[AArch64] Skip storing of stack arguments when lowering tail calls

When possible, do not emit trivial load and stores to the same offset on
the stack.

Author: guy-david

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

@@ -8968,6 +8968,46 @@ getSMToggleCondition(const SMECallAttrs &CallAttrs) {
   llvm_unreachable("Unsupported attributes");
 }
 
+/// Check whether a stack argument requires lowering in a tail call.
+static bool shouldLowerTailCallStackArg(const MachineFunction &MF,
+                                        const CCValAssign &VA, SDValue Arg,
+                                        ISD::ArgFlagsTy Flags, int CallOffset) {
+  // FIXME: We should be able to handle this case, but it's not clear how to.
+  if (Flags.isZExt() || Flags.isSExt())
+    return true;
+
+  for (;;) {
+    // Look through nodes that don't alter the bits of the incoming value.
+    unsigned Op = Arg.getOpcode();
+    if (Op == ISD::ZERO_EXTEND || Op == ISD::ANY_EXTEND || Op == ISD::BITCAST ||
+        Op == ISD::AssertZext || Op == ISD::AssertSext ||
+        Op == AArch64ISD::ASSERT_ZEXT_BOOL) {
+      Arg = Arg.getOperand(0);
+      continue;
+    }
+    break;
+  }
+
+  // If the argument is a load from the same immutable stack slot, we can reuse
+  // it.
+  if (auto *LoadNode = dyn_cast<LoadSDNode>(Arg)) {
+    if (auto *FINode = dyn_cast<FrameIndexSDNode>(LoadNode->getBasePtr())) {
+      const MachineFrameInfo &MFI = MF.getFrameInfo();
+      int FI = FINode->getIndex();
+      if (!MFI.isImmutableObjectIndex(FI))
+        return true;
+      if (CallOffset != MFI.getObjectOffset(FI))
+        return true;
+      uint64_t SizeInBits = LoadNode->getMemoryVT().getFixedSizeInBits();
+      if (SizeInBits / 8 != MFI.getObjectSize(FI))
+        return true;
+      return false;
+    }
+  }
+
+  return true;
+}
+
 /// LowerCall - Lower a call to a callseq_start + CALL + callseq_end chain,
 /// and add input and output parameter nodes.
 SDValue
@@ -9391,10 +9431,13 @@ AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI,
       }
       unsigned LocMemOffset = VA.getLocMemOffset();
       int32_t Offset = LocMemOffset + BEAlign;
-      SDValue PtrOff = DAG.getIntPtrConstant(Offset, DL);
-      PtrOff = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr, PtrOff);
 
       if (IsTailCall) {
+        // When the frame pointer is perfectly aligned for the tail call and the
+        // same stack argument is passed down intact, we can reuse it.
+        if (!FPDiff && !shouldLowerTailCallStackArg(MF, VA, Arg, Flags, Offset))
+          continue;
+
         Offset = Offset + FPDiff;
         int FI = MF.getFrameInfo().CreateFixedObject(OpSize, Offset, true);
 

llvm/lib/Target/AArch64/GISel/AArch64CallLowering.cpp

@@ -25,6 +25,7 @@
 #include "llvm/CodeGen/Analysis.h"
 #include "llvm/CodeGen/CallingConvLower.h"
 #include "llvm/CodeGen/FunctionLoweringInfo.h"
+#include "llvm/CodeGen/GlobalISel/GenericMachineInstrs.h"
 #include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h"
 #include "llvm/CodeGen/GlobalISel/Utils.h"
 #include "llvm/CodeGen/LowLevelTypeUtils.h"
@@ -35,6 +36,7 @@
 #include "llvm/CodeGen/MachineMemOperand.h"
 #include "llvm/CodeGen/MachineOperand.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/TargetOpcodes.h"
 #include "llvm/CodeGen/TargetRegisterInfo.h"
 #include "llvm/CodeGen/TargetSubtargetInfo.h"
 #include "llvm/CodeGen/ValueTypes.h"
@@ -296,10 +298,61 @@ struct OutgoingArgHandler : public CallLowering::OutgoingValueHandler {
     MIRBuilder.buildCopy(PhysReg, ExtReg);
   }
 
