[X86] Use movq for i64 atomic load on 32-bit targets when sse2 is enable

[craig.topper]

Summary

We used a lock cmpxchg8b to do i64 atomic loads. But if we have SSE2 we can do better and use a plain movq to do the load instead.

I tried to just use an f64 atomic load and add isel patterns to MOVSD(which the domain fixing pass can turn to MOVQ), but the atomic_load SDNode in TargetSelectionDAG.td requires the type to be integer.

So I've emitted VZEXT_LOAD instead which should be selected by isel to a MOVQ. Hopefully we don't need a specific atomic flavor of this. I kept the memory operand from the original AtomicSDNode. I wasn't sure if I might need to set the MOVolatile flag?

I've left some FIXMEs for improvements we can do without SSE2.

[efriedma]

From what I understand, the atomic marking in the memory operand should be enough (assuming there aren't any relevant target-specific DAGCombines).

Maybe also add CHECK lines to test/CodeGen/X86/atomic-non-integer.ll?

llvm/test/CodeGen/X86/atomic-load-store-wide.ll

90

Maybe we should keep test coverage for old CPUs? Or is that unlikely to be useful?

[craig.topper]

Update atomic-non-integer.ll too. Add an old CPU command line to atomic-load-store-wide.ll

[efriedma]

Do we need to check for NoImplicitFloat somewhere?

[craig.topper]

Block the transform if noimplicitfloat is set on the function.

[efriedma]

LGTM

llvm/lib/Target/X86/X86ISelLowering.cpp

for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
     setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
     setOperationAction(ISD::ATOMIC_LOAD_ADD, VT, Custom);
     setOperationAction(ISD::ATOMIC_LOAD_OR, VT, Custom);
     setOperationAction(ISD::ATOMIC_LOAD_XOR, VT, Custom);
     setOperationAction(ISD::ATOMIC_LOAD_AND, VT, Custom);
     setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
   }
 
+  if (!Subtarget.is64Bit())
+    setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Custom);
+
   if (Subtarget.hasCmpxchg16b()) {
     setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom);
   }
 
   // FIXME - use subtarget debug flags
   if (!Subtarget.isTargetDarwin() && !Subtarget.isTargetELF() &&
       !Subtarget.isTargetCygMing() && !Subtarget.isTargetWin64() &&
       TM.Options.ExceptionModel != ExceptionHandling::SjLj) {
bool X86TargetLowering::needsCmpXchgNb(Type *MemType) const {
   return false;
 }
 
 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
   return needsCmpXchgNb(SI->getValueOperand()->getType());
 }
 
 // Note: this turns large loads into lock cmpxchg8b/16b.
-// FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
+// TODO: In 32-bit mode, use MOVLPS when SSE1 is available?
+// TODO: In 32-bit mode, use FILD/FISTP when X87 is available?
 TargetLowering::AtomicExpansionKind
 X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
-  return needsCmpXchgNb(LI->getType()) ? AtomicExpansionKind::CmpXChg
-                                       : AtomicExpansionKind::None;
+  Type *MemType = LI->getType();
+
+  // If this a 64 bit atomic load on a 32-bit target and SSE2 is enabled, we
+  // can use movq to do the load.
+  bool NoImplicitFloatOps =
+      LI->getFunction()->hasFnAttribute(Attribute::NoImplicitFloat);
+  if (MemType->getPrimitiveSizeInBits() == 64 && !Subtarget.is64Bit() &&
+      !Subtarget.useSoftFloat() && !NoImplicitFloatOps && Subtarget.hasSSE2())
+    return AtomicExpansionKind::None;
+
+  return needsCmpXchgNb(MemType) ? AtomicExpansionKind::CmpXChg
 }
 
 TargetLowering::AtomicExpansionKind
 X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
   unsigned NativeWidth = Subtarget.is64Bit() ? 64 : 32;
   Type *MemType = AI->getType();
 
   // If the operand is too big, we must see if cmpxchg8/16b is available
case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
     SDValue Success = getSETCC(X86::COND_E, EFLAGS, dl, DAG);
     Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
 
     Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
     Results.push_back(Success);
     Results.push_back(EFLAGS.getValue(1));
     return;
   }
+  case ISD::ATOMIC_LOAD: {
+    assert(N->getValueType(0) == MVT::i64 && "Unexpected VT!");
+    bool NoImplicitFloatOps =
+        DAG.getMachineFunction().getFunction().hasFnAttribute(
+            Attribute::NoImplicitFloat);
+    if (!Subtarget.useSoftFloat() && !NoImplicitFloatOps &&
+        Subtarget.hasSSE2()) {
+      auto *Node = cast<AtomicSDNode>(N);
+      // Use a VZEXT_LOAD which will be selected as MOVQ. Then extract the lower
+      // 64-bits.
+      SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
+      SDValue Ops[] = { Node->getChain(), Node->getBasePtr() };
+      SDValue Ld = DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
+                                           MVT::i64, Node->getMemOperand());
+      SDValue Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Ld,
+                                DAG.getIntPtrConstant(0, dl));
+      Results.push_back(Res);
+      Results.push_back(Ld.getValue(1));
+      return;
+    }
+    // TODO: Use MOVLPS when SSE1 is available?
+    // TODO: Use FILD/FISTP when X87 is available?
+    // Delegate to generic TypeLegalization. Situations we can really handle
+    // should have already been dealt with by AtomicExpandPass.cpp.
+    break;
+  }
   case ISD::ATOMIC_SWAP:
   case ISD::ATOMIC_LOAD_ADD:
   case ISD::ATOMIC_LOAD_SUB:
   case ISD::ATOMIC_LOAD_AND:
   case ISD::ATOMIC_LOAD_OR:
   case ISD::ATOMIC_LOAD_XOR:
   case ISD::ATOMIC_LOAD_NAND:
   case ISD::ATOMIC_LOAD_MIN:
   case ISD::ATOMIC_LOAD_MAX:
   case ISD::ATOMIC_LOAD_UMIN:
   case ISD::ATOMIC_LOAD_UMAX:
-  case ISD::ATOMIC_LOAD: {
     // Delegate to generic TypeLegalization. Situations we can really handle
     // should have already been dealt with by AtomicExpandPass.cpp.
     break;
-  }
+
   case ISD::BITCAST: {
     assert(Subtarget.hasSSE2() && "Requires at least SSE2!");
     EVT DstVT = N->getValueType(0);
     EVT SrcVT = N->getOperand(0).getValueType();
 
     // If this is a bitcast from a v64i1 k-register to a i64 on a 32-bit target
     // we can split using the k-register rather than memory.
     if (SrcVT == MVT::v64i1 && DstVT == MVT::i64 && Subtarget.hasBWI()) {

llvm/test/CodeGen/X86/atomic-load-store-wide.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc < %s -mcpu=corei7 -mtriple=i686-- -verify-machineinstrs | FileCheck %s
+; RUN: llc < %s -mcpu=corei7 -mtriple=i686-- -verify-machineinstrs | FileCheck %s --check-prefix=CHECK --check-prefix=SSE42
+; RUN: llc < %s -mtriple=i686-- -verify-machineinstrs | FileCheck %s --check-prefix=CHECK --check-prefix=NOSSE
 
 ; 64-bit load/store on x86-32
 ; FIXME: The generated code can be substantially improved.
 
 define void @test1(i64* %ptr, i64 %val1) {
 ; CHECK-LABEL: test1:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    pushl %ebx

