Fully support multiple returns in Wasmtime (#2806)

* Fully support multiple returns in Wasmtime

For quite some time now Wasmtime has "supported" multiple return values,
but only in the mose bare bones ways. Up until recently you couldn't get
a typed version of functions with multiple return values, and never have
you been able to use `Func::wrap` with functions that return multiple
values. Even recently where `Func::typed` can call functions that return
multiple values it uses a double-indirection by calling a trampoline
which calls the real function.

The underlying reason for this lack of support is that cranelift's ABI
for returning multiple values is not possible to write in Rust. For
example if a wasm function returns two `i32` values there is no Rust (or
C!) function you can write to correspond to that. This commit, however
fixes that.

This commit adds two new ABIs to Cranelift: `WasmtimeSystemV` and
`WasmtimeFastcall`. The intention is that these Wasmtime-specific ABIs
match their corresponding ABI (e.g. `SystemV` or `WindowsFastcall`) for
everything *except* how multiple values are returned. For multiple
return values we simply define our own version of the ABI which Wasmtime
implements, which is that for N return values the first is returned as
if the function only returned that and the latter N-1 return values are
returned via an out-ptr that's the last parameter to the function.

These custom ABIs provides the ability for Wasmtime to bind these in
Rust meaning that `Func::wrap` can now wrap functions that return
multiple values and `Func::typed` no longer uses trampolines when
calling functions that return multiple values. Although there's lots of
internal changes there's no actual changes in the API surface area of
Wasmtime, just a few more impls of more public traits which means that
more types are supported in more places!

Another change made with this PR is a consolidation of how the ABI of
each function in a wasm module is selected. The native `SystemV` ABI,
for example, is more efficient at returning multiple values than the
wasmtime version of the ABI (since more things are in more registers).
To continue to take advantage of this Wasmtime will now classify some
functions in a wasm module with the "fast" ABI. Only functions that are
not reachable externally from the module are classified with the fast
ABI (e.g. those not exported, used in tables, or used with `ref.func`).
This should enable purely internal functions of modules to have a faster
calling convention than those which might be exposed to Wasmtime itself.

Closes #1178

* Tweak some names and add docs

* "fix" lightbeam compile

* Fix TODO with dummy environ

* Unwind info is a property of the target, not the ABI

* Remove lightbeam unused imports

* Attempt to fix arm64

* Document new ABIs aren't stable

* Fix filetests to use the right target

* Don't always do 64-bit stores with cranelift

This was overwriting upper bits when 32-bit registers were being stored
into return values, so fix the code inline to do a sized store instead
of one-size-fits-all store.

* At least get tests passing on the old backend

* Fix a typo

* Add some filetests with mixed abi calls

* Get `multi` example working

* Fix doctests on old x86 backend

* Add a mixture of wasmtime/system_v tests

Cargo.toml

@@ -98,7 +98,7 @@ all-arch = ["wasmtime/all-arch"]
 experimental_x64 = []
 
 # Use the old x86 backend.
-old-x86-backend = ["wasmtime-jit/old-x86-backend"]
+old-x86-backend = ["wasmtime/old-x86-backend"]
 
 [badges]
 maintenance = { status = "actively-developed" }

cranelift/codegen/src/context.rs

@@ -267,13 +267,7 @@ impl Context {
         isa: &dyn TargetIsa,
     ) -> CodegenResult<Option<crate::isa::unwind::UnwindInfo>> {
         if let Some(backend) = isa.get_mach_backend() {
-            use crate::isa::CallConv;
-            use crate::machinst::UnwindInfoKind;
-            let unwind_info_kind = match self.func.signature.call_conv {
-                CallConv::Fast | CallConv::Cold | CallConv::SystemV => UnwindInfoKind::SystemV,
-                CallConv::WindowsFastcall => UnwindInfoKind::Windows,
-                _ => UnwindInfoKind::None,
-            };
+            let unwind_info_kind = isa.unwind_info_kind();
             let result = self.mach_compile_result.as_ref().unwrap();
             return backend.emit_unwind_info(result, unwind_info_kind);
         }

