Skip to content

Commit 1946240

Browse files
authored
Unrolled build for rust-lang#127565
Rollup merge of rust-lang#127565 - esp-rs:xtensa-vaargs, r=workingjubilee Teach rustc about the Xtensa VaListImpl Following on from the target Xtensa target PRs (rust-lang#125141, rust-lang#126380), this PR teaches rustc about the structure of the VA list on the Xtensa arch, as well as adding the required lowering to be able to actually use it.
2 parents 0e98766 + 059f627 commit 1946240

File tree

3 files changed

+131
-2
lines changed

3 files changed

+131
-2
lines changed

compiler/rustc_codegen_llvm/src/va_arg.rs

+111-2
Original file line numberDiff line numberDiff line change
@@ -10,15 +10,23 @@ use crate::type_::Type;
1010
use crate::type_of::LayoutLlvmExt;
1111
use crate::value::Value;
1212

13+
fn round_up_to_alignment<'ll>(
14+
bx: &mut Builder<'_, 'll, '_>,
15+
mut value: &'ll Value,
16+
align: Align,
17+
) -> &'ll Value {
18+
value = bx.add(value, bx.cx().const_i32(align.bytes() as i32 - 1));
19+
return bx.and(value, bx.cx().const_i32(-(align.bytes() as i32)));
20+
}
21+
1322
fn round_pointer_up_to_alignment<'ll>(
1423
bx: &mut Builder<'_, 'll, '_>,
1524
addr: &'ll Value,
1625
align: Align,
1726
ptr_ty: &'ll Type,
1827
) -> &'ll Value {
1928
let mut ptr_as_int = bx.ptrtoint(addr, bx.cx().type_isize());
20-
ptr_as_int = bx.add(ptr_as_int, bx.cx().const_i32(align.bytes() as i32 - 1));
21-
ptr_as_int = bx.and(ptr_as_int, bx.cx().const_i32(-(align.bytes() as i32)));
29+
ptr_as_int = round_up_to_alignment(bx, ptr_as_int, align);
2230
bx.inttoptr(ptr_as_int, ptr_ty)
2331
}
2432

@@ -270,6 +278,106 @@ fn emit_s390x_va_arg<'ll, 'tcx>(
270278
bx.load(val_type, val_addr, layout.align.abi)
271279
}
272280

