[LoongArch] Add emergency spill slot for GPR for large frames

An emergency spill slot is created when the stack size cannot be
represented by an 11-bit signed number.

This patch also modifies how the `sp` is adjusted in the prologue.

`RegScavenger` will place the spill instruction before the prologue
if a VReg is created in the prologue. This will pollute the caller's
stack data. Therefore, until there is better way, we just use the
`addi.w/d` instruction for stack adjustment to ensure that VReg will
not be created. (RISCV has the same issue #58286)

Due to the addition of emergency spill slot, some test cases that use
exact stacksize need to be updated.

Differential Revision: https://reviews.llvm.org/D135757

diff --git a/llvm/lib/Target/LoongArch/LoongArchFrameLowering.cpp b/llvm/lib/Target/LoongArch/LoongArchFrameLowering.cpp
index e8985d9..f45ad53 100644 (file)
--- a/llvm/lib/Target/LoongArch/LoongArchFrameLowering.cpp
+++ b/llvm/lib/Target/LoongArch/LoongArchFrameLowering.cpp
@@ -118,6 +118,26 @@ void LoongArchFrameLowering::determineFrameLayout(MachineFunction &MF) const {
   MFI.setStackSize(FrameSize);
 }
 
+void LoongArchFrameLowering::processFunctionBeforeFrameFinalized(
+    MachineFunction &MF, RegScavenger *RS) const {
+  const LoongArchRegisterInfo *RI = STI.getRegisterInfo();
+  const TargetRegisterClass &RC = LoongArch::GPRRegClass;
+  MachineFrameInfo &MFI = MF.getFrameInfo();
+
+  // estimateStackSize has been observed to under-estimate the final stack
+  // size, so give ourselves wiggle-room by checking for stack size
+  // representable an 11-bit signed field rather than 12-bits.
+  if (isInt<11>(MFI.estimateStackSize(MF)))
+    return;
+
+  // Create an emergency spill slot.
+  int FI =
+      MFI.CreateStackObject(RI->getSpillSize(RC), RI->getSpillAlign(RC), false);
+  RS->addScavengingFrameIndex(FI);
+  LLVM_DEBUG(dbgs() << "Allocated FI(" << FI
+                    << ") as the emergency spill slot.\n");
+}
+
 void LoongArchFrameLowering::emitPrologue(MachineFunction &MF,
                                           MachineBasicBlock &MBB) const {
   MachineFrameInfo &MFI = MF.getFrameInfo();
@@ -125,6 +145,7 @@ void LoongArchFrameLowering::emitPrologue(MachineFunction &MF,
   const LoongArchRegisterInfo *RI = STI.getRegisterInfo();
   const LoongArchInstrInfo *TII = STI.getInstrInfo();
   MachineBasicBlock::iterator MBBI = MBB.begin();
+  bool IsLA64 = STI.is64Bit();
 
   Register SPReg = LoongArch::R3;
   Register FPReg = LoongArch::R22;
@@ -144,19 +165,22 @@ void LoongArchFrameLowering::emitPrologue(MachineFunction &MF,
   if (StackSize == 0 && !MFI.adjustsStack())
     return;
 
-  uint64_t FirstSPAdjustAmount = getFirstSPAdjustAmount(MF);
+  uint64_t FirstSPAdjustAmount = getFirstSPAdjustAmount(MF, true);
+  uint64_t SecondSPAdjustAmount = RealStackSize - FirstSPAdjustAmount;
   // Split the SP adjustment to reduce the offsets of callee saved spill.
   if (FirstSPAdjustAmount)
     StackSize = FirstSPAdjustAmount;
 
   // Adjust stack.
   adjustReg(MBB, MBBI, DL, SPReg, SPReg, -StackSize, MachineInstr::FrameSetup);
-  // Emit ".cfi_def_cfa_offset StackSize".
-  unsigned CFIIndex =
-      MF.addFrameInst(MCCFIInstruction::cfiDefCfaOffset(nullptr, StackSize));
-  BuildMI(MBB, MBBI, DL, TII->get(TargetOpcode::CFI_INSTRUCTION))
-      .addCFIIndex(CFIIndex)
-      .setMIFlag(MachineInstr::FrameSetup);
+  if (FirstSPAdjustAmount != 2048 || SecondSPAdjustAmount == 0) {
+    // Emit ".cfi_def_cfa_offset StackSize".
+    unsigned CFIIndex =
+        MF.addFrameInst(MCCFIInstruction::cfiDefCfaOffset(nullptr, StackSize));
+    BuildMI(MBB, MBBI, DL, TII->get(TargetOpcode::CFI_INSTRUCTION))
+        .addCFIIndex(CFIIndex)
+        .setMIFlag(MachineInstr::FrameSetup);
+  }
 
   const auto &CSI = MFI.getCalleeSavedInfo();
 
