]>
Commit | Line | Data |
---|---|---|
60c5eb7d XL |
1 | use super::operand::{OperandRef, OperandValue}; |
2 | use super::place::PlaceRef; | |
dfeec247 | 3 | use super::{FunctionCx, LocalRef}; |
60c5eb7d XL |
4 | |
5 | use crate::base; | |
dfeec247 | 6 | use crate::common::{self, IntPredicate, RealPredicate}; |
60c5eb7d | 7 | use crate::traits::*; |
dfeec247 | 8 | use crate::MemFlags; |
60c5eb7d | 9 | |
dfeec247 | 10 | use rustc_apfloat::{ieee, Float, Round, Status}; |
3dfed10e | 11 | use rustc_hir::lang_items::LangItem; |
ba9703b0 XL |
12 | use rustc_middle::mir; |
13 | use rustc_middle::ty::cast::{CastTy, IntTy}; | |
3dfed10e | 14 | use rustc_middle::ty::layout::{HasTyCtxt, TyAndLayout}; |
ba9703b0 | 15 | use rustc_middle::ty::{self, adjustment::PointerCast, Instance, Ty, TyCtxt}; |
dfeec247 XL |
16 | use rustc_span::source_map::{Span, DUMMY_SP}; |
17 | use rustc_span::symbol::sym; | |
ba9703b0 | 18 | use rustc_target::abi::{Abi, Int, LayoutOf, Variants}; |
92a42be0 | 19 | |
dc9dc135 | 20 | impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> { |
a1dfa0c6 XL |
21 | pub fn codegen_rvalue( |
22 | &mut self, | |
23 | mut bx: Bx, | |
24 | dest: PlaceRef<'tcx, Bx::Value>, | |
dfeec247 | 25 | rvalue: &mir::Rvalue<'tcx>, |
a1dfa0c6 | 26 | ) -> Bx { |
dfeec247 | 27 | debug!("codegen_rvalue(dest.llval={:?}, rvalue={:?})", dest.llval, rvalue); |
92a42be0 SL |
28 | |
29 | match *rvalue { | |
dfeec247 XL |
30 | mir::Rvalue::Use(ref operand) => { |
31 | let cg_operand = self.codegen_operand(&mut bx, operand); | |
32 | // FIXME: consider not copying constants through stack. (Fixable by codegen'ing | |
33 | // constants into `OperandValue::Ref`; why don’t we do that yet if we don’t?) | |
34 | cg_operand.val.store(&mut bx, dest); | |
35 | bx | |
36 | } | |
92a42be0 | 37 | |
48663c56 | 38 | mir::Rvalue::Cast(mir::CastKind::Pointer(PointerCast::Unsize), ref source, _) => { |
ff7c6d11 XL |
39 | // The destination necessarily contains a fat pointer, so if |
40 | // it's a scalar pair, it's a fat pointer or newtype thereof. | |
a1dfa0c6 | 41 | if bx.cx().is_backend_scalar_pair(dest.layout) { |
60c5eb7d | 42 | // Into-coerce of a thin pointer to a fat pointer -- just |
92a42be0 | 43 | // use the operand path. |
a1dfa0c6 XL |
44 | let (mut bx, temp) = self.codegen_rvalue_operand(bx, rvalue); |
45 | temp.val.store(&mut bx, dest); | |
2c00a5a8 | 46 | return bx; |
92a42be0 SL |
47 | } |
48 | ||
49 | // Unsize of a nontrivial struct. I would prefer for | |
94b46f34 | 50 | // this to be eliminated by MIR building, but |
92a42be0 SL |
51 | // `CoerceUnsized` can be passed by a where-clause, |
52 | // so the (generic) MIR may not be able to expand it. | |
a1dfa0c6 | 53 | let operand = self.codegen_operand(&mut bx, source); |
ff7c6d11 | 54 | match operand.val { |
dfeec247 | 55 | OperandValue::Pair(..) | OperandValue::Immediate(_) => { |
60c5eb7d | 56 | // Unsize from an immediate structure. We don't |
32a655c1 SL |
57 | // really need a temporary alloca here, but |
58 | // avoiding it would require us to have | |
60c5eb7d | 59 | // `coerce_unsized_into` use `extractvalue` to |
32a655c1 SL |
60 | // index into the struct, and this case isn't |
61 | // important enough for it. | |
94b46f34 | 62 | debug!("codegen_rvalue: creating ugly alloca"); |
e1599b0c | 63 | let scratch = PlaceRef::alloca(&mut bx, operand.layout); |
a1dfa0c6 XL |
64 | scratch.storage_live(&mut bx); |
65 | operand.val.store(&mut bx, scratch); | |
66 | base::coerce_unsized_into(&mut bx, scratch, dest); | |
67 | scratch.storage_dead(&mut bx); | |
32a655c1 | 68 | } |
b7449926 | 69 | OperandValue::Ref(llref, None, align) => { |
e1599b0c | 70 | let source = PlaceRef::new_sized_aligned(llref, operand.layout, align); |
a1dfa0c6 | 71 | base::coerce_unsized_into(&mut bx, source, dest); |
92a42be0 | 72 | } |
b7449926 | 73 | OperandValue::Ref(_, Some(_), _) => { |
60c5eb7d | 74 | bug!("unsized coercion on an unsized rvalue"); |
b7449926 | 75 | } |
ff7c6d11 | 76 | } |
2c00a5a8 | 77 | bx |
92a42be0 SL |
78 | } |
79 | ||
ea8adc8c | 80 | mir::Rvalue::Repeat(ref elem, count) => { |
a1dfa0c6 | 81 | let cg_elem = self.codegen_operand(&mut bx, elem); |
3b2f2976 | 82 | |
ff7c6d11 XL |
83 | // Do not generate the loop for zero-sized elements or empty arrays. |
84 | if dest.layout.is_zst() { | |
2c00a5a8 | 85 | return bx; |
3b2f2976 | 86 | } |
3b2f2976 | 87 | |
94b46f34 | 88 | if let OperandValue::Immediate(v) = cg_elem.val { |
532ac7d7 XL |
89 | let zero = bx.const_usize(0); |
90 | let start = dest.project_index(&mut bx, zero).llval; | |
91 | let size = bx.const_usize(dest.layout.size.bytes()); | |
ff7c6d11 | 92 | |
3b2f2976 | 93 | // Use llvm.memset.p0i8.* to initialize all zero arrays |
e74abb32 | 94 | if bx.cx().const_to_opt_uint(v) == Some(0) { |
a1dfa0c6 XL |
95 | let fill = bx.cx().const_u8(0); |
96 | bx.memset(start, fill, size, dest.align, MemFlags::empty()); | |
2c00a5a8 | 97 | return bx; |
3b2f2976 XL |
98 | } |
99 | ||
100 | // Use llvm.memset.p0i8.* to initialize byte arrays | |
1b1a35ee | 101 | let v = bx.from_immediate(v); |
a1dfa0c6 XL |
102 | if bx.cx().val_ty(v) == bx.cx().type_i8() { |
103 | bx.memset(start, v, size, dest.align, MemFlags::empty()); | |
2c00a5a8 | 104 | return bx; |
3b2f2976 XL |
105 | } |
106 | } | |
107 | ||
ba9703b0 | 108 | let count = |
fc512014 | 109 | self.monomorphize(count).eval_usize(bx.cx().tcx(), ty::ParamEnv::reveal_all()); |
ba9703b0 | 110 | |
532ac7d7 | 111 | bx.write_operand_repeatedly(cg_elem, count, dest) |
92a42be0 SL |
112 | } |
113 | ||
9cc50fc6 | 114 | mir::Rvalue::Aggregate(ref kind, ref operands) => { |
ff7c6d11 | 115 | let (dest, active_field_index) = match **kind { |
b7449926 | 116 | mir::AggregateKind::Adt(adt_def, variant_index, _, _, active_field_index) => { |
a1dfa0c6 | 117 | dest.codegen_set_discr(&mut bx, variant_index); |
ff7c6d11 | 118 | if adt_def.is_enum() { |
a1dfa0c6 | 119 | (dest.project_downcast(&mut bx, variant_index), active_field_index) |
