[LoongArch] Add support for emulated TLS model

[wangleiat]

Summary

This will ensure that the emulated TLS code should be generated when
useEmulatedTLS() is true.

[MaskRay]

Emulated TLS is legacy code that we should not support. If ELF TLS works well, stick with ELF TLS.

[wangleiat]

In D141183#4042149, @MaskRay wrote:

Emulated TLS is legacy code that we should not support. If ELF TLS works well, stick with ELF TLS.

Thank you for your comments.
For this support, the main purposes are:

1. Make the emulated-tls option work properly.
2. Make lli handle TLV correctly under LoongArch.

For the second point, it may be necessary at present. I try to remove EmulatedTLS support from the targets(x86_64/aarch64) that supports jit, and jit will not be able to handle TLV correctly.

[wangleiat]

It is a pity that emulated tls cannot be supported.

[MaskRay]

In D141183#4088560, @wangleiat wrote:

It is a pity that emulated tls cannot be supported.

Disagree. Nearly two years ago I objected to RISC-V adding emulated TLS as well.
Cutting corners is easy for porters but is not good for the long-term maintanance of the project.
I am unfamiliar with lib/ExecutionEngine/ but it appears to support ELF TLS for some other ports.

llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp

LoongArchTargetLowering::lowerGlobalTLSAddress(SDValue Op,
                                                SelectionDAG &DAG) const {
   if (DAG.getMachineFunction().getFunction().getCallingConv() ==
       CallingConv::GHC)
     report_fatal_error("In GHC calling convention TLS is not supported");
 
   GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Op);
   assert(N->getOffset() == 0 && "unexpected offset in global node");
 
+  if (DAG.getTarget().useEmulatedTLS())
+    return LowerToTLSEmulatedModel(N, DAG);
+
   SDValue Addr;
   switch (getTargetMachine().getTLSModel(N->getGlobal())) {
   case TLSModel::GeneralDynamic:
     // In this model, application code calls the dynamic linker function
     // __tls_get_addr to locate TLS offsets into the dynamic thread vector at
     // runtime.
     Addr = getDynamicTLSAddr(N, DAG, LoongArch::PseudoLA_TLS_GD);
     break;

llvm/test/CodeGen/LoongArch/emutls_generic.ll

+; RUN: llc < %s --emulated-tls --mtriple=loongarch64 --relocation-model=pic \
+; RUN:     | FileCheck %s
+
+; RUN: llc < %s --mtriple=loongarch64 --relocation-model=pic \
+; RUN:     | FileCheck --check-prefix=NoEMU %s
+
+;; NoEMU-NOT: __emutls
+
+;; Make sure that TLS symbols are emitted in expected order.
+
+@external_x = external thread_local global i32, align 8
+@external_y = thread_local global i8 7, align 2
+@internal_y = internal thread_local global i64 9, align 16
+
+define ptr @get_external_x() {
+; CHECK-LABEL: get_external_x:
+; CHECK-NOT:   _tls_get_address
+; CHECK:       __emutls_get_address
+entry:
+  ret ptr @external_x
+}
+
+define ptr @get_external_y() {
+; CHECK-LABEL: get_external_y:
+; CHECK:       __emutls_get_address
+; CHECK-NOT:   _tls_get_address
+entry:
+  ret ptr @external_y
+}
+
+define ptr @get_internal_y() {
+; CHECK-LABEL: get_internal_y:
+; CHECK-NOT:   __emutls_t.external_x:
+; CHECK-NOT:   __emutls_v.external_x:
+; CHECK-LABEL: __emutls_v.external_y:
+; CHECK-LABEL: __emutls_t.external_y:
+; CHECK:       __emutls_t.external_y
+; CHECK-LABEL: __emutls_v.internal_y:
+; CHECK-LABEL: __emutls_t.internal_y:
+; CHECK:       __emutls_t.internal_y
+entry:
+  ret ptr @internal_y
+}