Implement SystemV struct argument passing

Author: bjorn3

cranelift/codegen/meta/src/cdsl/types.rs

@@ -407,6 +407,8 @@ impl fmt::Debug for VectorType {
 #[derive(Clone, Copy, PartialEq, Eq, Hash)]
 pub(crate) enum SpecialType {
     Flag(shared_types::Flag),
+    // FIXME remove once the old style backends are removed.
+    StructArgument,
 }
 
 impl SpecialType {
@@ -421,13 +423,17 @@ impl SpecialType {
                 "CPU flags representing the result of a floating point comparison. These
                 flags can be tested with a :type:`floatcc` condition code.",
             ),
+            SpecialType::StructArgument => {
+                String::from("After legalization sarg__ arguments will get this type.")
+            }
         }
     }
 
     /// Return the number of bits in a lane.
     pub fn lane_bits(self) -> u64 {
         match self {
             SpecialType::Flag(_) => 0,
+            SpecialType::StructArgument => 0,
         }
     }
 
@@ -436,6 +442,7 @@ impl SpecialType {
         match self {
             SpecialType::Flag(shared_types::Flag::IFlags) => 1,
             SpecialType::Flag(shared_types::Flag::FFlags) => 2,
+            SpecialType::StructArgument => 3,
         }
     }
 }
@@ -445,6 +452,7 @@ impl fmt::Display for SpecialType {
         match *self {
             SpecialType::Flag(shared_types::Flag::IFlags) => write!(f, "iflags"),
             SpecialType::Flag(shared_types::Flag::FFlags) => write!(f, "fflags"),
+            SpecialType::StructArgument => write!(f, "sarg__"),
         }
     }
 }
@@ -456,6 +464,7 @@ impl fmt::Debug for SpecialType {
             "{}",
             match *self {
                 SpecialType::Flag(_) => format!("FlagsType({})", self),
+                SpecialType::StructArgument => format!("StructArgument"),
             }
         )
     }
@@ -469,12 +478,14 @@ impl From<shared_types::Flag> for SpecialType {
 
 pub(crate) struct SpecialTypeIterator {
     flag_iter: shared_types::FlagIterator,
+    done: bool,
 }
 
 impl SpecialTypeIterator {
     fn new() -> Self {
         Self {
             flag_iter: shared_types::FlagIterator::new(),
+            done: false,
         }
     }
 }
@@ -485,7 +496,12 @@ impl Iterator for SpecialTypeIterator {
         if let Some(f) = self.flag_iter.next() {
             Some(SpecialType::from(f))
         } else {
-            None
+            if !self.done {
+                self.done = true;
+                Some(SpecialType::StructArgument)
+            } else {
+                None
+            }
         }
     }
 }

cranelift/codegen/meta/src/isa/x86/encodings.rs

@@ -439,6 +439,7 @@ fn define_moves(e: &mut PerCpuModeEncodings, shared_defs: &SharedDefinitions, r:
     let sextend = shared.by_name("sextend");
     let set_pinned_reg = shared.by_name("set_pinned_reg");
     let uextend = shared.by_name("uextend");
+    let dummy_sarg__ = shared.by_name("dummy_sarg__");
 
     // Shorthands for recipes.
     let rec_copysp = r.template("copysp");
@@ -456,6 +457,7 @@ fn define_moves(e: &mut PerCpuModeEncodings, shared_defs: &SharedDefinitions, r:
     let rec_umr_reg_to_ssa = r.template("umr_reg_to_ssa");
     let rec_urm_noflags = r.template("urm_noflags");
     let rec_urm_noflags_abcd = r.template("urm_noflags_abcd");
+    let rec_dummy_sarg__ = r.recipe("dummy_sarg__");
 
     // The pinned reg is fixed to a certain value entirely user-controlled, so it generates nothing!
     e.enc64_rec(get_pinned_reg.bind(I64), rec_get_pinned_reg, 0);
@@ -721,6 +723,8 @@ fn define_moves(e: &mut PerCpuModeEncodings, shared_defs: &SharedDefinitions, r:
         copy_to_ssa.bind(F32),
         rec_furm_reg_to_ssa.opcodes(&MOVSS_LOAD),
     );
+
+    e.enc_32_64_rec(dummy_sarg__, rec_dummy_sarg__, 0);
 }
 