 ; CHECK-NEXT:    popl %ebx
 ; CHECK-NEXT:    .cfi_def_cfa_offset 4
 ; CHECK-NEXT:    retl
   store atomic i64 %val1, i64* %ptr seq_cst, align 8
   ret void
 }
 
 define i64 @test2(i64* %ptr) {
-; CHECK-LABEL: test2:
+; SSE42-LABEL: test2:
+; SSE42:       # %bb.0:
+; SSE42-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; SSE42-NEXT:    movq {{.*#+}} xmm0 = mem[0],zero
+; SSE42-NEXT:    movd %xmm0, %eax
+; SSE42-NEXT:    pextrd $1, %xmm0, %edx
+; SSE42-NEXT:    retl
+;
+; NOSSE-LABEL: test2:
+; NOSSE:       # %bb.0:
+; NOSSE-NEXT:    pushl %ebx
+; NOSSE-NEXT:    .cfi_def_cfa_offset 8
+; NOSSE-NEXT:    pushl %esi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 12
+; NOSSE-NEXT:    .cfi_offset %esi, -12
+; NOSSE-NEXT:    .cfi_offset %ebx, -8
+; NOSSE-NEXT:    movl {{[0-9]+}}(%esp), %esi
+; NOSSE-NEXT:    xorl %eax, %eax
+; NOSSE-NEXT:    xorl %edx, %edx
+; NOSSE-NEXT:    xorl %ecx, %ecx
+; NOSSE-NEXT:    xorl %ebx, %ebx
+; NOSSE-NEXT:    lock cmpxchg8b (%esi)
+; NOSSE-NEXT:    popl %esi
+; NOSSE-NEXT:    .cfi_def_cfa_offset 8
+; NOSSE-NEXT:    popl %ebx
+; NOSSE-NEXT:    .cfi_def_cfa_offset 4
+; NOSSE-NEXT:    retl
+  %val = load atomic i64, i64* %ptr seq_cst, align 8
+  ret i64 %val
+}
+
+; Same as test2, but with noimplicitfloat.
+define i64 @test3(i64* %ptr) noimplicitfloat {
+; CHECK-LABEL: test3:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    pushl %ebx
 ; CHECK-NEXT:    .cfi_def_cfa_offset 8
 ; CHECK-NEXT:    pushl %esi
 ; CHECK-NEXT:    .cfi_def_cfa_offset 12
 ; CHECK-NEXT:    .cfi_offset %esi, -12
 ; CHECK-NEXT:    .cfi_offset %ebx, -8
 ; CHECK-NEXT:    movl {{[0-9]+}}(%esp), %esi
 ; CHECK-NEXT:    xorl %eax, %eax
 ; CHECK-NEXT:    xorl %edx, %edx
 ; CHECK-NEXT:    xorl %ecx, %ecx
 ; CHECK-NEXT:    xorl %ebx, %ebx
 ; CHECK-NEXT:    lock cmpxchg8b (%esi)
 ; CHECK-NEXT:    popl %esi
 ; CHECK-NEXT:    .cfi_def_cfa_offset 8
 ; CHECK-NEXT:    popl %ebx
 ; CHECK-NEXT:    .cfi_def_cfa_offset 4
 ; CHECK-NEXT:    retl
   %val = load atomic i64, i64* %ptr seq_cst, align 8
   ret i64 %val
 }

llvm/test/CodeGen/X86/atomic-non-integer.ll

 ; X64-NEXT:    retq
   %v = load atomic float, float* %fptr unordered, align 4
   ret float %v
 }
 
