Possible improvements to DimensionMismatch errors

There are 14 hardcoded expressions like:

throw(DimensionMismatch("dimensions must match"))

in the code base. Also, as shown in

julia> x + y
ERROR: DimensionMismatch("dimensions must match")
...

this doesn't print nicely as an error (there is some ´(...)` syntax markers) and it doesn't say why the dimensions doesn't match.

---

There are also some kinda bad strings given to DimensionMismatch like:

throw(DimensionMismatch("transpose"))

in transpose.jl in LinearAlgebra.

---

In some cases, it is possible that allocating the string inside the functions for throwing the error can lead to a bunch of code bloat due to GC frames:

julia> @code_llvm debuginfo=:none promote_shape((1,), (2,))

define [1 x i64] @julia_promote_shape_15929([1 x i64] addrspace(11)* nocapture nonnull readonly dereferenceable(8), [1 x i64] addrspace(11)* nocapture nonnull readonly dereferenceable(8)) {
top:
  %gcframe = alloca %jl_value_t addrspace(10)*, i32 3
  %2 = bitcast %jl_value_t addrspace(10)** %gcframe to i8*
  call void @llvm.memset.p0i8.i32(i8* %2, i8 0, i32 24, i32 0, i1 false)
  %3 = call %jl_value_t*** inttoptr (i64 4435471600 to %jl_value_t*** ()*)() #5
  %4 = getelementptr %jl_value_t addrspace(10)*, %jl_value_t addrspace(10)** %gcframe, i32 0
  %5 = bitcast %jl_value_t addrspace(10)** %4 to i64*
  store i64 2, i64* %5
  %6 = getelementptr %jl_value_t**, %jl_value_t*** %3, i32 0
  %7 = load %jl_value_t**, %jl_value_t*** %6
  %8 = getelementptr %jl_value_t addrspace(10)*, %jl_value_t addrspace(10)** %gcframe, i32 1
  %9 = bitcast %jl_value_t addrspace(10)** %8 to %jl_value_t***
  store %jl_value_t** %7, %jl_value_t*** %9
  %10 = bitcast %jl_value_t*** %6 to %jl_value_t addrspace(10)***
  store %jl_value_t addrspace(10)** %gcframe, %jl_value_t addrspace(10)*** %10
  %11 = getelementptr [1 x i64], [1 x i64] addrspace(11)* %0, i64 0, i64 0
  %12 = getelementptr [1 x i64], [1 x i64] addrspace(11)* %1, i64 0, i64 0
  %13 = load i64, i64 addrspace(11)* %11, align 8
  %14 = load i64, i64 addrspace(11)* %12, align 8
  %15 = icmp eq i64 %13, %14
  br i1 %15, label %L9, label %L6

L6:                                               ; preds = %top
  %16 = bitcast %jl_value_t*** %3 to i8*
  %17 = call noalias nonnull %jl_value_t addrspace(10)* @jl_gc_pool_alloc(i8* %16, i32 1624, i32 16) #1
  %18 = bitcast %jl_value_t addrspace(10)* %17 to %jl_value_t addrspace(10)* addrspace(10)*
  %19 = getelementptr %jl_value_t addrspace(10)*, %jl_value_t addrspace(10)* addrspace(10)* %18, i64 -1
  store %jl_value_t addrspace(10)* addrspacecast (%jl_value_t* inttoptr (i64 4539098032 to %jl_value_t*) to %jl_value_t addrspace(10)*), %jl_value_t addrspace(10)* addrspace(10)* %19
  %20 = bitcast %jl_value_t addrspace(10)* %17 to %jl_value_t addrspace(10)* addrspace(10)*
  store %jl_value_t addrspace(10)* addrspacecast (%jl_value_t* inttoptr (i64 4554031888 to %jl_value_t*) to %jl_value_t addrspace(10)*), %jl_value_t addrspace(10)* addrspace(10)* %20, align 8
  %21 = addrspacecast %jl_value_t addrspace(10)* %17 to %jl_value_t addrspace(12)*
  %22 = getelementptr %jl_value_t addrspace(10)*, %jl_value_t addrspace(10)** %gcframe, i32 2
  store %jl_value_t addrspace(10)* %17, %jl_value_t addrspace(10)** %22
  call void @jl_throw(%jl_value_t addrspace(12)* %21)
  unreachable

L9:                                               ; preds = %top
  %23 = insertvalue [1 x i64] undef, i64 %13, 0
  %24 = getelementptr %jl_value_t addrspace(10)*, %jl_value_t addrspace(10)** %gcframe, i32 1
  %25 = load %jl_value_t addrspace(10)*, %jl_value_t addrspace(10)** %24
  %26 = getelementptr %jl_value_t**, %jl_value_t*** %3, i32 0
  %27 = bitcast %jl_value_t*** %26 to %jl_value_t addrspace(10)**
  store %jl_value_t addrspace(10)* %25, %jl_value_t addrspace(10)** %27
  ret [1 x i64] %23
}

While using something like

@noinline throw_dimensionmismatch() = throw(DimensionMismatch("dimensions must match"))

gives a nicer

julia> @code_llvm debuginfo=:none promote_shape((1,), (2,))

define [1 x i64] @julia_promote_shape_17082([1 x i64] addrspace(11)* nocapture nonnull readonly dereferenceable(8), [1 x i64] addrspace(11)* nocapture nonnull readonly dereferenceable(8)) {
top:
  %2 = getelementptr [1 x i64], [1 x i64] addrspace(11)* %0, i64 0, i64 0
  %3 = getelementptr [1 x i64], [1 x i64] addrspace(11)* %1, i64 0, i64 0
  %4 = load i64, i64 addrspace(11)* %2, align 8
  %5 = load i64, i64 addrspace(11)* %3, align 8
  %6 = icmp eq i64 %4, %5
  br i1 %6, label %L8, label %L6

L6:                                               ; preds = %top
  call void @julia_throw_dimensionmismatch_17029()
  call void @llvm.trap()
  unreachable

L8:                                               ; preds = %top
  %7 = insertvalue [1 x i64] undef, i64 %4, 0
  ret [1 x i64] %7
}

This might not be a problem in practice though but promote_shape might be called in quite tight places so maybe.