 #[inline(never)]

cranelift/codegen/meta/src/isa/x86/recipes.rs

@@ -1270,6 +1270,12 @@ pub(crate) fn define<'shared>(
         );
     }
 
+    recipes.add_recipe(
+        EncodingRecipeBuilder::new("dummy_sarg__", &formats.nullary, 0)
+            .operands_out(vec![Stack::new(gpr)])
+            .emit(""),
+    );
+
     // XX+rd id with Abs4 function relocation.
     recipes.add_template_recipe(
         EncodingRecipeBuilder::new("fnaddr4", &formats.func_addr, 4)

cranelift/codegen/meta/src/shared/instructions.rs

@@ -1835,6 +1835,31 @@ pub(crate) fn define(
         .can_load(true),
     );
 
+    let Sarg = &TypeVar::new(
+        "Sarg",
+        "Any scalar or vector type with as most 128 lanes",
+        TypeSetBuilder::new()
+            .specials(vec![crate::cdsl::types::SpecialType::StructArgument])
+            .build(),
+    );
+    let sarg__ = &Operand::new("sarg__", Sarg);
+
+    // FIXME remove once the old style codegen backends are removed.
+    ig.push(
+        Inst::new(
+            "dummy_sarg__",
+            r#"
+        This creates a sarg__
+
+        This instruction is internal and should not be created by
+        Cranelift users.
+        "#,
+            &formats.nullary,
+        )
+        .operands_in(vec![])
+        .operands_out(vec![sarg__]),
+    );
+
     let src = &Operand::new("src", &imm.regunit);
     let dst = &Operand::new("dst", &imm.regunit);
 

cranelift/codegen/src/abi.rs

@@ -18,6 +18,10 @@ pub enum ArgAction {
     /// Assign the argument to the given location.
     Assign(ArgumentLoc),
 
+    /// Assign the argument to the given location and change the type to the specified type.
+    /// This is used by [`ArgumentPurpose::StructArgument`].
+    AssignAndChangeType(ArgumentLoc, Type),
+
     /// Convert the argument, then call again.
     ///
     /// This action can split an integer type into two smaller integer arguments, or it can split a
@@ -119,6 +123,13 @@ pub fn legalize_args<AA: ArgAssigner>(args: &[AbiParam], aa: &mut AA) -> Option<
                 args.to_mut()[argno].location = loc;
                 argno += 1;
             }
+            // Assign argument to a location, change type to `INVALID` and move on to the next one.
+            ArgAction::AssignAndChangeType(loc, ty) => {
+                let arg = &mut args.to_mut()[argno];
+                arg.location = loc;
+                arg.value_type = ty;
+                argno += 1;
+            }
             // Split this argument into two smaller ones. Then revisit both.
             ArgAction::Convert(conv) => {
                 debug_assert!(

cranelift/codegen/src/ir/extfunc.rs

@@ -276,6 +276,9 @@ pub enum ArgumentPurpose {
     /// A normal user program value passed to or from a function.
     Normal,
 
+    /// A C struct passed as argument.
+    StructArgument(u32),
+
     /// Struct return pointer.
     ///
     /// When a function needs to return more data than will fit in registers, the caller passes a
@@ -328,21 +331,19 @@ pub enum ArgumentPurpose {
     StackLimit,
 }
 
-/// Text format names of the `ArgumentPurpose` variants.
-static PURPOSE_NAMES: [&str; 8] = [
-    "normal",
-    "sret",
-    "link",
-    "fp",
-    "csr",
-    "vmctx",
-    "sigid",
-    "stack_limit",
-];
-
 impl fmt::Display for ArgumentPurpose {
     fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
-        f.write_str(PURPOSE_NAMES[*self as usize])
+        f.write_str(match self {
+            Self::Normal => "normal",
+            Self::StructArgument(size) => return write!(f, "sarg({})", size),
+            Self::StructReturn => "sret",
+            Self::Link => "link",
+            Self::FramePointer => "fp",
+            Self::CalleeSaved => "csr",
+            Self::VMContext => "vmctx",
+            Self::SignatureId => "sigid",
+            Self::StackLimit => "stack_limit",
+        })
     }
 }
 
