1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
use rustc_apfloat::ieee::Double;
use rustc_span::Symbol;
use rustc_target::spec::abi::Abi;

use super::{
    bin_op_simd_float_all, bin_op_simd_float_first, convert_float_to_int, packssdw, packsswb,
    packuswb, shift_simd_by_scalar, FloatBinOp, ShiftOp,
};
use crate::*;

impl<'tcx> EvalContextExt<'tcx> for crate::MiriInterpCx<'tcx> {}
pub(super) trait EvalContextExt<'tcx>: crate::MiriInterpCxExt<'tcx> {
    fn emulate_x86_sse2_intrinsic(
        &mut self,
        link_name: Symbol,
        abi: Abi,
        args: &[OpTy<'tcx>],
        dest: &MPlaceTy<'tcx>,
    ) -> InterpResult<'tcx, EmulateItemResult> {
        let this = self.eval_context_mut();
        this.expect_target_feature_for_intrinsic(link_name, "sse2")?;
        // Prefix should have already been checked.
        let unprefixed_name = link_name.as_str().strip_prefix("llvm.x86.sse2.").unwrap();

        // These intrinsics operate on 128-bit (f32x4, f64x2, i8x16, i16x8, i32x4, i64x2) SIMD
        // vectors unless stated otherwise.
        // Many intrinsic names are sufixed with "ps" (packed single), "ss" (scalar signle),
        // "pd" (packed double) or "sd" (scalar double), where single means single precision
        // floating point (f32) and double means double precision floating point (f64). "ps"
        // and "pd" means thet the operation is performed on each element of the vector, while
        // "ss" and "sd" means that the operation is performed only on the first element, copying
        // the remaining elements from the input vector (for binary operations, from the left-hand
        // side).
        // Intrinsincs sufixed with "epiX" or "epuX" operate with X-bit signed or unsigned
        // vectors.
        match unprefixed_name {
            // Used to implement the _mm_madd_epi16 function.
            // Multiplies packed signed 16-bit integers in `left` and `right`, producing
            // intermediate signed 32-bit integers. Horizontally add adjacent pairs of
            // intermediate 32-bit integers, and pack the results in `dest`.
            "pmadd.wd" => {
                let [left, right] =
                    this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?;

                let (left, left_len) = this.operand_to_simd(left)?;
                let (right, right_len) = this.operand_to_simd(right)?;
                let (dest, dest_len) = this.mplace_to_simd(dest)?;

                assert_eq!(left_len, right_len);
                assert_eq!(dest_len.strict_mul(2), left_len);

                for i in 0..dest_len {
                    let j1 = i.strict_mul(2);
                    let left1 = this.read_scalar(&this.project_index(&left, j1)?)?.to_i16()?;
                    let right1 = this.read_scalar(&this.project_index(&right, j1)?)?.to_i16()?;

                    let j2 = j1.strict_add(1);
                    let left2 = this.read_scalar(&this.project_index(&left, j2)?)?.to_i16()?;
                    let right2 = this.read_scalar(&this.project_index(&right, j2)?)?.to_i16()?;

                    let dest = this.project_index(&dest, i)?;

                    // Multiplications are i16*i16->i32, which will not overflow.
                    let mul1 = i32::from(left1).strict_mul(right1.into());
                    let mul2 = i32::from(left2).strict_mul(right2.into());
                    // However, this addition can overflow in the most extreme case
                    // (-0x8000)*(-0x8000)+(-0x8000)*(-0x8000) = 0x80000000
                    let res = mul1.wrapping_add(mul2);

                    this.write_scalar(Scalar::from_i32(res), &dest)?;
                }
            }
            // Used to implement the _mm_sad_epu8 function.
            // Computes the absolute differences of packed unsigned 8-bit integers in `a`
            // and `b`, then horizontally sum each consecutive 8 differences to produce
            // two unsigned 16-bit integers, and pack these unsigned 16-bit integers in
            // the low 16 bits of 64-bit elements returned.
            //
            // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_sad_epu8
            "psad.bw" => {
                let [left, right] =
                    this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?;

                let (left, left_len) = this.operand_to_simd(left)?;
                let (right, right_len) = this.operand_to_simd(right)?;
                let (dest, dest_len) = this.mplace_to_simd(dest)?;