ff7c6d11 XL |
120 | } else { |
121 | (dest, active_field_index) | |
9cc50fc6 SL |
122 | } |
123 | } | |
dfeec247 | 124 | _ => (dest, None), |
ff7c6d11 XL |
125 | }; |
126 | for (i, operand) in operands.iter().enumerate() { | |
a1dfa0c6 | 127 | let op = self.codegen_operand(&mut bx, operand); |
ff7c6d11 XL |
128 | // Do not generate stores and GEPis for zero-sized fields. |
129 | if !op.layout.is_zst() { | |
130 | let field_index = active_field_index.unwrap_or(i); | |
a1dfa0c6 XL |
131 | let field = dest.project_field(&mut bx, field_index); |
132 | op.val.store(&mut bx, field); | |
ff7c6d11 | 133 | } |
92a42be0 | 134 | } |
2c00a5a8 | 135 | bx |
92a42be0 SL |
136 | } |
137 | ||
92a42be0 | 138 | _ => { |
416331ca | 139 | assert!(self.rvalue_creates_operand(rvalue, DUMMY_SP)); |
a1dfa0c6 XL |
140 | let (mut bx, temp) = self.codegen_rvalue_operand(bx, rvalue); |
141 | temp.val.store(&mut bx, dest); | |
2c00a5a8 | 142 | bx |
92a42be0 SL |
143 | } |
144 | } | |
145 | } | |
146 | ||
a1dfa0c6 XL |
147 | pub fn codegen_rvalue_unsized( |
148 | &mut self, | |
149 | mut bx: Bx, | |
150 | indirect_dest: PlaceRef<'tcx, Bx::Value>, | |
151 | rvalue: &mir::Rvalue<'tcx>, | |
152 | ) -> Bx { | |
dfeec247 XL |
153 | debug!( |
154 | "codegen_rvalue_unsized(indirect_dest.llval={:?}, rvalue={:?})", | |
155 | indirect_dest.llval, rvalue | |
156 | ); | |
b7449926 XL |
157 | |
158 | match *rvalue { | |
159 | mir::Rvalue::Use(ref operand) => { | |
a1dfa0c6 XL |
160 | let cg_operand = self.codegen_operand(&mut bx, operand); |
161 | cg_operand.val.store_unsized(&mut bx, indirect_dest); | |
b7449926 XL |
162 | bx |
163 | } | |
164 | ||
60c5eb7d | 165 | _ => bug!("unsized assignment other than `Rvalue::Use`"), |
b7449926 XL |
166 | } |
167 | } | |
168 | ||
a1dfa0c6 XL |
169 | pub fn codegen_rvalue_operand( |
170 | &mut self, | |
171 | mut bx: Bx, | |
dfeec247 | 172 | rvalue: &mir::Rvalue<'tcx>, |
a1dfa0c6 | 173 | ) -> (Bx, OperandRef<'tcx, Bx::Value>) { |
416331ca XL |
174 | assert!( |
175 | self.rvalue_creates_operand(rvalue, DUMMY_SP), | |
176 | "cannot codegen {:?} to operand", | |
177 | rvalue, | |
178 | ); | |
92a42be0 SL |
179 | |
180 | match *rvalue { | |
ff7c6d11 | 181 | mir::Rvalue::Cast(ref kind, ref source, mir_cast_ty) => { |
a1dfa0c6 | 182 | let operand = self.codegen_operand(&mut bx, source); |
54a0048b | 183 | debug!("cast operand is {:?}", operand); |
fc512014 | 184 | let cast = bx.cx().layout_of(self.monomorphize(mir_cast_ty)); |
92a42be0 SL |
185 | |
186 | let val = match *kind { | |
48663c56 | 187 | mir::CastKind::Pointer(PointerCast::ReifyFnPointer) => { |
1b1a35ee | 188 | match *operand.layout.ty.kind() { |
b7449926 | 189 | ty::FnDef(def_id, substs) => { |
48663c56 XL |
190 | if bx.cx().tcx().has_attr(def_id, sym::rustc_args_required_const) { |
191 | bug!("reifying a fn ptr that requires const arguments"); | |
2c00a5a8 | 192 | } |
3dfed10e XL |
193 | let instance = ty::Instance::resolve_for_fn_ptr( |
194 | bx.tcx(), | |
195 | ty::ParamEnv::reveal_all(), | |
196 | def_id, | |
197 | substs, | |
e74abb32 | 198 | ) |
3dfed10e XL |
199 | .unwrap() |
200 | .polymorphize(bx.cx().tcx()); | |
201 | OperandValue::Immediate(bx.get_fn_addr(instance)) | |
54a0048b | 202 | } |
dfeec247 | 203 | _ => bug!("{} cannot be reified to a fn ptr", operand.layout.ty), |
54a0048b SL |
204 | } |
205 | } | |
48663c56 | 206 | mir::CastKind::Pointer(PointerCast::ClosureFnPointer(_)) => { |
1b1a35ee | 207 | match *operand.layout.ty.kind() { |
b7449926 | 208 | ty::Closure(def_id, substs) => { |
dc9dc135 | 209 | let instance = Instance::resolve_closure( |
e74abb32 XL |
210 | bx.cx().tcx(), |
211 | def_id, | |
212 | substs, | |
dfeec247 | 213 | ty::ClosureKind::FnOnce, |
3dfed10e XL |
214 | ) |
215 | .polymorphize(bx.cx().tcx()); | |
e74abb32 | 216 | OperandValue::Immediate(bx.cx().get_fn_addr(instance)) |
8bb4bdeb | 217 | } |
dfeec247 | 218 | _ => bug!("{} cannot be cast to a fn ptr", operand.layout.ty), |
8bb4bdeb XL |
219 | } |
220 | } | |
48663c56 | 221 | mir::CastKind::Pointer(PointerCast::UnsafeFnPointer) => { |
60c5eb7d | 222 | // This is a no-op at the LLVM level. |
92a42be0 SL |
223 | operand.val |
224 | } | |
48663c56 | 225 | mir::CastKind::Pointer(PointerCast::Unsize) => { |
a1dfa0c6 | 226 | assert!(bx.cx().is_backend_scalar_pair(cast)); |
92a42be0 | 227 | match operand.val { |
3157f602 | 228 | OperandValue::Pair(lldata, llextra) => { |
60c5eb7d | 229 | // unsize from a fat pointer -- this is a |
92a42be0 | 230 | // "trait-object-to-supertrait" coercion, for |
60c5eb7d | 231 | // example, `&'a fmt::Debug + Send => &'a fmt::Debug`. |
ff7c6d11 XL |
232 | |
233 | // HACK(eddyb) have to bitcast pointers | |
234 | // until LLVM removes pointee types. | |
dfeec247 XL |
235 | let lldata = bx.pointercast( |
236 | lldata, | |
237 | bx.cx().scalar_pair_element_backend_type(cast, 0, true), | |
238 | ); | |
3157f602 | 239 | OperandValue::Pair(lldata, llextra) |
92a42be0 SL |
240 | } |
241 | OperandValue::Immediate(lldata) => { | |
242 | // "standard" unsize | |
dfeec247 XL |
243 | let (lldata, llextra) = base::unsize_thin_ptr( |
244 | &mut bx, | |
245 | lldata, | |
246 | operand.layout.ty, | |
247 | cast.ty, | |
248 | ); | |
3157f602 | 249 | OperandValue::Pair(lldata, llextra) |
92a42be0 | 250 | } |
32a655c1 | 251 | OperandValue::Ref(..) => { |
dfeec247 | 252 | bug!("by-ref operand {:?} in `codegen_rvalue_operand`", operand); |
92a42be0 SL |
253 | } |
254 | } | |
255 | } | |
dfeec247 XL |
256 | mir::CastKind::Pointer(PointerCast::MutToConstPointer) |
257 | | mir::CastKind::Misc | |
258 | if bx.cx().is_backend_scalar_pair(operand.layout) => | |
259 | { | |
ff7c6d11 | 260 | if let OperandValue::Pair(data_ptr, meta) = operand.val { |
a1dfa0c6 | 261 | if bx.cx().is_backend_scalar_pair(cast) { |
dfeec247 XL |
262 | let data_cast = bx.pointercast( |
263 | data_ptr, | |
264 | bx.cx().scalar_pair_element_backend_type(cast, 0, true), | |