@@ -193,14 +217,25 @@ void LoongArchFrameLowering::emitPrologue(MachineFunction &MF,
   }
 
   // Emit the second SP adjustment after saving callee saved registers.
-  if (FirstSPAdjustAmount) {
-    uint64_t SecondSPAdjustAmount = RealStackSize - FirstSPAdjustAmount;
-    assert(SecondSPAdjustAmount > 0 &&
-           "SecondSPAdjustAmount should be greater than zero");
-    adjustReg(MBB, MBBI, DL, SPReg, SPReg, -SecondSPAdjustAmount,
-              MachineInstr::FrameSetup);
+  if (FirstSPAdjustAmount && SecondSPAdjustAmount) {
+    if (hasFP(MF)) {
+      assert(SecondSPAdjustAmount > 0 &&
+             "SecondSPAdjustAmount should be greater than zero");
+      adjustReg(MBB, MBBI, DL, SPReg, SPReg, -SecondSPAdjustAmount,
+                MachineInstr::FrameSetup);
+    } else {
+      // FIXME: RegScavenger will place the spill instruction before the
+      // prologue if a VReg is created in the prologue. This will pollute the
+      // caller's stack data. Therefore, until there is better way, we just use
+      // the `addi.w/d` instruction for stack adjustment to ensure that VReg
+      // will not be created.
+      for (int Val = SecondSPAdjustAmount; Val > 0; Val -= 2048)
+        BuildMI(MBB, MBBI, DL,
+                TII->get(IsLA64 ? LoongArch::ADDI_D : LoongArch::ADDI_W), SPReg)
+            .addReg(SPReg)
+            .addImm(Val < 2048 ? -Val : -2048)
+            .setMIFlag(MachineInstr::FrameSetup);
 
-    if (!hasFP(MF)) {
       // If we are using a frame-pointer, and thus emitted ".cfi_def_cfa fp, 0",
       // don't emit an sp-based .cfi_def_cfa_offset
       // Emit ".cfi_def_cfa_offset RealStackSize"
@@ -219,14 +254,12 @@ void LoongArchFrameLowering::emitPrologue(MachineFunction &MF,
       Register VR =
           MF.getRegInfo().createVirtualRegister(&LoongArch::GPRRegClass);
       BuildMI(MBB, MBBI, DL,
-              TII->get(STI.is64Bit() ? LoongArch::SRLI_D : LoongArch::SRLI_W),
-              VR)
+              TII->get(IsLA64 ? LoongArch::SRLI_D : LoongArch::SRLI_W), VR)
           .addReg(SPReg)
           .addImm(ShiftAmount)
           .setMIFlag(MachineInstr::FrameSetup);
       BuildMI(MBB, MBBI, DL,
-              TII->get(STI.is64Bit() ? LoongArch::SLLI_D : LoongArch::SLLI_W),
-              SPReg)
+              TII->get(IsLA64 ? LoongArch::SLLI_D : LoongArch::SLLI_W), SPReg)
           .addReg(VR)
           .addImm(ShiftAmount)
           .setMIFlag(MachineInstr::FrameSetup);
@@ -295,20 +328,27 @@ void LoongArchFrameLowering::emitEpilogue(MachineFunction &MF,
 //   st.d    $ra, $sp,  2024
 //   st.d    $fp, $sp,  2016
 //   addi.d  $sp, $sp,   -16
-uint64_t LoongArchFrameLowering::getFirstSPAdjustAmount(
-    const MachineFunction &MF) const {
+uint64_t
+LoongArchFrameLowering::getFirstSPAdjustAmount(const MachineFunction &MF,
+                                               bool IsPrologue) const {
   const MachineFrameInfo &MFI = MF.getFrameInfo();
   const std::vector<CalleeSavedInfo> &CSI = MFI.getCalleeSavedInfo();
 
   // Return the FirstSPAdjustAmount if the StackSize can not fit in a signed
   // 12-bit and there exists a callee-saved register needing to be pushed.
-  if (!isInt<12>(MFI.getStackSize()) && (CSI.size() > 0)) {
+  if (!isInt<12>(MFI.getStackSize())) {
     // FirstSPAdjustAmount is chosen as (2048 - StackAlign) because 2048 will
     // cause sp = sp + 2048 in the epilogue to be split into multiple
     // instructions. Offsets smaller than 2048 can fit in a single load/store
     // instruction, and we have to stick with the stack alignment.
     // So (2048 - StackAlign) will satisfy the stack alignment.
-    return 2048 - getStackAlign().value();
+    //
+    // FIXME: This place may seem odd. When using multiple ADDI instructions to
+    // adjust the stack in Prologue, and there are no callee-saved registers, we
+    // can take advantage of the logic of split sp ajustment to reduce code
+    // changes.
+    return CSI.size() > 0 ? 2048 - getStackAlign().value()
+                          : (IsPrologue ? 2048 : 0);
   }
   return 0;
 }

diff --git a/llvm/lib/Target/LoongArch/LoongArchFrameLowering.h b/llvm/lib/Target/LoongArch/LoongArchFrameLowering.h
index 7ef79aa..f14d9b8 100644 (file)
--- a/llvm/lib/Target/LoongArch/LoongArchFrameLowering.h
+++ b/llvm/lib/Target/LoongArch/LoongArchFrameLowering.h
@@ -34,6 +34,9 @@ public:
   void determineCalleeSaves(MachineFunction &MF, BitVector &SavedRegs,
                             RegScavenger *RS) const override;
 
