[PATCH, PR 42704] Do not remove scalar assignments in SRA

[Martin Jambor <mjambor at suse dot cz>]

Hi,

SRA sometimes removes assignment statements when it knows that on the
RHS all live data (ranges that have stores to) have scalar
replacements.  This caused a verification issue PR 42704 when a
removal of a scalar load from uninitialized memory (and therefore
without any live data) resulted in an SSA_NAME without a definition.
(This actually happened in in dead code.)

This happens only in very rare cases (scalar accesses with
sub-accesses) and the statements are most likely to be eliminated
later on anyway so I believe the solution is simply not to do this if
the type of the assignment is is_gimple_reg_type.

Bootstrapped and tested on x86_64-linux.  OK for trunk?

Thanks,

Martin


2010-01-13  Martin Jambor  <mjambor@suse.cz>

	PR tree-optimization/42704
	* tree-sra.c (sra_modify_assign): Do not delete assignments of
	gimple register types.

	* testsuite/g++.dg/torture/pr42704.C: New test.

Index: mine/gcc/testsuite/g++.dg/torture/pr42704.C
===================================================================
--- /dev/null
+++ mine/gcc/testsuite/g++.dg/torture/pr42704.C
@@ -0,0 +1,44 @@
+/* { dg-do compile } */
+
+typedef int PRInt32;
+class nsTreeRows {
+    class Subtree { };
+    enum { kMaxDepth = 32 };
+    struct Link {
+        Subtree* mParent;
+        PRInt32 mChildIndex;
+        Link&         operator=(const Link& aLink) {
+            mParent = aLink.mParent;
+            mChildIndex = aLink.mChildIndex;
+        }
+    };
+    class iterator {
+        PRInt32 mTop;
+        PRInt32 mRowIndex;
+        Link mLink[kMaxDepth];
+    public:
+        iterator() : mTop(-1), mRowIndex(-1) { }
+        iterator& operator=(const iterator& aIterator);
+    };
+    Subtree*     EnsureSubtreeFor(Subtree* aParent, PRInt32 aChildIndex);
+    Subtree*     GetSubtreeFor(const Subtree* aParent,                  
+PRInt32 aChildIndex,                   PRInt32* aSubtreeSize = 0);
+    void     InvalidateCachedRow() {
+        mLastRow = iterator();
+    }
+    iterator mLastRow;
+};
+nsTreeRows::Subtree* nsTreeRows::EnsureSubtreeFor(Subtree* aParent,            
+                     PRInt32 aChildIndex) {
+    Subtree* subtree = GetSubtreeFor(aParent, aChildIndex);
+    if (! subtree) {
+        InvalidateCachedRow();
+    }
+}
+nsTreeRows::iterator& nsTreeRows::iterator::operator=(const iterator&
+aIterator) {
+    mTop = aIterator.mTop;
+    for (PRInt32 i = mTop;
+         i >= 0;
+         --i)         mLink[i] = aIterator.mLink[i];
+}
Index: mine/gcc/tree-sra.c
===================================================================
--- mine.orig/gcc/tree-sra.c
+++ mine/gcc/tree-sra.c
@@ -2533,7 +2533,11 @@ sra_modify_assign (gimple *stmt, gimple_
 	{
 	  if (access_has_children_p (racc))
 	    {
-	      if (!racc->grp_unscalarized_data)
+	      if (!racc->grp_unscalarized_data
+		  /* When this is a register-type assignment, the LHS is an
+		     SSA_NAME definition that we have to keep even when loading
+		     uninitialized data (PR 42704).  */
+		  && !is_gimple_reg_type (TREE_TYPE (lhs)))
 		{
 		  generate_subtree_copies (racc->first_child, lhs,
 					   racc->offset, 0, 0, gsi,