[llvm] [SDAG] Avoid creating redundant stack slots when lowering FSINCOS (PR #108401)

Benjamin Maxwell via llvm-commits llvm-commits at lists.llvm.org
Thu Sep 12 11:49:25 PDT 2024


https://github.com/MacDue updated https://github.com/llvm/llvm-project/pull/108401

>From ce4e988ab90a064a6d6c5202a75f9226491e75e0 Mon Sep 17 00:00:00 2001
From: Benjamin Maxwell <benjamin.maxwell at arm.com>
Date: Thu, 12 Sep 2024 13:52:24 +0000
Subject: [PATCH 1/3] Precommit test for sincos stack slots

---
 .../CodeGen/AArch64/sincos-stack-slots.ll     | 152 ++++++++++++++++++
 1 file changed, 152 insertions(+)
 create mode 100644 llvm/test/CodeGen/AArch64/sincos-stack-slots.ll

diff --git a/llvm/test/CodeGen/AArch64/sincos-stack-slots.ll b/llvm/test/CodeGen/AArch64/sincos-stack-slots.ll
new file mode 100644
index 00000000000000..9c362ba117fefa
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/sincos-stack-slots.ll
@@ -0,0 +1,152 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=aarch64-linux-gnu -o - %s | FileCheck %s
+
+define { float, float } @sincos_f32_value_return(float %x) {
+; CHECK-LABEL: sincos_f32_value_return:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    str x30, [sp, #-16]! // 8-byte Folded Spill
+; CHECK-NEXT:    .cfi_def_cfa_offset 16
+; CHECK-NEXT:    .cfi_offset w30, -16
+; CHECK-NEXT:    add x0, sp, #12
+; CHECK-NEXT:    add x1, sp, #8
+; CHECK-NEXT:    bl sincosf
+; CHECK-NEXT:    ldp s1, s0, [sp, #8]
+; CHECK-NEXT:    ldr x30, [sp], #16 // 8-byte Folded Reload
+; CHECK-NEXT:    ret
+entry:
+  %sin = tail call float @llvm.sin.f32(float %x)
+  %cos = tail call float @llvm.cos.f32(float %x)
+  %ret_0 = insertvalue { float, float } poison, float %sin, 0
+  %ret_1 = insertvalue { float, float } %ret_0, float %cos, 1
+  ret { float, float } %ret_1
+}
+
+define void @sincos_f32_ptr_return(float %x, ptr %out_sin, ptr %out_cos) {
+; CHECK-LABEL: sincos_f32_ptr_return:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    str x30, [sp, #-32]! // 8-byte Folded Spill
+; CHECK-NEXT:    stp x20, x19, [sp, #16] // 16-byte Folded Spill
+; CHECK-NEXT:    .cfi_def_cfa_offset 32
+; CHECK-NEXT:    .cfi_offset w19, -8
+; CHECK-NEXT:    .cfi_offset w20, -16
+; CHECK-NEXT:    .cfi_offset w30, -32
+; CHECK-NEXT:    mov x19, x1
+; CHECK-NEXT:    mov x20, x0
+; CHECK-NEXT:    add x0, sp, #12
+; CHECK-NEXT:    add x1, sp, #8
+; CHECK-NEXT:    bl sincosf
+; CHECK-NEXT:    ldp s1, s0, [sp, #8]
+; CHECK-NEXT:    str s0, [x20]
+; CHECK-NEXT:    str s1, [x19]
+; CHECK-NEXT:    ldp x20, x19, [sp, #16] // 16-byte Folded Reload
+; CHECK-NEXT:    ldr x30, [sp], #32 // 8-byte Folded Reload
+; CHECK-NEXT:    ret
+entry:
+  %sin = tail call float @llvm.sin.f32(float %x)
+  %cos = tail call float @llvm.cos.f32(float %x)
+  store float %sin, ptr %out_sin, align 4
+  store float %cos, ptr %out_cos, align 4
+  ret void
+}
+
+define float @sincos_f32_mixed_return(float %x, ptr %out_sin) {
+; CHECK-LABEL: sincos_f32_mixed_return:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    sub sp, sp, #32
+; CHECK-NEXT:    stp x30, x19, [sp, #16] // 16-byte Folded Spill
+; CHECK-NEXT:    .cfi_def_cfa_offset 32
+; CHECK-NEXT:    .cfi_offset w19, -8
+; CHECK-NEXT:    .cfi_offset w30, -16