                // left and right are u8x16, dest is u64x2
                assert_eq!(left_len, right_len);
                assert_eq!(left_len, 16);
                assert_eq!(dest_len, 2);

                for i in 0..dest_len {
                    let dest = this.project_index(&dest, i)?;

                    let mut res: u16 = 0;
                    let n = left_len.strict_div(dest_len);
                    for j in 0..n {
                        let op_i = j.strict_add(i.strict_mul(n));
                        let left = this.read_scalar(&this.project_index(&left, op_i)?)?.to_u8()?;
                        let right =
                            this.read_scalar(&this.project_index(&right, op_i)?)?.to_u8()?;

                        res = res.strict_add(left.abs_diff(right).into());
                    }

                    this.write_scalar(Scalar::from_u64(res.into()), &dest)?;
                }
            }
            // Used to implement the _mm_{sll,srl,sra}_epi{16,32,64} functions
            // (except _mm_sra_epi64, which is not available in SSE2).
            // Shifts N-bit packed integers in left by the amount in right.
            // Both operands are 128-bit vectors. However, right is interpreted as
            // a single 64-bit integer (remaining bits are ignored).
            // For logic shifts, when right is larger than N - 1, zero is produced.
            // For arithmetic shifts, when right is larger than N - 1, the sign bit
            // is copied to remaining bits.
            "psll.w" | "psrl.w" | "psra.w" | "psll.d" | "psrl.d" | "psra.d" | "psll.q"
            | "psrl.q" => {
                let [left, right] =
                    this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?;

                let which = match unprefixed_name {
                    "psll.w" | "psll.d" | "psll.q" => ShiftOp::Left,
                    "psrl.w" | "psrl.d" | "psrl.q" => ShiftOp::RightLogic,
                    "psra.w" | "psra.d" => ShiftOp::RightArith,
                    _ => unreachable!(),
                };

                shift_simd_by_scalar(this, left, right, which, dest)?;
            }
            // Used to implement the _mm_cvtps_epi32, _mm_cvttps_epi32, _mm_cvtpd_epi32
            // and _mm_cvttpd_epi32 functions.
            // Converts packed f32/f64 to packed i32.
            "cvtps2dq" | "cvttps2dq" | "cvtpd2dq" | "cvttpd2dq" => {
                let [op] = this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?;

                let (op_len, _) = op.layout.ty.simd_size_and_type(*this.tcx);
                let (dest_len, _) = dest.layout.ty.simd_size_and_type(*this.tcx);
                match unprefixed_name {
                    "cvtps2dq" | "cvttps2dq" => {
                        // f32x4 to i32x4 conversion
                        assert_eq!(op_len, 4);
                        assert_eq!(dest_len, op_len);
                    }
                    "cvtpd2dq" | "cvttpd2dq" => {
                        // f64x2 to i32x4 conversion
                        // the last two values are filled with zeros
                        assert_eq!(op_len, 2);
                        assert_eq!(dest_len, 4);
                    }
                    _ => unreachable!(),
                }

                let rnd = match unprefixed_name {
                    // "current SSE rounding mode", assume nearest
                    // https://www.felixcloutier.com/x86/cvtps2dq
                    // https://www.felixcloutier.com/x86/cvtpd2dq
                    "cvtps2dq" | "cvtpd2dq" => rustc_apfloat::Round::NearestTiesToEven,
                    // always truncate
                    // https://www.felixcloutier.com/x86/cvttps2dq
                    // https://www.felixcloutier.com/x86/cvttpd2dq
                    "cvttps2dq" | "cvttpd2dq" => rustc_apfloat::Round::TowardZero,
                    _ => unreachable!(),
                };

