Merge pull request #3026 from afonso360/aarch64-elf-tls

aarch64: Implement TLS ELF GD Relocations

Author: cfallin

cranelift/codegen/src/binemit/mod.rs

@@ -68,6 +68,16 @@ pub enum Reloc {
 
     /// Mach-O x86_64 32 bit signed PC relative offset to a `__thread_vars` entry.
     MachOX86_64Tlv,
+
+    /// AArch64 TLS GD
+    /// Set an ADRP immediate field to the top 21 bits of the final address. Checks for overflow.
+    /// This is equivalent to `R_AARCH64_TLSGD_ADR_PAGE21` in the [aaelf64](https://github.com/ARM-software/abi-aa/blob/2bcab1e3b22d55170c563c3c7940134089176746/aaelf64/aaelf64.rst#relocations-for-thread-local-storage)
+    Aarch64TlsGdAdrPage21,
+
+    /// AArch64 TLS GD
+    /// Set the add immediate field to the low 12 bits of the final address. Does not check for overflow.
+    /// This is equivalent to `R_AARCH64_TLSGD_ADD_LO12_NC` in the [aaelf64](https://github.com/ARM-software/abi-aa/blob/2bcab1e3b22d55170c563c3c7940134089176746/aaelf64/aaelf64.rst#relocations-for-thread-local-storage)
+    Aarch64TlsGdAddLo12Nc,
 }
 
 impl fmt::Display for Reloc {
@@ -87,6 +97,8 @@ impl fmt::Display for Reloc {
 
             Self::ElfX86_64TlsGd => write!(f, "ElfX86_64TlsGd"),
             Self::MachOX86_64Tlv => write!(f, "MachOX86_64Tlv"),
+            Self::Aarch64TlsGdAdrPage21 => write!(f, "Aarch64TlsGdAdrPage21"),
+            Self::Aarch64TlsGdAddLo12Nc => write!(f, "Aarch64TlsGdAddLo12Nc"),
         }
     }
 }

cranelift/codegen/src/isa/aarch64/inst/emit.rs

@@ -3,7 +3,7 @@
 use crate::binemit::{CodeOffset, Reloc, StackMap};
 use crate::ir::constant::ConstantData;
 use crate::ir::types::*;
-use crate::ir::{MemFlags, TrapCode};
+use crate::ir::{LibCall, MemFlags, TrapCode};
 use crate::isa::aarch64::inst::*;
 use crate::machinst::ty_bits;
 
@@ -2563,6 +2563,32 @@ impl MachInstEmit for Inst {
                     sink.bind_label(jump_around_label);
                 }
             }
+
+            &Inst::ElfTlsGetAddr { ref symbol } => {
+                // This is the instruction sequence that GCC emits for ELF GD TLS Relocations in aarch64
+                // See: https://gcc.godbolt.org/z/KhMh5Gvra
+
+                // adrp x0, <label>
+                sink.add_reloc(state.cur_srcloc(), Reloc::Aarch64TlsGdAdrPage21, symbol, 0);
+                sink.put4(0x90000000);
+
+                // add x0, x0, <label>
+                sink.add_reloc(state.cur_srcloc(), Reloc::Aarch64TlsGdAddLo12Nc, symbol, 0);
+                sink.put4(0x91000000);
+
+                // bl __tls_get_addr
+                sink.add_reloc(
+                    state.cur_srcloc(),
+                    Reloc::Arm64Call,
+                    &ExternalName::LibCall(LibCall::ElfTlsGetAddr),
+                    0,
+                );
+                sink.put4(0x94000000);
+
+                // nop
+                sink.put4(0xd503201f);
+            }
+
             &Inst::ValueLabelMarker { .. } => {
                 // Nothing; this is only used to compute debug info.
             }

cranelift/codegen/src/isa/aarch64/inst/mod.rs

@@ -30,6 +30,8 @@ pub mod args;
 pub use self::args::*;
 pub mod emit;
 pub use self::emit::*;
+use crate::isa::aarch64::abi::AArch64MachineDeps;
+
 pub mod unwind;
 
 #[cfg(test)]
@@ -1276,6 +1278,11 @@ pub enum Inst {
         needed_space: CodeOffset,
     },
 
+    /// A call to the `ElfTlsGetAddr` libcall. Returns address of TLS symbol in x0.
+    ElfTlsGetAddr {
+        symbol: ExternalName,
+    },
+
     /// A definition of a value label.
     ValueLabelMarker {
         reg: Reg,
@@ -2150,6 +2157,12 @@ fn aarch64_get_regs(inst: &Inst, collector: &mut RegUsageCollector) {
         &Inst::ValueLabelMarker { reg, .. } => {
             collector.add_use(reg);
         }
+
+        &Inst::ElfTlsGetAddr { .. } => {
+            for reg in AArch64MachineDeps::get_regs_clobbered_by_call(CallConv::SystemV) {
+                collector.add_def(reg);
+            }
+        }
         &Inst::Unwind { .. } => {}
         &Inst::EmitIsland { .. } => {}
     }
@@ -2959,6 +2972,7 @@ fn aarch64_map_regs<RUM: RegUsageMapper>(inst: &mut Inst, mapper: &RUM) {
         }
         &mut Inst::VirtualSPOffsetAdj { .. } => {}
         &mut Inst::EmitIsland { .. } => {}
+        &mut Inst::ElfTlsGetAddr { .. } => {}
         &mut Inst::ValueLabelMarker { ref mut reg, .. } => {
             map_use(mapper, reg);
         }
@@ -4333,6 +4347,10 @@ impl Inst {
             }
             &Inst::EmitIsland { needed_space } => format!("emit_island {}", needed_space),
 