+; CHECK-NEXT:    mov x19, x0
+; CHECK-NEXT:    add x0, sp, #12
+; CHECK-NEXT:    add x1, sp, #8
+; CHECK-NEXT:    bl sincosf
+; CHECK-NEXT:    ldp s0, s1, [sp, #8]
+; CHECK-NEXT:    str s1, [x19]
+; CHECK-NEXT:    ldp x30, x19, [sp, #16] // 16-byte Folded Reload
+; CHECK-NEXT:    add sp, sp, #32
+; CHECK-NEXT:    ret
+entry:
+  %sin = tail call float @llvm.sin.f32(float %x)
+  %cos = tail call float @llvm.cos.f32(float %x)
+  store float %sin, ptr %out_sin, align 4
+  ret float %cos
+}
+
+define { double, double } @sincos_f64_value_return(double %x) {
+; CHECK-LABEL: sincos_f64_value_return:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    sub sp, sp, #32
+; CHECK-NEXT:    str x30, [sp, #16] // 8-byte Folded Spill
+; CHECK-NEXT:    .cfi_def_cfa_offset 32
+; CHECK-NEXT:    .cfi_offset w30, -16
+; CHECK-NEXT:    add x0, sp, #24
+; CHECK-NEXT:    add x1, sp, #8
+; CHECK-NEXT:    bl sincos
+; CHECK-NEXT:    ldr d0, [sp, #24]
+; CHECK-NEXT:    ldr d1, [sp, #8]
+; CHECK-NEXT:    ldr x30, [sp, #16] // 8-byte Folded Reload
+; CHECK-NEXT:    add sp, sp, #32
+; CHECK-NEXT:    ret
+entry:
+  %sin = tail call double @llvm.sin.f64(double %x)
+  %cos = tail call double @llvm.cos.f64(double %x)
+  %ret_0 = insertvalue { double, double } poison, double %sin, 0
+  %ret_1 = insertvalue { double, double } %ret_0, double %cos, 1
+  ret { double, double } %ret_1
+}
+
+define void @sincos_f64_ptr_return(double %x, ptr %out_sin, ptr %out_cos) {
+; CHECK-LABEL: sincos_f64_ptr_return:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    sub sp, sp, #48
+; CHECK-NEXT:    str x30, [sp, #16] // 8-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 w30, -32
+; CHECK-NEXT:    mov x19, x1
+; CHECK-NEXT:    mov x20, x0
+; CHECK-NEXT:    add x0, sp, #24
+; CHECK-NEXT:    add x1, sp, #8
+; CHECK-NEXT:    bl sincos
+; CHECK-NEXT:    ldr d0, [sp, #24]
+; CHECK-NEXT:    ldr d1, [sp, #8]
+; CHECK-NEXT:    ldr x30, [sp, #16] // 8-byte Folded Reload
+; CHECK-NEXT:    str d0, [x20]
+; CHECK-NEXT:    str d1, [x19]
+; CHECK-NEXT:    ldp x20, x19, [sp, #32] // 16-byte Folded Reload
+; CHECK-NEXT:    add sp, sp, #48
+; CHECK-NEXT:    ret
+entry:
+  %sin = tail call double @llvm.sin.f64(double %x)
+  %cos = tail call double @llvm.cos.f64(double %x)
+  store double %sin, ptr %out_sin, align 4
+  store double %cos, ptr %out_cos, align 4
+  ret void
+}
+
+define double @sincos_f64_mixed_return(double %x, ptr %out_sin) {
+; CHECK-LABEL: sincos_f64_mixed_return:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    sub sp, sp, #32
+; CHECK-NEXT:    stp x30, x19, [sp, #16] // 16-byte Folded Spill
+; CHECK-NEXT:    .cfi_def_cfa_offset 32
+; CHECK-NEXT:    .cfi_offset w19, -8
+; CHECK-NEXT:    .cfi_offset w30, -16
+; CHECK-NEXT:    mov x19, x0
+; CHECK-NEXT:    add x0, sp, #8
+; CHECK-NEXT:    mov x1, sp
+; CHECK-NEXT:    bl sincos
+; CHECK-NEXT:    ldp d0, d1, [sp]
+; CHECK-NEXT:    str d1, [x19]
+; CHECK-NEXT:    ldp x30, x19, [sp, #16] // 16-byte Folded Reload
+; CHECK-NEXT:    add sp, sp, #32
+; CHECK-NEXT:    ret
+entry:
+  %sin = tail call double @llvm.sin.f64(double %x)
+  %cos = tail call double @llvm.cos.f64(double %x)
+  store double %sin, ptr %out_sin, align 4
+  ret double %cos
+}

