re PR tree-optimization/42833 (sra miscompiles qemu)

	PR tree-opt/42833
	* tree-sra.c (sra_modify_assign): Delay re-gimplification of
	the RHS until after generate_subtree_copies has insertted its
	code before the current statement.

From-SVN: r156176

3 files changed, 191 insertions, 9 deletions

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index e799c94..10a3626 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,10 @@
+2010-01-22  Richard Henderson  <rth@redhat.com>
+
+	PR tree-opt/42833
+	* tree-sra.c (sra_modify_assign): Delay re-gimplification of
+	the RHS until after generate_subtree_copies has insertted its
+	code before the current statement.
+
 2010-01-22  Joern Rennecke  <amylaar@spamcop.net>
 
 	* doc/tm.texi (TARGET_MIN_DIVISIONS_FOR_RECIP_MUL): Fix return type.
diff --git a/gcc/testsuite/gcc.c-torture/execute/pr42833.c b/gcc/testsuite/gcc.c-torture/execute/pr42833.c
new file mode 100644
index 0000000..5494f98
--- /dev/null
+++ b/gcc/testsuite/gcc.c-torture/execute/pr42833.c
@@ -0,0 +1,171 @@
+typedef __INT_LEAST8_TYPE__ int8_t;
+typedef __UINT_LEAST32_TYPE__ uint32_t;
+typedef int ssize_t;
+typedef struct { int8_t v1; int8_t v2; int8_t v3; int8_t v4; } neon_s8;
+
+uint32_t helper_neon_rshl_s8 (uint32_t arg1, uint32_t arg2);
+
+uint32_t
+helper_neon_rshl_s8 (uint32_t arg1, uint32_t arg2)
+{
+  uint32_t res;
+  neon_s8 vsrc1;
+  neon_s8 vsrc2;
+  neon_s8 vdest;
+  do
+    {
+      union
+      {
+	neon_s8 v;
+	uint32_t i;
+      } conv_u;
+      conv_u.i = (arg1);
+      vsrc1 = conv_u.v;
+    }
+  while (0);
+  do
+    {
+      union
+      {
+	neon_s8 v;
+	uint32_t i;
+      } conv_u;
+      conv_u.i = (arg2);
+      vsrc2 = conv_u.v;
+    }
+  while (0);
+  do
+    {
+      int8_t tmp;
+      tmp = (int8_t) vsrc2.v1;
+      if (tmp >= (ssize_t) sizeof (vsrc1.v1) * 8)
+	{
+	  vdest.v1 = 0;
+	}
+      else if (tmp < -(ssize_t) sizeof (vsrc1.v1) * 8)
+	{
+	  vdest.v1 = vsrc1.v1 >> (sizeof (vsrc1.v1) * 8 - 1);
+	}
+      else if (tmp == -(ssize_t) sizeof (vsrc1.v1) * 8)
+	{
+	  vdest.v1 = vsrc1.v1 >> (tmp - 1);
+	  vdest.v1++;
+	  vdest.v1 >>= 1;
+	}
+      else if (tmp < 0)
+	{
+	  vdest.v1 = (vsrc1.v1 + (1 << (-1 - tmp))) >> -tmp;
+	}
+      else
+	{
+	  vdest.v1 = vsrc1.v1 << tmp;
+	}
+    }
+  while (0);
+  do
+    {
+      int8_t tmp;
+      tmp = (int8_t) vsrc2.v2;
+      if (tmp >= (ssize_t) sizeof (vsrc1.v2) * 8)
+	{
+	  vdest.v2 = 0;
+	}
+      else if (tmp < -(ssize_t) sizeof (vsrc1.v2) * 8)
+	{
+	  vdest.v2 = vsrc1.v2 >> (sizeof (vsrc1.v2) * 8 - 1);
+	}
+      else if (tmp == -(ssize_t) sizeof (vsrc1.v2) * 8)
+	{
+	  vdest.v2 = vsrc1.v2 >> (tmp - 1);
+	  vdest.v2++;
+	  vdest.v2 >>= 1;
+	}
+      else if (tmp < 0)
+	{
+	  vdest.v2 = (vsrc1.v2 + (1 << (-1 - tmp))) >> -tmp;
+	}
+      else
+	{
+	  vdest.v2 = vsrc1.v2 << tmp;
+	}
+    }
+  while (0);
+  do
+    {
+      int8_t tmp;
+      tmp = (int8_t) vsrc2.v3;
+      if (tmp >= (ssize_t) sizeof (vsrc1.v3) * 8)
+	{
+	  vdest.v3 = 0;
+	}
+      else if (tmp < -(ssize_t) sizeof (vsrc1.v3) * 8)
+	{
+	  vdest.v3 = vsrc1.v3 >> (sizeof (vsrc1.v3) * 8 - 1);
+	}
+      else if (tmp == -(ssize_t) sizeof (vsrc1.v3) * 8)
+	{
+	  vdest.v3 = vsrc1.v3 >> (tmp - 1);
+	  vdest.v3++;
+	  vdest.v3 >>= 1;
+	}
+      else if (tmp < 0)
+	{
+	  vdest.v3 = (vsrc1.v3 + (1 << (-1 - tmp))) >> -tmp;
+	}
+      else
+	{
+	  vdest.v3 = vsrc1.v3 << tmp;
+	}
+    }
+  while (0);
+  do
+    {
+      int8_t tmp;
+      tmp = (int8_t) vsrc2.v4;
+      if (tmp >= (ssize_t) sizeof (vsrc1.v4) * 8)
+	{
+	  vdest.v4 = 0;
+	}
+      else if (tmp < -(ssize_t) sizeof (vsrc1.v4) * 8)
+	{
+	  vdest.v4 = vsrc1.v4 >> (sizeof (vsrc1.v4) * 8 - 1);
+	}
+      else if (tmp == -(ssize_t) sizeof (vsrc1.v4) * 8)
+	{
+	  vdest.v4 = vsrc1.v4 >> (tmp - 1);
+	  vdest.v4++;
+	  vdest.v4 >>= 1;
+	}
+      else if (tmp < 0)
+	{
+	  vdest.v4 = (vsrc1.v4 + (1 << (-1 - tmp))) >> -tmp;
+	}
+      else
+	{
+	  vdest.v4 = vsrc1.v4 << tmp;
+	}
+    }
+  while (0);;
+  do
+    {
+      union
+      {
+	neon_s8 v;
+	uint32_t i;
+      } conv_u;
+      conv_u.v = (vdest);
+      res = conv_u.i;
+    }
+  while (0);
+  return res;
+}
+
+extern void abort(void);
+
+int main()
+{
+  uint32_t r = helper_neon_rshl_s8 (0x05050505, 0x01010101);
+  if (r != 0x0a0a0a0a)
+    abort ();
+  return 0;
+}
diff --git a/gcc/tree-sra.c b/gcc/tree-sra.c
index 07658bd..79c5c76 100644
--- a/gcc/tree-sra.c
+++ b/gcc/tree-sra.c
@@ -2533,6 +2533,7 @@ sra_modify_assign (gimple *stmt, gimple_stmt_iterator *gsi,
   bool modify_this_stmt = false;
   bool force_gimple_rhs = false;
   location_t loc = gimple_location (*stmt);
+  gimple_stmt_iterator orig_gsi = *gsi;
 