281+
fn emit_xtensa_va_arg<'ll, 'tcx>(
282+
bx: &mut Builder<'_, 'll, 'tcx>,
283+
list: OperandRef<'tcx, &'ll Value>,
284+
target_ty: Ty<'tcx>,
285+
) -> &'ll Value {
286+
// Implementation of va_arg for Xtensa. There doesn't seem to be an authoritative source for
287+
// this, other than "what GCC does".
288+
//
289+
// The va_list type has three fields:
290+
// struct __va_list_tag {
291+
// int32_t *va_stk; // Arguments passed on the stack
292+
// int32_t *va_reg; // Arguments passed in registers, saved to memory by the prologue.
293+
// int32_t va_ndx; // Offset into the arguments, in bytes
294+
// };
295+
//
296+
// The first 24 bytes (equivalent to 6 registers) come from va_reg, the rest from va_stk.
297+
// Thus if va_ndx is less than 24, the next va_arg *may* read from va_reg,
298+
// otherwise it must come from va_stk.
299+
//
300+
// Primitive arguments are never split between registers and the stack. For example, if loading an 8 byte
301+
// primitive value and va_ndx = 20, we instead bump the offset and read everything from va_stk.
302+
let va_list_addr = list.immediate();
303+
// FIXME: handle multi-field structs that split across regsave/stack?
304+
let layout = bx.cx.layout_of(target_ty);
305+
let from_stack = bx.append_sibling_block("va_arg.from_stack");
306+
let from_regsave = bx.append_sibling_block("va_arg.from_regsave");
307+
let end = bx.append_sibling_block("va_arg.end");
308+
309+
// (*va).va_ndx
310+
let va_reg_offset = 4;
311+
let va_ndx_offset = va_reg_offset + 4;
312+
let offset_ptr =
313+
bx.inbounds_gep(bx.type_i8(), va_list_addr, &[bx.cx.const_usize(va_ndx_offset)]);
314+
315+
let offset = bx.load(bx.type_i32(), offset_ptr, bx.tcx().data_layout.i32_align.abi);
316+
let offset = round_up_to_alignment(bx, offset, layout.align.abi);
317+
318+
let slot_size = layout.size.align_to(Align::from_bytes(4).unwrap()).bytes() as i32;
319+
320+
// Update the offset in va_list, by adding the slot's size.
321+
let offset_next = bx.add(offset, bx.const_i32(slot_size));
322+
323+
// Figure out where to look for our value. We do that by checking the end of our slot (offset_next).
324+
// If that is within the regsave area, then load from there. Otherwise load from the stack area.
325+
let regsave_size = bx.const_i32(24);
326+
let use_regsave = bx.icmp(IntPredicate::IntULE, offset_next, regsave_size);
327+
bx.cond_br(use_regsave, from_regsave, from_stack);
328+
329+
bx.switch_to_block(from_regsave);
330+
// update va_ndx
331+
bx.store(offset_next, offset_ptr, bx.tcx().data_layout.pointer_align.abi);
332+
333+
// (*va).va_reg
334+
let regsave_area_ptr =
335+
bx.inbounds_gep(bx.type_i8(), va_list_addr, &[bx.cx.const_usize(va_reg_offset)]);
336+
let regsave_area =
337+
bx.load(bx.type_ptr(), regsave_area_ptr, bx.tcx().data_layout.pointer_align.abi);
338+
let regsave_value_ptr = bx.inbounds_gep(bx.type_i8(), regsave_area, &[offset]);
339+
bx.br(end);
340+
341+
bx.switch_to_block(from_stack);
342+
343+
// The first time we switch from regsave to stack we needs to adjust our offsets a bit.
344+
// va_stk is set up such that the first stack argument is always at va_stk + 32.
345+
// The corrected offset is written back into the va_list struct.
346+
347+
// let offset_corrected = cmp::max(offset, 32);
348+
let stack_offset_start = bx.const_i32(32);
349+
let needs_correction = bx.icmp(IntPredicate::IntULE, offset, stack_offset_start);
350+
let offset_corrected = bx.select(needs_correction, stack_offset_start, offset);
351+
352+
// let offset_next_corrected = offset_corrected + slot_size;
353+
// va_ndx = offset_next_corrected;
354+
let offset_next_corrected = bx.add(offset_next, bx.const_i32(slot_size));
355+
// update va_ndx
356+
bx.store(offset_next_corrected, offset_ptr, bx.tcx().data_layout.pointer_align.abi);
357+
358+
// let stack_value_ptr = unsafe { (*va).va_stk.byte_add(offset_corrected) };
359+
let stack_area_ptr = bx.inbounds_gep(bx.type_i8(), va_list_addr, &[bx.cx.const_usize(0)]);
360+
let stack_area = bx.load(bx.type_ptr(), stack_area_ptr, bx.tcx().data_layout.pointer_align.abi);
361+
let stack_value_ptr = bx.inbounds_gep(bx.type_i8(), stack_area, &[offset_corrected]);
362+
bx.br(end);
363+
364+
bx.switch_to_block(end);
365+
366+
// On big-endian, for values smaller than the slot size we'd have to align the read to the end
367+
// of the slot rather than the start. While the ISA and GCC support big-endian, all the Xtensa
368+
// targets supported by rustc are litte-endian so don't worry about it.
369+
370+
// if from_regsave {
371+
// unsafe { *regsave_value_ptr }
372+
// } else {
373+
// unsafe { *stack_value_ptr }
374+
// }
375+
assert!(bx.tcx().sess.target.endian == Endian::Little);
376+
let value_ptr =
377+
bx.phi(bx.type_ptr(), &[regsave_value_ptr, stack_value_ptr], &[from_regsave, from_stack]);
378+
return bx.load(layout.llvm_type(bx), value_ptr, layout.align.abi);
379+
}
380+
273381
pub(super) fn emit_va_arg<'ll, 'tcx>(
274382
bx: &mut Builder<'_, 'll, 'tcx>,
275383
addr: OperandRef<'tcx, &'ll Value>,
@@ -302,6 +410,7 @@ pub(super) fn emit_va_arg<'ll, 'tcx>(
302410
let indirect: bool = target_ty_size > 8 || !target_ty_size.is_power_of_two();
303411
emit_ptr_va_arg(bx, addr, target_ty, indirect, Align::from_bytes(8).unwrap(), false)
304412
}
413+
"xtensa" => emit_xtensa_va_arg(bx, addr, target_ty),
305414
// For all other architecture/OS combinations fall back to using
306415
// the LLVM va_arg instruction.
307416
// https://llvm.org/docs/LangRef.html#va-arg-instruction