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

@@ -197,18 +197,26 @@ impl ABIMachineSpec for AArch64MachineDeps {
             next_stack = 16;
         }
 
-        // Note on return values: on the regular non-baldrdash ABI, we may return values in 8
-        // registers for V128 and I64 registers independently of the number of register values
-        // returned in the other class. That is, we can return values in up to 8 integer and 8
-        // vector registers at once.
-        // In Baldrdash, we can only use one register for return value for all the register
-        // classes. That is, we can't return values in both one integer and one vector register;
-        // only one return value may be in a register.
-
-        let (max_per_class_reg_vals, mut remaining_reg_vals) = match (args_or_rets, is_baldrdash) {
-            (ArgsOrRets::Args, _) => (8, 16),     // x0-x7 and v0-v7
-            (ArgsOrRets::Rets, false) => (8, 16), // x0-x7 and v0-v7
-            (ArgsOrRets::Rets, true) => (1, 1),   // x0 or v0, but not both
+        let (max_per_class_reg_vals, mut remaining_reg_vals) = match args_or_rets {
+            ArgsOrRets::Args => (8, 16), // x0-x7 and v0-v7
+
+            // Note on return values: on the regular ABI, we may return values
+            // in 8 registers for V128 and I64 registers independently of the
+            // number of register values returned in the other class. That is,
+            // we can return values in up to 8 integer and
+            // 8 vector registers at once.
+            //
+            // In Baldrdash and Wasmtime, we can only use one register for
+            // return value for all the register classes. That is, we can't
+            // return values in both one integer and one vector register; only
+            // one return value may be in a register.
+            ArgsOrRets::Rets => {
+                if is_baldrdash || call_conv.extends_wasmtime() {
+                    (1, 1) // x0 or v0, but not both
+                } else {
+                    (8, 16) // x0-x7 and v0-v7
+                }
+            }
         };
 
         for i in 0..params.len() {
@@ -282,15 +290,18 @@ impl ABIMachineSpec for AArch64MachineDeps {
                 // Compute the stack slot's size.
                 let size = (ty_bits(param.value_type) / 8) as u64;
 
-                let size = if call_conv != isa::CallConv::AppleAarch64 {
+                let size = if call_conv == isa::CallConv::AppleAarch64
+                    || (call_conv.extends_wasmtime() && args_or_rets == ArgsOrRets::Rets)
+                {
+                    // MacOS aarch64 and Wasmtime allow stack slots with
+                    // sizes less than 8 bytes. They still need to be
+                    // properly aligned on their natural data alignment,
+                    // though.
+                    size
+                } else {
                     // Every arg takes a minimum slot of 8 bytes. (16-byte stack
                     // alignment happens separately after all args.)
                     std::cmp::max(size, 8)
-                } else {
-                    // MacOS aarch64 allows stack slots with sizes less than 8
-                    // bytes. They still need to be properly aligned on their
-                    // natural data alignment, though.
-                    size
                 };
 
                 // Align the stack slot.

cranelift/codegen/src/isa/call_conv.rs

@@ -29,6 +29,16 @@ pub enum CallConv {
     Baldrdash2020,
     /// Specialized convention for the probestack function.
     Probestack,
+    /// Wasmtime equivalent of SystemV, not ABI-stable.
+    ///
+    /// Currently only differs in how multiple return values are handled,
+    /// returning the first return value in a register and everything else
+    /// through a return-pointer.
+    WasmtimeSystemV,
+    /// Wasmtime equivalent of WindowsFastcall, not ABI-stable.
+    ///
+    /// Differs from fastcall in the same way as `WasmtimeSystemV`.
+    WasmtimeFastcall,
 }
 
 impl CallConv {
@@ -63,7 +73,7 @@ impl CallConv {
     /// Is the calling convention extending the Windows Fastcall ABI?
     pub fn extends_windows_fastcall(self) -> bool {
         match self {
-            Self::WindowsFastcall | Self::BaldrdashWindows => true,
+            Self::WindowsFastcall | Self::BaldrdashWindows | Self::WasmtimeFastcall => true,
             _ => false,
         }
     }
@@ -75,6 +85,14 @@ impl CallConv {
             _ => false,
         }
     }
+
+    /// Is the calling convention extending the Wasmtime ABI?
+    pub fn extends_wasmtime(self) -> bool {
+        match self {
+            Self::WasmtimeSystemV | Self::WasmtimeFastcall => true,
+            _ => false,
+        }
+    }
 }
 