>From 8033c5caf3a98504ced4f312d843ca5a5baa3ddd Mon Sep 17 00:00:00 2001
From: Benjamin Maxwell <benjamin.maxwell at arm.com>
Date: Thu, 12 Sep 2024 13:55:25 +0000
Subject: [PATCH 2/3] [SDAG] Avoid creating redundant stack slots when lowering
 FSINCOS

When lowering `FSINCOS` to a library call (that takes output pointers)
we can avoid creating new stack allocations if the results of the
`FSINCOS` are being stored. Instead, we can take the destination
pointers from the stores and pass those to the library call.
---
 .../include/llvm/CodeGen/RuntimeLibcallUtil.h |  4 +
 llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp | 99 +++++++++----------
 llvm/lib/CodeGen/TargetLoweringBase.cpp       |  5 +
 .../CodeGen/AArch64/sincos-stack-slots.ll     | 73 ++++----------
 4 files changed, 75 insertions(+), 106 deletions(-)

diff --git a/llvm/include/llvm/CodeGen/RuntimeLibcallUtil.h b/llvm/include/llvm/CodeGen/RuntimeLibcallUtil.h
index 7a131645893921..045ec7d3653119 100644
--- a/llvm/include/llvm/CodeGen/RuntimeLibcallUtil.h
+++ b/llvm/include/llvm/CodeGen/RuntimeLibcallUtil.h
@@ -62,6 +62,10 @@ Libcall getLDEXP(EVT RetVT);
 /// UNKNOWN_LIBCALL if there is none.
 Libcall getFREXP(EVT RetVT);
 
+/// getFSINCOS - Return the FSINCOS_* value for the given types, or
+/// UNKNOWN_LIBCALL if there is none.
+Libcall getFSINCOS(EVT RetVT);
+
 /// Return the SYNC_FETCH_AND_* value for the given opcode and type, or
 /// UNKNOWN_LIBCALL if there is none.
 Libcall getSYNC(unsigned Opc, MVT VT);
diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
index f5fbc01cd95e96..7b0dc63c473d25 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
@@ -2326,15 +2326,7 @@ SelectionDAGLegalize::ExpandDivRemLibCall(SDNode *Node,
 
 /// Return true if sincos libcall is available.
 static bool isSinCosLibcallAvailable(SDNode *Node, const TargetLowering &TLI) {
-  RTLIB::Libcall LC;
-  switch (Node->getSimpleValueType(0).SimpleTy) {
-  default: llvm_unreachable("Unexpected request for libcall!");
-  case MVT::f32:     LC = RTLIB::SINCOS_F32; break;
-  case MVT::f64:     LC = RTLIB::SINCOS_F64; break;
-  case MVT::f80:     LC = RTLIB::SINCOS_F80; break;
-  case MVT::f128:    LC = RTLIB::SINCOS_F128; break;
-  case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break;
-  }
+  RTLIB::Libcall LC = RTLIB::getFSINCOS(Node->getSimpleValueType(0).SimpleTy);
   return TLI.getLibcallName(LC) != nullptr;
 }
 
@@ -2355,68 +2347,73 @@ static bool useSinCos(SDNode *Node) {
 }
 
 /// Issue libcalls to sincos to compute sin / cos pairs.
-void
-SelectionDAGLegalize::ExpandSinCosLibCall(SDNode *Node,
-                                          SmallVectorImpl<SDValue> &Results) {
-  RTLIB::Libcall LC;
-  switch (Node->getSimpleValueType(0).SimpleTy) {
-  default: llvm_unreachable("Unexpected request for libcall!");
-  case MVT::f32:     LC = RTLIB::SINCOS_F32; break;
-  case MVT::f64:     LC = RTLIB::SINCOS_F64; break;
-  case MVT::f80:     LC = RTLIB::SINCOS_F80; break;
-  case MVT::f128:    LC = RTLIB::SINCOS_F128; break;
-  case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break;
-  }
-
-  // The input chain to this libcall is the entry node of the function.
-  // Legalizing the call will automatically add the previous call to the
-  // dependence.
-  SDValue InChain = DAG.getEntryNode();
-
+void SelectionDAGLegalize::ExpandSinCosLibCall(
+    SDNode *Node, SmallVectorImpl<SDValue> &Results) {
   EVT RetVT = Node->getValueType(0);
   Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
 
