On December 21, 2017 6:50:45 PM GMT+01:00, Jakub Jelinek <ja...@redhat.com>
wrote:
>Hi!
>
>The problem here is that the code expects fold_build1 will actually not
>fold, because using gimple_build_assign (..., VIEW_CONVERT_EXPR, temp);
>is valid only if TREE_CODE (temp) == VIEW_CONVERT_EXPR.
>So, either we can replace the fold_build1 with build1, or we should
>just
>use the gimple_build_assign overload that will handle even what
><view_convert_expr <SSA_NAME_result_of_phi>> will fold to (in this case
>a <nop_expr <SSA_NAME_result_of_phi>>).
>
>Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
OK.
Richard.
>2017-12-21 Jakub Jelinek <ja...@redhat.com>
>
> PR tree-optimization/83521
> * tree-ssa-phiopt.c (factor_out_conditional_conversion): Use
> gimple_build_assign without code on result of
> fold_build1 (VIEW_CONVERT_EXPR, ...), as it might not create
> a VIEW_CONVERT_EXPR.
>
> * gcc.dg/pr83521.c: New test.
>
>--- gcc/tree-ssa-phiopt.c.jj 2017-12-20 20:40:18.000000000 +0100
>+++ gcc/tree-ssa-phiopt.c 2017-12-21 11:25:33.335004838 +0100
>@@ -548,8 +548,12 @@ factor_out_conditional_conversion (edge
>
> /* Create the conversion stmt and insert it. */
> if (convert_code == VIEW_CONVERT_EXPR)
>- temp = fold_build1 (VIEW_CONVERT_EXPR, TREE_TYPE (result), temp);
>- new_stmt = gimple_build_assign (result, convert_code, temp);
>+ {
>+ temp = fold_build1 (VIEW_CONVERT_EXPR, TREE_TYPE (result),
>temp);
>+ new_stmt = gimple_build_assign (result, temp);
>+ }
>+ else
>+ new_stmt = gimple_build_assign (result, convert_code, temp);
> gsi = gsi_after_labels (gimple_bb (phi));
> gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT);
>
>--- gcc/testsuite/gcc.dg/pr83521.c.jj 2017-12-21 11:29:54.348662949
>+0100
>+++ gcc/testsuite/gcc.dg/pr83521.c 2017-12-21 11:29:41.000000000 +0100
>@@ -0,0 +1,10 @@
>+/* PR tree-optimization/83521 */
>+/* { dg-do compile } */
>+/* { dg-options "-O1 -fno-tree-forwprop" } */
>+
>+int
>+foo (unsigned int x, int y)
>+{
>+ int *z = (int *)&x;
>+ return (y == 0) ? y : *z;
>+}
>
> Jakub