+  /// Check whether a stack argument requires lowering in a tail call.
+  static bool shouldLowerTailCallStackArg(const MachineFunction &MF,
+                                          const CCValAssign &VA,
+                                          Register ValVReg,
+                                          Register StoreAddr) {
+    // Print the defining instruction for the value.
+    auto *DefMI = MF.getRegInfo().getVRegDef(ValVReg);
+    assert(DefMI && "No defining instruction");
+    for (;;) {
+      // Look through nodes that don't alter the bits of the incoming value.
+      unsigned Op = DefMI->getOpcode();
+      if (Op == TargetOpcode::G_ZEXT || Op == TargetOpcode::G_ANYEXT ||
+          Op == TargetOpcode::G_TRUNC || Op == TargetOpcode::G_BITCAST ||
+          Op == TargetOpcode::G_ASSERT_ZEXT ||
+          Op == TargetOpcode::G_ASSERT_SEXT) {
+        DefMI = MF.getRegInfo().getVRegDef(DefMI->getOperand(1).getReg());
+        continue;
+      }
+      break;
+    }
+
+    auto *Load = dyn_cast<GLoad>(DefMI);
+    if (!Load)
+      return true;
+    Register LoadReg = Load->getPointerReg();
+    auto *LoadAddrDef = MF.getRegInfo().getVRegDef(LoadReg);
+    assert(LoadAddrDef && "No defining instruction");
+    if (LoadAddrDef->getOpcode() != TargetOpcode::G_FRAME_INDEX)
+      return true;
+    assert(LoadAddrDef && "No defining instruction");
+    const MachineFrameInfo &MFI = MF.getFrameInfo();
+    int LoadFI = LoadAddrDef->getOperand(1).getIndex();
+
+    auto *StoreAddrDef = MF.getRegInfo().getVRegDef(StoreAddr);
+    assert(StoreAddrDef && "No defining instruction");
+    if (StoreAddrDef->getOpcode() != TargetOpcode::G_FRAME_INDEX)
+      return true;
+    int StoreFI = StoreAddrDef->getOperand(1).getIndex();
+
+    if (!MFI.isImmutableObjectIndex(LoadFI))
+      return true;
+    if (MFI.getObjectOffset(LoadFI) != MFI.getObjectOffset(StoreFI))
+      return true;
+    if (Load->getMemSize() != MFI.getObjectSize(StoreFI))
+      return true;
+
+    return false;
+  }
+
   void assignValueToAddress(Register ValVReg, Register Addr, LLT MemTy,
                             const MachinePointerInfo &MPO,
                             const CCValAssign &VA) override {
     MachineFunction &MF = MIRBuilder.getMF();
+    if (!shouldLowerTailCallStackArg(MF, VA, ValVReg, Addr))
+      return;
     auto MMO = MF.getMachineMemOperand(MPO, MachineMemOperand::MOStore, MemTy,
                                        inferAlignFromPtrInfo(MF, MPO));
     MIRBuilder.buildStore(ValVReg, Addr, *MMO);

llvm/test/CodeGen/AArch64/darwinpcs-tail.ll

@@ -5,11 +5,11 @@
 
 ; CHECK-LABEL: __ZThn16_N1C3addEPKcz:
 ; CHECK:       b __ZN1C3addEPKcz
+
 ; CHECK-LABEL: _tailTest:
 ; CHECK:       b __ZN1C3addEPKcz
+
 ; CHECK-LABEL: __ZThn8_N1C1fEiiiiiiiiiz:
-; CHECK:       ldr     w9, [sp, #4]
-; CHECK:       str     w9, [sp, #4]
 ; CHECK:       b __ZN1C1fEiiiiiiiiiz
 
 %class.C = type { %class.A.base, [4 x i8], %class.B.base, [4 x i8] }

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

@@ -4,7 +4,7 @@
 ; 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)
+  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
 }
 