   TargetLowering::ArgListTy Args;
-  TargetLowering::ArgListEntry Entry;
+  TargetLowering::ArgListEntry Entry{};
+
+  // Find users of the node that store the results. The destination pointers
+  // can be used instead of creating stack allocations.
+  StoreSDNode *SinST = nullptr;
+  StoreSDNode *CosST = nullptr;
+  for (SDNode::use_iterator UI = Node->use_begin(), UE = Node->use_end();
+       UI != UE; ++UI) {
+    SDUse &Use = UI.getUse();
+    SDNode *User = Use.getUser();
+    if (!ISD::isNormalStore(User))
+      continue;
+    auto *ST = cast<StoreSDNode>(User);
+    if (Use.getResNo() == 0)
+      SinST = ST;
+    if (Use.getResNo() == 1)
+      CosST = ST;
+  }
 
   // Pass the argument.
   Entry.Node = Node->getOperand(0);
   Entry.Ty = RetTy;
-  Entry.IsSExt = false;
-  Entry.IsZExt = false;
   Args.push_back(Entry);
 
+  auto GetOrCreateOutPointer = [&](StoreSDNode *MaybeStore) {
+    if (MaybeStore)
+      return std::make_pair(MaybeStore->getBasePtr(),
+                            MaybeStore->getPointerInfo());
+    SDValue StackSlot = DAG.CreateStackTemporary(RetVT);
+    auto PtrInfo = MachinePointerInfo::getFixedStack(
+        DAG.getMachineFunction(),
+        cast<FrameIndexSDNode>(StackSlot)->getIndex());
+    return std::make_pair(StackSlot, PtrInfo);
+  };
+
   // Pass the return address of sin.
-  SDValue SinPtr = DAG.CreateStackTemporary(RetVT);
-  Entry.Node = SinPtr;
+  auto SinPtr = GetOrCreateOutPointer(SinST);
+  Entry.Node = SinPtr.first;
   Entry.Ty = PointerType::getUnqual(RetTy->getContext());
-  Entry.IsSExt = false;
-  Entry.IsZExt = false;
   Args.push_back(Entry);
 
   // Also pass the return address of the cos.
-  SDValue CosPtr = DAG.CreateStackTemporary(RetVT);
-  Entry.Node = CosPtr;
+  auto CosPtr = GetOrCreateOutPointer(CosST);
+  Entry.Node = CosPtr.first;
   Entry.Ty = PointerType::getUnqual(RetTy->getContext());
-  Entry.IsSExt = false;
-  Entry.IsZExt = false;
   Args.push_back(Entry);
 
-  SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
-                                         TLI.getPointerTy(DAG.getDataLayout()));
-
-  SDLoc dl(Node);
-  TargetLowering::CallLoweringInfo CLI(DAG);
-  CLI.setDebugLoc(dl).setChain(InChain).setLibCallee(
-      TLI.getLibcallCallingConv(LC), Type::getVoidTy(*DAG.getContext()), Callee,
-      std::move(Args));
+  RTLIB::Libcall LC = RTLIB::getFSINCOS(RetVT);
+  auto [Call, Chain] = ExpandLibCall(LC, Node, std::move(Args), false);
 
-  std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
+  // Replace explict stores with the library call.
+  for (StoreSDNode *ST : {SinST, CosST}) {
+    if (ST)
+      DAG.ReplaceAllUsesOfValueWith(SDValue(ST, 0), Chain);
+  }
 
-  Results.push_back(
-      DAG.getLoad(RetVT, dl, CallInfo.second, SinPtr, MachinePointerInfo()));
-  Results.push_back(
-      DAG.getLoad(RetVT, dl, CallInfo.second, CosPtr, MachinePointerInfo()));
+  SDLoc DL(Node);
+  for (auto [Ptr, PtrInfo] : {SinPtr, CosPtr}) {
+    SDValue LoadExp = DAG.getLoad(RetVT, DL, Chain, Ptr, PtrInfo);
+    Results.push_back(LoadExp);
+  }
 }
 
 SDValue SelectionDAGLegalize::expandLdexp(SDNode *Node) const {
diff --git a/llvm/lib/CodeGen/TargetLoweringBase.cpp b/llvm/lib/CodeGen/TargetLoweringBase.cpp
index eb3190c7cd247a..8e5a0a0ca06082 100644
--- a/llvm/lib/CodeGen/TargetLoweringBase.cpp
+++ b/llvm/lib/CodeGen/TargetLoweringBase.cpp
@@ -398,6 +398,11 @@ RTLIB::Libcall RTLIB::getFREXP(EVT RetVT) {
                       FREXP_PPCF128);
 }
 