+  void processFunctionBeforeFrameFinalized(MachineFunction &MF,
+                                           RegScavenger *RS) const override;
+
   bool hasReservedCallFrame(const MachineFunction &MF) const override;
   MachineBasicBlock::iterator
   eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB,
@@ -45,7 +48,8 @@ public:
   bool hasFP(const MachineFunction &MF) const override;
   bool hasBP(const MachineFunction &MF) const;
 
-  uint64_t getFirstSPAdjustAmount(const MachineFunction &MF) const;
+  uint64_t getFirstSPAdjustAmount(const MachineFunction &MF,
+                                  bool IsPrologue = false) const;
 
 private:
   void determineFrameLayout(MachineFunction &MF) const;

diff --git a/llvm/test/CodeGen/LoongArch/emergency-spill-slot.ll b/llvm/test/CodeGen/LoongArch/emergency-spill-slot.ll
new file mode 100644 (file)
index 0000000..08426b0
--- /dev/null
+++ b/llvm/test/CodeGen/LoongArch/emergency-spill-slot.ll
@@ -0,0 +1,101 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc --mtriple=loongarch64 -O0 < %s | FileCheck %s
+
+@var = external global i32
+
+define void @func() {
+; CHECK-LABEL: func:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    addi.d $sp, $sp, -2048
+; CHECK-NEXT:    addi.d $sp, $sp, -2048
+; CHECK-NEXT:    addi.d $sp, $sp, -16
+; CHECK-NEXT:    .cfi_def_cfa_offset 4112
+; CHECK-NEXT:    pcalau12i $a0, %got_pc_hi20(var)
+; CHECK-NEXT:    ld.d $a1, $a0, %got_pc_lo12(var)
+; CHECK-NEXT:    ld.w $t8, $a1, 0
+; CHECK-NEXT:    ld.w $t7, $a1, 0
+; CHECK-NEXT:    ld.w $t6, $a1, 0
+; CHECK-NEXT:    ld.w $t5, $a1, 0
+; CHECK-NEXT:    ld.w $t4, $a1, 0
+; CHECK-NEXT:    ld.w $t3, $a1, 0
+; CHECK-NEXT:    ld.w $t2, $a1, 0
+; CHECK-NEXT:    ld.w $t1, $a1, 0
+; CHECK-NEXT:    ld.w $t0, $a1, 0
+; CHECK-NEXT:    ld.w $a7, $a1, 0
+; CHECK-NEXT:    ld.w $a6, $a1, 0
+; CHECK-NEXT:    ld.w $a5, $a1, 0
+; CHECK-NEXT:    ld.w $a4, $a1, 0
+; CHECK-NEXT:    ld.w $a3, $a1, 0
+; CHECK-NEXT:    ld.w $a2, $a1, 0
+; CHECK-NEXT:    ld.w $a0, $a1, 0
+; CHECK-NEXT:    st.d $fp, $sp, 0
+; CHECK-NEXT:    lu12i.w $fp, 1
+; CHECK-NEXT:    ori $fp, $fp, 12
+; CHECK-NEXT:    add.d $fp, $sp, $fp
+; CHECK-NEXT:    st.w $t8, $fp, 0
+; CHECK-NEXT:    ld.d $fp, $sp, 0
+; CHECK-NEXT:    st.w $t8, $a1, 0
+; CHECK-NEXT:    st.w $t7, $a1, 0
+; CHECK-NEXT:    st.w $t6, $a1, 0
+; CHECK-NEXT:    st.w $t5, $a1, 0
+; CHECK-NEXT:    st.w $t4, $a1, 0
+; CHECK-NEXT:    st.w $t3, $a1, 0
+; CHECK-NEXT:    st.w $t2, $a1, 0
+; CHECK-NEXT:    st.w $t1, $a1, 0
+; CHECK-NEXT:    st.w $t0, $a1, 0
+; CHECK-NEXT:    st.w $a7, $a1, 0
+; CHECK-NEXT:    st.w $a6, $a1, 0
+; CHECK-NEXT:    st.w $a5, $a1, 0
+; CHECK-NEXT:    st.w $a4, $a1, 0
+; CHECK-NEXT:    st.w $a3, $a1, 0
+; CHECK-NEXT:    st.w $a2, $a1, 0
+; CHECK-NEXT:    st.w $a0, $a1, 0
+; CHECK-NEXT:    lu12i.w $a0, 1
+; CHECK-NEXT:    ori $a0, $a0, 16
+; CHECK-NEXT:    add.d $sp, $sp, $a0
+; CHECK-NEXT:    ret
+  %space = alloca i32, align 4
+  %stackspace = alloca[1024 x i32], align 4
+
+  ;; Load values to increase register pressure.
+  %v0 = load volatile i32, ptr @var
+  %v1 = load volatile i32, ptr @var
+  %v2 = load volatile i32, ptr @var
+  %v3 = load volatile i32, ptr @var
+  %v4 = load volatile i32, ptr @var
+  %v5 = load volatile i32, ptr @var
+  %v6 = load volatile i32, ptr @var
+  %v7 = load volatile i32, ptr @var
+  %v8 = load volatile i32, ptr @var
+  %v9 = load volatile i32, ptr @var
+  %v10 = load volatile i32, ptr @var
+  %v11 = load volatile i32, ptr @var
+  %v12 = load volatile i32, ptr @var
+  %v13 = load volatile i32, ptr @var
+  %v14 = load volatile i32, ptr @var
+  %v15 = load volatile i32, ptr @var
+
+  ;; Computing a stack-relative values needs an additional register.
+  ;; We should get an emergency spill/reload for this.
+  store volatile i32 %v0, ptr %space
+
+  ;; store values so they are used.
+  store volatile i32 %v0, ptr @var
+  store volatile i32 %v1, ptr @var
+  store volatile i32 %v2, ptr @var
+  store volatile i32 %v3, ptr @var
+  store volatile i32 %v4, ptr @var
+  store volatile i32 %v5, ptr @var
+  store volatile i32 %v6, ptr @var
+  store volatile i32 %v7, ptr @var
+  store volatile i32 %v8, ptr @var
+  store volatile i32 %v9, ptr @var
+  store volatile i32 %v10, ptr @var
+  store volatile i32 %v11, ptr @var
+  store volatile i32 %v12, ptr @var
+  store volatile i32 %v13, ptr @var
+  store volatile i32 %v14, ptr @var
+  store volatile i32 %v15, ptr @var
+
+  ret void
+}

