[llvm] [AArch64][SME] Exclude runtime defined liveins when computing liveouts (PR #154325)
via llvm-commits
llvm-commits at lists.llvm.org
Tue Aug 19 05:57:03 PDT 2025
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-llvm-regalloc
Author: Benjamin Maxwell (MacDue)
<details>
<summary>Changes</summary>
These liveins are not defined by predecessors, so attempting to preserve them results in generating invalid MIR (as the registers may not be defined). This takes a similar approach to: https://github.com/llvm/llvm-project/commit/c1dc267258e06bb69e1ca217d1d8ce2d15b8757f
This resolves:
- https://github.com/llvm/llvm-project/pull/149062#discussion_r2285072001
- https://github.com/llvm/llvm-project/pull/153417#issuecomment-3199972351
---
Full diff: https://github.com/llvm/llvm-project/pull/154325.diff
5 Files Affected:
- (modified) llvm/include/llvm/CodeGen/LiveRegUnits.h (+6-2)
- (modified) llvm/lib/CodeGen/LiveRegUnits.cpp (+22-4)
- (modified) llvm/lib/Target/AArch64/MachineSMEABIPass.cpp (+1-1)
- (added) llvm/test/CodeGen/AArch64/sme-abi-eh-liveins.mir (+82)
- (modified) llvm/test/CodeGen/AArch64/sme-za-exceptions.ll (+3-5)
``````````diff
diff --git a/llvm/include/llvm/CodeGen/LiveRegUnits.h b/llvm/include/llvm/CodeGen/LiveRegUnits.h
index 37c31cc6f4ac5..f58188377b520 100644
--- a/llvm/include/llvm/CodeGen/LiveRegUnits.h
+++ b/llvm/include/llvm/CodeGen/LiveRegUnits.h
@@ -26,6 +26,7 @@ namespace llvm {
class MachineInstr;
class MachineBasicBlock;
+class TargetLowering;
/// A set of register units used to track register liveness.
class LiveRegUnits {
@@ -135,8 +136,11 @@ class LiveRegUnits {
/// Adds registers living out of block \p MBB.
/// Live out registers are the union of the live-in registers of the successor
/// blocks and pristine registers. Live out registers of the end block are the
- /// callee saved registers.
- LLVM_ABI void addLiveOuts(const MachineBasicBlock &MBB);
+ /// callee saved registers. If the target lowering information \p TLI is
+ /// provided, runtime-defined live ins of successors will be excluded from the
+ /// live outs.
+ LLVM_ABI void addLiveOuts(const MachineBasicBlock &MBB,
+ const TargetLowering *TLI = nullptr);
/// Adds registers living into block \p MBB.
LLVM_ABI void addLiveIns(const MachineBasicBlock &MBB);
diff --git a/llvm/lib/CodeGen/LiveRegUnits.cpp b/llvm/lib/CodeGen/LiveRegUnits.cpp
index 34de09dd2944b..c034d5f3925fc 100644
--- a/llvm/lib/CodeGen/LiveRegUnits.cpp
+++ b/llvm/lib/CodeGen/LiveRegUnits.cpp
@@ -16,6 +16,7 @@
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/TargetLowering.h"
using namespace llvm;
@@ -86,9 +87,15 @@ void LiveRegUnits::accumulate(const MachineInstr &MI) {
/// Add live-in registers of basic block \p MBB to \p LiveUnits.
static void addBlockLiveIns(LiveRegUnits &LiveUnits,
- const MachineBasicBlock &MBB) {
- for (const auto &LI : MBB.liveins())
+ const MachineBasicBlock &MBB,
+ MCPhysReg ExceptionPointer = {},
+ MCPhysReg ExceptionSelector = {}) {
+ for (const auto &LI : MBB.liveins()) {
+ if (MBB.isEHPad() &&
+ (LI.PhysReg == ExceptionPointer || LI.PhysReg == ExceptionSelector))
+ continue;
LiveUnits.addRegMasked(LI.PhysReg, LI.LaneMask);
+ }
}
/// Adds all callee saved registers to \p LiveUnits.
@@ -135,14 +142,25 @@ void LiveRegUnits::addPristines(const MachineFunction &MF) {
addUnits(Pristine.getBitVector());
}
-void LiveRegUnits::addLiveOuts(const MachineBasicBlock &MBB) {
+void LiveRegUnits::addLiveOuts(const MachineBasicBlock &MBB,
+ const TargetLowering *TLI) {
const MachineFunction &MF = *MBB.getParent();
addPristines(MF);
+ MCPhysReg ExceptionPointer;
+ MCPhysReg ExceptionSelector;
+
+ // Remove live-ins from successors that are defined by the runtime.