 impl fmt::Display for CallConv {
@@ -89,6 +107,8 @@ impl fmt::Display for CallConv {
             Self::BaldrdashWindows => "baldrdash_windows",
             Self::Baldrdash2020 => "baldrdash_2020",
             Self::Probestack => "probestack",
+            Self::WasmtimeSystemV => "wasmtime_system_v",
+            Self::WasmtimeFastcall => "wasmtime_fastcall",
         })
     }
 }
@@ -106,6 +126,8 @@ impl str::FromStr for CallConv {
             "baldrdash_windows" => Ok(Self::BaldrdashWindows),
             "baldrdash_2020" => Ok(Self::Baldrdash2020),
             "probestack" => Ok(Self::Probestack),
+            "wasmtime_system_v" => Ok(Self::WasmtimeSystemV),
+            "wasmtime_fastcall" => Ok(Self::WasmtimeFastcall),
             _ => Err(()),
         }
     }

cranelift/codegen/src/isa/mod.rs

@@ -57,7 +57,7 @@ use crate::flowgraph;
 use crate::ir;
 #[cfg(feature = "unwind")]
 use crate::isa::unwind::systemv::RegisterMappingError;
-use crate::machinst::MachBackend;
+use crate::machinst::{MachBackend, UnwindInfoKind};
 use crate::regalloc;
 use crate::result::CodegenResult;
 use crate::settings;
@@ -68,7 +68,7 @@ use core::any::Any;
 use core::fmt;
 use core::fmt::{Debug, Formatter};
 use core::hash::Hasher;
-use target_lexicon::{triple, Architecture, PointerWidth, Triple};
+use target_lexicon::{triple, Architecture, OperatingSystem, PointerWidth, Triple};
 use thiserror::Error;
 
 #[cfg(feature = "riscv")]
@@ -476,6 +476,18 @@ pub trait TargetIsa: fmt::Display + Send + Sync {
     /// IntCC condition for Unsigned Subtraction Overflow (Borrow/Carry).
     fn unsigned_sub_overflow_condition(&self) -> ir::condcodes::IntCC;
 
+    /// Returns the flavor of unwind information emitted for this target.
+    fn unwind_info_kind(&self) -> UnwindInfoKind {
+        match self.triple().operating_system {
+            #[cfg(feature = "unwind")]
+            OperatingSystem::Windows => UnwindInfoKind::Windows,
+            #[cfg(feature = "unwind")]
+            _ => UnwindInfoKind::SystemV,
+            #[cfg(not(feature = "unwind"))]
+            _ => UnwindInfoKind::None,
+        }
+    }
+
     /// Creates unwind information for the function.
     ///
     /// Returns `None` if there is no unwind information for the function.

cranelift/codegen/src/isa/x64/abi.rs

@@ -237,10 +237,20 @@ impl ABIMachineSpec for X64ABIMachineSpec {
                         extension: param.extension,
                     });
                 } else {
-                    // Compute size. Every arg takes a minimum slot of 8 bytes. (16-byte
-                    // stack alignment happens separately after all args.)
+                    // Compute size. For the wasmtime ABI it differs from native
+                    // ABIs in how multiple values are returned, so we take a
+                    // leaf out of arm64's book by not rounding everything up to
+                    // 8 bytes. For all ABI arguments, and other ABI returns,
+                    // though, each slot takes a minimum of 8 bytes.
+                    //
+                    // Note that in all cases 16-byte stack alignment happens
+                    // separately after all args.
                     let size = (reg_ty.bits() / 8) as u64;
-                    let size = std::cmp::max(size, 8);
+                    let size = if args_or_rets == ArgsOrRets::Rets && call_conv.extends_wasmtime() {
+                        size
+                    } else {
+                        std::cmp::max(size, 8)
+                    };
                     // Align.
                     debug_assert!(size.is_power_of_two());
                     next_stack = align_to(next_stack, size);
@@ -824,15 +834,7 @@ impl From<StackAMode> for SyntheticAmode {
 }
 
