implement va_arg for x86_64 systemv

Turns out LLVM's `va_arg` is also unreliable for this target.

Author: folkertdev

compiler/rustc_codegen_llvm/src/va_arg.rs

@@ -1,4 +1,4 @@
-use rustc_abi::{Align, Endian, HasDataLayout, Size};
+use rustc_abi::{Align, BackendRepr, Endian, HasDataLayout, Primitive, Size};
 use rustc_codegen_ssa::common::IntPredicate;
 use rustc_codegen_ssa::mir::operand::OperandRef;
 use rustc_codegen_ssa::traits::{BaseTypeCodegenMethods, BuilderMethods, ConstCodegenMethods};
@@ -278,6 +278,179 @@ fn emit_s390x_va_arg<'ll, 'tcx>(
     bx.load(val_type, val_addr, layout.align.abi)
 }
 
+fn emit_x86_64_sysv64_va_arg<'ll, 'tcx>(
+    bx: &mut Builder<'_, 'll, 'tcx>,
+    list: OperandRef<'tcx, &'ll Value>,
+    target_ty: Ty<'tcx>,
+) -> &'ll Value {
+    let dl = bx.cx.data_layout();
+
+    // Implementation of the systemv x86_64 ABI calling convention for va_args, see
+    // https://gitlab.com/x86-psABIs/x86-64-ABI (section 3.5.7). This implementation is heavily
+    // based on the one in clang.
+
+    // We're able to take some shortcuts because the return type of `va_arg` must implement the
+    // `VaArgSafe` trait. Currently, only pointers, f64, i32, u32, i64 and u64 implement this trait.
+
+    // typedef struct __va_list_tag {
+    //     unsigned int gp_offset;
+    //     unsigned int fp_offset;
+    //     void *overflow_arg_area;
+    //     void *reg_save_area;
+    // } va_list[1];
+    let va_list_addr = list.immediate();
+
+    let unsigned_int_offset = 4;
+    let ptr_offset = 8;
+    let gp_offset_ptr = va_list_addr;
+    let fp_offset_ptr = bx.inbounds_ptradd(va_list_addr, bx.cx.const_usize(unsigned_int_offset));
+
+    let layout = bx.cx.layout_of(target_ty);
+
+    // AMD64-ABI 3.5.7p5: Step 1. Determine whether type may be passed
+    // in the registers. If not go to step 7.
+
+    // AMD64-ABI 3.5.7p5: Step 2. Compute num_gp to hold the number of
+    // general purpose registers needed to pass type and num_fp to hold
+    // the number of floating point registers needed.
+
+    let mut num_gp_registers = 0;
+    let mut num_fp_registers = 0;
+
+    match layout.layout.backend_repr() {
+        BackendRepr::Scalar(scalar) => match scalar.primitive() {
+            Primitive::Int(integer, _is_signed) => {
+                num_gp_registers += integer.size().bytes().div_ceil(8) as u32;
+            }
+            Primitive::Float(float) => {
+                num_fp_registers += float.size().bytes().div_ceil(16) as u32;
+            }
+            Primitive::Pointer(_) => {
+                num_gp_registers += 1;
+            }
+        },
+        BackendRepr::ScalarPair(..)
+        | BackendRepr::SimdVector { .. }
+        | BackendRepr::Memory { .. } => {
+            // Because no instance of VaArgSafe uses a non-scalar `BackendRepr`.
+            unreachable!(
+                "No x86-64 SysV va_arg implementation for {:?}",
+                layout.layout.backend_repr()
+            )
+        }
+    };
+
+    if num_gp_registers == 0 && num_fp_registers == 0 {
+        unreachable!("VaArgSafe is not implemented for ZSTs")
+    }
+
+    // AMD64-ABI 3.5.7p5: Step 3. Verify whether arguments fit into
+    // registers. In the case: l->gp_offset > 48 - num_gp * 8 or
+    // l->fp_offset > 176 - num_fp * 16 go to step 7.
+
+    let gp_offset_v = bx.load(bx.type_i32(), gp_offset_ptr, Align::from_bytes(8).unwrap());
+    let fp_offset_v = bx.load(bx.type_i32(), fp_offset_ptr, Align::from_bytes(4).unwrap());
+
+    let mut use_regs = bx.const_bool(false);
+
+    if num_gp_registers > 0 {
+        let max_offset_val = 48u32 - num_gp_registers * 8;
+        let fits_in_gp = bx.icmp(IntPredicate::IntULE, gp_offset_v, bx.const_u32(max_offset_val));
+        use_regs = fits_in_gp;
+    }
+
+    if num_fp_registers > 0 {
+        let max_offset_val = 176u32 - num_fp_registers * 16;
+        let fits_in_fp = bx.icmp(IntPredicate::IntULE, fp_offset_v, bx.const_u32(max_offset_val));
+        use_regs = if num_gp_registers > 0 { bx.and(use_regs, fits_in_fp) } else { fits_in_fp };
+    }
+
+    let in_reg = bx.append_sibling_block("va_arg.in_reg");
+    let in_mem = bx.append_sibling_block("va_arg.in_mem");
+    let end = bx.append_sibling_block("va_arg.end");
+
+    bx.cond_br(use_regs, in_reg, in_mem);