diff --git a/llvm/test/CodeGen/LoongArch/frame.ll b/llvm/test/CodeGen/LoongArch/frame.ll
index d52a946..2a97005 100644 (file)
--- a/llvm/test/CodeGen/LoongArch/frame.ll
+++ b/llvm/test/CodeGen/LoongArch/frame.ll
@@ -27,6 +27,7 @@ define i32 @test() nounwind {
   ret i32 0
 }
 
+;; Note: will create an emergency spill slot, if (!isInt<11>(StackSize)).
 ;; Should involve only one SP-adjusting addi per adjustment.
 define void @test_large_frame_size_2032() {
 ; CHECK-LABEL: test_large_frame_size_2032:
@@ -35,7 +36,7 @@ define void @test_large_frame_size_2032() {
 ; CHECK-NEXT:    .cfi_def_cfa_offset 2032
 ; CHECK-NEXT:    addi.d $sp, $sp, 2032
 ; CHECK-NEXT:    ret
-  %1 = alloca i8, i32 2032
+  %1 = alloca i8, i32 2016 ; + 16(emergency slot) = 2032
   ret void
 }
 
@@ -49,7 +50,7 @@ define void @test_large_frame_size_2048() {
 ; CHECK-NEXT:    addi.d $sp, $sp, 2032
 ; CHECK-NEXT:    addi.d $sp, $sp, 16
 ; CHECK-NEXT:    ret
-  %1 = alloca i8, i32 2048
+  %1 = alloca i8, i32 2032 ; + 16(emergency slot) = 2048
   ret void
 }
 
@@ -63,21 +64,35 @@ define void @test_large_frame_size_2064() {
 ; CHECK-NEXT:    addi.d $sp, $sp, 2032
 ; CHECK-NEXT:    addi.d $sp, $sp, 32
 ; CHECK-NEXT:    ret
-  %1 = alloca i8, i32 2064
+  %1 = alloca i8, i32 2048 ; + 16(emergency slot) = 2064
   ret void
 }
 
