auto merge of rust-lang#19294 : huonw/rust/transmute-inplace, r=nikom…

…atsakis

This detects (a subset of) the cases when `transmute::<T, U>(x)` can be
lowered to a direct `bitcast T x to U` in LLVM. This assists with
efficiently handling a SIMD vector as multiple different types,
e.g. swapping bytes/words/double words around inside some larger vector
type.

C compilers like GCC and Clang handle integer vector types as `__m128i`
for all widths, and implicitly insert bitcasts as required. This patch
allows Rust to express this, even if it takes a bit of `unsafe`, whereas
previously it was impossible to do at all without inline assembly.

Example:

    pub fn reverse_u32s(u: u64x2) -> u64x2 {
        unsafe {
            let tmp = mem::transmute::<_, u32x4>(u);
            let swapped = u32x4(tmp.3, tmp.2, tmp.1, tmp.0);
            mem::transmute::<_, u64x2>(swapped)
        }
    }

Compiling with `--opt-level=3` gives:

Before

    define <2 x i64> @_ZN12reverse_u32s20hbdb206aba18a03d8tbaE(<2 x i64>) unnamed_addr #0 {
    entry-block:
      %1 = bitcast <2 x i64> %0 to i128
      %u.0.extract.trunc = trunc i128 %1 to i32
      %u.4.extract.shift = lshr i128 %1, 32
      %u.4.extract.trunc = trunc i128 %u.4.extract.shift to i32
      %u.8.extract.shift = lshr i128 %1, 64
      %u.8.extract.trunc = trunc i128 %u.8.extract.shift to i32
      %u.12.extract.shift = lshr i128 %1, 96
      %u.12.extract.trunc = trunc i128 %u.12.extract.shift to i32
      %2 = insertelement <4 x i32> undef, i32 %u.12.extract.trunc, i64 0
      %3 = insertelement <4 x i32> %2, i32 %u.8.extract.trunc, i64 1
      %4 = insertelement <4 x i32> %3, i32 %u.4.extract.trunc, i64 2
      %5 = insertelement <4 x i32> %4, i32 %u.0.extract.trunc, i64 3
      %6 = bitcast <4 x i32> %5 to <2 x i64>
      ret <2 x i64> %6
    }

    _ZN12reverse_u32s20hbdb206aba18a03d8tbaE:
    	.cfi_startproc
    	movd	%xmm0, %rax
    	punpckhqdq	%xmm0, %xmm0
    	movd	%xmm0, %rcx
    	movq	%rcx, %rdx
    	shrq	$32, %rdx
    	movq	%rax, %rsi
    	shrq	$32, %rsi
    	movd	%eax, %xmm0
    	movd	%ecx, %xmm1
    	punpckldq	%xmm0, %xmm1
    	movd	%esi, %xmm2
    	movd	%edx, %xmm0
    	punpckldq	%xmm2, %xmm0
    	punpckldq	%xmm1, %xmm0
    	retq

After

    define <2 x i64> @_ZN12reverse_u32s20hbdb206aba18a03d8tbaE(<2 x i64>) unnamed_addr #0 {
    entry-block:
      %1 = bitcast <2 x i64> %0 to <4 x i32>
      %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
      %3 = bitcast <4 x i32> %2 to <2 x i64>
      ret <2 x i64> %3
    }

    _ZN12reverse_u32s20hbdb206aba18a03d8tbaE:
    	.cfi_startproc
    	pshufd	$27, %xmm0, %xmm0
    	retq

src/librustc_trans/trans/intrinsic.rs

@@ -11,7 +11,7 @@
 #![allow(non_upper_case_globals)]
 
 use llvm;
-use llvm::{SequentiallyConsistent, Acquire, Release, AtomicXchg, ValueRef};
+use llvm::{SequentiallyConsistent, Acquire, Release, AtomicXchg, ValueRef, TypeKind};
 use middle::subst;
 use middle::subst::FnSpace;
 use trans::base::*;
@@ -174,12 +174,65 @@ pub fn trans_intrinsic_call<'a, 'blk, 'tcx>(mut bcx: Block<'blk, 'tcx>,
                 // This should be caught by the intrinsicck pass
                 assert_eq!(in_type_size, out_type_size);
 
-                // We need to cast the dest so the types work out
-                let dest = match dest {
-                    expr::SaveIn(d) => expr::SaveIn(PointerCast(bcx, d, llintype.ptr_to())),
-                    expr::Ignore => expr::Ignore
+                let nonpointer_nonaggregate = |llkind: TypeKind| -> bool {
+                    use llvm::TypeKind::*;
+                    match llkind {
+                        Half | Float | Double | X86_FP80 | FP128 |
+                            PPC_FP128 | Integer | Vector | X86_MMX => true,
+                        _ => false
+                    }
+                };
+
+                // An approximation to which types can be directly cast via
+                // LLVM's bitcast.  This doesn't cover pointer -> pointer casts,
+                // but does, importantly, cover SIMD types.
+                let in_kind = llintype.kind();
+                let ret_kind = llret_ty.kind();
+                let bitcast_compatible =
+                    (nonpointer_nonaggregate(in_kind) && nonpointer_nonaggregate(ret_kind)) || {
+                        in_kind == TypeKind::Pointer && ret_kind == TypeKind::Pointer
+                    };
+
+                let dest = if bitcast_compatible {
+                    // if we're here, the type is scalar-like (a primitive, a
+                    // SIMD type or a pointer), and so can be handled as a
+                    // by-value ValueRef and can also be directly bitcast to the
+                    // target type.  Doing this special case makes conversions
+                    // like `u32x4` -> `u64x2` much nicer for LLVM and so more
+                    // efficient (these are done efficiently implicitly in C
+                    // with the `__m128i` type and so this means Rust doesn't
+                    // lose out there).
+                    let expr = &*arg_exprs[0];
+                    let datum = unpack_datum!(bcx, expr::trans(bcx, expr));
+                    let datum = unpack_datum!(bcx, datum.to_rvalue_datum(bcx, "transmute_temp"));
+                    let val = if datum.kind.is_by_ref() {
+                        load_ty(bcx, datum.val, datum.ty)
+                    } else {
+                        datum.val
+                    };
+
+                    let cast_val = BitCast(bcx, val, llret_ty);
+
+                    match dest {
+                        expr::SaveIn(d) => {
+                            // this often occurs in a sequence like `Store(val,
+                            // d); val2 = Load(d)`, so disappears easily.
+                            Store(bcx, cast_val, d);
+                        }
+                        expr::Ignore => {}
+                    }
+                    dest
+                } else {
+                    // The types are too complicated to do with a by-value
+                    // bitcast, so pointer cast instead. We need to cast the
+                    // dest so the types work out.
+                    let dest = match dest {
+                        expr::SaveIn(d) => expr::SaveIn(PointerCast(bcx, d, llintype.ptr_to())),
+                        expr::Ignore => expr::Ignore
+                    };
+                    bcx = expr::trans_into(bcx, &*arg_exprs[0], dest);
+                    dest
                 };
-                bcx = expr::trans_into(bcx, &*arg_exprs[0], dest);
 
                 fcx.pop_custom_cleanup_scope(cleanup_scope);