 fn get_intreg_for_arg(call_conv: &CallConv, idx: usize, arg_idx: usize) -> Option<Reg> {
-    let is_fastcall = match call_conv {
-        CallConv::Fast
-        | CallConv::Cold
-        | CallConv::SystemV
-        | CallConv::BaldrdashSystemV
-        | CallConv::Baldrdash2020 => false,
-        CallConv::WindowsFastcall => true,
-        _ => panic!("int args only supported for SysV or Fastcall calling convention"),
-    };
+    let is_fastcall = call_conv.extends_windows_fastcall();
 
     // Fastcall counts by absolute argument number; SysV counts by argument of
     // this (integer) class.
@@ -853,15 +855,7 @@ fn get_intreg_for_arg(call_conv: &CallConv, idx: usize, arg_idx: usize) -> Optio
 }
 
 fn get_fltreg_for_arg(call_conv: &CallConv, idx: usize, arg_idx: usize) -> Option<Reg> {
-    let is_fastcall = match call_conv {
-        CallConv::Fast
-        | CallConv::Cold
-        | CallConv::SystemV
-        | CallConv::BaldrdashSystemV
-        | CallConv::Baldrdash2020 => false,
-        CallConv::WindowsFastcall => true,
-        _ => panic!("float args only supported for SysV or Fastcall calling convention"),
-    };
+    let is_fastcall = call_conv.extends_windows_fastcall();
 
     // Fastcall counts by absolute argument number; SysV counts by argument of
     // this (floating-point) class.
@@ -894,7 +888,10 @@ fn get_intreg_for_retval(
             1 => Some(regs::rdx()),
             _ => None,
         },
-        CallConv::BaldrdashSystemV | CallConv::Baldrdash2020 => {
+        CallConv::BaldrdashSystemV
+        | CallConv::Baldrdash2020
+        | CallConv::WasmtimeSystemV
+        | CallConv::WasmtimeFastcall => {
             if intreg_idx == 0 && retval_idx == 0 {
                 Some(regs::rax())
             } else {
@@ -922,7 +919,10 @@ fn get_fltreg_for_retval(
             1 => Some(regs::xmm1()),
             _ => None,
         },
-        CallConv::BaldrdashSystemV | CallConv::Baldrdash2020 => {
+        CallConv::BaldrdashSystemV
+        | CallConv::Baldrdash2020
+        | CallConv::WasmtimeFastcall
+        | CallConv::WasmtimeSystemV => {
             if fltreg_idx == 0 && retval_idx == 0 {
                 Some(regs::xmm0())
             } else {
@@ -992,12 +992,12 @@ fn get_callee_saves(call_conv: &CallConv, regs: &Set<Writable<RealReg>>) -> Vec<
         CallConv::BaldrdashWindows => {
             todo!("baldrdash windows");
         }
-        CallConv::Fast | CallConv::Cold | CallConv::SystemV => regs
+        CallConv::Fast | CallConv::Cold | CallConv::SystemV | CallConv::WasmtimeSystemV => regs
             .iter()
             .cloned()
             .filter(|r| is_callee_save_systemv(r.to_reg()))
             .collect(),
-        CallConv::WindowsFastcall => regs
+        CallConv::WindowsFastcall | CallConv::WasmtimeFastcall => regs
             .iter()
             .cloned()
             .filter(|r| is_callee_save_fastcall(r.to_reg()))

cranelift/codegen/src/isa/x64/inst/mod.rs

@@ -1122,11 +1122,16 @@ impl Inst {
     pub(crate) fn store(ty: Type, from_reg: Reg, to_addr: impl Into<SyntheticAmode>) -> Inst {
         let rc = from_reg.get_class();
         match rc {
-            RegClass::I64 => {
-                // Always store the full register, to ensure that the high bits are properly set
-                // when doing a full reload.
-                Inst::mov_r_m(OperandSize::Size64, from_reg, to_addr)
-            }
+            RegClass::I64 => Inst::mov_r_m(
+                match ty {
+                    types::B1 => OperandSize::Size8,
+                    types::I32 | types::R32 => OperandSize::Size32,
+                    types::I64 | types::R64 => OperandSize::Size64,
+                    _ => unimplemented!("integer store of type: {}", ty),
+                },
+                from_reg,
+                to_addr,
+            ),
             RegClass::V128 => {
                 let opcode = match ty {
                     types::F32 => SseOpcode::Movss,