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

@@ -11,66 +11,64 @@ target triple = "aarch64-unknown-linux-gnu"
 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 x29, [sp, #-48]! // 8-byte Folded Spill
-; CHECK-NEXT:    stp x22, x21, [sp, #16] // 16-byte Folded Spill
-; CHECK-NEXT:    stp x20, x19, [sp, #32] // 16-byte Folded Spill
-; CHECK-NEXT:    .cfi_def_cfa_offset 48
-; 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 w29, -48
-; CHECK-NEXT:    add x10, sp, #176
-; CHECK-NEXT:    add x8, sp, #48
-; CHECK-NEXT:    add x9, sp, #144
-; CHECK-NEXT:    ldr z3, [x10]
+; CHECK-NEXT:    sub sp, sp, #368
+; CHECK-NEXT:    stp x29, x30, [sp, #336] // 16-byte Folded Spill
+; CHECK-NEXT:    str x28, [sp, #352] // 8-byte Folded Spill
+; CHECK-NEXT:    add x29, sp, #336
+; CHECK-NEXT:    .cfi_def_cfa w29, 32
+; CHECK-NEXT:    .cfi_offset w28, -16
+; CHECK-NEXT:    .cfi_offset w30, -24
+; CHECK-NEXT:    .cfi_offset w29, -32
+; CHECK-NEXT:    add x8, x29, #32
+; CHECK-NEXT:    add x9, x29, #72
+; CHECK-NEXT:    ptrue p0.d
 ; CHECK-NEXT:    ldr z0, [x8]
-; CHECK-NEXT:    add x8, sp, #112
-; CHECK-NEXT:    ldr z2, [x9]
+; CHECK-NEXT:    add x8, x29, #256
+; CHECK-NEXT:    ldr z3, [x9]
 ; CHECK-NEXT:    ldr z1, [x8]
-; CHECK-NEXT:    add x20, sp, #176
-; CHECK-NEXT:    ldp x9, x8, [sp, #328]
-; CHECK-NEXT:    ldr x15, [sp, #104]
-; CHECK-NEXT:    ldp x11, x10, [sp, #312]
-; CHECK-NEXT:    ldur q4, [sp, #88]
-; CHECK-NEXT:    ldp x13, x12, [sp, #296]
-; CHECK-NEXT:    ldr x19, [sp, #272]
-; CHECK-NEXT:    ldp x18, x14, [sp, #280]
-; CHECK-NEXT:    ldp x16, x17, [sp, #208]
-; CHECK-NEXT:    ldp x21, x22, [sp, #352]
-; CHECK-NEXT:    str z3, [x20]
-; CHECK-NEXT:    add x20, sp, #144
-; CHECK-NEXT:    str z2, [x20]
-; CHECK-NEXT:    add x20, sp, #112
-; CHECK-NEXT:    str z1, [x20]
-; CHECK-NEXT:    add x20, sp, #48
-; CHECK-NEXT:    str z0, [x20]
-; CHECK-NEXT:    stp x21, x22, [sp, #352]
-; CHECK-NEXT:    ldp x22, x21, [sp, #16] // 16-byte Folded Reload
-; CHECK-NEXT:    stp x19, x18, [sp, #272]
-; CHECK-NEXT:    ldp x20, x19, [sp, #32] // 16-byte Folded Reload
-; CHECK-NEXT:    stp x16, x17, [sp, #208]
-; CHECK-NEXT:    stur q4, [sp, #88]
-; CHECK-NEXT:    str x15, [sp, #104]
-; CHECK-NEXT:    stp x14, x13, [sp, #288]
-; CHECK-NEXT:    stp x12, x11, [sp, #304]
-; CHECK-NEXT:    stp x10, x9, [sp, #320]
-; CHECK-NEXT:    str x8, [sp, #336]
-; CHECK-NEXT:    ldr x29, [sp], #48 // 8-byte Folded Reload
-; CHECK-NEXT:    b func2
+; CHECK-NEXT:    add x8, x29, #288
+; CHECK-NEXT:    add x9, x29, #168
+; CHECK-NEXT:    ldr z2, [x8]
+; CHECK-NEXT:    add x8, x29, #104
+; CHECK-NEXT:    ldr z6, [x9]
+; CHECK-NEXT:    ldr z4, [x8]
+; CHECK-NEXT:    add x8, x29, #136