+ if (TLI && MF.getFunction().hasPersonalityFn()) {
+ auto PersonalityFn = MF.getFunction().getPersonalityFn();
+ ExceptionPointer = TLI->getExceptionPointerRegister(PersonalityFn);
+ ExceptionSelector = TLI->getExceptionSelectorRegister(PersonalityFn);
+ }
+
// To get the live-outs we simply merge the live-ins of all successors.
for (const MachineBasicBlock *Succ : MBB.successors())
- addBlockLiveIns(*this, *Succ);
+ addBlockLiveIns(*this, *Succ, ExceptionPointer, ExceptionSelector);
// For the return block: Add all callee saved registers.
if (MBB.isReturnBlock()) {
diff --git a/llvm/lib/Target/AArch64/MachineSMEABIPass.cpp b/llvm/lib/Target/AArch64/MachineSMEABIPass.cpp
index b58dfdf32e4ab..2f00977e9a1cc 100644
--- a/llvm/lib/Target/AArch64/MachineSMEABIPass.cpp
+++ b/llvm/lib/Target/AArch64/MachineSMEABIPass.cpp
@@ -277,7 +277,7 @@ void MachineSMEABI::collectNeededZAStates(SMEAttrs SMEFnAttrs) {
}
LiveRegUnits LiveUnits(*TRI);
- LiveUnits.addLiveOuts(MBB);
+ LiveUnits.addLiveOuts(MBB, Subtarget->getTargetLowering());
auto GetPhysLiveRegs = [&] {
LiveRegs PhysLiveRegs = LiveRegs::None;
diff --git a/llvm/test/CodeGen/AArch64/sme-abi-eh-liveins.mir b/llvm/test/CodeGen/AArch64/sme-abi-eh-liveins.mir
new file mode 100644
index 0000000000000..711745e528bbf
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/sme-abi-eh-liveins.mir
@@ -0,0 +1,82 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 5
+# RUN: llc -mtriple=aarch64 -mattr=+sve -mattr=+sme -run-pass=aarch64-machine-sme-abi -verify-machineinstrs %s -o - | FileCheck %s
+
+# This test verifies that runtime defined live-ins are not included in the live
+# outs of predecessors in the MachineSMEABIPass, as including them would result
+# in copies of undefined registers.
+
+--- |
+ define void @sme_abi_eh_liveins() "aarch64_inout_za" personality ptr @__gxx_personality_v0 { entry: unreachable }
+
+ declare i32 @__gxx_personality_v0(...)
+...
+---
+name: sme_abi_eh_liveins
+tracksRegLiveness: true
+isSSA: true
+noVRegs: false
+
+body: |
+ ; CHECK-LABEL: name: sme_abi_eh_liveins
+ ; CHECK: bb.0:
+ ; CHECK-NEXT: successors: %bb.2(0x00000000), %bb.1(0x80000000)
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: [[RDSVLI_XI:%[0-9]+]]:gpr64 = RDSVLI_XI 1, implicit $vg
+ ; CHECK-NEXT: [[COPY:%[0-9]+]]:gpr64 = COPY $sp
+ ; CHECK-NEXT: [[MSUBXrrr:%[0-9]+]]:gpr64 = MSUBXrrr [[RDSVLI_XI]], [[RDSVLI_XI]], [[COPY]]
+ ; CHECK-NEXT: $sp = COPY [[MSUBXrrr]]
+ ; CHECK-NEXT: STPXi [[MSUBXrrr]], [[RDSVLI_XI]], %stack.0, 0
+ ; CHECK-NEXT: ADJCALLSTACKDOWN 0, 0, implicit-def dead $sp, implicit $sp
+ ; CHECK-NEXT: InOutZAUsePseudo
+ ; CHECK-NEXT: ADJCALLSTACKUP 0, 0, implicit-def dead $sp, implicit $sp
+ ; CHECK-NEXT: [[ADDXri:%[0-9]+]]:gpr64sp = ADDXri %stack.0, 0, 0
+ ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gpr64 = COPY [[ADDXri]]
+ ; CHECK-NEXT: MSR 56965, [[COPY1]]
+ ; CHECK-NEXT: ADJCALLSTACKDOWN 0, 0, implicit-def dead $sp, implicit $sp
+ ; CHECK-NEXT: RequiresZASavePseudo
+ ; CHECK-NEXT: ADJCALLSTACKUP 0, 0, implicit-def dead $sp, implicit $sp
+ ; CHECK-NEXT: MSRpstatesvcrImm1 2, 1, implicit-def $nzcv
+ ; CHECK-NEXT: [[MRS:%[0-9]+]]:gpr64 = MRS 56965, implicit-def $nzcv