265 | ); | |
ff7c6d11 | 266 | OperandValue::Pair(data_cast, meta) |
dfeec247 XL |
267 | } else { |
268 | // cast to thin-ptr | |
5bcae85e SL |
269 | // Cast of fat-ptr to thin-ptr is an extraction of data-ptr and |
270 | // pointer-cast of that pointer to desired pointer type. | |
a1dfa0c6 | 271 | let llcast_ty = bx.cx().immediate_backend_type(cast); |
2c00a5a8 | 272 | let llval = bx.pointercast(data_ptr, llcast_ty); |
5bcae85e SL |
273 | OperandValue::Immediate(llval) |
274 | } | |
275 | } else { | |
60c5eb7d | 276 | bug!("unexpected non-pair operand"); |
5bcae85e SL |
277 | } |
278 | } | |
ba9703b0 XL |
279 | mir::CastKind::Pointer( |
280 | PointerCast::MutToConstPointer | PointerCast::ArrayToPointer, | |
281 | ) | |
dfeec247 | 282 | | mir::CastKind::Misc => { |
a1dfa0c6 XL |
283 | assert!(bx.cx().is_backend_immediate(cast)); |
284 | let ll_t_out = bx.cx().immediate_backend_type(cast); | |
0bf4aa26 | 285 | if operand.layout.abi.is_uninhabited() { |
a1dfa0c6 | 286 | let val = OperandValue::Immediate(bx.cx().const_undef(ll_t_out)); |
dfeec247 | 287 | return (bx, OperandRef { val, layout: cast }); |
83c7162d | 288 | } |
dfeec247 XL |
289 | let r_t_in = |
290 | CastTy::from_ty(operand.layout.ty).expect("bad input type for cast"); | |
ff7c6d11 | 291 | let r_t_out = CastTy::from_ty(cast.ty).expect("bad output type for cast"); |
a1dfa0c6 | 292 | let ll_t_in = bx.cx().immediate_backend_type(operand.layout); |
83c7162d | 293 | match operand.layout.variants { |
ba9703b0 | 294 | Variants::Single { index } => { |
48663c56 XL |
295 | if let Some(discr) = |
296 | operand.layout.ty.discriminant_for_variant(bx.tcx(), index) | |
297 | { | |
ba9703b0 XL |
298 | let discr_layout = bx.cx().layout_of(discr.ty); |
299 | let discr_t = bx.cx().immediate_backend_type(discr_layout); | |
300 | let discr_val = bx.cx().const_uint_big(discr_t, discr.val); | |
301 | let discr_val = | |
302 | bx.intcast(discr_val, ll_t_out, discr.ty.is_signed()); | |
303 | ||
dfeec247 XL |
304 | return ( |
305 | bx, | |
306 | OperandRef { | |
307 | val: OperandValue::Immediate(discr_val), | |
308 | layout: cast, | |
309 | }, | |
310 | ); | |
83c7162d XL |
311 | } |
312 | } | |
ba9703b0 | 313 | Variants::Multiple { .. } => {} |
83c7162d | 314 | } |
32a655c1 | 315 | let llval = operand.immediate(); |
c30ab7b3 | 316 | |
ff7c6d11 | 317 | let mut signed = false; |
ba9703b0 XL |
318 | if let Abi::Scalar(ref scalar) = operand.layout.abi { |
319 | if let Int(_, s) = scalar.value { | |
94b46f34 | 320 | // We use `i1` for bytes that are always `0` or `1`, |
0731742a | 321 | // e.g., `#[repr(i8)] enum E { A, B }`, but we can't |
94b46f34 | 322 | // let LLVM interpret the `i1` as signed, because |
0731742a | 323 | // then `i1 1` (i.e., E::B) is effectively `i8 -1`. |
94b46f34 | 324 | signed = !scalar.is_bool() && s; |
ff7c6d11 | 325 | |
a1dfa0c6 | 326 | let er = scalar.valid_range_exclusive(bx.cx()); |
dfeec247 | 327 | if er.end != er.start |
6a06907d | 328 | && scalar.valid_range.end() >= scalar.valid_range.start() |
dfeec247 | 329 | { |
1b1a35ee | 330 | // We want `table[e as usize ± k]` to not |
ff7c6d11 | 331 | // have bound checks, and this is the most |
1b1a35ee XL |
332 | // convenient place to put the `assume`s. |
333 | if *scalar.valid_range.start() > 0 { | |
334 | let enum_value_lower_bound = bx | |
335 | .cx() | |
336 | .const_uint_big(ll_t_in, *scalar.valid_range.start()); | |
337 | let cmp_start = bx.icmp( | |
338 | IntPredicate::IntUGE, | |
339 | llval, | |
340 | enum_value_lower_bound, | |
341 | ); | |
342 | bx.assume(cmp_start); | |
343 | } | |
344 | ||
345 | let enum_value_upper_bound = | |
a1dfa0c6 | 346 | bx.cx().const_uint_big(ll_t_in, *scalar.valid_range.end()); |
1b1a35ee XL |
347 | let cmp_end = bx.icmp( |
348 | IntPredicate::IntULE, | |
349 | llval, | |
350 | enum_value_upper_bound, | |
351 | ); | |
352 | bx.assume(cmp_end); | |
ff7c6d11 XL |
353 | } |
354 | } | |
355 | } | |
9cc50fc6 SL |
356 | |
357 | let newval = match (r_t_in, r_t_out) { | |
dfeec247 | 358 | (CastTy::Int(_), CastTy::Int(_)) => bx.intcast(llval, ll_t_out, signed), |
9cc50fc6 | 359 | (CastTy::Float, CastTy::Float) => { |
a1dfa0c6 XL |
360 | let srcsz = bx.cx().float_width(ll_t_in); |
361 | let dstsz = bx.cx().float_width(ll_t_out); | |
9cc50fc6 | 362 | if dstsz > srcsz { |
2c00a5a8 | 363 | bx.fpext(llval, ll_t_out) |
9cc50fc6 | 364 | } else if srcsz > dstsz { |
2c00a5a8 | 365 | bx.fptrunc(llval, ll_t_out) |
9cc50fc6 SL |
366 | } else { |
367 | llval | |
368 | } | |
369 | } | |
dfeec247 XL |
370 | (CastTy::Int(_), CastTy::Float) => { |
371 | if signed { | |
372 | bx.sitofp(llval, ll_t_out) | |
373 | } else { | |
374 | bx.uitofp(llval, ll_t_out) | |
375 | } | |
376 | } | |
ba9703b0 | 377 | (CastTy::Ptr(_) | CastTy::FnPtr, CastTy::Ptr(_)) => { |
dfeec247 XL |
378 | bx.pointercast(llval, ll_t_out) |
379 | } | |
ba9703b0 | 380 | (CastTy::Ptr(_) | CastTy::FnPtr, CastTy::Int(_)) => { |
dfeec247 XL |
381 | bx.ptrtoint(llval, ll_t_out) |
382 | } | |
2c00a5a8 | 383 | (CastTy::Int(_), CastTy::Ptr(_)) => { |
a1dfa0c6 | 384 | let usize_llval = bx.intcast(llval, bx.cx().type_isize(), signed); |
2c00a5a8 XL |
385 | bx.inttoptr(usize_llval, ll_t_out) |
386 | } | |
dfeec247 | 387 | (CastTy::Float, CastTy::Int(IntTy::I)) => { |
3dfed10e | 388 | cast_float_to_int(&mut bx, true, llval, ll_t_in, ll_t_out, cast) |
dfeec247 XL |
389 | } |
390 | (CastTy::Float, CastTy::Int(_)) => { | |
3dfed10e | 391 | cast_float_to_int(&mut bx, false, llval, ll_t_in, ll_t_out, cast) |
dfeec247 XL |
392 | } |
393 | _ => bug!("unsupported cast: {:?} to {:?}", operand.layout.ty, cast.ty), | |
9cc50fc6 SL |
394 | }; |
395 | OperandValue::Immediate(newval) | |
396 | } | |
92a42be0 | 397 | }; |
dfeec247 | 398 | (bx, OperandRef { val, layout: cast }) |
92a42be0 SL |
399 | } |
400 | ||
ba9703b0 | 401 | mir::Rvalue::Ref(_, bk, place) => { |
dfeec247 XL |
402 | let mk_ref = move |tcx: TyCtxt<'tcx>, ty: Ty<'tcx>| { |