+
+    // Emit code to load the value if it was passed in a register.
+    bx.switch_to_block(in_reg);
+
+    // AMD64-ABI 3.5.7p5: Step 4. Fetch type from l->reg_save_area with
+    // an offset of l->gp_offset and/or l->fp_offset. This may require
+    // copying to a temporary location in case the parameter is passed
+    // in different register classes or requires an alignment greater
+    // than 8 for general purpose registers and 16 for XMM registers.
+    //
+    // FIXME(llvm): This really results in shameful code when we end up needing to
+    // collect arguments from different places; often what should result in a
+    // simple assembling of a structure from scattered addresses has many more
+    // loads than necessary. Can we clean this up?
+    let reg_save_area_ptr =
+        bx.inbounds_ptradd(va_list_addr, bx.cx.const_usize(2 * unsigned_int_offset + ptr_offset));
+    let reg_save_area_v = bx.load(bx.type_ptr(), reg_save_area_ptr, dl.pointer_align.abi);
+
+    let reg_addr = if num_gp_registers > 0 && num_fp_registers > 0 {
+        unreachable!("instances of VaArgSafe cannot use both int and sse registers");
+    } else if num_gp_registers > 0 || num_fp_registers == 1 {
+        let gp_or_fp_offset = if num_gp_registers > 0 { gp_offset_v } else { fp_offset_v };
+        bx.gep(bx.type_i8(), reg_save_area_v, &[gp_or_fp_offset])
+    } else {
+        // assert_eq!(num_sse_registers, 2);
+        unreachable!("all instances of VaArgSafe have an alignment <= 8");
+    };
+
+    // AMD64-ABI 3.5.7p5: Step 5. Set:
+    // l->gp_offset = l->gp_offset + num_gp * 8
+    if num_gp_registers > 0 {
+        let offset = bx.const_u32(num_gp_registers * 8);
+        let sum = bx.add(gp_offset_v, offset);
+        bx.store(sum, gp_offset_ptr, Align::from_bytes(8).unwrap());
+    }
+
+    // l->fp_offset = l->fp_offset + num_fp * 16.
+    if num_fp_registers > 0 {
+        let offset = bx.const_u32(num_fp_registers * 16);
+        let sum = bx.add(fp_offset_v, offset);
+        bx.store(sum, fp_offset_ptr, Align::from_bytes(4).unwrap());
+    }
+
+    bx.br(end);
+
+    bx.switch_to_block(in_mem);
+
+    let overflow_arg_area_ptr =
+        bx.inbounds_ptradd(va_list_addr, bx.cx.const_usize(2 * unsigned_int_offset));
+
+    let overflow_arg_area_v = bx.load(bx.type_ptr(), overflow_arg_area_ptr, dl.pointer_align.abi);
+    // AMD64-ABI 3.5.7p5: Step 7. Align l->overflow_arg_area upwards to a 16
+    // byte boundary if alignment needed by type exceeds 8 byte boundary.
+    // It isn't stated explicitly in the standard, but in practice we use
+    // alignment greater than 16 where necessary.
+    if layout.layout.align.abi.bytes() > 8 {
+        unreachable!("all instances of VaArgSafe have an alignment <= 8");
+    }
+
+    // AMD64-ABI 3.5.7p5: Step 8. Fetch type from l->overflow_arg_area.
+    let mem_addr = overflow_arg_area_v;
+
+    // AMD64-ABI 3.5.7p5: Step 9. Set l->overflow_arg_area to:
+    // l->overflow_arg_area + sizeof(type).
+    // AMD64-ABI 3.5.7p5: Step 10. Align l->overflow_arg_area upwards to
+    // an 8 byte boundary.
+    let size_in_bytes = layout.layout.size().bytes();
+    let offset = bx.const_i32(size_in_bytes.next_multiple_of(8) as i32);
+    let overflow_arg_area = bx.gep(bx.type_i8(), overflow_arg_area_v, &[offset]);
+    bx.store(overflow_arg_area, overflow_arg_area_ptr, dl.pointer_align.abi);
+
+    bx.br(end);
+
+    bx.switch_to_block(end);
+
+    let val_type = layout.llvm_type(bx);
+    let val_addr = bx.phi(bx.type_ptr(), &[reg_addr, mem_addr], &[in_reg, in_mem]);
+
+    bx.load(val_type, val_addr, layout.align.abi)
+}
+
 fn emit_xtensa_va_arg<'ll, 'tcx>(
     bx: &mut Builder<'_, 'll, 'tcx>,
     list: OperandRef<'tcx, &'ll Value>,
@@ -410,6 +583,7 @@ pub(super) fn emit_va_arg<'ll, 'tcx>(
             let indirect: bool = target_ty_size > 8 || !target_ty_size.is_power_of_two();
             emit_ptr_va_arg(bx, addr, target_ty, indirect, Align::from_bytes(8).unwrap(), false)
         }
+        "x86_64" if !target.is_like_darwin => emit_x86_64_sysv64_va_arg(bx, addr, target_ty),
         "xtensa" => emit_xtensa_va_arg(bx, addr, target_ty),
         // For all other architecture/OS combinations fall back to using
         // the LLVM va_arg instruction.

