Add TLS support for ELF and MachO (#1174)

* Add TLS support
* Add binemit and legalize tests
* Spill all caller-saved registers when necessary

cranelift/codegen/src/binemit/mod.rs

@@ -56,6 +56,12 @@ pub enum Reloc {
     Arm64Call,
     /// RISC-V call target
     RiscvCall,
+
+    /// Elf x86_64 32 bit signed PC relative offset to two GOT entries for GD symbol.
+    ElfX86_64TlsGd,
+
+    /// Mach-O x86_64 32 bit signed PC relative offset to a `__thread_vars` entry.
+    MachOX86_64Tlv,
 }
 
 impl fmt::Display for Reloc {
@@ -71,6 +77,9 @@ impl fmt::Display for Reloc {
             Self::X86CallPLTRel4 => write!(f, "CallPLTRel4"),
             Self::X86GOTPCRel4 => write!(f, "GOTPCRel4"),
             Self::Arm32Call | Self::Arm64Call | Self::RiscvCall => write!(f, "Call"),
+
+            Self::ElfX86_64TlsGd => write!(f, "ElfX86_64TlsGd"),
+            Self::MachOX86_64Tlv => write!(f, "MachOX86_64Tlv"),
         }
     }
 }

cranelift/codegen/src/ir/globalvalue.rs

@@ -63,6 +63,9 @@ pub enum GlobalValueData {
         /// away, after linking? If so, references to it can avoid going through a GOT. Note that
         /// symbols meant to be preemptible cannot be colocated.
         colocated: bool,
+
+        /// Does this symbol refer to a thread local storage value?
+        tls: bool,
     },
 }
 
@@ -110,11 +113,13 @@ impl fmt::Display for GlobalValueData {
                 ref name,
                 offset,
                 colocated,
+                tls,
             } => {
                 write!(
                     f,
-                    "symbol {}{}",
+                    "symbol {}{}{}",
                     if colocated { "colocated " } else { "" },
+                    if tls { "tls " } else { "" },
                     name
                 )?;
                 let offset_val: i64 = offset.into();

cranelift/codegen/src/ir/libcall.rs

@@ -46,6 +46,9 @@ pub enum LibCall {
     Memset,
     /// libc.memmove
     Memmove,
+
+    /// Elf __tls_get_addr
+    ElfTlsGetAddr,
 }
 
 impl fmt::Display for LibCall {
@@ -71,6 +74,8 @@ impl FromStr for LibCall {
             "Memcpy" => Ok(Self::Memcpy),
             "Memset" => Ok(Self::Memset),
             "Memmove" => Ok(Self::Memmove),
+
+            "ElfTlsGetAddr" => Ok(Self::ElfTlsGetAddr),
             _ => Err(()),
         }
     }

cranelift/codegen/src/isa/x86/binemit.rs

@@ -4,7 +4,10 @@ use super::enc_tables::{needs_offset, needs_sib_byte};
 use super::registers::RU;
 use crate::binemit::{bad_encoding, CodeSink, Reloc};
 use crate::ir::condcodes::{CondCode, FloatCC, IntCC};
-use crate::ir::{Block, Constant, Function, Inst, InstructionData, JumpTable, Opcode, TrapCode};
+use crate::ir::{
+    Block, Constant, ExternalName, Function, Inst, InstructionData, JumpTable, LibCall, Opcode,
+    TrapCode,
+};
 use crate::isa::{RegUnit, StackBase, StackBaseMask, StackRef, TargetIsa};
 use crate::regalloc::RegDiversions;
 

cranelift/codegen/src/isa/x86/enc_tables.rs

@@ -1288,3 +1288,40 @@ fn convert_ineg(
         unreachable!()
     }
 }
+
+fn expand_tls_value(
+    inst: ir::Inst,
+    func: &mut ir::Function,
+    _cfg: &mut ControlFlowGraph,
+    isa: &dyn TargetIsa,
+) {
+    use crate::settings::TlsModel;
+
+    assert!(
+        isa.triple().architecture == target_lexicon::Architecture::X86_64,
+        "Not yet implemented for {:?}",
+        isa.triple(),
+    );
+
+    if let ir::InstructionData::UnaryGlobalValue {
+        opcode: ir::Opcode::TlsValue,
+        global_value,
+    } = func.dfg[inst]
+    {
+        let ctrl_typevar = func.dfg.ctrl_typevar(inst);
+        assert_eq!(ctrl_typevar, ir::types::I64);
+
+        match isa.flags().tls_model() {
+            TlsModel::None => panic!("tls_model flag is not set."),
+            TlsModel::ElfGd => {
+                func.dfg.replace(inst).x86_elf_tls_get_addr(global_value);
+            }
+            TlsModel::Macho => {
+                func.dfg.replace(inst).x86_macho_tls_get_addr(global_value);
+            }
+            model => unimplemented!("tls_value for tls model {:?}", model),
+        }
+    } else {
+        unreachable!();
+    }
+}

cranelift/codegen/src/legalizer/globalvalue.rs

@@ -40,7 +40,7 @@ pub fn expand_global_value(
             global_type,
             readonly,
         } => load_addr(inst, func, base, offset, global_type, readonly, isa),
-        ir::GlobalValueData::Symbol { .. } => symbol(inst, func, gv, isa),
+        ir::GlobalValueData::Symbol { tls, .. } => symbol(inst, func, gv, isa, tls),
     }
 }
 
@@ -123,7 +123,18 @@ fn load_addr(
 }
 
 /// Expand a `global_value` instruction for a symbolic name global.
-fn symbol(inst: ir::Inst, func: &mut ir::Function, gv: ir::GlobalValue, isa: &dyn TargetIsa) {
+fn symbol(
+    inst: ir::Inst,
+    func: &mut ir::Function,
+    gv: ir::GlobalValue,
+    isa: &dyn TargetIsa,
+    tls: bool,
+) {
     let ptr_ty = isa.pointer_type();
-    func.dfg.replace(inst).symbol_value(ptr_ty, gv);
+
+    if tls {
+        func.dfg.replace(inst).tls_value(ptr_ty, gv);
+    } else {
+        func.dfg.replace(inst).symbol_value(ptr_ty, gv);
+    }
 }

cranelift/codegen/src/regalloc/spilling.rs

@@ -267,7 +267,11 @@ impl<'a> Context<'a> {
         // If inst is a call, spill all register values that are live across the call.
         // This means that we don't currently take advantage of callee-saved registers.
         // TODO: Be more sophisticated.
-        if call_sig.is_some() {
+        let opcode = self.cur.func.dfg[inst].opcode();
+        if call_sig.is_some()
+            || opcode == crate::ir::Opcode::X86ElfTlsGetAddr
+            || opcode == crate::ir::Opcode::X86MachoTlsGetAddr
+        {
             for lv in throughs {
                 if lv.affinity.is_reg() && !self.spills.contains(&lv.value) {
                     self.spill_reg(lv.value);

cranelift/codegen/src/settings.rs

@@ -379,6 +379,7 @@ mod tests {
             f.to_string(),
             "[shared]\n\
              opt_level = \"none\"\n\
+             tls_model = \"none\"\n\
              libcall_call_conv = \"isa_default\"\n\
              baldrdash_prologue_words = 0\n\
              probestack_size_log2 = 12\n\