+;; NOTE: Due to the problem with the emegency spill slot, the scratch register
+;; will not be used when the fp is eliminated. To make this test valid, add the
+;; attribute "frame-pointer=all".
+
 ;; SP should be adjusted with help of a scratch register.
-define void @test_large_frame_size_1234576() {
+define void @test_large_frame_size_1234576() "frame-pointer"="all" {
 ; CHECK-LABEL: test_large_frame_size_1234576:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    lu12i.w $a0, 301
-; CHECK-NEXT:    ori $a0, $a0, 1680
+; CHECK-NEXT:    addi.d $sp, $sp, -2032
+; CHECK-NEXT:    .cfi_def_cfa_offset 2032
+; CHECK-NEXT:    st.d $ra, $sp, 2024 # 8-byte Folded Spill
+; CHECK-NEXT:    st.d $fp, $sp, 2016 # 8-byte Folded Spill
+; CHECK-NEXT:    .cfi_offset 1, -8
+; CHECK-NEXT:    .cfi_offset 22, -16
+; CHECK-NEXT:    addi.d $fp, $sp, 2032
+; CHECK-NEXT:    .cfi_def_cfa 22, 0
+; CHECK-NEXT:    lu12i.w $a0, 300
+; CHECK-NEXT:    ori $a0, $a0, 3760
 ; CHECK-NEXT:    sub.d $sp, $sp, $a0
-; CHECK-NEXT:    .cfi_def_cfa_offset 1234576
-; CHECK-NEXT:    lu12i.w $a0, 301
-; CHECK-NEXT:    ori $a0, $a0, 1680
+; CHECK-NEXT:    lu12i.w $a0, 300
+; CHECK-NEXT:    ori $a0, $a0, 3760
 ; CHECK-NEXT:    add.d $sp, $sp, $a0
+; CHECK-NEXT:    ld.d $fp, $sp, 2016 # 8-byte Folded Reload
+; CHECK-NEXT:    ld.d $ra, $sp, 2024 # 8-byte Folded Reload
+; CHECK-NEXT:    addi.d $sp, $sp, 2032
 ; CHECK-NEXT:    ret
   %1 = alloca i8, i32 1234567
   ret void

diff --git a/llvm/test/CodeGen/LoongArch/split-sp-adjust.ll b/llvm/test/CodeGen/LoongArch/split-sp-adjust.ll
index 093c92b..8217336 100644 (file)
--- a/llvm/test/CodeGen/LoongArch/split-sp-adjust.ll
+++ b/llvm/test/CodeGen/LoongArch/split-sp-adjust.ll
@@ -24,19 +24,20 @@ entry:
 }
 
 ;; The stack size is 2032 and the SP adjustment will not be split.
+;; 2016 + 8(RA) + 8(emergency spill slot) = 2032
 define i32 @NoSplitSP() nounwind {
 ; CHECK-LABEL: NoSplitSP:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    addi.d $sp, $sp, -2032
 ; CHECK-NEXT:    st.d $ra, $sp, 2024 # 8-byte Folded Spill
-; CHECK-NEXT:    addi.d $a0, $sp, 0
+; CHECK-NEXT:    addi.d $a0, $sp, 8
 ; CHECK-NEXT:    bl %plt(foo)
 ; CHECK-NEXT:    move $a0, $zero
 ; CHECK-NEXT:    ld.d $ra, $sp, 2024 # 8-byte Folded Reload
 ; CHECK-NEXT:    addi.d $sp, $sp, 2032
 ; CHECK-NEXT:    ret
 entry:
-  %xx = alloca [2024 x i8], align 1
+  %xx = alloca [2016 x i8], align 1
   %0 = getelementptr inbounds [2024 x i8], ptr %xx, i32 0, i32 0
   %call = call i32 @foo(ptr nonnull %0)
   ret i32 0