                convert_float_to_int(this, op, rnd, dest)?;
            }
            // Used to implement the _mm_packs_epi16 function.
            // Converts two 16-bit integer vectors to a single 8-bit integer
            // vector with signed saturation.
            "packsswb.128" => {
                let [left, right] =
                    this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?;

                packsswb(this, left, right, dest)?;
            }
            // Used to implement the _mm_packus_epi16 function.
            // Converts two 16-bit signed integer vectors to a single 8-bit
            // unsigned integer vector with saturation.
            "packuswb.128" => {
                let [left, right] =
                    this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?;

                packuswb(this, left, right, dest)?;
            }
            // Used to implement the _mm_packs_epi32 function.
            // Converts two 32-bit integer vectors to a single 16-bit integer
            // vector with signed saturation.
            "packssdw.128" => {
                let [left, right] =
                    this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?;

                packssdw(this, left, right, dest)?;
            }
            // Used to implement _mm_min_sd and _mm_max_sd functions.
            // Note that the semantics are a bit different from Rust simd_min
            // and simd_max intrinsics regarding handling of NaN and -0.0: Rust
            // matches the IEEE min/max operations, while x86 has different
            // semantics.
            "min.sd" | "max.sd" => {
                let [left, right] =
                    this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?;

                let which = match unprefixed_name {
                    "min.sd" => FloatBinOp::Min,
                    "max.sd" => FloatBinOp::Max,
                    _ => unreachable!(),
                };

                bin_op_simd_float_first::<Double>(this, which, left, right, dest)?;
            }
            // Used to implement _mm_min_pd and _mm_max_pd functions.
            // Note that the semantics are a bit different from Rust simd_min
            // and simd_max intrinsics regarding handling of NaN and -0.0: Rust
            // matches the IEEE min/max operations, while x86 has different
            // semantics.
            "min.pd" | "max.pd" => {
                let [left, right] =
                    this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?;

                let which = match unprefixed_name {
                    "min.pd" => FloatBinOp::Min,
                    "max.pd" => FloatBinOp::Max,
                    _ => unreachable!(),
                };

                bin_op_simd_float_all::<Double>(this, which, left, right, dest)?;
            }
            // Used to implement the _mm_cmp*_sd functions.
            // Performs a comparison operation on the first component of `left`
            // and `right`, returning 0 if false or `u64::MAX` if true. The remaining
            // components are copied from `left`.
            // _mm_cmp_sd is actually an AVX function where the operation is specified
            // by a const parameter.
            // _mm_cmp{eq,lt,le,gt,ge,neq,nlt,nle,ngt,nge,ord,unord}_sd are SSE2 functions
            // with hard-coded operations.
            "cmp.sd" => {
                let [left, right, imm] =
                    this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?;

                let which =
                    FloatBinOp::cmp_from_imm(this, this.read_scalar(imm)?.to_i8()?, link_name)?;

                bin_op_simd_float_first::<Double>(this, which, left, right, dest)?;
            }
            // Used to implement the _mm_cmp*_pd functions.
            // Performs a comparison operation on each component of `left`
            // and `right`. For each component, returns 0 if false or `u64::MAX`
            // if true.
            // _mm_cmp_pd is actually an AVX function where the operation is specified
            // by a const parameter.
            // _mm_cmp{eq,lt,le,gt,ge,neq,nlt,nle,ngt,nge,ord,unord}_pd are SSE2 functions
            // with hard-coded operations.
            "cmp.pd" => {
                let [left, right, imm] =
                    this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?;

                let which =
                    FloatBinOp::cmp_from_imm(this, this.read_scalar(imm)?.to_i8()?, link_name)?;

                bin_op_simd_float_all::<Double>(this, which, left, right, dest)?;
            }
            // Used to implement _mm_{,u}comi{eq,lt,le,gt,ge,neq}_sd functions.
            // Compares the first component of `left` and `right` and returns
            // a scalar value (0 or 1).
            "comieq.sd" | "comilt.sd" | "comile.sd" | "comigt.sd" | "comige.sd" | "comineq.sd"
            | "ucomieq.sd" | "ucomilt.sd" | "ucomile.sd" | "ucomigt.sd" | "ucomige.sd"
            | "ucomineq.sd" => {
                let [left, right] =
                    this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?;