+; CHECK-NEXT:    mov x12, #17 // =0x11
+; CHECK-NEXT:    ldr z5, [x8]
+; CHECK-NEXT:    ldp x10, x11, [x29, #336]
+; CHECK-NEXT:    st1d { z6.d }, p0, [sp, x12, lsl #3]
+; CHECK-NEXT:    mov x12, #13 // =0xd
+; CHECK-NEXT:    ldr x8, [x29, #200]
+; CHECK-NEXT:    ldr x9, [x29, #320]
+; CHECK-NEXT:    st1d { z5.d }, p0, [sp, x12, lsl #3]
+; CHECK-NEXT:    mov x12, #9 // =0x9
+; CHECK-NEXT:    st1d { z4.d }, p0, [sp, x12, lsl #3]
+; CHECK-NEXT:    mov x12, #5 // =0x5
+; CHECK-NEXT:    st1d { z3.d }, p0, [sp, x12, lsl #3]
+; CHECK-NEXT:    stp x10, x11, [sp, #304]
+; CHECK-NEXT:    str x9, [sp, #288]
+; CHECK-NEXT:    str z2, [sp, #8, mul vl]
+; CHECK-NEXT:    str z1, [sp, #7, mul vl]
+; CHECK-NEXT:    str x8, [sp, #168]
+; CHECK-NEXT:    str z0, [sp]
+; CHECK-NEXT:    bl func2
+; CHECK-NEXT:    ldp x29, x30, [sp, #336] // 16-byte Folded Reload
+; CHECK-NEXT:    ldr x28, [sp, #352] // 8-byte Folded Reload
+; CHECK-NEXT:    add sp, sp, #368
+; CHECK-NEXT:    ret
                              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,
-                        ptr %v9, ptr %v10, ptr %v11, ptr %v12, ptr undef, 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 undef, ptr undef,
-                        ptr undef, ptr undef, ptr undef, ptr undef, ptr %v37, ptr %v38, ptr %v39, ptr %v40,
-                        ptr %v41, ptr %v42, ptr %v43, ptr %v44, ptr %v45, ptr undef, ptr %v47, ptr %v48,
-                        i64 undef)
+  call void @func2(ptr %v1, ptr %v2, ptr %v3, ptr %v4, ptr %v5, ptr %v6, ptr %v7, ptr %v8,
+                   ptr %v9, ptr %v10, ptr %v11, ptr %v12, ptr undef, 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 undef, ptr undef,
+                   ptr undef, ptr undef, ptr undef, ptr undef, ptr %v37, ptr %v38, ptr %v39, ptr %v40,
+                   ptr %v41, ptr %v42, ptr %v43, ptr %v44, ptr %v45, ptr undef, ptr %v47, ptr %v48,
+                   i64 undef)
   ret void
 }
 

llvm/test/CodeGen/AArch64/tail-call-stack-args.ll

@@ -0,0 +1,15 @@
+; RUN: llc %s -mtriple=aarch64 -o - | FileCheck %s
+; RUN: llc %s -mtriple=aarch64 -global-isel -o - | FileCheck %s
+
+; Tail calls which have stack arguments in the same offsets as the caller do not
+; need to load and store the arguments from the stack.
+
+declare void @func_i1(i32 %a, i32 %b, i32 %c, i32 %d, i32 %e, i32 %f, i32 %g, i32 %h, i32 %i, i1 %j)
+
+; CHECK-LABEL: wrapper_func_i1:
+define void @wrapper_func_i1(i32 %a, i32 %b, i32 %c, i32 %d, i32 %e, i32 %f, i32 %g, i32 %h, i32 %i, i1 %j) {
+  ; CHECK: // %bb.
+  ; CHECK-NEXT: b func_i1
+  tail call void @func_i1(i32 %a, i32 %b, i32 %c, i32 %d, i32 %e, i32 %f, i32 %g, i32 %h, i32 %i, i1 %j)
+  ret void
+}