diff --git a/llvm/test/CodeGen/LoongArch/stack-realignment.ll b/llvm/test/CodeGen/LoongArch/stack-realignment.ll
index 89672fb..b49fe35 100644 (file)
--- a/llvm/test/CodeGen/LoongArch/stack-realignment.ll
+++ b/llvm/test/CodeGen/LoongArch/stack-realignment.ll
@@ -305,42 +305,42 @@ define void @caller_no_realign256() "no-realign-stack" {
 define void @caller512() {
 ; LA32-LABEL: caller512:
 ; LA32:       # %bb.0:
-; LA32-NEXT:    addi.w $sp, $sp, -512
-; LA32-NEXT:    .cfi_def_cfa_offset 512
-; LA32-NEXT:    st.w $ra, $sp, 508 # 4-byte Folded Spill
-; LA32-NEXT:    st.w $fp, $sp, 504 # 4-byte Folded Spill
+; LA32-NEXT:    addi.w $sp, $sp, -1024
+; LA32-NEXT:    .cfi_def_cfa_offset 1024
+; LA32-NEXT:    st.w $ra, $sp, 1020 # 4-byte Folded Spill
+; LA32-NEXT:    st.w $fp, $sp, 1016 # 4-byte Folded Spill
 ; LA32-NEXT:    .cfi_offset 1, -4
 ; LA32-NEXT:    .cfi_offset 22, -8
-; LA32-NEXT:    addi.w $fp, $sp, 512
+; LA32-NEXT:    addi.w $fp, $sp, 1024
 ; LA32-NEXT:    .cfi_def_cfa 22, 0
 ; LA32-NEXT:    srli.w $a0, $sp, 9
 ; LA32-NEXT:    slli.w $sp, $a0, 9
-; LA32-NEXT:    addi.w $a0, $sp, 0
+; LA32-NEXT:    addi.w $a0, $sp, 512
 ; LA32-NEXT:    bl %plt(callee)
-; LA32-NEXT:    addi.w $sp, $fp, -512
-; LA32-NEXT:    ld.w $fp, $sp, 504 # 4-byte Folded Reload
-; LA32-NEXT:    ld.w $ra, $sp, 508 # 4-byte Folded Reload
-; LA32-NEXT:    addi.w $sp, $sp, 512
+; LA32-NEXT:    addi.w $sp, $fp, -1024
+; LA32-NEXT:    ld.w $fp, $sp, 1016 # 4-byte Folded Reload
+; LA32-NEXT:    ld.w $ra, $sp, 1020 # 4-byte Folded Reload
+; LA32-NEXT:    addi.w $sp, $sp, 1024
 ; LA32-NEXT:    ret
 ;
 ; LA64-LABEL: caller512:
 ; LA64:       # %bb.0:
-; LA64-NEXT:    addi.d $sp, $sp, -512
-; LA64-NEXT:    .cfi_def_cfa_offset 512
-; LA64-NEXT:    st.d $ra, $sp, 504 # 8-byte Folded Spill
-; LA64-NEXT:    st.d $fp, $sp, 496 # 8-byte Folded Spill
+; LA64-NEXT:    addi.d $sp, $sp, -1024
+; LA64-NEXT:    .cfi_def_cfa_offset 1024
+; LA64-NEXT:    st.d $ra, $sp, 1016 # 8-byte Folded Spill
+; LA64-NEXT:    st.d $fp, $sp, 1008 # 8-byte Folded Spill
 ; LA64-NEXT:    .cfi_offset 1, -8
 ; LA64-NEXT:    .cfi_offset 22, -16
-; LA64-NEXT:    addi.d $fp, $sp, 512
+; LA64-NEXT:    addi.d $fp, $sp, 1024
 ; LA64-NEXT:    .cfi_def_cfa 22, 0
 ; LA64-NEXT:    srli.d $a0, $sp, 9
 ; LA64-NEXT:    slli.d $sp, $a0, 9
-; LA64-NEXT:    addi.d $a0, $sp, 0
+; LA64-NEXT:    addi.d $a0, $sp, 512
 ; LA64-NEXT:    bl %plt(callee)
-; LA64-NEXT:    addi.d $sp, $fp, -512
-; LA64-NEXT:    ld.d $fp, $sp, 496 # 8-byte Folded Reload
-; LA64-NEXT:    ld.d $ra, $sp, 504 # 8-byte Folded Reload
-; LA64-NEXT:    addi.d $sp, $sp, 512
+; LA64-NEXT:    addi.d $sp, $fp, -1024
+; LA64-NEXT:    ld.d $fp, $sp, 1008 # 8-byte Folded Reload
+; LA64-NEXT:    ld.d $ra, $sp, 1016 # 8-byte Folded Reload
+; LA64-NEXT:    addi.d $sp, $sp, 1024
 ; LA64-NEXT:    ret
   %1 = alloca i8, align 512
   call void @callee(i8* %1)
@@ -379,42 +379,46 @@ define void @caller_no_realign512() "no-realign-stack" {
 define void @caller1024() {
 ; LA32-LABEL: caller1024:
 ; LA32:       # %bb.0:
-; LA32-NEXT:    addi.w $sp, $sp, -1024
-; LA32-NEXT:    .cfi_def_cfa_offset 1024