                let (left, left_len) = this.operand_to_simd(left)?;
                let (right, right_len) = this.operand_to_simd(right)?;

                assert_eq!(left_len, right_len);

                let left = this.read_scalar(&this.project_index(&left, 0)?)?.to_f64()?;
                let right = this.read_scalar(&this.project_index(&right, 0)?)?.to_f64()?;
                // The difference between the com* and ucom* variants is signaling
                // of exceptions when either argument is a quiet NaN. We do not
                // support accessing the SSE status register from miri (or from Rust,
                // for that matter), so we treat both variants equally.
                let res = match unprefixed_name {
                    "comieq.sd" | "ucomieq.sd" => left == right,
                    "comilt.sd" | "ucomilt.sd" => left < right,
                    "comile.sd" | "ucomile.sd" => left <= right,
                    "comigt.sd" | "ucomigt.sd" => left > right,
                    "comige.sd" | "ucomige.sd" => left >= right,
                    "comineq.sd" | "ucomineq.sd" => left != right,
                    _ => unreachable!(),
                };
                this.write_scalar(Scalar::from_i32(i32::from(res)), dest)?;
            }
            // Use to implement the _mm_cvtsd_si32, _mm_cvttsd_si32,
            // _mm_cvtsd_si64 and _mm_cvttsd_si64 functions.
            // Converts the first component of `op` from f64 to i32/i64.
            "cvtsd2si" | "cvttsd2si" | "cvtsd2si64" | "cvttsd2si64" => {
                let [op] = this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?;
                let (op, _) = this.operand_to_simd(op)?;

                let op = this.read_immediate(&this.project_index(&op, 0)?)?;

                let rnd = match unprefixed_name {
                    // "current SSE rounding mode", assume nearest
                    // https://www.felixcloutier.com/x86/cvtsd2si
                    "cvtsd2si" | "cvtsd2si64" => rustc_apfloat::Round::NearestTiesToEven,
                    // always truncate
                    // https://www.felixcloutier.com/x86/cvttsd2si
                    "cvttsd2si" | "cvttsd2si64" => rustc_apfloat::Round::TowardZero,
                    _ => unreachable!(),
                };

                let res = this.float_to_int_checked(&op, dest.layout, rnd)?.unwrap_or_else(|| {
                    // Fallback to minimum according to SSE semantics.
                    ImmTy::from_int(dest.layout.size.signed_int_min(), dest.layout)
                });

                this.write_immediate(*res, dest)?;
            }
            // Used to implement the _mm_cvtsd_ss and _mm_cvtss_sd functions.
            // Converts the first f64/f32 from `right` to f32/f64 and copies
            // the remaining elements from `left`
            "cvtsd2ss" | "cvtss2sd" => {
                let [left, right] =
                    this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?;

                let (left, left_len) = this.operand_to_simd(left)?;
                let (right, _) = this.operand_to_simd(right)?;
                let (dest, dest_len) = this.mplace_to_simd(dest)?;

                assert_eq!(dest_len, left_len);

                // Convert first element of `right`
                let right0 = this.read_immediate(&this.project_index(&right, 0)?)?;
                let dest0 = this.project_index(&dest, 0)?;
                // `float_to_float_or_int` here will convert from f64 to f32 (cvtsd2ss) or
                // from f32 to f64 (cvtss2sd).
                let res0 = this.float_to_float_or_int(&right0, dest0.layout)?;
                this.write_immediate(*res0, &dest0)?;

                // Copy remaining from `left`
                for i in 1..dest_len {
                    this.copy_op(&this.project_index(&left, i)?, &this.project_index(&dest, i)?)?;
                }
            }
            _ => return Ok(EmulateItemResult::NotSupported),
        }
        Ok(EmulateItemResult::NeedsReturn)
    }
}