On Tue, Aug 3, 2010 at 6:11 PM, Bingfeng Mei <b...@broadcom.com> wrote:
> Richard,
> I applied the patch. The simple example in my previous mail is
> compiled as expected. However, for a bit more complex example,
> restrict qualifier still doesn't work as expected. This happens
> even on trunk compiler so it is not due to some missing patches on 4.5.
>
> void foo (int * restrict a, int * restrict b, int * restrict c)
> {
> int i;
> for(i = 0; i < 100; i+=4)
> {
> a[i] = b[i] * c[i];
> a[i+1] = b[i+1] * c[i+1];
> a[i+2] = b[i+2] * c[i+2];
> a[i+3] = b[i+3] * c[i+3];
> }
> }
>
> ~/work/install-x86/bin/gcc tst3.c -O2 -S -std=c99 -da -fschedule-insns
> -frename-register
>
> .L2:
> movl (%rdx,%rax), %r10d
> imull (%rsi,%rax), %r10d
> movl %r10d, (%rdi,%rax)
> movl 4(%rdx,%rax), %r9d
> imull 4(%rsi,%rax), %r9d
> movl %r9d, 4(%rdi,%rax)
> movl 8(%rdx,%rax), %r8d
> imull 8(%rsi,%rax), %r8d
> movl %r8d, 8(%rdi,%rax)
> movl 12(%rdx,%rax), %ecx
> imull 12(%rsi,%rax), %ecx
> movl %ecx, 12(%rdi,%rax)
> addq $16, %rax
> cmpq $400, %rax
> jne .L2
> rep
>
> This used to compile efficiently on our 4.4 port. Any comments?
It's due to TMR_ORIGINAL being used for MEM_EXPRs during
expansion (and TMRs not being handled by the alias oracles
well). I can look at this if you file a bugreport, so I remember.
A patch as simple as
Index: expr.c
===================================================================
--- expr.c (revision 162841)
+++ expr.c (working copy)
@@ -8665,7 +8665,7 @@ expand_expr_real_1 (tree exp, rtx target
set_mem_addr_space (temp, as);
base = get_base_address (TMR_ORIGINAL (exp));
if (base
- && INDIRECT_REF_P (base)
+ && (INDIRECT_REF_P (base) || TREE_CODE (base) == MEM_REF)
&& TMR_BASE (exp)
&& TREE_CODE (TMR_BASE (exp)) == SSA_NAME
&& POINTER_TYPE_P (TREE_TYPE (TMR_BASE (exp))))
might help. On the 4.5 branch you need to backport the various
changes to retain points-to info during IVOPTs (or use -fno-tree-ivopts).
Thanks,
Richard.