-; LA32-NEXT:    st.w $ra, $sp, 1020 # 4-byte Folded Spill
-; LA32-NEXT:    st.w $fp, $sp, 1016 # 4-byte Folded Spill
+; LA32-NEXT:    addi.w $sp, $sp, -2032
+; LA32-NEXT:    .cfi_def_cfa_offset 2032
+; LA32-NEXT:    st.w $ra, $sp, 2028 # 4-byte Folded Spill
+; LA32-NEXT:    st.w $fp, $sp, 2024 # 4-byte Folded Spill
 ; LA32-NEXT:    .cfi_offset 1, -4
 ; LA32-NEXT:    .cfi_offset 22, -8
-; LA32-NEXT:    addi.w $fp, $sp, 1024
+; LA32-NEXT:    addi.w $fp, $sp, 2032
 ; LA32-NEXT:    .cfi_def_cfa 22, 0
+; LA32-NEXT:    addi.w $sp, $sp, -16
 ; LA32-NEXT:    srli.w $a0, $sp, 10
 ; LA32-NEXT:    slli.w $sp, $a0, 10
-; LA32-NEXT:    addi.w $a0, $sp, 0
+; LA32-NEXT:    addi.w $a0, $sp, 1024
 ; LA32-NEXT:    bl %plt(callee)
-; LA32-NEXT:    addi.w $sp, $fp, -1024
-; LA32-NEXT:    ld.w $fp, $sp, 1016 # 4-byte Folded Reload
-; LA32-NEXT:    ld.w $ra, $sp, 1020 # 4-byte Folded Reload
-; LA32-NEXT:    addi.w $sp, $sp, 1024
+; LA32-NEXT:    addi.w $sp, $fp, -2048
+; LA32-NEXT:    addi.w $sp, $sp, 16
+; LA32-NEXT:    ld.w $fp, $sp, 2024 # 4-byte Folded Reload
+; LA32-NEXT:    ld.w $ra, $sp, 2028 # 4-byte Folded Reload
+; LA32-NEXT:    addi.w $sp, $sp, 2032
 ; LA32-NEXT:    ret
 ;
 ; LA64-LABEL: caller1024:
 ; LA64:       # %bb.0:
-; LA64-NEXT:    addi.d $sp, $sp, -1024
-; LA64-NEXT:    .cfi_def_cfa_offset 1024
-; LA64-NEXT:    st.d $ra, $sp, 1016 # 8-byte Folded Spill
-; LA64-NEXT:    st.d $fp, $sp, 1008 # 8-byte Folded Spill
+; LA64-NEXT:    addi.d $sp, $sp, -2032
+; LA64-NEXT:    .cfi_def_cfa_offset 2032
+; LA64-NEXT:    st.d $ra, $sp, 2024 # 8-byte Folded Spill
+; LA64-NEXT:    st.d $fp, $sp, 2016 # 8-byte Folded Spill
 ; LA64-NEXT:    .cfi_offset 1, -8
 ; LA64-NEXT:    .cfi_offset 22, -16
-; LA64-NEXT:    addi.d $fp, $sp, 1024
+; LA64-NEXT:    addi.d $fp, $sp, 2032
 ; LA64-NEXT:    .cfi_def_cfa 22, 0
+; LA64-NEXT:    addi.d $sp, $sp, -16
 ; LA64-NEXT:    srli.d $a0, $sp, 10
 ; LA64-NEXT:    slli.d $sp, $a0, 10
-; LA64-NEXT:    addi.d $a0, $sp, 0
+; LA64-NEXT:    addi.d $a0, $sp, 1024
 ; LA64-NEXT:    bl %plt(callee)
-; LA64-NEXT:    addi.d $sp, $fp, -1024
-; LA64-NEXT:    ld.d $fp, $sp, 1008 # 8-byte Folded Reload
-; LA64-NEXT:    ld.d $ra, $sp, 1016 # 8-byte Folded Reload
-; LA64-NEXT:    addi.d $sp, $sp, 1024
+; LA64-NEXT:    addi.d $sp, $fp, -2048
+; LA64-NEXT:    addi.d $sp, $sp, 16
+; LA64-NEXT:    ld.d $fp, $sp, 2016 # 8-byte Folded Reload
+; LA64-NEXT:    ld.d $ra, $sp, 2024 # 8-byte Folded Reload
+; LA64-NEXT:    addi.d $sp, $sp, 2032
 ; LA64-NEXT:    ret
   %1 = alloca i8, align 1024
   call void @callee(i8* %1)
@@ -461,13 +465,17 @@ define void @caller2048() {
 ; LA32-NEXT:    .cfi_offset 22, -8
 ; LA32-NEXT:    addi.w $fp, $sp, 2032
 ; LA32-NEXT:    .cfi_def_cfa 22, 0
+; LA32-NEXT:    addi.w $sp, $sp, -2048
 ; LA32-NEXT:    addi.w $sp, $sp, -16
 ; LA32-NEXT:    srli.w $a0, $sp, 11
 ; LA32-NEXT:    slli.w $sp, $a0, 11
-; LA32-NEXT:    addi.w $a0, $sp, 0
+; LA32-NEXT:    ori $a0, $zero, 2048
+; LA32-NEXT:    add.w $a0, $sp, $a0
 ; LA32-NEXT:    bl %plt(callee)
-; LA32-NEXT:    addi.w $sp, $fp, -2048
-; LA32-NEXT:    addi.w $sp, $sp, 16
+; LA32-NEXT:    lu12i.w $a0, 1
