Fix i386 Fiber.swapcontext

[bcardiff]

Regression introduced by #9829

Before #9829 the asm emitted by Fiber.swapcontext was:

"*Fiber::swapcontext<Pointer(Fiber::Context), Pointer(Fiber::Context)>:Nil":
	.cfi_startproc
	movl	8(%esp), %eax
	movl	4(%esp), %ecx
	#APP

	pushl	%edi	# push 1st argument (because of initial resume)
	pushl	%ebx	# push callee-saved registers on the stack
	pushl	%ebp
	pushl	%esi
	movl	%esp, (%ecx)	# current_context.stack_top = %esp
	movl	$1, 4(%ecx)	# current_context.resumable = 1

	movl	$0, 4(%eax)	# new_context.resumable = 0
	movl	(%eax), %esp	# %esp = new_context.stack_top
	popl	%esi	# pop callee-saved registers from the stack
	popl	%ebp
	popl	%ebx
	popl	%edi	# pop first argument (for initial resume)

	#NO_APP
	retl

i386 is the only target that has this prelude in the swapcontext. It seems I didn't notice it before but test-ecosystem in 32 bits platforms was complaining.

[asterite]

Awesome!

[straight-shoota]

Where did those movl instructions come from before #9829? They've not been in i386.cr:

crystal/src/fiber/context/i386.cr

Lines 20 to 24 in 902060d

	def self.swapcontext(current_context, new_context) : Nil
	asm("
	pushl %edi // push 1st argument (because of initial resume)
	pushl %ebx // push callee-saved registers on the stack
	pushl %ebp

Is this related to switching from LLVMConstInlineAsm to LLVMGetInlineAsm? If so, does this patch work for LLVM < 7.0?

[bcardiff]

It's because on i386 the arguments are passed entirely on the stack.

The inline asm before #9829 was getting the arguments as registers thanks to the :: "r"(current_context), "r"(new_context). In the rest of the swapcontext implementation having the argument values in registers is directly.

This PR simulates what the :: "r"(current_context), "r"(new_context) was doing on i386 accurately. The LLVM < 7.0 branch has arguments so we are fine, although argument in inline asm in naked functions are not actually supported and it was part of the fix to support llvm 11.0.