|
| 1 | +use rustc_middle::mir; |
| 2 | +use rustc_span::Symbol; |
| 3 | +use rustc_target::abi::Align; |
| 4 | +use rustc_target::spec::abi::Abi; |
| 5 | + |
| 6 | +use super::horizontal_bin_op; |
| 7 | +use crate::*; |
| 8 | +use shims::foreign_items::EmulateByNameResult; |
| 9 | + |
| 10 | +impl<'mir, 'tcx: 'mir> EvalContextExt<'mir, 'tcx> for crate::MiriInterpCx<'mir, 'tcx> {} |
| 11 | +pub(super) trait EvalContextExt<'mir, 'tcx: 'mir>: |
| 12 | + crate::MiriInterpCxExt<'mir, 'tcx> |
| 13 | +{ |
| 14 | + fn emulate_x86_sse3_intrinsic( |
| 15 | + &mut self, |
| 16 | + link_name: Symbol, |
| 17 | + abi: Abi, |
| 18 | + args: &[OpTy<'tcx, Provenance>], |
| 19 | + dest: &PlaceTy<'tcx, Provenance>, |
| 20 | + ) -> InterpResult<'tcx, EmulateByNameResult<'mir, 'tcx>> { |
| 21 | + let this = self.eval_context_mut(); |
| 22 | + // Prefix should have already been checked. |
| 23 | + let unprefixed_name = link_name.as_str().strip_prefix("llvm.x86.sse3.").unwrap(); |
| 24 | + |
| 25 | + match unprefixed_name { |
| 26 | + // Used to implement the _mm_addsub_ps and _mm_addsub_pd functions. |
| 27 | + // Alternatingly add and subtract floating point (f32 or f64) from |
| 28 | + // `left` and `right` |
| 29 | + "addsub.ps" | "addsub.pd" => { |
| 30 | + let [left, right] = |
| 31 | + this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?; |
| 32 | + |
| 33 | + let (left, left_len) = this.operand_to_simd(left)?; |
| 34 | + let (right, right_len) = this.operand_to_simd(right)?; |
| 35 | + let (dest, dest_len) = this.place_to_simd(dest)?; |
| 36 | + |
| 37 | + assert_eq!(dest_len, left_len); |
| 38 | + assert_eq!(dest_len, right_len); |
| 39 | + |
| 40 | + for i in 0..dest_len { |
| 41 | + let left = this.read_immediate(&this.project_index(&left, i)?)?; |
| 42 | + let right = this.read_immediate(&this.project_index(&right, i)?)?; |
| 43 | + let dest = this.project_index(&dest, i)?; |
| 44 | + |
| 45 | + // Even elements are subtracted and odd elements are added. |
| 46 | + let op = if i % 2 == 0 { mir::BinOp::Sub } else { mir::BinOp::Add }; |
| 47 | + let res = this.wrapping_binary_op(op, &left, &right)?; |
| 48 | + |
| 49 | + this.write_immediate(*res, &dest)?; |
| 50 | + } |
| 51 | + } |
| 52 | + // Used to implement the _mm_h{add,sub}_p{s,d} functions. |
| 53 | + // Horizontally add/subtract adjacent floating point values |
| 54 | + // in `left` and `right`. |
| 55 | + "hadd.ps" | "hadd.pd" | "hsub.ps" | "hsub.pd" => { |
| 56 | + let [left, right] = |
| 57 | + this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?; |
| 58 | + |
| 59 | + let which = match unprefixed_name { |
| 60 | + "hadd.ps" | "hadd.pd" => mir::BinOp::Add, |
| 61 | + "hsub.ps" | "hsub.pd" => mir::BinOp::Sub, |
| 62 | + _ => unreachable!(), |
| 63 | + }; |
| 64 | + |
| 65 | + horizontal_bin_op(this, which, /*saturating*/ false, left, right, dest)?; |
| 66 | + } |
| 67 | + // Used to implement the _mm_lddqu_si128 function. |
| 68 | + // Reads a 128-bit vector from an unaligned pointer. This intrinsic |
| 69 | + // is expected to perform better than a regular unaligned read when |
| 70 | + // the data crosses a cache line, but for Miri this is just a regular |
| 71 | + // unaligned read. |
| 72 | + "ldu.dq" => { |
| 73 | + let [src_ptr] = this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?; |
| 74 | + let src_ptr = this.read_pointer(src_ptr)?; |
| 75 | + let dest = dest.force_mplace(this)?; |
| 76 | + |
| 77 | + this.mem_copy( |
| 78 | + src_ptr, |
| 79 | + Align::ONE, |
| 80 | + dest.ptr(), |
| 81 | + Align::ONE, |
| 82 | + dest.layout.size, |
| 83 | + /*nonoverlapping*/ true, |
| 84 | + )?; |
| 85 | + } |
| 86 | + _ => return Ok(EmulateByNameResult::NotSupported), |
| 87 | + } |
| 88 | + Ok(EmulateByNameResult::NeedsJumping) |
| 89 | + } |
| 90 | +} |
0 commit comments