tests/run-make/c-link-to-rust-va-list-fn/checkrust.rs

@@ -112,6 +112,9 @@ pub unsafe extern "C" fn check_varargs_4(_: c_double, mut ap: ...) -> usize {
     continue_if!(ap.arg::<c_double>() == 8.0);
     continue_if!(ap.arg::<c_double>() == 9.0);
     continue_if!(ap.arg::<c_double>() == 10.0);
+    continue_if!(ap.arg::<c_double>() == 11.0);
+    continue_if!(ap.arg::<c_double>() == 12.0);
+    continue_if!(ap.arg::<c_double>() == 13.0);
     0
 }
 
@@ -137,5 +140,11 @@ pub unsafe extern "C" fn check_varargs_5(_: c_int, mut ap: ...) -> usize {
     continue_if!(ap.arg::<c_double>() == 9.0);
     continue_if!(ap.arg::<c_int>() == 10);
     continue_if!(ap.arg::<c_double>() == 10.0);
+    continue_if!(ap.arg::<c_int>() == 11);
+    continue_if!(ap.arg::<c_double>() == 11.0);
+    continue_if!(ap.arg::<c_int>() == 12);
+    continue_if!(ap.arg::<c_double>() == 12.0);
+    continue_if!(ap.arg::<c_int>() == 13);
+    continue_if!(ap.arg::<c_double>() == 13.0);
     0
 }

tests/run-make/c-link-to-rust-va-list-fn/test.c

@@ -41,10 +41,11 @@ int main(int argc, char* argv[]) {
 
     assert(check_varargs_3(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10) == 0);
 
-    assert(check_varargs_4(0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0) == 0);
+    assert(check_varargs_4(0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0,
+                           13.0) == 0);
 
     assert(check_varargs_5(0, 1.0, 1, 2.0, 2, 3.0, 3, 4.0, 4, 5, 5.0, 6, 6.0, 7, 7.0, 8, 8.0,
-                           9, 9.0, 10, 10.0) == 0);
+                           9, 9.0, 10, 10.0, 11, 11.0, 12, 12.0, 13, 13.0) == 0);
 
     return 0;
 }