+RTLIB::Libcall RTLIB::getFSINCOS(EVT RetVT) {
+  return getFPLibCall(RetVT, SINCOS_F32, SINCOS_F64, SINCOS_F80, SINCOS_F128,
+                      SINCOS_PPCF128);
+}
+
 RTLIB::Libcall RTLIB::getOutlineAtomicHelper(const Libcall (&LC)[5][4],
                                              AtomicOrdering Order,
                                              uint64_t MemSize) {
diff --git a/llvm/test/CodeGen/AArch64/sincos-stack-slots.ll b/llvm/test/CodeGen/AArch64/sincos-stack-slots.ll
index 9c362ba117fefa..afd054a83a5014 100644
--- a/llvm/test/CodeGen/AArch64/sincos-stack-slots.ll
+++ b/llvm/test/CodeGen/AArch64/sincos-stack-slots.ll
@@ -24,22 +24,11 @@ entry:
 define void @sincos_f32_ptr_return(float %x, ptr %out_sin, ptr %out_cos) {
 ; CHECK-LABEL: sincos_f32_ptr_return:
 ; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    str x30, [sp, #-32]! // 8-byte Folded Spill
-; CHECK-NEXT:    stp x20, x19, [sp, #16] // 16-byte Folded Spill
-; CHECK-NEXT:    .cfi_def_cfa_offset 32
-; CHECK-NEXT:    .cfi_offset w19, -8
-; CHECK-NEXT:    .cfi_offset w20, -16
-; CHECK-NEXT:    .cfi_offset w30, -32
-; CHECK-NEXT:    mov x19, x1
-; CHECK-NEXT:    mov x20, x0
-; CHECK-NEXT:    add x0, sp, #12
-; CHECK-NEXT:    add x1, sp, #8
+; CHECK-NEXT:    str x30, [sp, #-16]! // 8-byte Folded Spill
+; CHECK-NEXT:    .cfi_def_cfa_offset 16
+; CHECK-NEXT:    .cfi_offset w30, -16
 ; CHECK-NEXT:    bl sincosf
-; CHECK-NEXT:    ldp s1, s0, [sp, #8]
-; CHECK-NEXT:    str s0, [x20]
-; CHECK-NEXT:    str s1, [x19]
-; CHECK-NEXT:    ldp x20, x19, [sp, #16] // 16-byte Folded Reload
-; CHECK-NEXT:    ldr x30, [sp], #32 // 8-byte Folded Reload
+; CHECK-NEXT:    ldr x30, [sp], #16 // 8-byte Folded Reload
 ; CHECK-NEXT:    ret
 entry:
   %sin = tail call float @llvm.sin.f32(float %x)
@@ -52,19 +41,13 @@ entry:
 define float @sincos_f32_mixed_return(float %x, ptr %out_sin) {
 ; CHECK-LABEL: sincos_f32_mixed_return:
 ; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    sub sp, sp, #32
-; CHECK-NEXT:    stp x30, x19, [sp, #16] // 16-byte Folded Spill
-; CHECK-NEXT:    .cfi_def_cfa_offset 32
-; CHECK-NEXT:    .cfi_offset w19, -8
+; CHECK-NEXT:    str x30, [sp, #-16]! // 8-byte Folded Spill
+; CHECK-NEXT:    .cfi_def_cfa_offset 16
 ; CHECK-NEXT:    .cfi_offset w30, -16
-; CHECK-NEXT:    mov x19, x0
-; CHECK-NEXT:    add x0, sp, #12
-; CHECK-NEXT:    add x1, sp, #8
+; CHECK-NEXT:    add x1, sp, #12
 ; CHECK-NEXT:    bl sincosf
-; CHECK-NEXT:    ldp s0, s1, [sp, #8]
-; CHECK-NEXT:    str s1, [x19]
-; CHECK-NEXT:    ldp x30, x19, [sp, #16] // 16-byte Folded Reload
-; CHECK-NEXT:    add sp, sp, #32
+; CHECK-NEXT:    ldr s0, [sp, #12]
+; CHECK-NEXT:    ldr x30, [sp], #16 // 8-byte Folded Reload
 ; CHECK-NEXT:    ret
 entry:
   %sin = tail call float @llvm.sin.f32(float %x)
@@ -99,25 +82,11 @@ entry:
 define void @sincos_f64_ptr_return(double %x, ptr %out_sin, ptr %out_cos) {
 ; CHECK-LABEL: sincos_f64_ptr_return:
 ; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    sub sp, sp, #48
-; CHECK-NEXT:    str x30, [sp, #16] // 8-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 w30, -32
-; CHECK-NEXT:    mov x19, x1
-; CHECK-NEXT:    mov x20, x0
-; CHECK-NEXT:    add x0, sp, #24
-; CHECK-NEXT:    add x1, sp, #8
+; CHECK-NEXT:    str x30, [sp, #-16]! // 8-byte Folded Spill
+; CHECK-NEXT:    .cfi_def_cfa_offset 16
+; CHECK-NEXT:    .cfi_offset w30, -16
 ; CHECK-NEXT:    bl sincos
-; CHECK-NEXT:    ldr d0, [sp, #24]
-; CHECK-NEXT:    ldr d1, [sp, #8]
-; CHECK-NEXT:    ldr x30, [sp, #16] // 8-byte Folded Reload
-; CHECK-NEXT:    str d0, [x20]
-; CHECK-NEXT:    str d1, [x19]
-; CHECK-NEXT:    ldp x20, x19, [sp, #32] // 16-byte Folded Reload
-; CHECK-NEXT:    add sp, sp, #48
+; CHECK-NEXT:    ldr x30, [sp], #16 // 8-byte Folded Reload
 ; CHECK-NEXT:    ret
 entry:
   %sin = tail call double @llvm.sin.f64(double %x)
@@ -130,19 +99,13 @@ entry:
 define double @sincos_f64_mixed_return(double %x, ptr %out_sin) {
 ; CHECK-LABEL: sincos_f64_mixed_return:
 ; CHECK:       // %bb.0: // %entry
-; CHECK-NEXT:    sub sp, sp, #32
-; CHECK-NEXT:    stp x30, x19, [sp, #16] // 16-byte Folded Spill
-; CHECK-NEXT:    .cfi_def_cfa_offset 32
-; CHECK-NEXT:    .cfi_offset w19, -8
+; CHECK-NEXT:    str x30, [sp, #-16]! // 8-byte Folded Spill
+; CHECK-NEXT:    .cfi_def_cfa_offset 16
 ; CHECK-NEXT:    .cfi_offset w30, -16
-; CHECK-NEXT:    mov x19, x0
-; CHECK-NEXT:    add x0, sp, #8
-; CHECK-NEXT:    mov x1, sp
+; CHECK-NEXT:    add x1, sp, #8
 ; CHECK-NEXT:    bl sincos
-; CHECK-NEXT:    ldp d0, d1, [sp]
-; CHECK-NEXT:    str d1, [x19]
-; CHECK-NEXT:    ldp x30, x19, [sp, #16] // 16-byte Folded Reload
-; CHECK-NEXT:    add sp, sp, #32
+; CHECK-NEXT:    ldr d0, [sp, #8]
+; CHECK-NEXT:    ldr x30, [sp], #16 // 8-byte Folded Reload
 ; CHECK-NEXT:    ret
 entry:
   %sin = tail call double @llvm.sin.f64(double %x)