403 | tcx.mk_ref( | |
404 | tcx.lifetimes.re_erased, | |
405 | ty::TypeAndMut { ty, mutbl: bk.to_mutbl_lossy() }, | |
406 | ) | |
407 | }; | |
408 | self.codegen_place_to_pointer(bx, place, mk_ref) | |
409 | } | |
92a42be0 | 410 | |
ba9703b0 | 411 | mir::Rvalue::AddressOf(mutability, place) => { |
dfeec247 | 412 | let mk_ptr = move |tcx: TyCtxt<'tcx>, ty: Ty<'tcx>| { |
74b04a01 | 413 | tcx.mk_ptr(ty::TypeAndMut { ty, mutbl: mutability }) |
7453a54e | 414 | }; |
dfeec247 | 415 | self.codegen_place_to_pointer(bx, place, mk_ptr) |
92a42be0 SL |
416 | } |
417 | ||
ba9703b0 | 418 | mir::Rvalue::Len(place) => { |
a1dfa0c6 | 419 | let size = self.evaluate_array_len(&mut bx, place); |
7453a54e | 420 | let operand = OperandRef { |
3b2f2976 | 421 | val: OperandValue::Immediate(size), |
a1dfa0c6 | 422 | layout: bx.cx().layout_of(bx.tcx().types.usize), |
7453a54e | 423 | }; |
2c00a5a8 | 424 | (bx, operand) |
92a42be0 SL |
425 | } |
426 | ||
6a06907d | 427 | mir::Rvalue::BinaryOp(op, box (ref lhs, ref rhs)) => { |
a1dfa0c6 XL |
428 | let lhs = self.codegen_operand(&mut bx, lhs); |
429 | let rhs = self.codegen_operand(&mut bx, rhs); | |
ff7c6d11 | 430 | let llresult = match (lhs.val, rhs.val) { |
dfeec247 XL |
431 | ( |
432 | OperandValue::Pair(lhs_addr, lhs_extra), | |
433 | OperandValue::Pair(rhs_addr, rhs_extra), | |
434 | ) => self.codegen_fat_ptr_binop( | |
435 | &mut bx, | |
436 | op, | |
437 | lhs_addr, | |
438 | lhs_extra, | |
439 | rhs_addr, | |
440 | rhs_extra, | |
441 | lhs.layout.ty, | |
442 | ), | |
443 | ||
444 | (OperandValue::Immediate(lhs_val), OperandValue::Immediate(rhs_val)) => { | |
a1dfa0c6 | 445 | self.codegen_scalar_binop(&mut bx, op, lhs_val, rhs_val, lhs.layout.ty) |
ff7c6d11 XL |
446 | } |
447 | ||
dfeec247 | 448 | _ => bug!(), |
92a42be0 | 449 | }; |
7453a54e | 450 | let operand = OperandRef { |
92a42be0 | 451 | val: OperandValue::Immediate(llresult), |
dfeec247 | 452 | layout: bx.cx().layout_of(op.ty(bx.tcx(), lhs.layout.ty, rhs.layout.ty)), |
7453a54e | 453 | }; |
2c00a5a8 | 454 | (bx, operand) |
92a42be0 | 455 | } |
6a06907d | 456 | mir::Rvalue::CheckedBinaryOp(op, box (ref lhs, ref rhs)) => { |
a1dfa0c6 XL |
457 | let lhs = self.codegen_operand(&mut bx, lhs); |
458 | let rhs = self.codegen_operand(&mut bx, rhs); | |
dfeec247 XL |
459 | let result = self.codegen_scalar_checked_binop( |
460 | &mut bx, | |
461 | op, | |
462 | lhs.immediate(), | |
463 | rhs.immediate(), | |
464 | lhs.layout.ty, | |
465 | ); | |
2c00a5a8 | 466 | let val_ty = op.ty(bx.tcx(), lhs.layout.ty, rhs.layout.ty); |
0531ce1d | 467 | let operand_ty = bx.tcx().intern_tup(&[val_ty, bx.tcx().types.bool]); |
dfeec247 | 468 | let operand = OperandRef { val: result, layout: bx.cx().layout_of(operand_ty) }; |
3157f602 | 469 | |
2c00a5a8 | 470 | (bx, operand) |
3157f602 | 471 | } |
92a42be0 SL |
472 | |
473 | mir::Rvalue::UnaryOp(op, ref operand) => { | |
a1dfa0c6 | 474 | let operand = self.codegen_operand(&mut bx, operand); |
92a42be0 | 475 | let lloperand = operand.immediate(); |
dc9dc135 | 476 | let is_float = operand.layout.ty.is_floating_point(); |
92a42be0 | 477 | let llval = match op { |
2c00a5a8 | 478 | mir::UnOp::Not => bx.not(lloperand), |
dfeec247 XL |
479 | mir::UnOp::Neg => { |
480 | if is_float { | |
481 | bx.fneg(lloperand) | |
482 | } else { | |
483 | bx.neg(lloperand) | |
484 | } | |
92a42be0 SL |
485 | } |
486 | }; | |
dfeec247 | 487 | (bx, OperandRef { val: OperandValue::Immediate(llval), layout: operand.layout }) |
92a42be0 SL |
488 | } |
489 | ||
ff7c6d11 | 490 | mir::Rvalue::Discriminant(ref place) => { |
f9f354fc | 491 | let discr_ty = rvalue.ty(self.mir, bx.tcx()); |
fc512014 | 492 | let discr_ty = self.monomorphize(discr_ty); |
dfeec247 | 493 | let discr = self |
74b04a01 | 494 | .codegen_place(&mut bx, place.as_ref()) |
a1dfa0c6 | 495 | .codegen_get_discr(&mut bx, discr_ty); |
dfeec247 XL |
496 | ( |
497 | bx, | |
498 | OperandRef { | |
499 | val: OperandValue::Immediate(discr), | |
500 | layout: self.cx.layout_of(discr_ty), | |
501 | }, | |
502 | ) | |
8bb4bdeb XL |
503 | } |
504 | ||
7cac9316 | 505 | mir::Rvalue::NullaryOp(mir::NullOp::SizeOf, ty) => { |
fc512014 | 506 | let ty = self.monomorphize(ty); |
a1dfa0c6 XL |
507 | assert!(bx.cx().type_is_sized(ty)); |
508 | let val = bx.cx().const_usize(bx.cx().layout_of(ty).size.bytes()); | |
509 | let tcx = self.cx.tcx(); | |
dfeec247 XL |
510 | ( |
511 | bx, | |
512 | OperandRef { | |
513 | val: OperandValue::Immediate(val), | |
514 | layout: self.cx.layout_of(tcx.types.usize), | |
515 | }, | |
516 | ) | |
7cac9316 XL |
517 | } |
518 | ||
519 | mir::Rvalue::NullaryOp(mir::NullOp::Box, content_ty) => { | |
fc512014 | 520 | let content_ty = self.monomorphize(content_ty); |
a1dfa0c6 XL |
521 | let content_layout = bx.cx().layout_of(content_ty); |
522 | let llsize = bx.cx().const_usize(content_layout.size.bytes()); | |
523 | let llalign = bx.cx().const_usize(content_layout.align.abi.bytes()); | |
524 | let box_layout = bx.cx().layout_of(bx.tcx().mk_box(content_ty)); | |
525 | let llty_ptr = bx.cx().backend_type(box_layout); | |
32a655c1 SL |
526 | |
527 | // Allocate space: | |
3dfed10e | 528 | let def_id = match bx.tcx().lang_items().require(LangItem::ExchangeMalloc) { |
32a655c1 SL |
529 | Ok(id) => id, |
530 | Err(s) => { | |
a1dfa0c6 | 531 | bx.cx().sess().fatal(&format!("allocation of `{}` {}", box_layout.ty, s)); |
32a655c1 SL |
532 | } |
533 | }; | |
2c00a5a8 | 534 | let instance = ty::Instance::mono(bx.tcx(), def_id); |
e74abb32 | 535 | let r = bx.cx().get_fn_addr(instance); |
a1dfa0c6 XL |
536 | let call = bx.call(r, &[llsize, llalign], None); |
537 | let val = bx.pointercast(call, llty_ptr); | |
32a655c1 | 538 | |
dfeec247 | 539 | let operand = OperandRef { val: OperandValue::Immediate(val), layout: box_layout }; |
2c00a5a8 | 540 | (bx, operand) |
92a42be0 | 541 | } |
f9f354fc XL |
542 | mir::Rvalue::ThreadLocalRef(def_id) => { |
