Only add begin/end region primitives for [Ctrywith] when stack allocation is enabled

[azewierzejew]

For each try with block the stack for local allocations is stored at the beginning of the block and in the beginning of the handler the stack pointer is restored.
That is because the exception can be raised at a point where there are some additional allocation compared to the stored pointer.

The current "selectgen.ml" always inserts those Ibeginregion and Iendregion but they aren't necessary if the stack allocation is disabled.

With this change the operations will only be inserted if the stack allocation is enabled.

For amd64, flambda middle end and upstream register allocator.
The code

let foo x = try bar x with _ -> x

produces following assembly on the main branch with or without stack allocation and on fixed branch with stack allocation only:

0000000000000010 <camlTest__foo_14>:
  10:	48 83 ec 18          	sub    rsp,0x18
  14:	48 89 44 24 08       	mov    QWORD PTR [rsp+0x8],rax
  19:	49 8b 9e 88 00 00 00 	mov    rbx,QWORD PTR [r14+0x88]
  20:	48 89 1c 24          	mov    QWORD PTR [rsp],rbx
  24:	4c 8d 1d 1d 00 00 00 	lea    r11,[rip+0x1d]        # 48 <camlTest__foo_14+0x38>
  2b:	41 53                	push   r11
  2d:	41 ff 76 10          	push   QWORD PTR [r14+0x10]
  31:	49 89 66 10          	mov    QWORD PTR [r14+0x10],rsp
  35:	e8 00 00 00 00       	call   3a <camlTest__foo_14+0x2a>
			36: R_X86_64_PLT32	camlTest__bar_8-0x4
  3a:	41 8f 46 10          	pop    QWORD PTR [r14+0x10]
  3e:	48 83 c4 08          	add    rsp,0x8
  42:	48 83 c4 18          	add    rsp,0x18
  46:	c3                   	ret    
  47:	90                   	nop
  48:	48 8b 04 24          	mov    rax,QWORD PTR [rsp]
  4c:	49 89 86 88 00 00 00 	mov    QWORD PTR [r14+0x88],rax
  53:	48 8b 44 24 08       	mov    rax,QWORD PTR [rsp+0x8]
  58:	48 83 c4 18          	add    rsp,0x18
  5c:	c3                   	ret    
  5d:	0f 1f 00             	nop    DWORD PTR [rax]

The fix branch with stack allocation disabled produces following assembly:

0000000000000010 <camlTest__foo_14>:
  10:	48 83 ec 08          	sub    rsp,0x8
  14:	48 89 04 24          	mov    QWORD PTR [rsp],rax
  18:	4c 8d 1d 1d 00 00 00 	lea    r11,[rip+0x1d]        # 3c <camlTest__foo_14+0x2c>
  1f:	41 53                	push   r11
  21:	41 ff 76 10          	push   QWORD PTR [r14+0x10]
  25:	49 89 66 10          	mov    QWORD PTR [r14+0x10],rsp
  29:	e8 00 00 00 00       	call   2e <camlTest__foo_14+0x1e>
			2a: R_X86_64_PLT32	camlTest__bar_8-0x4
  2e:	41 8f 46 10          	pop    QWORD PTR [r14+0x10]
  32:	48 83 c4 08          	add    rsp,0x8
  36:	48 83 c4 08          	add    rsp,0x8
  3a:	c3                   	ret    
  3b:	90                   	nop
  3c:	48 8b 04 24          	mov    rax,QWORD PTR [rsp]
  40:	48 83 c4 08          	add    rsp,0x8
  44:	c3                   	ret    
  45:	90                   	nop
  46:	66 2e 0f 1f 84 00 00 	nop    WORD PTR cs:[rax+rax*1+0x0]
  4d:	00 00 00 

The new code doesn't have moves at offsets 19 and 20 which fetch the local allocation stack pointer.
It also doesn't have moves at offsets 48 and 4cwhich restores the pointer in the handler.

Apart from that the code remains the same.

[xclerc]

I don't know that part of the backend
very well, but the change looks correct.

[gretay-js]

@azewierzejew can you make the same change in ocaml/asmcomp directory please?

[mshinwell]

#809 contains the analogous logic (and more) for Flambda 2.