+; LA32-NEXT:    sub.w $sp, $fp, $a0
+; LA32-NEXT:    addi.w $sp, $sp, 2032
+; LA32-NEXT:    addi.w $sp, $sp, 32
 ; LA32-NEXT:    ld.w $fp, $sp, 2024 # 4-byte Folded Reload
 ; LA32-NEXT:    ld.w $ra, $sp, 2028 # 4-byte Folded Reload
 ; LA32-NEXT:    addi.w $sp, $sp, 2032
@@ -483,13 +491,17 @@ define void @caller2048() {
 ; LA64-NEXT:    .cfi_offset 22, -16
 ; LA64-NEXT:    addi.d $fp, $sp, 2032
 ; LA64-NEXT:    .cfi_def_cfa 22, 0
+; LA64-NEXT:    addi.d $sp, $sp, -2048
 ; LA64-NEXT:    addi.d $sp, $sp, -16
 ; LA64-NEXT:    srli.d $a0, $sp, 11
 ; LA64-NEXT:    slli.d $sp, $a0, 11
-; LA64-NEXT:    addi.d $a0, $sp, 0
+; LA64-NEXT:    ori $a0, $zero, 2048
+; LA64-NEXT:    add.d $a0, $sp, $a0
 ; LA64-NEXT:    bl %plt(callee)
-; LA64-NEXT:    addi.d $sp, $fp, -2048
-; LA64-NEXT:    addi.d $sp, $sp, 16
+; LA64-NEXT:    lu12i.w $a0, 1
+; LA64-NEXT:    sub.d $sp, $fp, $a0
+; LA64-NEXT:    addi.d $sp, $sp, 2032
+; LA64-NEXT:    addi.d $sp, $sp, 32
 ; LA64-NEXT:    ld.d $fp, $sp, 2016 # 8-byte Folded Reload
 ; LA64-NEXT:    ld.d $ra, $sp, 2024 # 8-byte Folded Reload
 ; LA64-NEXT:    addi.d $sp, $sp, 2032
@@ -539,16 +551,19 @@ define void @caller4096() {
 ; LA32-NEXT:    .cfi_offset 22, -8
 ; LA32-NEXT:    addi.w $fp, $sp, 2032
 ; LA32-NEXT:    .cfi_def_cfa 22, 0
-; LA32-NEXT:    addi.w $sp, $sp, -2048
-; LA32-NEXT:    addi.w $sp, $sp, -16
+; LA32-NEXT:    lu12i.w $a0, 1
+; LA32-NEXT:    ori $a0, $a0, 2064
+; LA32-NEXT:    sub.w $sp, $sp, $a0
 ; LA32-NEXT:    srli.w $a0, $sp, 12
 ; LA32-NEXT:    slli.w $sp, $a0, 12
-; LA32-NEXT:    addi.w $a0, $sp, 0
-; LA32-NEXT:    bl %plt(callee)
 ; LA32-NEXT:    lu12i.w $a0, 1
+; LA32-NEXT:    add.w $a0, $sp, $a0
+; LA32-NEXT:    bl %plt(callee)
+; LA32-NEXT:    lu12i.w $a0, 2
 ; LA32-NEXT:    sub.w $sp, $fp, $a0
-; LA32-NEXT:    addi.w $sp, $sp, 2032
-; LA32-NEXT:    addi.w $sp, $sp, 32
+; LA32-NEXT:    lu12i.w $a0, 1
+; LA32-NEXT:    ori $a0, $a0, 2064
+; LA32-NEXT:    add.w $sp, $sp, $a0
 ; LA32-NEXT:    ld.w $fp, $sp, 2024 # 4-byte Folded Reload
 ; LA32-NEXT:    ld.w $ra, $sp, 2028 # 4-byte Folded Reload
 ; LA32-NEXT:    addi.w $sp, $sp, 2032
@@ -564,16 +579,19 @@ define void @caller4096() {
 ; LA64-NEXT:    .cfi_offset 22, -16
 ; LA64-NEXT:    addi.d $fp, $sp, 2032
 ; LA64-NEXT:    .cfi_def_cfa 22, 0
-; LA64-NEXT:    addi.d $sp, $sp, -2048
-; LA64-NEXT:    addi.d $sp, $sp, -16
+; LA64-NEXT:    lu12i.w $a0, 1
+; LA64-NEXT:    ori $a0, $a0, 2064
+; LA64-NEXT:    sub.d $sp, $sp, $a0
 ; LA64-NEXT:    srli.d $a0, $sp, 12
 ; LA64-NEXT:    slli.d $sp, $a0, 12
-; LA64-NEXT:    addi.d $a0, $sp, 0
-; LA64-NEXT:    bl %plt(callee)
 ; LA64-NEXT:    lu12i.w $a0, 1
+; LA64-NEXT:    add.d $a0, $sp, $a0
+; LA64-NEXT:    bl %plt(callee)
+; LA64-NEXT:    lu12i.w $a0, 2
 ; LA64-NEXT:    sub.d $sp, $fp, $a0
-; LA64-NEXT:    addi.d $sp, $sp, 2032
-; LA64-NEXT:    addi.d $sp, $sp, 32
+; LA64-NEXT:    lu12i.w $a0, 1
+; LA64-NEXT:    ori $a0, $a0, 2064
+; LA64-NEXT:    add.d $sp, $sp, $a0
 ; LA64-NEXT:    ld.d $fp, $sp, 2016 # 8-byte Folded Reload
 ; LA64-NEXT:    ld.d $ra, $sp, 2024 # 8-byte Folded Reload
 ; LA64-NEXT:    addi.d $sp, $sp, 2032