[AArch64] Ensure the LR is preserved if we must call __arm_get_current_vg

[MacDue]

Fixes #145635

[llvmbot]

@llvm/pr-subscribers-backend-aarch64

Author: Benjamin Maxwell (MacDue)

Changes

Fixes #145635

---

Full diff: https://github.com/llvm/llvm-project/pull/145760.diff

2 Files Affected:

• (modified) llvm/lib/Target/AArch64/AArch64FrameLowering.cpp (+4)
• (added) llvm/test/CodeGen/AArch64/sme-must-save-lr-for-vg.ll (+51)

diff --git a/llvm/lib/Target/AArch64/AArch64FrameLowering.cpp b/llvm/lib/Target/AArch64/AArch64FrameLowering.cpp
index a71668e71c235..791f77956462c 100644
--- a/llvm/lib/Target/AArch64/AArch64FrameLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64FrameLowering.cpp
@@ -3955,6 +3955,10 @@ void AArch64FrameLowering::determineCalleeSaves(MachineFunction &MF,
   // Save number of saved regs, so we can easily update CSStackSize later.
   unsigned NumSavedRegs = SavedRegs.count();
 
+  // If we must call __arm_get_current_vg in the prologue preserve the LR.
+  if (requiresSaveVG(MF) && !Subtarget.hasSVE())
+    SavedRegs.set(AArch64::LR);
+
   // The frame record needs to be created by saving the appropriate registers
   uint64_t EstimatedStackSize = MFI.estimateStackSize(MF);
   if (hasFP(MF) ||
diff --git a/llvm/test/CodeGen/AArch64/sme-must-save-lr-for-vg.ll b/llvm/test/CodeGen/AArch64/sme-must-save-lr-for-vg.ll
new file mode 100644
index 0000000000000..487d87a84549b
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/sme-must-save-lr-for-vg.ll
@@ -0,0 +1,51 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sme -O0 < %s | FileCheck %s
+
+; Example of locally streaming function that (at -O0) must preserve the LR (X30)
+; before calling __arm_get_current_vg.
+define void @foo() "aarch64_pstate_sm_body" {
+; CHECK-LABEL: foo:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    stp d15, d14, [sp, #-96]! // 16-byte Folded Spill
+; CHECK-NEXT:    .cfi_def_cfa_offset 96
+; CHECK-NEXT:    stp d13, d12, [sp, #16] // 16-byte Folded Spill
+; CHECK-NEXT:    stp d11, d10, [sp, #32] // 16-byte Folded Spill
+; CHECK-NEXT:    stp d9, d8, [sp, #48] // 16-byte Folded Spill
+; CHECK-NEXT:    stp x29, x30, [sp, #64] // 16-byte Folded Spill
+; CHECK-NEXT:    rdsvl x9, #1
+; CHECK-NEXT:    lsr x9, x9, #3
+; CHECK-NEXT:    str x9, [sp, #80] // 8-byte Folded Spill
+; CHECK-NEXT:    bl __arm_get_current_vg
+; CHECK-NEXT:    str x0, [sp, #88] // 8-byte Folded Spill
+; CHECK-NEXT:    .cfi_offset vg, -8
+; CHECK-NEXT:    .cfi_offset w30, -24
+; CHECK-NEXT:    .cfi_offset w29, -32
+; CHECK-NEXT:    .cfi_offset b8, -40
+; CHECK-NEXT:    .cfi_offset b9, -48
+; CHECK-NEXT:    .cfi_offset b10, -56
+; CHECK-NEXT:    .cfi_offset b11, -64
+; CHECK-NEXT:    .cfi_offset b12, -72
+; CHECK-NEXT:    .cfi_offset b13, -80
+; CHECK-NEXT:    .cfi_offset b14, -88
+; CHECK-NEXT:    .cfi_offset b15, -96
+; CHECK-NEXT:    smstart sm
+; CHECK-NEXT:    smstop sm
+; CHECK-NEXT:    ldp x29, x30, [sp, #64] // 16-byte Folded Reload
+; CHECK-NEXT:    ldp d9, d8, [sp, #48] // 16-byte Folded Reload
+; CHECK-NEXT:    ldp d11, d10, [sp, #32] // 16-byte Folded Reload
+; CHECK-NEXT:    ldp d13, d12, [sp, #16] // 16-byte Folded Reload
+; CHECK-NEXT:    ldp d15, d14, [sp], #96 // 16-byte Folded Reload
+; CHECK-NEXT:    .cfi_def_cfa_offset 0
+; CHECK-NEXT:    .cfi_restore w30
+; CHECK-NEXT:    .cfi_restore w29
+; CHECK-NEXT:    .cfi_restore b8
+; CHECK-NEXT:    .cfi_restore b9
+; CHECK-NEXT:    .cfi_restore b10
+; CHECK-NEXT:    .cfi_restore b11
+; CHECK-NEXT:    .cfi_restore b12
+; CHECK-NEXT:    .cfi_restore b13
+; CHECK-NEXT:    .cfi_restore b14
+; CHECK-NEXT:    .cfi_restore b15
+; CHECK-NEXT:    ret
+  ret void
+}

[MacDue]

Note: I also explored the idea of moving the VG saves from frame lowering to a pre-RA pass in #145748 (which I think generally simplifies the implementation), but it's probably too much churn to be worth considering.

[efriedma-quic]

Are you intentionally doing this after we compute NumSavedRegs?

[MacDue]

Yes, as later it's used to update the CSStackSize to include the additional saves:

  // Adding the size of additional 64bit GPR saves.
  CSStackSize += 8 * (SavedRegs.count() - NumSavedRegs);

[efriedma-quic]

Oh, hmm, I guess that works.

Maybe we should move the declaration of NumSavedRegs immediately after the for (unsigned Reg : SavedRegs.set_bits()) { loop, to make the connection more clear?

[MacDue]

Sure, I've also updated the comment to make this a little more obvious 👍

[efriedma-quic]

LGTM