 define double @load_double(double* %fptr) {
 ; X86-SSE-LABEL: load_double:
 ; X86-SSE:       # %bb.0:
-; X86-SSE-NEXT:    pushl %ebx
-; X86-SSE-NEXT:    .cfi_def_cfa_offset 8
-; X86-SSE-NEXT:    pushl %esi
-; X86-SSE-NEXT:    .cfi_def_cfa_offset 12
 ; X86-SSE-NEXT:    subl $12, %esp
-; X86-SSE-NEXT:    .cfi_def_cfa_offset 24
-; X86-SSE-NEXT:    .cfi_offset %esi, -12
-; X86-SSE-NEXT:    .cfi_offset %ebx, -8
-; X86-SSE-NEXT:    movl {{[0-9]+}}(%esp), %esi
-; X86-SSE-NEXT:    xorl %eax, %eax
-; X86-SSE-NEXT:    xorl %edx, %edx
-; X86-SSE-NEXT:    xorl %ecx, %ecx
-; X86-SSE-NEXT:    xorl %ebx, %ebx
-; X86-SSE-NEXT:    lock cmpxchg8b (%esi)
-; X86-SSE-NEXT:    movd %edx, %xmm0
-; X86-SSE-NEXT:    movd %eax, %xmm1
-; X86-SSE-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; X86-SSE-NEXT:    movq %xmm1, (%esp)
+; X86-SSE-NEXT:    .cfi_def_cfa_offset 16
+; X86-SSE-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-SSE-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
+; X86-SSE-NEXT:    movlps %xmm0, (%esp)
 ; X86-SSE-NEXT:    fldl (%esp)
 ; X86-SSE-NEXT:    addl $12, %esp
-; X86-SSE-NEXT:    .cfi_def_cfa_offset 12
-; X86-SSE-NEXT:    popl %esi
-; X86-SSE-NEXT:    .cfi_def_cfa_offset 8
-; X86-SSE-NEXT:    popl %ebx
 ; X86-SSE-NEXT:    .cfi_def_cfa_offset 4
 ; X86-SSE-NEXT:    retl
 ;
 ; X86-NOSSE-LABEL: load_double:
 ; X86-NOSSE:       # %bb.0:
 ; X86-NOSSE-NEXT:    pushl %ebx
 ; X86-NOSSE-NEXT:    .cfi_def_cfa_offset 8
 ; X86-NOSSE-NEXT:    pushl %esi

 ; X64-NEXT:    retq
   %v = load atomic float, float* %fptr seq_cst, align 4
   ret float %v
 }
 
 define double @load_double_seq_cst(double* %fptr) {
 ; X86-SSE-LABEL: load_double_seq_cst:
 ; X86-SSE:       # %bb.0:
-; X86-SSE-NEXT:    pushl %ebx
-; X86-SSE-NEXT:    .cfi_def_cfa_offset 8
-; X86-SSE-NEXT:    pushl %esi
-; X86-SSE-NEXT:    .cfi_def_cfa_offset 12
 ; X86-SSE-NEXT:    subl $12, %esp
-; X86-SSE-NEXT:    .cfi_def_cfa_offset 24
-; X86-SSE-NEXT:    .cfi_offset %esi, -12
-; X86-SSE-NEXT:    .cfi_offset %ebx, -8
-; X86-SSE-NEXT:    movl {{[0-9]+}}(%esp), %esi
-; X86-SSE-NEXT:    xorl %eax, %eax
-; X86-SSE-NEXT:    xorl %edx, %edx
-; X86-SSE-NEXT:    xorl %ecx, %ecx
-; X86-SSE-NEXT:    xorl %ebx, %ebx
-; X86-SSE-NEXT:    lock cmpxchg8b (%esi)
-; X86-SSE-NEXT:    movd %edx, %xmm0
-; X86-SSE-NEXT:    movd %eax, %xmm1
-; X86-SSE-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; X86-SSE-NEXT:    movq %xmm1, (%esp)
+; X86-SSE-NEXT:    .cfi_def_cfa_offset 16
+; X86-SSE-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-SSE-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
+; X86-SSE-NEXT:    movlps %xmm0, (%esp)
 ; X86-SSE-NEXT:    fldl (%esp)
 ; X86-SSE-NEXT:    addl $12, %esp
-; X86-SSE-NEXT:    .cfi_def_cfa_offset 12
-; X86-SSE-NEXT:    popl %esi
-; X86-SSE-NEXT:    .cfi_def_cfa_offset 8
-; X86-SSE-NEXT:    popl %ebx
 ; X86-SSE-NEXT:    .cfi_def_cfa_offset 4
 ; X86-SSE-NEXT:    retl
 ;
 ; X86-NOSSE-LABEL: load_double_seq_cst:
 ; X86-NOSSE:       # %bb.0:
 ; X86-NOSSE-NEXT:    pushl %ebx
 ; X86-NOSSE-NEXT:    .cfi_def_cfa_offset 8
 ; X86-NOSSE-NEXT:    pushl %esi