x86: Expanded MM block move to include YMM registers under AVX

This merge request expands the XMM block move feature to use YMM registers if they're available.

Criteria

Compile under -Cfavx -OpCOREAVX -CpCOREAVX and confirm correct compilation

Notes

Currently, the RTL can't take advantage of this extension, and there's only one example in the compiler in TRealConstNode.GetCopy, so it might provide a token speed boost - before:

.section .text.n_ncon$_$trealconstnode_$__$$_dogetcopy$$tnode,"ax"
	.balign 16,0x90
.globl	NCON$_$TREALCONSTNODE_$__$$_DOGETCOPY$$TNODE
NCON$_$TREALCONSTNODE_$__$$_DOGETCOPY$$TNODE:
.seh_proc NCON$_$TREALCONSTNODE_$__$$_DOGETCOPY$$TNODE
	pushq	%rbx
.seh_pushreg %rbx
	leaq	-32(%rsp),%rsp
.seh_stackalloc 32
.seh_endprologue
	movq	%rcx,%rbx
# Peephole Optimization: %rbx = %rcx; changed to minimise pipeline stall (MovXXX2MovXXX)
# Peephole Optimization: Mov2Nop 4 done
	call	NODE$_$TNODE_$__$$_DOGETCOPY$$TNODE
# Peephole Optimization: Used %xmm0 to merge a pair of memory moves (MovMovMovMov2MovdqMovdq 1)
	vmovdqu	144(%rbx),%xmm0
	vmovdqu	%xmm0,144(%rax)
# Peephole Optimization: Used %xmm0 to merge a pair of memory moves (MovMovMovMov2MovdqMovdq 1)
	vmovdqu	160(%rbx),%xmm0
	vmovdqu	%xmm0,160(%rax)
	nop
	leaq	32(%rsp),%rsp
	popq	%rbx
	ret
.seh_endproc

After:

.section .text.n_ncon$_$trealconstnode_$__$$_dogetcopy$$tnode,"ax"
	.balign 16,0x90
.globl	NCON$_$TREALCONSTNODE_$__$$_DOGETCOPY$$TNODE
NCON$_$TREALCONSTNODE_$__$$_DOGETCOPY$$TNODE:
.seh_proc NCON$_$TREALCONSTNODE_$__$$_DOGETCOPY$$TNODE
	pushq	%rbx
.seh_pushreg %rbx
	leaq	-32(%rsp),%rsp
.seh_stackalloc 32
.seh_endprologue
	movq	%rcx,%rbx
# Peephole Optimization: %rbx = %rcx; changed to minimise pipeline stall (MovXXX2MovXXX)
# Peephole Optimization: Mov2Nop 4 done
	call	NODE$_$TNODE_$__$$_DOGETCOPY$$TNODE
# Peephole Optimization: Used %xmm0 to merge a pair of memory moves (MovMovMovMov2MovdqMovdq 1)
# Peephole Optimization: Used %ymm0 to merge a pair of memory moves (VmovdqxVmovdqxVmovdqxVmovdqx2VmovdqyVmovdqy 1)
	vmovdqu	144(%rbx),%ymm0
	vmovdqu	%ymm0,144(%rax)
# Peephole Optimization: Used %xmm0 to merge a pair of memory moves (MovMovMovMov2MovdqMovdq 1)
	nop
	leaq	32(%rsp),%rsp
	popq	%rbx
	ret
.seh_endproc