543 | assert!(bx.cx().tcx().is_static(def_id)); | |
544 | let static_ = bx.get_static(def_id); | |
545 | let layout = bx.layout_of(bx.cx().tcx().static_ptr_ty(def_id)); | |
546 | let operand = OperandRef::from_immediate_or_packed_pair(&mut bx, static_, layout); | |
547 | (bx, operand) | |
548 | } | |
a7813a04 | 549 | mir::Rvalue::Use(ref operand) => { |
a1dfa0c6 | 550 | let operand = self.codegen_operand(&mut bx, operand); |
2c00a5a8 | 551 | (bx, operand) |
a7813a04 | 552 | } |
dfeec247 | 553 | mir::Rvalue::Repeat(..) | mir::Rvalue::Aggregate(..) => { |
cc61c64b XL |
554 | // According to `rvalue_creates_operand`, only ZST |
555 | // aggregate rvalues are allowed to be operands. | |
f9f354fc | 556 | let ty = rvalue.ty(self.mir, self.cx.tcx()); |
dfeec247 | 557 | let operand = |
fc512014 | 558 | OperandRef::new_zst(&mut bx, self.cx.layout_of(self.monomorphize(ty))); |
532ac7d7 | 559 | (bx, operand) |
92a42be0 SL |
560 | } |
561 | } | |
562 | } | |
563 | ||
ba9703b0 | 564 | fn evaluate_array_len(&mut self, bx: &mut Bx, place: mir::Place<'tcx>) -> Bx::Value { |
3b2f2976 | 565 | // ZST are passed as operands and require special handling |
94b46f34 | 566 | // because codegen_place() panics if Local is operand. |
e74abb32 | 567 | if let Some(index) = place.as_local() { |
3b2f2976 | 568 | if let LocalRef::Operand(Some(op)) = self.locals[index] { |
1b1a35ee | 569 | if let ty::Array(_, n) = op.layout.ty.kind() { |
416331ca | 570 | let n = n.eval_usize(bx.cx().tcx(), ty::ParamEnv::reveal_all()); |
a1dfa0c6 | 571 | return bx.cx().const_usize(n); |
3b2f2976 XL |
572 | } |
573 | } | |
574 | } | |
575 | // use common size calculation for non zero-sized types | |
74b04a01 | 576 | let cg_value = self.codegen_place(bx, place.as_ref()); |
416331ca | 577 | cg_value.len(bx.cx()) |
3b2f2976 XL |
578 | } |
579 | ||
dfeec247 XL |
580 | /// Codegen an `Rvalue::AddressOf` or `Rvalue::Ref` |
581 | fn codegen_place_to_pointer( | |
582 | &mut self, | |
583 | mut bx: Bx, | |
ba9703b0 | 584 | place: mir::Place<'tcx>, |
dfeec247 XL |
585 | mk_ptr_ty: impl FnOnce(TyCtxt<'tcx>, Ty<'tcx>) -> Ty<'tcx>, |
586 | ) -> (Bx, OperandRef<'tcx, Bx::Value>) { | |
74b04a01 | 587 | let cg_place = self.codegen_place(&mut bx, place.as_ref()); |
dfeec247 XL |
588 | |
589 | let ty = cg_place.layout.ty; | |
590 | ||
591 | // Note: places are indirect, so storing the `llval` into the | |
592 | // destination effectively creates a reference. | |
593 | let val = if !bx.cx().type_has_metadata(ty) { | |
594 | OperandValue::Immediate(cg_place.llval) | |
595 | } else { | |
596 | OperandValue::Pair(cg_place.llval, cg_place.llextra.unwrap()) | |
597 | }; | |
598 | (bx, OperandRef { val, layout: self.cx.layout_of(mk_ptr_ty(self.cx.tcx(), ty)) }) | |
599 | } | |
600 | ||
b7449926 XL |
601 | pub fn codegen_scalar_binop( |
602 | &mut self, | |
a1dfa0c6 | 603 | bx: &mut Bx, |
b7449926 | 604 | op: mir::BinOp, |
a1dfa0c6 XL |
605 | lhs: Bx::Value, |
606 | rhs: Bx::Value, | |
b7449926 | 607 | input_ty: Ty<'tcx>, |
a1dfa0c6 | 608 | ) -> Bx::Value { |
dc9dc135 | 609 | let is_float = input_ty.is_floating_point(); |
92a42be0 SL |
610 | let is_signed = input_ty.is_signed(); |
611 | match op { | |
dfeec247 XL |
612 | mir::BinOp::Add => { |
613 | if is_float { | |
614 | bx.fadd(lhs, rhs) | |
615 | } else { | |
616 | bx.add(lhs, rhs) | |
617 | } | |
618 | } | |
619 | mir::BinOp::Sub => { | |
620 | if is_float { | |
621 | bx.fsub(lhs, rhs) | |
622 | } else { | |
623 | bx.sub(lhs, rhs) | |
624 | } | |
625 | } | |
626 | mir::BinOp::Mul => { | |
627 | if is_float { | |
628 | bx.fmul(lhs, rhs) | |
629 | } else { | |
630 | bx.mul(lhs, rhs) | |
631 | } | |
632 | } | |
633 | mir::BinOp::Div => { | |
634 | if is_float { | |
635 | bx.fdiv(lhs, rhs) | |
636 | } else if is_signed { | |
637 | bx.sdiv(lhs, rhs) | |
638 | } else { | |
639 | bx.udiv(lhs, rhs) | |
640 | } | |
641 | } | |
642 | mir::BinOp::Rem => { | |
643 | if is_float { | |
644 | bx.frem(lhs, rhs) | |
645 | } else if is_signed { | |
646 | bx.srem(lhs, rhs) | |
647 | } else { | |
648 | bx.urem(lhs, rhs) | |
649 | } | |
650 | } | |
2c00a5a8 XL |
651 | mir::BinOp::BitOr => bx.or(lhs, rhs), |
652 | mir::BinOp::BitAnd => bx.and(lhs, rhs), | |
653 | mir::BinOp::BitXor => bx.xor(lhs, rhs), | |
654 | mir::BinOp::Offset => bx.inbounds_gep(lhs, &[rhs]), | |
655 | mir::BinOp::Shl => common::build_unchecked_lshift(bx, lhs, rhs), | |
656 | mir::BinOp::Shr => common::build_unchecked_rshift(bx, input_ty, lhs, rhs), | |
dfeec247 XL |
657 | mir::BinOp::Ne |
658 | | mir::BinOp::Lt | |
659 | | mir::BinOp::Gt | |
660 | | mir::BinOp::Eq | |
661 | | mir::BinOp::Le | |
662 | | mir::BinOp::Ge => { | |
663 | if is_float { | |
664 | bx.fcmp(base::bin_op_to_fcmp_predicate(op.to_hir_binop()), lhs, rhs) | |
665 | } else { | |
666 | bx.icmp(base::bin_op_to_icmp_predicate(op.to_hir_binop(), is_signed), lhs, rhs) | |
667 | } | |
c30ab7b3 SL |
668 | } |
669 | } | |
670 | } | |
671 | ||
b7449926 XL |
672 | pub fn codegen_fat_ptr_binop( |
673 | &mut self, | |
a1dfa0c6 | 674 | bx: &mut Bx, |
b7449926 | 675 | op: mir::BinOp, |
a1dfa0c6 XL |
676 | lhs_addr: Bx::Value, |
677 | lhs_extra: Bx::Value, | |
678 | rhs_addr: Bx::Value, | |
679 | rhs_extra: Bx::Value, | |
b7449926 | 680 | _input_ty: Ty<'tcx>, |
a1dfa0c6 | 681 | ) -> Bx::Value { |
c30ab7b3 SL |
682 | match op { |
683 | mir::BinOp::Eq => { | |
a1dfa0c6 XL |
684 | let lhs = bx.icmp(IntPredicate::IntEQ, lhs_addr, rhs_addr); |
685 | let rhs = bx.icmp(IntPredicate::IntEQ, lhs_extra, rhs_extra); | |
686 | bx.and(lhs, rhs) | |
c30ab7b3 SL |
687 | } |
688 | mir::BinOp::Ne => { | |
a1dfa0c6 XL |
689 | let lhs = bx.icmp(IntPredicate::IntNE, lhs_addr, rhs_addr); |
690 | let rhs = bx.icmp(IntPredicate::IntNE, lhs_extra, rhs_extra); | |
691 | bx.or(lhs, rhs) | |
c30ab7b3 | 692 | } |
dfeec247 | 693 | mir::BinOp::Le | mir::BinOp::Lt | mir::BinOp::Ge | mir::BinOp::Gt => { |
c30ab7b3 SL |
694 | // a OP b ~ a.0 STRICT(OP) b.0 | (a.0 == b.0 && a.1 OP a.1) |