>From f46d132f1f1d383a5dc838e256c344ce43d5e52d Mon Sep 17 00:00:00 2001
From: Benjamin Maxwell <benjamin.maxwell at arm.com>
Date: Thu, 12 Sep 2024 18:16:46 +0000
Subject: [PATCH 3/3] Fixups

---
 llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp | 49 ++++++---
 .../CodeGen/AArch64/sincos-stack-slots.ll     | 99 ++++++++++++++++++-
 2 files changed, 131 insertions(+), 17 deletions(-)

diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
index 7b0dc63c473d25..aa91d064db48e7 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
@@ -2351,9 +2351,7 @@ void SelectionDAGLegalize::ExpandSinCosLibCall(
     SDNode *Node, SmallVectorImpl<SDValue> &Results) {
   EVT RetVT = Node->getValueType(0);
   Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
-
-  TargetLowering::ArgListTy Args;
-  TargetLowering::ArgListEntry Entry{};
+  RTLIB::Libcall LC = RTLIB::getFSINCOS(RetVT);
 
   // Find users of the node that store the results. The destination pointers
   // can be used instead of creating stack allocations.
@@ -2366,17 +2364,15 @@ void SelectionDAGLegalize::ExpandSinCosLibCall(
     if (!ISD::isNormalStore(User))
       continue;
     auto *ST = cast<StoreSDNode>(User);
+    if (!ST->isSimple() || ST->getPointerInfo().getAddrSpace() != 0 ||
+        ST->getAlign() < DAG.getDataLayout().getABITypeAlign(RetTy))
+      continue;
     if (Use.getResNo() == 0)
       SinST = ST;
     if (Use.getResNo() == 1)
       CosST = ST;
   }
 
-  // Pass the argument.
-  Entry.Node = Node->getOperand(0);
-  Entry.Ty = RetTy;
-  Args.push_back(Entry);
-
   auto GetOrCreateOutPointer = [&](StoreSDNode *MaybeStore) {
     if (MaybeStore)
       return std::make_pair(MaybeStore->getBasePtr(),
@@ -2388,6 +2384,14 @@ void SelectionDAGLegalize::ExpandSinCosLibCall(
     return std::make_pair(StackSlot, PtrInfo);
   };
 
