Re: [PATCH] lower-subreg: Decompose multiword shifts

[Rask Ingemann Lambertsen <rask at sygehus dot dk>]

On Thu, Aug 02, 2007 at 05:34:34PM +0200, Andreas Krebbel wrote:
> Hello,
> 
> the attached patch enhances the lower subreg pass to be able to
> decompose multiword shift instructions.
> 
> Currently GCC generates quite ugly code (on s390 31bit) for a function
> like the first of the attached testcase:
> 
> u64
> foo (u32 high, u32 low)
> {
>   return ((u64)high << 32) | low;
> }

   On x86_64 with -O2 -march=k6 -m32, I get:

foo:
	pushl	%ebp		# 39	*pushsi2	[length = 1]
	movl	%esp, %ebp	# 40	*movsi_1/1	[length = 2]
	movl	8(%ebp), %edx	# 37	*movsi_1/1	[length = 3]
	movl	12(%ebp), %eax	# 10	*movsi_1/1	[length = 3]
	leave			# 43	leave	[length = 1]
	movl	%edx, %ecx	# 46	*movsi_1/1	[length = 2]
	movl	%ecx, %edx	# 11	*movsi_1/1	[length = 2]
	ret			# 44	return_internal	[length = 1]

   This appears sort of fine, i.e. no shifts, with just two useless "movl"
instructions. But the shift remains until the split4 pass, leaving the two
"movl" instructions behind. So I tried your patch:

foo:
	pushl	%ebp		# 43	*pushsi2	[length = 1]
	movl	%esp, %ebp	# 44	*movsi_1/1	[length = 2]
	pushl	%esi		# 45	*pushsi2	[length = 1]
	pushl	%ebx		# 46	*pushsi2	[length = 1]
	movl	12(%ebp), %eax	# 39	*movsi_1/1	[length = 3]
	popl	%ebx		# 49	popsi1	[length = 1]
	movl	8(%ebp), %ecx	# 41	*movsi_1/1	[length = 3]
	popl	%esi		# 50	popsi1	[length = 1]
	leave			# 51	leave	[length = 1]
	movl	%ecx, %edx	# 36	*movsi_1/1	[length = 2]
	ret			# 52	return_internal	[length = 1]

I don't know if it's your fault, but it's worse!

-- 
Rask Ingemann Lambertsen