[Patch PR 45260]: [4.5/4.6 Regression] g++4.5: -prefetch-loop-arrays internal compiler error: in verify_expr, at tree-cfg.c:2541

[Patch PR 45260]: [4.5/4.6 Regression] g++4.5: -prefetch-loop-arrays internal compiler error: in verify_expr, at tree-cfg.c:2541

[Fang, Changpeng]

[-- Attachment #1: Type: text/plain, Size: 762 bytes --]

Hi,

The attached patch fixes PR 45260:  [4.5/4.6 Regression] g++4.5: -prefetch-loop-arrays internal compiler error: in verify_expr, at tree-cfg.c:2541

Problem: The prefetch pass generates an ADDR_EXPR for COMPONENT_REF(VIEW_CONVERT_EXPR(SSA_NAME)) whose
address could not be taken, and thus invokes an assertion failure in verify_expr.

Solution: Don't generate prefetch for references whose base address could not be taken.

In this patch, we make may_be_nonaddressable_p (tree expr) global and use it in the prefetch pass.

This patch passed bootstrapping and gcc regression tests on x86_64-unknown-linux-gnu systems ib both
trunk and 4.5 branch.

Is it OK to commit to trunk and 4.5 branch?

Thanks and have a wonderful weekend!

Changpeng

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #2: 0001-pr45260-Don-t-generate-prefetch-if-the-address-of-ba.patch --]
[-- Type: text/x-patch; name="0001-pr45260-Don-t-generate-prefetch-if-the-address-of-ba.patch", Size: 2260 bytes --]

From ac383eec351bab858f0be30b4a89947a566853e2 Mon Sep 17 00:00:00 2001
From: Changpeng Fang <chfang@houghton.(none)>
Date: Fri, 20 Aug 2010 12:10:43 -0700
Subject: [PATCH] pr45260 Don't generate prefetch if the address of base could not be taken.

	* tree-flow.h (may_be_nonaddressable_p): New definition. Make the
	existing static function global.

	*tree-ssa-loop-ivopts.c (may_be_nonaddressable_p): This function
	is changed to global.

	*tree-ssa-loop-prefetch.c (gather_memory_references_ref): Call
	may_be_nonaddressable_p on base, and don't collect this reference
	if the address of the base could not be taken.
---
 gcc/tree-flow.h              |    1 +
 gcc/tree-ssa-loop-ivopts.c   |    2 +-
 gcc/tree-ssa-loop-prefetch.c |    4 ++++
 3 files changed, 6 insertions(+), 1 deletions(-)

diff --git a/gcc/tree-flow.h b/gcc/tree-flow.h
index 04ba532..6731308 100644
--- a/gcc/tree-flow.h
+++ b/gcc/tree-flow.h
@@ -817,6 +817,7 @@ bool stmt_invariant_in_loop_p (struct loop *, gimple);
 bool multiplier_allowed_in_address_p (HOST_WIDE_INT, enum machine_mode,
 				      addr_space_t);
 unsigned multiply_by_cost (HOST_WIDE_INT, enum machine_mode, bool);
+bool may_be_nonaddressable_p (tree expr);
 
 /* In tree-ssa-threadupdate.c.  */
 extern bool thread_through_all_blocks (bool);
diff --git a/gcc/tree-ssa-loop-ivopts.c b/gcc/tree-ssa-loop-ivopts.c
index a347c86..0029c76 100644
--- a/gcc/tree-ssa-loop-ivopts.c
+++ b/gcc/tree-ssa-loop-ivopts.c
@@ -1640,7 +1640,7 @@ may_be_unaligned_p (tree ref, tree step)
 
 /* Return true if EXPR may be non-addressable.   */
 
-static bool
+bool
 may_be_nonaddressable_p (tree expr)
 {
   switch (TREE_CODE (expr))
diff --git a/gcc/tree-ssa-loop-prefetch.c b/gcc/tree-ssa-loop-prefetch.c
index 9ad6cd2..4e7068f 100644
--- a/gcc/tree-ssa-loop-prefetch.c
+++ b/gcc/tree-ssa-loop-prefetch.c
@@ -539,6 +539,10 @@ gather_memory_references_ref (struct loop *loop, struct mem_ref_group **refs,
   if (step == NULL_TREE)
     return false;
 
+  /* Stop if the address of BASE could not taken.  */
+  if (may_be_nonaddressable_p (base))
+    return false;
+
   /* Limit non-constant step prefetching only to the innermost loops.  */
   if (!cst_and_fits_in_hwi (step) && loop->inner != NULL)
     return false;
-- 
1.6.3.3

Re: [Patch PR 45260]: [4.5/4.6 Regression] g++4.5: -prefetch-loop-arrays internal compiler error: in verify_expr, at tree-cfg.c:2541

[Richard Guenther]

On Sat, Aug 21, 2010 at 12:44 AM, Fang, Changpeng
<Changpeng.Fang@amd.com> wrote:
> Hi,
>
> The attached patch fixes PR 45260:  [4.5/4.6 Regression] g++4.5: -prefetch-loop-arrays internal compiler error: in verify_expr, at tree-cfg.c:2541
>
> Problem: The prefetch pass generates an ADDR_EXPR for COMPONENT_REF(VIEW_CONVERT_EXPR(SSA_NAME)) whose
> address could not be taken, and thus invokes an assertion failure in verify_expr.
>
> Solution: Don't generate prefetch for references whose base address could not be taken.
>
> In this patch, we make may_be_nonaddressable_p (tree expr) global and use it in the prefetch pass.
>
> This patch passed bootstrapping and gcc regression tests on x86_64-unknown-linux-gnu systems ib both
> trunk and 4.5 branch.
>
> Is it OK to commit to trunk and 4.5 branch?

Ok.

Thanks,
Richard.

> Thanks and have a wonderful weekend!
>
> Changpeng