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

@@ -503,10 +503,12 @@ fn callee_saved_regs_used(isa: &dyn TargetIsa, func: &ir::Function) -> RegisterS
 pub fn prologue_epilogue(func: &mut ir::Function, isa: &dyn TargetIsa) -> CodegenResult<()> {
     match func.signature.call_conv {
         // For now, just translate fast and cold as system_v.
-        CallConv::Fast | CallConv::Cold | CallConv::SystemV => {
+        CallConv::Fast | CallConv::Cold | CallConv::SystemV | CallConv::WasmtimeSystemV => {
             system_v_prologue_epilogue(func, isa)
         }
-        CallConv::WindowsFastcall => fastcall_prologue_epilogue(func, isa),
+        CallConv::WindowsFastcall | CallConv::WasmtimeFastcall => {
+            fastcall_prologue_epilogue(func, isa)
+        }
         CallConv::BaldrdashSystemV | CallConv::BaldrdashWindows => {
             baldrdash_prologue_epilogue(func, isa)
         }
@@ -1084,16 +1086,17 @@ pub fn create_unwind_info(
     isa: &dyn TargetIsa,
 ) -> CodegenResult<Option<crate::isa::unwind::UnwindInfo>> {
     use crate::isa::unwind::UnwindInfo;
+    use crate::machinst::UnwindInfoKind;
 
     // Assumption: RBP is being used as the frame pointer for both calling conventions
     // In the future, we should be omitting frame pointer as an optimization, so this will change
-    Ok(match func.signature.call_conv {
-        CallConv::Fast | CallConv::Cold | CallConv::SystemV => {
+    Ok(match isa.unwind_info_kind() {
+        UnwindInfoKind::SystemV => {
             super::unwind::systemv::create_unwind_info(func, isa)?.map(|u| UnwindInfo::SystemV(u))
         }
-        CallConv::WindowsFastcall => {
+        UnwindInfoKind::Windows => {
             super::unwind::winx64::create_unwind_info(func, isa)?.map(|u| UnwindInfo::WindowsX64(u))
         }
-        _ => None,
+        UnwindInfoKind::None => None,
     })
 }

cranelift/codegen/src/isa/x86/unwind/systemv.rs

@@ -3,7 +3,7 @@
 use crate::ir::Function;
 use crate::isa::{
     unwind::systemv::{RegisterMappingError, UnwindInfo},
-    CallConv, RegUnit, TargetIsa,
+    RegUnit, TargetIsa,
 };
 use crate::result::CodegenResult;
 use gimli::{write::CommonInformationEntry, Encoding, Format, Register, X86_64};
@@ -97,8 +97,8 @@ pub(crate) fn create_unwind_info(
     isa: &dyn TargetIsa,
 ) -> CodegenResult<Option<UnwindInfo>> {
     // Only System V-like calling conventions are supported
-    match func.signature.call_conv {
-        CallConv::Fast | CallConv::Cold | CallConv::SystemV => {}
+    match isa.unwind_info_kind() {
+        crate::machinst::UnwindInfoKind::SystemV => {}
         _ => return Ok(None),
     }
 

cranelift/codegen/src/machinst/abi.rs

@@ -189,9 +189,6 @@ pub trait ABICallee {
         from_slot: SpillSlot,
         ty: Option<Type>,
     ) -> Self::I;
-
-    /// Desired unwind info type.
-    fn unwind_info_kind(&self) -> UnwindInfoKind;
 }
 
 /// Trait implemented by an object that tracks ABI-related state and can