@@ -358,6 +359,13 @@ impl FromStr for ArgumentPurpose {
             "vmctx" => Ok(Self::VMContext),
             "sigid" => Ok(Self::SignatureId),
             "stack_limit" => Ok(Self::StackLimit),
+            _ if s.starts_with("sarg(") => {
+                if !s.ends_with(")") {
+                    return Err(());
+                }
+                let size: u32 = s["sarg(".len()..s.len() - 1].parse().map_err(|_| ())?;
+                Ok(Self::StructArgument(size))
+            }
             _ => Err(()),
         }
     }
@@ -430,16 +438,17 @@ mod tests {
     #[test]
     fn argument_purpose() {
         let all_purpose = [
-            ArgumentPurpose::Normal,
-            ArgumentPurpose::StructReturn,
-            ArgumentPurpose::Link,
-            ArgumentPurpose::FramePointer,
-            ArgumentPurpose::CalleeSaved,
-            ArgumentPurpose::VMContext,
-            ArgumentPurpose::SignatureId,
-            ArgumentPurpose::StackLimit,
+            (ArgumentPurpose::Normal, "normal"),
+            (ArgumentPurpose::StructReturn, "sret"),
+            (ArgumentPurpose::Link, "link"),
+            (ArgumentPurpose::FramePointer, "fp"),
+            (ArgumentPurpose::CalleeSaved, "csr"),
+            (ArgumentPurpose::VMContext, "vmctx"),
+            (ArgumentPurpose::SignatureId, "sigid"),
+            (ArgumentPurpose::StackLimit, "stack_limit"),
+            (ArgumentPurpose::StructArgument(42), "sarg(42)"),
         ];
-        for (&e, &n) in all_purpose.iter().zip(PURPOSE_NAMES.iter()) {
+        for &(e, n) in &all_purpose {
             assert_eq!(e.to_string(), n);
             assert_eq!(Ok(e), n.parse());
         }

cranelift/codegen/src/ir/stackslot.rs

@@ -287,7 +287,12 @@ impl StackSlots {
 
     /// Create a stack slot representing an incoming function argument.
     pub fn make_incoming_arg(&mut self, ty: Type, offset: StackOffset) -> StackSlot {
-        let mut data = StackSlotData::new(StackSlotKind::IncomingArg, ty.bytes());
+        self.make_incoming_struct_arg(ty.bytes(), offset)
+    }
+
+    /// Create a stack slot representing an incoming struct function argument.
+    pub fn make_incoming_struct_arg(&mut self, size: u32, offset: StackOffset) -> StackSlot {
+        let mut data = StackSlotData::new(StackSlotKind::IncomingArg, size);
         debug_assert!(offset <= StackOffset::max_value() - data.size as StackOffset);
         data.offset = Some(offset);
         self.push(data)
@@ -301,8 +306,11 @@ impl StackSlots {
     /// The requested offset is relative to this function's stack pointer immediately before making
     /// the call.
     pub fn get_outgoing_arg(&mut self, ty: Type, offset: StackOffset) -> StackSlot {
-        let size = ty.bytes();
+        self.get_outgoing_struct_arg(ty.bytes(), offset)
+    }
 
+    /// FIXME
+    pub fn get_outgoing_struct_arg(&mut self, size: u32, offset: StackOffset) -> StackSlot {
         // Look for an existing outgoing stack slot with the same offset and size.
         let inspos = match self.outgoing.binary_search_by_key(&(offset, size), |&ss| {
             (self[ss].offset.unwrap(), self[ss].size)

cranelift/codegen/src/ir/types.rs

@@ -329,6 +329,7 @@ impl Display for Type {
             f.write_str(match *self {
                 IFLAGS => "iflags",
                 FFLAGS => "fflags",
+                SARG__ => "sarg__",
                 INVALID => panic!("INVALID encountered"),
                 _ => panic!("Unknown Type(0x{:x})", self.0),
             })

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

@@ -2069,6 +2069,8 @@ pub(crate) fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
             panic!("x86-specific opcode in supposedly arch-neutral IR!");
         }
 
+        Opcode::DummySarg => unreachable!(),
+
         Opcode::AvgRound => unimplemented!(),
         Opcode::TlsValue => unimplemented!(),
     }

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

@@ -108,6 +108,19 @@ impl Args {
 
 impl ArgAssigner for Args {
     fn assign(&mut self, arg: &AbiParam) -> ArgAction {
+        if let ArgumentPurpose::StructArgument(size) = arg.purpose {
+            if self.call_conv != CallConv::SystemV {
+                panic!(
+                    "The sarg argument purpose is not yet implemented for non-systemv call conv {:?}",
+                    self.call_conv,
+                );
+            }
+            let loc = ArgumentLoc::Stack(self.offset as i32);
+            self.offset += size;
+            debug_assert!(self.offset <= i32::MAX as u32);
+            return ArgAction::AssignAndChangeType(loc, types::SARG__);
+        }
+
         let ty = arg.value_type;
 
         if ty.bits() > u16::from(self.pointer_bits) {