Update to 2.1.5.

New in version 2.1.5

    * minor incompatible change: on x86-64, the backend instruction
      encoders for movzx and for string opcodes have changed their
      semantics.
    * platform support:
        * compatibility: support the latest MinGW on x86. (#1923325,
          thanks to Alexis Rivera)
        * bug fix: on x86-64, fix instruction encoding for TEST on
          RIP-relative addresses. (#1925808, reported by Shinmera on
          #sbcl, thanks also to 3b)
        * bug fix: on x86-64, loading all-1s into an AVX2 register no
          longer causes an error. (thanks to Marco Heisig)
        * bug fix: on arm64, improve disassembly of ADD with constant
          0 as MOV
        * enhancement: on arm64, support debugger commands
          RETURN-FROM-FRAME and RESTART-FRAME more efficiently.
        * enhancement: on x86-64, add support for vshuf* AVX2
          instructions. (reported by Bela Pecsek)
        * optimization: faster function calls on arm64.
        * optimization: (SETF SBIT) is faster on x86-64.
    * bug fix: INTEGER-DECODE-FLOAT was computing the wrong answer for
      denormal double floats. (#1926383, reported by Stavros Macrakis)
    * bug fix: RANDOM on a floating point argument now does not
      cons. (reported by Tito Latini)
    * bug fix: fix a compiler crash in type derivation of
      LOGTEST. (#1928243)
    * bug fix: fix a compiler failure when a declared function type
      contains a literal structure with a valid MAKE-LOAD-FORM
      method. (#1929160, thanks to Yurii Hryhorenko)
    * optimization: FBOUNDP on a constant symbol is now faster.
    * optimization: file compilation now produces smaller fasls for
      files which reference package literals.
    * optimization: derive the type of calls to FLOAT-SIGN.

(rjs)