+            &Inst::ElfTlsGetAddr { ref symbol } => {
+                format!("elf_tls_get_addr {}", symbol)
+            }
+
             &Inst::ValueLabelMarker { label, reg } => {
                 format!("value_label {:?}, {}", label, reg.show_rru(mb_rru))
             }

cranelift/codegen/src/isa/aarch64/lower_inst.rs

@@ -8,7 +8,7 @@ use crate::ir::{InstructionData, Opcode, TrapCode};
 use crate::isa::aarch64::settings as aarch64_settings;
 use crate::machinst::lower::*;
 use crate::machinst::*;
-use crate::settings::Flags;
+use crate::settings::{Flags, TlsModel};
 use crate::{CodegenError, CodegenResult};
 
 use crate::isa::aarch64::abi::*;
@@ -3506,7 +3506,24 @@ pub(crate) fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
             });
         }
 
-        Opcode::TlsValue => unimplemented!("tls_value"),
+        Opcode::TlsValue => match flags.tls_model() {
+            TlsModel::ElfGd => {
+                let dst = get_output_reg(ctx, outputs[0]).only_reg().unwrap();
+                let (name, _, _) = ctx.symbol_value(insn).unwrap();
+                let symbol = name.clone();
+                ctx.emit(Inst::ElfTlsGetAddr { symbol });
+
+                let x0 = xreg(0);
+                ctx.emit(Inst::gen_move(dst, x0, I64));
+            }
+            _ => {
+                todo!(
+                    "Unimplemented TLS model in AArch64 backend: {:?}",
+                    flags.tls_model()
+                );
+            }
+        },
+
         Opcode::FcvtLowFromSint => unimplemented!("FcvtLowFromSint"),
         Opcode::FvpromoteLow => unimplemented!("FvpromoteLow"),
         Opcode::Fvdemote => unimplemented!("Fvdemote"),

cranelift/filetests/filetests/isa/aarch64/tls-elf-gd.clif

@@ -0,0 +1,29 @@
+test compile
+set tls_model=elf_gd
+target aarch64
+
+function u0:0(i32) -> i32, i64 {
+gv0 = symbol colocated tls u1:0
+
+block0(v0: i32):
+    v1 = global_value.i64 gv0
+    return v0, v1
+}
+; check: stp fp, lr, [sp, #-16]!
+; nextln: mov fp, sp
+; nextln: str x19, [sp, #-16]!
+; nextln: stp d14, d15, [sp, #-16]!
+; nextln: stp d12, d13, [sp, #-16]!
+; nextln: stp d10, d11, [sp, #-16]!
+; nextln: stp d8, d9, [sp, #-16]!
+; nextln: mov x19, x0
+; nextln: elf_tls_get_addr u1:0
+; nextln: mov x1, x0
+; nextln: mov x0, x19
+; nextln: ldp d8, d9, [sp], #16
+; nextln: ldp d10, d11, [sp], #16
+; nextln: ldp d12, d13, [sp], #16
+; nextln: ldp d14, d15, [sp], #16
+; nextln: ldr x19, [sp], #16
+; nextln: ldp fp, lr, [sp], #16
+; nextln: ret

cranelift/object/src/backend.rs

@@ -618,6 +618,30 @@ impl ObjectModule {
                     32,
                 )
             }
+            Reloc::Aarch64TlsGdAdrPage21 => {
+                assert_eq!(
+                    self.object.format(),
+                    object::BinaryFormat::Elf,
+                    "Aarch64TlsGdAdrPrel21 is not supported for this file format"
+                );
+                (
+                    RelocationKind::Elf(object::elf::R_AARCH64_TLSGD_ADR_PAGE21),
+                    RelocationEncoding::Generic,
+                    21,
+                )
+            }
+            Reloc::Aarch64TlsGdAddLo12Nc => {
+                assert_eq!(
+                    self.object.format(),
+                    object::BinaryFormat::Elf,
+                    "Aarch64TlsGdAddLo12Nc is not supported for this file format"
+                );
+                (
+                    RelocationKind::Elf(object::elf::R_AARCH64_TLSGD_ADD_LO12_NC),
+                    RelocationEncoding::Generic,
+                    12,
+                )
+            }
             // FIXME
             reloc => unimplemented!("{:?}", reloc),
         };