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

use rustc_apfloat::ieee::Single;
use rustc_middle::mir;
use rustc_span::Symbol;
use rustc_target::spec::abi::Abi;

use super::{
    bin_op_simd_float_all, bin_op_simd_float_first, unary_op_ps, unary_op_ss, FloatBinOp,
    FloatUnaryOp,
};
use crate::*;
use shims::foreign_items::EmulateForeignItemResult;

impl<'mir, 'tcx: 'mir> EvalContextExt<'mir, 'tcx> for crate::MiriInterpCx<'mir, 'tcx> {}
pub(super) trait EvalContextExt<'mir, 'tcx: 'mir>:
    crate::MiriInterpCxExt<'mir, 'tcx>
{
    fn emulate_x86_sse_intrinsic(
        &mut self,
        link_name: Symbol,
        abi: Abi,
        args: &[OpTy<'tcx, Provenance>],
        dest: &MPlaceTy<'tcx, Provenance>,
    ) -> InterpResult<'tcx, EmulateForeignItemResult> {
        let this = self.eval_context_mut();
        this.expect_target_feature_for_intrinsic(link_name, "sse")?;
        // Prefix should have already been checked.
        let unprefixed_name = link_name.as_str().strip_prefix("llvm.x86.sse.").unwrap();
        // All these intrinsics operate on 128-bit (f32x4) SIMD vectors unless stated otherwise.
        // Many intrinsic names are sufixed with "ps" (packed single) or "ss" (scalar single),
        // where single means single precision floating point (f32). "ps" means thet the operation
        // is performed on each element of the vector, while "ss" 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).
        match unprefixed_name {
            // Used to implement _mm_{add,sub,mul,div,min,max}_ss functions.
            // Performs the operations on the first component of `left` and
            // `right` and copies the remaining components from `left`.
            "add.ss" | "sub.ss" | "mul.ss" | "div.ss" | "min.ss" | "max.ss" => {
                let [left, right] =
                    this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?;

                let which = match unprefixed_name {
                    "add.ss" => FloatBinOp::Arith(mir::BinOp::Add),
                    "sub.ss" => FloatBinOp::Arith(mir::BinOp::Sub),
                    "mul.ss" => FloatBinOp::Arith(mir::BinOp::Mul),
                    "div.ss" => FloatBinOp::Arith(mir::BinOp::Div),
                    "min.ss" => FloatBinOp::Min,
                    "max.ss" => FloatBinOp::Max,
                    _ => unreachable!(),
                };

                bin_op_simd_float_first::<Single>(this, which, left, right, dest)?;
            }
            // Used to implement _mm_min_ps and _mm_max_ps 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.ps" | "max.ps" => {
                let [left, right] =
                    this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?;

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

                bin_op_simd_float_all::<Single>(this, which, left, right, dest)?;
            }
            // Used to implement _mm_{sqrt,rcp,rsqrt}_ss functions.
            // Performs the operations on the first component of `op` and
            // copies the remaining components from `op`.
            "sqrt.ss" | "rcp.ss" | "rsqrt.ss" => {
                let [op] = this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?;

                let which = match unprefixed_name {
                    "sqrt.ss" => FloatUnaryOp::Sqrt,
                    "rcp.ss" => FloatUnaryOp::Rcp,
                    "rsqrt.ss" => FloatUnaryOp::Rsqrt,
                    _ => unreachable!(),
                };

                unary_op_ss(this, which, op, dest)?;
            }
            // Used to implement _mm_{sqrt,rcp,rsqrt}_ps functions.
            // Performs the operations on all components of `op`.
            "sqrt.ps" | "rcp.ps" | "rsqrt.ps" => {
                let [op] = this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?;

                let which = match unprefixed_name {
                    "sqrt.ps" => FloatUnaryOp::Sqrt,
                    "rcp.ps" => FloatUnaryOp::Rcp,
                    "rsqrt.ps" => FloatUnaryOp::Rsqrt,
                    _ => unreachable!(),
                };

                unary_op_ps(this, which, op, dest)?;
            }
            // Used to implement the _mm_cmp*_ss functions.
            // Performs a comparison operation on the first component of `left`
            // and `right`, returning 0 if false or `u32::MAX` if true. The remaining
            // components are copied from `left`.
            // _mm_cmp_ss 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}_ss are SSE functions
            // with hard-coded operations.
            "cmp.ss" => {
                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::<Single>(this, which, left, right, dest)?;
            }
            // Used to implement the _mm_cmp*_ps functions.
            // Performs a comparison operation on each component of `left`
            // and `right`. For each component, returns 0 if false or u32::MAX
            // if true.
            // _mm_cmp_ps 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}_ps are SSE functions
            // with hard-coded operations.
            "cmp.ps" => {
                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::<Single>(this, which, left, right, dest)?;
            }
            // Used to implement _mm_{,u}comi{eq,lt,le,gt,ge,neq}_ss functions.
            // Compares the first component of `left` and `right` and returns
            // a scalar value (0 or 1).
            "comieq.ss" | "comilt.ss" | "comile.ss" | "comigt.ss" | "comige.ss" | "comineq.ss"
            | "ucomieq.ss" | "ucomilt.ss" | "ucomile.ss" | "ucomigt.ss" | "ucomige.ss"
            | "ucomineq.ss" => {
                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_f32()?;
                let right = this.read_scalar(&this.project_index(&right, 0)?)?.to_f32()?;
                // 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.ss" | "ucomieq.ss" => left == right,
                    "comilt.ss" | "ucomilt.ss" => left < right,
                    "comile.ss" | "ucomile.ss" => left <= right,
                    "comigt.ss" | "ucomigt.ss" => left > right,
                    "comige.ss" | "ucomige.ss" => left >= right,
                    "comineq.ss" | "ucomineq.ss" => left != right,
                    _ => unreachable!(),
                };
                this.write_scalar(Scalar::from_i32(i32::from(res)), dest)?;
            }
            // Use to implement the _mm_cvtss_si32, _mm_cvttss_si32,
            // _mm_cvtss_si64 and _mm_cvttss_si64 functions.
            // Converts the first component of `op` from f32 to i32/i64.
            "cvtss2si" | "cvttss2si" | "cvtss2si64" | "cvttss2si64" => {
                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/cvtss2si
                    "cvtss2si" | "cvtss2si64" => rustc_apfloat::Round::NearestTiesToEven,
                    // always truncate
                    // https://www.felixcloutier.com/x86/cvttss2si
                    "cvttss2si" | "cvttss2si64" => 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_cvtsi32_ss and _mm_cvtsi64_ss functions.
            // Converts `right` from i32/i64 to f32. Returns a SIMD vector with
            // the result in the first component and the remaining components
            // are copied from `left`.
            // https://www.felixcloutier.com/x86/cvtsi2ss
            "cvtsi2ss" | "cvtsi642ss" => {
                let [left, right] =
                    this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?;

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

                assert_eq!(dest_len, left_len);

                let right = this.read_immediate(right)?;
                let dest0 = this.project_index(&dest, 0)?;
                let res0 = this.int_to_int_or_float(&right, dest0.layout)?;
                this.write_immediate(*res0, &dest0)?;

                for i in 1..dest_len {
                    this.copy_op(&this.project_index(&left, i)?, &this.project_index(&dest, i)?)?;
                }
            }
            _ => return Ok(EmulateForeignItemResult::NotSupported),
        }
        Ok(EmulateForeignItemResult::NeedsJumping)
    }
}