+  TargetLowering::ArgListTy Args;
+  TargetLowering::ArgListEntry Entry{};
+
+  // Pass the argument.
+  Entry.Node = Node->getOperand(0);
+  Entry.Ty = RetTy;
+  Args.push_back(Entry);
+
   // Pass the return address of sin.
   auto SinPtr = GetOrCreateOutPointer(SinST);
   Entry.Node = SinPtr.first;
@@ -2400,18 +2404,35 @@ void SelectionDAGLegalize::ExpandSinCosLibCall(
   Entry.Ty = PointerType::getUnqual(RetTy->getContext());
   Args.push_back(Entry);
 
-  RTLIB::Libcall LC = RTLIB::getFSINCOS(RetVT);
-  auto [Call, Chain] = ExpandLibCall(LC, Node, std::move(Args), false);
-
-  // Replace explict stores with the library call.
+  // Combine any input chains from the stores.
+  SmallVector<SDValue, 2> InChains{};
   for (StoreSDNode *ST : {SinST, CosST}) {
     if (ST)
-      DAG.ReplaceAllUsesOfValueWith(SDValue(ST, 0), Chain);
+      InChains.push_back(ST->getChain());
   }
+  if (InChains.empty())
+    InChains.push_back(DAG.getEntryNode());
 
   SDLoc DL(Node);
+  SDValue InChain = DAG.getTokenFactor(DL, InChains);
+  SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
+                                         TLI.getPointerTy(DAG.getDataLayout()));
+  TargetLowering::CallLoweringInfo CLI(DAG);
+  CLI.setDebugLoc(DL).setChain(InChain).setLibCallee(
+      TLI.getLibcallCallingConv(LC), Type::getVoidTy(*DAG.getContext()), Callee,
+      std::move(Args));
+
+  auto [Call, OutChain] = TLI.LowerCallTo(CLI);
+
+  // Replace the stores with the library call.
+  for (StoreSDNode *ST : {SinST, CosST}) {
+    if (!ST)
+      continue;
+    DAG.ReplaceAllUsesOfValueWith(SDValue(ST, 0), OutChain);
+  }
+
   for (auto [Ptr, PtrInfo] : {SinPtr, CosPtr}) {
-    SDValue LoadExp = DAG.getLoad(RetVT, DL, Chain, Ptr, PtrInfo);
+    SDValue LoadExp = DAG.getLoad(RetVT, DL, OutChain, Ptr, PtrInfo);
     Results.push_back(LoadExp);
   }
 }
diff --git a/llvm/test/CodeGen/AArch64/sincos-stack-slots.ll b/llvm/test/CodeGen/AArch64/sincos-stack-slots.ll
index afd054a83a5014..c32ac58a6a851e 100644
--- a/llvm/test/CodeGen/AArch64/sincos-stack-slots.ll
+++ b/llvm/test/CodeGen/AArch64/sincos-stack-slots.ll
@@ -1,6 +1,8 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
 ; RUN: llc -mtriple=aarch64-linux-gnu -o - %s | FileCheck %s
 
+; This file tests eliding stack slots when lowering the FSINCOS ISD node.
+
 define { float, float } @sincos_f32_value_return(float %x) {
 ; CHECK-LABEL: sincos_f32_value_return:
 ; CHECK:       // %bb.0: // %entry
@@ -91,8 +93,8 @@ define void @sincos_f64_ptr_return(double %x, ptr %out_sin, ptr %out_cos) {
 entry:
   %sin = tail call double @llvm.sin.f64(double %x)
   %cos = tail call double @llvm.cos.f64(double %x)
-  store double %sin, ptr %out_sin, align 4
-  store double %cos, ptr %out_cos, align 4
+  store double %sin, ptr %out_sin, align 8
+  store double %cos, ptr %out_cos, align 8
   ret void
 }
 
