Fix building with gcc 8 and later

Describe the bug

As mentioned in #68 (closed), gcc 8.1.1 can't build a working lisp. Verified again that gcc 9.3.1 can't either.

To Reproduce

Steps to reproduce the behavior:

1. Modify Config.x86_linux to use -O2
2. Build lisp
3. See error as mentioned in #68 (closed).

Expected behavior

gcc 8.1.1 should be able to build a working lisp.

Desktop (please complete the following information):

• OS: Linux, possible others
• Version 21d
• gcc 9.3.1

Additional context

Did some debugging and determined that everything works if gencgc.c is compiled with -O2 but the alloc function is compiled with -O1 (using #pragma optimize ("-O1")). When compiled with -O2, we see the following disassembly of alloc:

alloc:
        pushl   %ebp
        movl    %esp, %ebp
        pushl   %ebx
        subl    $532, %esp
        movl    gc_assert_level, %eax
        movl    8(%ebp), %ebx
        testl   %eax, %eax
        jle     .L2073
        testb   $7, 671089780
        jne     .L2074
        testb   $7, %bl
        jne     .L2089
.L2076:
        movl    671089180, %edx
        testl   %edx, %edx
        je      .L2090
.L2073:
        movq    bytes_allocated_sum, %xmm1
        movl    %ebx, %eax
        movd    %ebx, %xmm0
        cltd
        movd    %edx, %xmm2
        punpckldq       %xmm2, %xmm0
        paddq   %xmm1, %xmm0
        movq    %xmm0, bytes_allocated_sum

So in computing bytes_allocated_sum, gcc uses xmm registers. And this happens before we can save the FPU state. Thus xmm0-xmm2 are trashed and Lisp doesn't know about it.

Saving the state on entry and restoring it on exit seems to make this work. But maybe we should revisit option 1 mentioned in #68 (closed).