suboptimal GCC code generation for unsigned long long on sparc

[Philippe De Muyter]

> 
> C code:
> 
> unsigned
> afunc( unsigned long long var1, unsigned long long var2) {
>    return ((unsigned)( (var1 ^ var2) >> 48)) ;
> }
> 
> Assembler code as follows:
> 
> gcc (2.7.2.1)
> 
> Xor %o0, %o2, %o0
> Xor %o1, %o3, %o1
> Srl %o0, 16, %g3
> Mov 0, %g2
> Retl
> Mov %g3, %o0
> 
> Sun C (4.2)
> 
> Xor %o0, %o2, %g1
> Retl
> Srl %g1, 16, %o0
> 
What should I say about m68k code generation for your testcase ?
This is what I get with the current snapshot.
	pea (%fp)
	move.l %sp,%fp
	mov.l %d3,-(%sp)
	mov.l %d2,-(%sp)
	mov.l 8(%fp),%d0 ;# 28 movdi+1/2
	mov.l 12(%fp),%d1
	mov.l 16(%fp),%d2 ;# 30 movdi+1/2
	mov.l 20(%fp),%d3
	eor.l %d2,%d0 ;# 9 xordi3
	eor.l %d3,%d1
	mov.l %d0,%d1 ;# 33 movsi+1/1
	clr.w %d1 ;# 10 lshrsi_16
	swap.w %d1
	clr.l %d0 ;# 12 movsi_const0
	mov.l %d1,%d0 ;# 18 movsi+1/1
	mov.l (%sp)+,%d2
	mov.l (%sp)+,%d3
	unlk %fp
	rts

If I modify m68k.md by enhancing the lshrdi3 pattern, I still get similar code.
	pea (%fp)
	move.l %sp,%fp
	mov.l %d3,-(%sp)
	mov.l %d2,-(%sp)
	mov.l 8(%fp),%d0 ;# 25 movdi+1/2
	mov.l 12(%fp),%d1
	mov.l 16(%fp),%d2 ;# 27 movdi+1/2
	mov.l 20(%fp),%d3
	eor.l %d2,%d0 ;# 9 xordi3
	eor.l %d3,%d1
	mov.l %d0,%d1 ;# 11 lshrdi_const
	clr.w %d1
	clr.l %d0
	swap.w %d1
	mov.l %d1,%d0 ;# 15 movsi+1/1
	mov.l (%sp)+,%d2
	mov.l (%sp)+,%d3
	unlk %fp
	rts

And even if I disable all the DImode patterns, (hoping to get simplifications
removal of useless instructions by using only SImode operands), I now get
worse code with the same amount of instructions, but using more registers.
	pea (%fp)
	move.l %sp,%fp
	movm.l &0x3c00,-(%sp)
	mov.l 8(%fp),%d4 ;# 5 movsi+1/1
	mov.l 12(%fp),%d5 ;# 7 movsi+1/1
	mov.l 16(%fp),%d0 ;# 10 movsi+1/1
	mov.l 20(%fp),%d1 ;# 12 movsi+1/1
	mov.l %d4,%d2 ;# 38 movsi+1/1
	eor.l %d0,%d2 ;# 17 xorsi3_internal/1
	mov.l %d5,%d3 ;# 41 movsi+1/1
	eor.l %d1,%d3 ;# 18 xorsi3_internal/1
	mov.l %d2,%d1 ;# 44 movsi+1/1
	clr.w %d1 ;# 20 lshrsi_16
	swap.w %d1
	clr.l %d0 ;# 22 movsi_const0
	mov.l %d1,%d0 ;# 27 movsi+1/1
	movm.l (%sp)+,&0x3c
	unlk %fp
	rts

Here is what I would expect gcc to generate (this is basically the same as
what the Sun cc compiler generates, except we pass the parameters
on the stack and we handle explicitely the frame-pointer) :
	pea (%fp)
	move.l %sp,%fp
	mov.l 8(%fp),%d0 ;# 5 movsi+1/1
	mov.l 16(%fp),%d1 ;# 10 movsi+1/1
	eor.l %d1,%d0 ;# 18 xorsi3_internal/1
	clr.w %d0 ;# 20 lshrsi_16
	swap.w %d0
	unlk %fp
	rts