library/core/Cargo.toml

+1
Original file line numberDiff line numberDiff line change
@@ -43,6 +43,7 @@ check-cfg = [
4343
'cfg(bootstrap)',
4444
'cfg(no_fp_fmt_parse)',
4545
'cfg(stdarch_intel_sde)',
46+
'cfg(target_arch, values("xtensa"))',
4647
# core use #[path] imports to portable-simd `core_simd` crate
4748
# and to stdarch `core_arch` crate which messes-up with Cargo list
4849
# of declared features, we therefor expect any feature cfg

library/core/src/ffi/va_list.rs

+19
Original file line numberDiff line numberDiff line change
@@ -15,6 +15,7 @@ use crate::ops::{Deref, DerefMut};
1515
not(target_arch = "aarch64"),
1616
not(target_arch = "powerpc"),
1717
not(target_arch = "s390x"),
18+
not(target_arch = "xtensa"),
1819
not(target_arch = "x86_64")
1920
),
2021
all(target_arch = "aarch64", target_vendor = "apple"),
@@ -37,6 +38,7 @@ pub struct VaListImpl<'f> {
3738
not(target_arch = "aarch64"),
3839
not(target_arch = "powerpc"),
3940
not(target_arch = "s390x"),
41+
not(target_arch = "xtensa"),
4042
not(target_arch = "x86_64")
4143
),
4244
all(target_arch = "aarch64", target_vendor = "apple"),
@@ -113,6 +115,18 @@ pub struct VaListImpl<'f> {
113115
_marker: PhantomData<&'f mut &'f c_void>,
114116
}
115117

118+
/// Xtensa ABI implementation of a `va_list`.
119+
#[cfg(target_arch = "xtensa")]
120+
#[repr(C)]
121+
#[derive(Debug)]
122+
#[lang = "va_list"]
123+
pub struct VaListImpl<'f> {
124+
stk: *mut i32,
125+
reg: *mut i32,
126+
ndx: i32,
127+
_marker: PhantomData<&'f mut &'f c_void>,
128+
}
129+
116130
/// A wrapper for a `va_list`
117131
#[repr(transparent)]
118132
#[derive(Debug)]
@@ -124,6 +138,7 @@ pub struct VaList<'a, 'f: 'a> {
124138
not(target_arch = "s390x"),
125139
not(target_arch = "x86_64")
126140
),
141+
target_arch = "xtensa",
127142
all(target_arch = "aarch64", target_vendor = "apple"),
128143
target_family = "wasm",
129144
target_os = "uefi",
@@ -138,6 +153,7 @@ pub struct VaList<'a, 'f: 'a> {
138153
target_arch = "s390x",
139154
target_arch = "x86_64"
140155
),
156+
not(target_arch = "xtensa"),
141157
any(not(target_arch = "aarch64"), not(target_vendor = "apple")),
142158
not(target_family = "wasm"),
143159
not(target_os = "uefi"),
@@ -155,6 +171,7 @@ pub struct VaList<'a, 'f: 'a> {
155171
not(target_arch = "s390x"),
156172
not(target_arch = "x86_64")
157173
),
174+
target_arch = "xtensa",
158175
all(target_arch = "aarch64", target_vendor = "apple"),
159176
target_family = "wasm",
160177
target_os = "uefi",
@@ -173,8 +190,10 @@ impl<'f> VaListImpl<'f> {
173190
target_arch = "aarch64",
174191
target_arch = "powerpc",
175192
target_arch = "s390x",
193+
target_arch = "xtensa",
176194
target_arch = "x86_64"
177195
),
196+
not(target_arch = "xtensa"),
178197
any(not(target_arch = "aarch64"), not(target_vendor = "apple")),
179198
not(target_family = "wasm"),
180199
not(target_os = "uefi"),

0 commit comments

Comments
 (0)