptr %v37, ptr %v38, ptr %v39, ptr %v40,
                              ptr %v41, ptr %v42, ptr %v43, ptr %v44, ptr %v45, ptr %v46, ptr %v47, ptr %v48,
                              i64 %v49) #0 {
   tail call void @func2(ptr %v1, ptr %v2, ptr %v3, ptr %v4, ptr %v5, ptr %v6, ptr %v7, ptr %v8,