@@ -110,6 +112,97 @@ define double @sincos_f64_mixed_return(double %x, ptr %out_sin) {
 entry:
   %sin = tail call double @llvm.sin.f64(double %x)
   %cos = tail call double @llvm.cos.f64(double %x)
-  store double %sin, ptr %out_sin, align 4
+  store double %sin, ptr %out_sin, align 8
   ret double %cos
 }
+
+; Negative test. We can't fold volatile stores into the library call.
+define void @sincos_volatile_result_stores(float %x, ptr %out_sin, ptr %out_cos) {
+; CHECK-LABEL: negative_fold_sincos_volatile_store:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    str x30, [sp, #-32]! // 8-byte Folded Spill
+; CHECK-NEXT:    stp x20, x19, [sp, #16] // 16-byte Folded Spill
+; CHECK-NEXT:    .cfi_def_cfa_offset 32
+; CHECK-NEXT:    .cfi_offset w19, -8
+; CHECK-NEXT:    .cfi_offset w20, -16
+; CHECK-NEXT:    .cfi_offset w30, -32
+; CHECK-NEXT:    mov x19, x1
+; CHECK-NEXT:    mov x20, x0
+; CHECK-NEXT:    add x0, sp, #12
+; CHECK-NEXT:    add x1, sp, #8
+; CHECK-NEXT:    bl sincosf
+; CHECK-NEXT:    ldp s1, s0, [sp, #8]
+; CHECK-NEXT:    str s0, [x20]
+; CHECK-NEXT:    str s1, [x19]
+; CHECK-NEXT:    ldp x20, x19, [sp, #16] // 16-byte Folded Reload
+; CHECK-NEXT:    ldr x30, [sp], #32 // 8-byte Folded Reload
+; CHECK-NEXT:    ret
+entry:
+  %sin = tail call float @llvm.sin.f32(float %x)
+  %cos = tail call float @llvm.cos.f32(float %x)
+  store volatile float %sin, ptr %out_sin, align 4
+  store volatile float %cos, ptr %out_cos, align 4
+  ret void
+}
+
+; Negative test. We can't fold atomic stores into the library call.
+define void @sincos_atomic_result_stores(float %x, ptr %out_sin, ptr %out_cos) {
+; CHECK-LABEL: negative_fold_sincos_atomic_store:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    str x30, [sp, #-32]! // 8-byte Folded Spill
+; CHECK-NEXT:    stp x20, x19, [sp, #16] // 16-byte Folded Spill
+; CHECK-NEXT:    .cfi_def_cfa_offset 32
+; CHECK-NEXT:    .cfi_offset w19, -8
+; CHECK-NEXT:    .cfi_offset w20, -16
+; CHECK-NEXT:    .cfi_offset w30, -32
+; CHECK-NEXT:    mov x19, x1
+; CHECK-NEXT:    mov x20, x0
+; CHECK-NEXT:    add x0, sp, #12
+; CHECK-NEXT:    add x1, sp, #8
+; CHECK-NEXT:    bl sincosf
+; CHECK-NEXT:    ldr w8, [sp, #12]
+; CHECK-NEXT:    str w8, [x20]
+; CHECK-NEXT:    ldr w8, [sp, #8]
+; CHECK-NEXT:    str w8, [x19]
+; CHECK-NEXT:    ldp x20, x19, [sp, #16] // 16-byte Folded Reload
+; CHECK-NEXT:    ldr x30, [sp], #32 // 8-byte Folded Reload
+; CHECK-NEXT:    ret
+entry:
+  %sin = tail call float @llvm.sin.f32(float %x)
+  %cos = tail call float @llvm.cos.f32(float %x)
+  store atomic float %sin, ptr %out_sin unordered, align 4
+  store atomic float %cos, ptr %out_cos unordered, align 4
+  ret void
+}
+
+; Negative test. We can't fold misaligned stores into the library call.
+define void @sincos_misaligned_result_stores(double %x, ptr %out_sin, ptr %out_cos) {
+; CHECK-LABEL: negative_sincos_bad_alignment:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    sub sp, sp, #48
+; CHECK-NEXT:    str x30, [sp, #16] // 8-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 w30, -32
+; CHECK-NEXT:    mov x19, x1
+; CHECK-NEXT:    mov x20, x0
+; CHECK-NEXT:    add x0, sp, #24
+; CHECK-NEXT:    add x1, sp, #8
+; CHECK-NEXT:    bl sincos
+; CHECK-NEXT:    ldr d0, [sp, #24]
+; CHECK-NEXT:    ldr d1, [sp, #8]
+; CHECK-NEXT:    ldr x30, [sp, #16] // 8-byte Folded Reload
+; CHECK-NEXT:    str d0, [x20]
+; CHECK-NEXT:    str d1, [x19]
+; CHECK-NEXT:    ldp x20, x19, [sp, #32] // 16-byte Folded Reload
+; CHECK-NEXT:    add sp, sp, #48
+; CHECK-NEXT:    ret
+entry:
+  %sin = tail call double @llvm.sin.f64(double %x)
+  %cos = tail call double @llvm.cos.f64(double %x)
+  store double %sin, ptr %out_sin, align 4
+  store double %cos, ptr %out_cos, align 4
+  ret void
+}



More information about the llvm-commits mailing list