695 | let (op, strict_op) = match op { | |
a1dfa0c6 XL |
696 | mir::BinOp::Lt => (IntPredicate::IntULT, IntPredicate::IntULT), |
697 | mir::BinOp::Le => (IntPredicate::IntULE, IntPredicate::IntULT), | |
698 | mir::BinOp::Gt => (IntPredicate::IntUGT, IntPredicate::IntUGT), | |
699 | mir::BinOp::Ge => (IntPredicate::IntUGE, IntPredicate::IntUGT), | |
c30ab7b3 SL |
700 | _ => bug!(), |
701 | }; | |
a1dfa0c6 XL |
702 | let lhs = bx.icmp(strict_op, lhs_addr, rhs_addr); |
703 | let and_lhs = bx.icmp(IntPredicate::IntEQ, lhs_addr, rhs_addr); | |
704 | let and_rhs = bx.icmp(op, lhs_extra, rhs_extra); | |
705 | let rhs = bx.and(and_lhs, and_rhs); | |
706 | bx.or(lhs, rhs) | |
c30ab7b3 SL |
707 | } |
708 | _ => { | |
709 | bug!("unexpected fat ptr binop"); | |
92a42be0 SL |
710 | } |
711 | } | |
712 | } | |
3157f602 | 713 | |
a1dfa0c6 XL |
714 | pub fn codegen_scalar_checked_binop( |
715 | &mut self, | |
716 | bx: &mut Bx, | |
717 | op: mir::BinOp, | |
718 | lhs: Bx::Value, | |
719 | rhs: Bx::Value, | |
dfeec247 | 720 | input_ty: Ty<'tcx>, |
a1dfa0c6 | 721 | ) -> OperandValue<Bx::Value> { |
3157f602 XL |
722 | // This case can currently arise only from functions marked |
723 | // with #[rustc_inherit_overflow_checks] and inlined from | |
724 | // another crate (mostly core::num generic/#[inline] fns), | |
725 | // while the current crate doesn't use overflow checks. | |
a1dfa0c6 | 726 | if !bx.cx().check_overflow() { |
94b46f34 | 727 | let val = self.codegen_scalar_binop(bx, op, lhs, rhs, input_ty); |
a1dfa0c6 | 728 | return OperandValue::Pair(val, bx.cx().const_bool(false)); |
3157f602 XL |
729 | } |
730 | ||
3157f602 XL |
731 | let (val, of) = match op { |
732 | // These are checked using intrinsics | |
733 | mir::BinOp::Add | mir::BinOp::Sub | mir::BinOp::Mul => { | |
734 | let oop = match op { | |
735 | mir::BinOp::Add => OverflowOp::Add, | |
736 | mir::BinOp::Sub => OverflowOp::Sub, | |
737 | mir::BinOp::Mul => OverflowOp::Mul, | |
dfeec247 | 738 | _ => unreachable!(), |
3157f602 | 739 | }; |
a1dfa0c6 | 740 | bx.checked_binop(oop, input_ty, lhs, rhs) |
3157f602 XL |
741 | } |
742 | mir::BinOp::Shl | mir::BinOp::Shr => { | |
a1dfa0c6 XL |
743 | let lhs_llty = bx.cx().val_ty(lhs); |
744 | let rhs_llty = bx.cx().val_ty(rhs); | |
745 | let invert_mask = common::shift_mask_val(bx, lhs_llty, rhs_llty, true); | |
2c00a5a8 | 746 | let outer_bits = bx.and(rhs, invert_mask); |
3157f602 | 747 | |
a1dfa0c6 | 748 | let of = bx.icmp(IntPredicate::IntNE, outer_bits, bx.cx().const_null(rhs_llty)); |
94b46f34 | 749 | let val = self.codegen_scalar_binop(bx, op, lhs, rhs, input_ty); |
3157f602 XL |
750 | |
751 | (val, of) | |
752 | } | |
dfeec247 | 753 | _ => bug!("Operator `{:?}` is not a checkable operator", op), |
3157f602 XL |
754 | }; |
755 | ||
756 | OperandValue::Pair(val, of) | |
757 | } | |
a1dfa0c6 | 758 | } |
92a42be0 | 759 | |
dc9dc135 | 760 | impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> { |
416331ca | 761 | pub fn rvalue_creates_operand(&self, rvalue: &mir::Rvalue<'tcx>, span: Span) -> bool { |
cc61c64b XL |
762 | match *rvalue { |
763 | mir::Rvalue::Ref(..) | | |
dfeec247 | 764 | mir::Rvalue::AddressOf(..) | |
cc61c64b XL |
765 | mir::Rvalue::Len(..) | |
766 | mir::Rvalue::Cast(..) | // (*) | |
767 | mir::Rvalue::BinaryOp(..) | | |
768 | mir::Rvalue::CheckedBinaryOp(..) | | |
769 | mir::Rvalue::UnaryOp(..) | | |
770 | mir::Rvalue::Discriminant(..) | | |
7cac9316 | 771 | mir::Rvalue::NullaryOp(..) | |
f9f354fc | 772 | mir::Rvalue::ThreadLocalRef(_) | |
cc61c64b XL |
773 | mir::Rvalue::Use(..) => // (*) |
774 | true, | |
775 | mir::Rvalue::Repeat(..) | | |
776 | mir::Rvalue::Aggregate(..) => { | |
f9f354fc | 777 | let ty = rvalue.ty(self.mir, self.cx.tcx()); |
fc512014 | 778 | let ty = self.monomorphize(ty); |
416331ca | 779 | self.cx.spanned_layout_of(ty, span).is_zst() |
cc61c64b XL |
780 | } |
781 | } | |
92a42be0 | 782 | |
cc61c64b XL |
783 | // (*) this is only true if the type is suitable |
784 | } | |
92a42be0 | 785 | } |
3157f602 | 786 | |
dc9dc135 | 787 | fn cast_float_to_int<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>( |
a1dfa0c6 XL |
788 | bx: &mut Bx, |
789 | signed: bool, | |
790 | x: Bx::Value, | |
791 | float_ty: Bx::Type, | |
dfeec247 | 792 | int_ty: Bx::Type, |
3dfed10e | 793 | int_layout: TyAndLayout<'tcx>, |
a1dfa0c6 | 794 | ) -> Bx::Value { |
f9f354fc | 795 | if let Some(false) = bx.cx().sess().opts.debugging_opts.saturating_float_casts { |
f035d41b XL |
796 | return if signed { bx.fptosi(x, int_ty) } else { bx.fptoui(x, int_ty) }; |
797 | } | |
798 | ||
799 | let try_sat_result = if signed { bx.fptosi_sat(x, int_ty) } else { bx.fptoui_sat(x, int_ty) }; | |
800 | if let Some(try_sat_result) = try_sat_result { | |
801 | return try_sat_result; | |
abe05a73 | 802 | } |
a1dfa0c6 XL |
803 | |
804 | let int_width = bx.cx().int_width(int_ty); | |
805 | let float_width = bx.cx().float_width(float_ty); | |
abe05a73 XL |
806 | // LLVM's fpto[su]i returns undef when the input x is infinite, NaN, or does not fit into the |
807 | // destination integer type after rounding towards zero. This `undef` value can cause UB in | |
808 | // safe code (see issue #10184), so we implement a saturating conversion on top of it: | |
809 | // Semantically, the mathematical value of the input is rounded towards zero to the next | |
810 | // mathematical integer, and then the result is clamped into the range of the destination | |
811 | // integer type. Positive and negative infinity are mapped to the maximum and minimum value of | |
812 | // the destination integer type. NaN is mapped to 0. | |
813 | // | |
814 | // Define f_min and f_max as the largest and smallest (finite) floats that are exactly equal to | |
815 | // a value representable in int_ty. | |
816 | // They are exactly equal to int_ty::{MIN,MAX} if float_ty has enough significand bits. | |