+ ; CHECK-NEXT: $x0 = ADDXri %stack.0, 0, 0
+ ; CHECK-NEXT: RestoreZAPseudo [[MRS]], $x0, &__arm_tpidr2_restore, csr_aarch64_sme_abi_support_routines_preservemost_from_x0
+ ; CHECK-NEXT: MSR 56965, $xzr
+ ; CHECK-NEXT: B %bb.2
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: bb.1 (landing-pad):
+ ; CHECK-NEXT: successors: %bb.2(0x80000000)
+ ; CHECK-NEXT: liveins: $x0, $x1
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: MSRpstatesvcrImm1 2, 1, implicit-def $nzcv
+ ; CHECK-NEXT: [[MRS1:%[0-9]+]]:gpr64 = MRS 56965, implicit-def $nzcv
+ ; CHECK-NEXT: $x0 = ADDXri %stack.0, 0, 0
+ ; CHECK-NEXT: RestoreZAPseudo [[MRS1]], $x0, &__arm_tpidr2_restore, csr_aarch64_sme_abi_support_routines_preservemost_from_x0
+ ; CHECK-NEXT: MSR 56965, $xzr
+ ; CHECK-NEXT: ADJCALLSTACKDOWN 0, 0, implicit-def dead $sp, implicit $sp
+ ; CHECK-NEXT: InOutZAUsePseudo
+ ; CHECK-NEXT: ADJCALLSTACKUP 0, 0, implicit-def dead $sp, implicit $sp
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: bb.2:
+ bb.0:
+ successors: %bb.2(0x00000000), %bb.1(0x80000000)
+
+ ; Simulate shared ZA call
+ ADJCALLSTACKDOWN 0, 0, implicit-def dead $sp, implicit $sp
+ InOutZAUsePseudo
+ ADJCALLSTACKUP 0, 0, implicit-def dead $sp, implicit $sp
+
+ ; Simulate private ZA call at the end of the block
+ ADJCALLSTACKDOWN 0, 0, implicit-def dead $sp, implicit $sp
+ RequiresZASavePseudo
+ ADJCALLSTACKUP 0, 0, implicit-def dead $sp, implicit $sp
+
+ B %bb.2
+
+ bb.1 (landing-pad):
+ liveins: $x0, $x1
+
+ ; Simulate shared ZA call
+ ADJCALLSTACKDOWN 0, 0, implicit-def dead $sp, implicit $sp
+ InOutZAUsePseudo
+ ADJCALLSTACKUP 0, 0, implicit-def dead $sp, implicit $sp
+
+ bb.2:
diff --git a/llvm/test/CodeGen/AArch64/sme-za-exceptions.ll b/llvm/test/CodeGen/AArch64/sme-za-exceptions.ll
index c497a95a58c8a..bb88142efa592 100644
--- a/llvm/test/CodeGen/AArch64/sme-za-exceptions.ll
+++ b/llvm/test/CodeGen/AArch64/sme-za-exceptions.ll
@@ -1,5 +1,5 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
-; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sme -aarch64-new-sme-abi < %s | FileCheck %s
+; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sme -aarch64-new-sme-abi -verify-machineinstrs < %s | FileCheck %s
; A simple EH test case that corresponds to the following C++ source:
;
@@ -62,16 +62,14 @@ define void @za_with_raii(i1 %fail) "aarch64_inout_za" personality ptr @__gxx_pe
; CHECK-NEXT: ldr x1, [x1, :got_lo12:typeinfo_for_char_const_ptr]
; CHECK-NEXT: bl __cxa_throw
; CHECK-NEXT: .Ltmp1:
-; CHECK-NEXT: mov x8, x0
; CHECK-NEXT: smstart za
-; CHECK-NEXT: mrs x9, TPIDR2_EL0
+; CHECK-NEXT: mrs x8, TPIDR2_EL0
; CHECK-NEXT: sub x0, x29, #16
-; CHECK-NEXT: cbnz x9, .LBB0_4
+; CHECK-NEXT: cbnz x8, .LBB0_4
; CHECK-NEXT: // %bb.3: // %throw_exception
; CHECK-NEXT: bl __arm_tpidr2_restore
; CHECK-NEXT: .LBB0_4: // %throw_exception
; CHECK-NEXT: msr TPIDR2_EL0, xzr
-; CHECK-NEXT: // kill: def $x0 killed $x8
; CHECK-NEXT: // %bb.5: // %throw_fail
; CHECK-NEXT: .LBB0_6: // %unwind_dtors
; CHECK-NEXT: .Ltmp2:
``````````
</details>
https://github.com/llvm/llvm-project/pull/154325
More information about the llvm-commits
mailing list