   if (!gimple_assign_single_p (*stmt))
     return SRA_SA_NONE;
@@ -2611,15 +2612,6 @@ sra_modify_assign (gimple *stmt, gimple_stmt_iterator *gsi,
 		force_gimple_rhs = true;
 	    }
 	}
-
-      if (force_gimple_rhs)
-	rhs = force_gimple_operand_gsi (gsi, rhs, true, NULL_TREE,
-					true, GSI_SAME_STMT);
-      if (gimple_assign_rhs1 (*stmt) != rhs)
-	{
-	  gimple_assign_set_rhs_from_tree (gsi, rhs);
-	  gcc_assert (*stmt == gsi_stmt (*gsi));
-	}
     }
 
   /* From this point on, the function deals with assignments in between
@@ -2721,6 +2713,18 @@ sra_modify_assign (gimple *stmt, gimple_stmt_iterator *gsi,
 				     0, 0, gsi, true, true);
 	}
     }
+
+  /* This gimplification must be done after generate_subtree_copies, lest we
+     insert the subtree copies in the middle of the gimplified sequence.  */
+  if (force_gimple_rhs)
+    rhs = force_gimple_operand_gsi (&orig_gsi, rhs, true, NULL_TREE,
+				    true, GSI_SAME_STMT);
+  if (gimple_assign_rhs1 (*stmt) != rhs)
+    {
+      gimple_assign_set_rhs_from_tree (&orig_gsi, rhs);
+      gcc_assert (*stmt == gsi_stmt (orig_gsi));
+    }
+
   return modify_this_stmt ? SRA_SA_PROCESSED : SRA_SA_NONE;
 }