817 | // Otherwise, int_ty::MAX must be rounded towards zero, as it is one less than a power of two. | |
818 | // int_ty::MIN, however, is either zero or a negative power of two and is thus exactly | |
0531ce1d | 819 | // representable. Note that this only works if float_ty's exponent range is sufficiently large. |
abe05a73 XL |
820 | // f16 or 256 bit integers would break this property. Right now the smallest float type is f32 |
821 | // with exponents ranging up to 127, which is barely enough for i128::MIN = -2^127. | |
822 | // On the other hand, f_max works even if int_ty::MAX is greater than float_ty::MAX. Because | |
823 | // we're rounding towards zero, we just get float_ty::MAX (which is always an integer). | |
824 | // This already happens today with u128::MAX = 2^128 - 1 > f32::MAX. | |
a1dfa0c6 XL |
825 | let int_max = |signed: bool, int_width: u64| -> u128 { |
826 | let shift_amount = 128 - int_width; | |
dfeec247 | 827 | if signed { i128::MAX as u128 >> shift_amount } else { u128::MAX >> shift_amount } |
a1dfa0c6 XL |
828 | }; |
829 | let int_min = |signed: bool, int_width: u64| -> i128 { | |
dfeec247 | 830 | if signed { i128::MIN >> (128 - int_width) } else { 0 } |
a1dfa0c6 XL |
831 | }; |
832 | ||
dfeec247 | 833 | let compute_clamp_bounds_single = |signed: bool, int_width: u64| -> (u128, u128) { |
a1dfa0c6 XL |
834 | let rounded_min = ieee::Single::from_i128_r(int_min(signed, int_width), Round::TowardZero); |
835 | assert_eq!(rounded_min.status, Status::OK); | |
836 | let rounded_max = ieee::Single::from_u128_r(int_max(signed, int_width), Round::TowardZero); | |
837 | assert!(rounded_max.value.is_finite()); | |
838 | (rounded_min.value.to_bits(), rounded_max.value.to_bits()) | |
839 | }; | |
dfeec247 | 840 | let compute_clamp_bounds_double = |signed: bool, int_width: u64| -> (u128, u128) { |
a1dfa0c6 XL |
841 | let rounded_min = ieee::Double::from_i128_r(int_min(signed, int_width), Round::TowardZero); |
842 | assert_eq!(rounded_min.status, Status::OK); | |
843 | let rounded_max = ieee::Double::from_u128_r(int_max(signed, int_width), Round::TowardZero); | |
844 | assert!(rounded_max.value.is_finite()); | |
845 | (rounded_min.value.to_bits(), rounded_max.value.to_bits()) | |
846 | }; | |
847 | ||
848 | let mut float_bits_to_llval = |bits| { | |
dfeec247 | 849 | let bits_llval = match float_width { |
a1dfa0c6 XL |
850 | 32 => bx.cx().const_u32(bits as u32), |
851 | 64 => bx.cx().const_u64(bits as u64), | |
abe05a73 XL |
852 | n => bug!("unsupported float width {}", n), |
853 | }; | |
a1dfa0c6 | 854 | bx.bitcast(bits_llval, float_ty) |
abe05a73 | 855 | }; |
a1dfa0c6 XL |
856 | let (f_min, f_max) = match float_width { |
857 | 32 => compute_clamp_bounds_single(signed, int_width), | |
858 | 64 => compute_clamp_bounds_double(signed, int_width), | |
abe05a73 XL |
859 | n => bug!("unsupported float width {}", n), |
860 | }; | |
861 | let f_min = float_bits_to_llval(f_min); | |
862 | let f_max = float_bits_to_llval(f_max); | |
863 | // To implement saturation, we perform the following steps: | |
864 | // | |
865 | // 1. Cast x to an integer with fpto[su]i. This may result in undef. | |
866 | // 2. Compare x to f_min and f_max, and use the comparison results to select: | |
867 | // a) int_ty::MIN if x < f_min or x is NaN | |
868 | // b) int_ty::MAX if x > f_max | |
869 | // c) the result of fpto[su]i otherwise | |
870 | // 3. If x is NaN, return 0.0, otherwise return the result of step 2. | |
871 | // | |
872 | // This avoids resulting undef because values in range [f_min, f_max] by definition fit into the | |
873 | // destination type. It creates an undef temporary, but *producing* undef is not UB. Our use of | |
874 | // undef does not introduce any non-determinism either. | |
875 | // More importantly, the above procedure correctly implements saturating conversion. | |
876 | // Proof (sketch): | |
877 | // If x is NaN, 0 is returned by definition. | |
878 | // Otherwise, x is finite or infinite and thus can be compared with f_min and f_max. | |
879 | // This yields three cases to consider: | |
880 | // (1) if x in [f_min, f_max], the result of fpto[su]i is returned, which agrees with | |
881 | // saturating conversion for inputs in that range. | |
882 | // (2) if x > f_max, then x is larger than int_ty::MAX. This holds even if f_max is rounded | |
883 | // (i.e., if f_max < int_ty::MAX) because in those cases, nextUp(f_max) is already larger | |
884 | // than int_ty::MAX. Because x is larger than int_ty::MAX, the return value of int_ty::MAX | |
885 | // is correct. | |
886 | // (3) if x < f_min, then x is smaller than int_ty::MIN. As shown earlier, f_min exactly equals | |
887 | // int_ty::MIN and therefore the return value of int_ty::MIN is correct. | |
888 | // QED. | |
889 | ||
a1dfa0c6 XL |
890 | let int_max = bx.cx().const_uint_big(int_ty, int_max(signed, int_width)); |
891 | let int_min = bx.cx().const_uint_big(int_ty, int_min(signed, int_width) as u128); | |
3dfed10e XL |
892 | let zero = bx.cx().const_uint(int_ty, 0); |
893 | ||
894 | // The codegen here differs quite a bit depending on whether our builder's | |
895 | // `fptosi` and `fptoui` instructions may trap for out-of-bounds values. If | |
896 | // they don't trap then we can start doing everything inline with a | |
897 | // `select` instruction because it's ok to execute `fptosi` and `fptoui` | |
898 | // even if we don't use the results. | |
899 | if !bx.fptosui_may_trap(x, int_ty) { | |
900 | // Step 1 ... | |
901 | let fptosui_result = if signed { bx.fptosi(x, int_ty) } else { bx.fptoui(x, int_ty) }; | |
902 | let less_or_nan = bx.fcmp(RealPredicate::RealULT, x, f_min); | |
903 | let greater = bx.fcmp(RealPredicate::RealOGT, x, f_max); | |
904 | ||
905 | // Step 2: We use two comparisons and two selects, with %s1 being the | |
906 | // result: | |
907 | // %less_or_nan = fcmp ult %x, %f_min | |
908 | // %greater = fcmp olt %x, %f_max | |
909 | // %s0 = select %less_or_nan, int_ty::MIN, %fptosi_result | |
910 | // %s1 = select %greater, int_ty::MAX, %s0 | |
911 | // Note that %less_or_nan uses an *unordered* comparison. This | |
912 | // comparison is true if the operands are not comparable (i.e., if x is | |
913 | // NaN). The unordered comparison ensures that s1 becomes int_ty::MIN if | |
914 | // x is NaN. | |
915 | // | |
916 | // Performance note: Unordered comparison can be lowered to a "flipped" | |
917 | // comparison and a negation, and the negation can be merged into the | |
918 | // select. Therefore, it not necessarily any more expensive than a | |
919 | // ordered ("normal") comparison. Whether these optimizations will be | |
920 | // performed is ultimately up to the backend, but at least x86 does | |
921 | // perform them. | |
922 | let s0 = bx.select(less_or_nan, int_min, fptosui_result); | |
923 | let s1 = bx.select(greater, int_max, s0); | |
924 | ||
925 | // Step 3: NaN replacement. | |
926 | // For unsigned types, the above step already yielded int_ty::MIN == 0 if x is NaN. | |
927 | // Therefore we only need to execute this step for signed integer types. | |
928 | if signed { | |
929 | // LLVM has no isNaN predicate, so we use (x == x) instead | |
930 | let cmp = bx.fcmp(RealPredicate::RealOEQ, x, x); | |
931 | bx.select(cmp, s1, zero) | |
932 | } else { | |
933 | s1 | |
934 | } | |
abe05a73 | 935 | } else { |
3dfed10e XL |
936 | // In this case we cannot execute `fptosi` or `fptoui` and then later |
937 | // discard the result. The builder is telling us that these instructions | |
938 | // will trap on out-of-bounds values, so we need to use basic blocks and | |
939 | // control flow to avoid executing the `fptosi` and `fptoui` | |
940 | // instructions. | |
941 | // | |
942 | // The general idea of what we're constructing here is, for f64 -> i32: | |
943 | // | |
944 | // ;; block so far... %0 is the argument | |
945 | // %result = alloca i32, align 4 | |
946 | // %inbound_lower = fcmp oge double %0, 0xC1E0000000000000 | |
947 | // %inbound_upper = fcmp ole double %0, 0x41DFFFFFFFC00000 | |
948 | // ;; match (inbound_lower, inbound_upper) { | |
949 | // ;; (true, true) => %0 can be converted without trapping | |
950 | // ;; (false, false) => %0 is a NaN | |
951 | // ;; (true, false) => %0 is too large | |
952 | // ;; (false, true) => %0 is too small | |
953 | // ;; } | |
954 | // ;; | |
955 | // ;; The (true, true) check, go to %convert if so. | |
956 | // %inbounds = and i1 %inbound_lower, %inbound_upper | |
957 | // br i1 %inbounds, label %convert, label %specialcase | |
958 | // | |
959 | // convert: | |
960 | // %cvt = call i32 @llvm.wasm.trunc.signed.i32.f64(double %0) | |
961 | // store i32 %cvt, i32* %result, align 4 | |
962 | // br label %done | |
963 | // | |
964 | // specialcase: | |
965 | // ;; Handle the cases where the number is NaN, too large or too small | |
966 | // | |
967 | // ;; Either (true, false) or (false, true) | |
968 | // %is_not_nan = or i1 %inbound_lower, %inbound_upper | |
969 | // ;; Figure out which saturated value we are interested in if not `NaN` | |
970 | // %saturated = select i1 %inbound_lower, i32 2147483647, i32 -2147483648 | |
971 | // ;; Figure out between saturated and NaN representations | |
972 | // %result_nan = select i1 %is_not_nan, i32 %saturated, i32 0 | |
973 | // store i32 %result_nan, i32* %result, align 4 | |
974 | // br label %done | |
975 | // | |
976 | // done: | |
977 | // %r = load i32, i32* %result, align 4 | |
978 | // ;; ... | |
979 | let done = bx.build_sibling_block("float_cast_done"); | |
980 | let mut convert = bx.build_sibling_block("float_cast_convert"); | |
981 | let mut specialcase = bx.build_sibling_block("float_cast_specialcase"); | |
982 | ||
983 | let result = PlaceRef::alloca(bx, int_layout); | |
984 | result.storage_live(bx); | |
985 | ||
986 | // Use control flow to figure out whether we can execute `fptosi` in a | |
987 | // basic block, or whether we go to a different basic block to implement | |
988 | // the saturating logic. | |
989 | let inbound_lower = bx.fcmp(RealPredicate::RealOGE, x, f_min); | |
990 | let inbound_upper = bx.fcmp(RealPredicate::RealOLE, x, f_max); | |
991 | let inbounds = bx.and(inbound_lower, inbound_upper); | |
992 | bx.cond_br(inbounds, convert.llbb(), specialcase.llbb()); | |
993 | ||
994 | // Translation of the `convert` basic block | |
995 | let cvt = if signed { convert.fptosi(x, int_ty) } else { convert.fptoui(x, int_ty) }; | |
996 | convert.store(cvt, result.llval, result.align); | |
997 | convert.br(done.llbb()); | |
998 | ||
999 | // Translation of the `specialcase` basic block. Note that like above | |
1000 | // we try to be a bit clever here for unsigned conversions. In those | |
1001 | // cases the `int_min` is zero so we don't need two select instructions, | |
1002 | // just one to choose whether we need `int_max` or not. If | |
1003 | // `inbound_lower` is true then we're guaranteed to not be `NaN` and | |
1004 | // since we're greater than zero we must be saturating to `int_max`. If | |
1005 | // `inbound_lower` is false then we're either NaN or less than zero, so | |
1006 | // we saturate to zero. | |
1007 | let result_nan = if signed { | |
1008 | let is_not_nan = specialcase.or(inbound_lower, inbound_upper); | |
1009 | let saturated = specialcase.select(inbound_lower, int_max, int_min); | |
1010 | specialcase.select(is_not_nan, saturated, zero) | |
1011 | } else { | |
1012 | specialcase.select(inbound_lower, int_max, int_min) | |
1013 | }; | |
1014 | specialcase.store(result_nan, result.llval, result.align); | |
1015 | specialcase.br(done.llbb()); | |
1016 | ||
1017 | // Translation of the `done` basic block, positioning ourselves to | |
1018 | // continue from that point as well. | |
1019 | *bx = done; | |
1020 | let ret = bx.load(result.llval, result.align); | |
1021 | result.storage_dead(bx); | |
1022 | ret | |
abe05a73 XL |
1023 | } |
1024 | } |