| Message ID | CAFULd4agWkmyJbts0wWatKYat+YQHFgbSfbjtWnSQj=gg-JLUg@mail.gmail.com |
|---|---|
| State | New |
| Headers | show |
On Wed, Nov 7, 2012 at 11:08 PM, Uros Bizjak <ubizjak@gmail.com> wrote: > The patch simply removes the call to emit_block_move, while still > calling copy_blkmode_from_reg when appropriate. The patch fixes the > testsuite failure and produces the same code as gcc-4.7. > > 2012-11-07 Uros Bizjak <ubizjak@gmail.com> > > PR middle-end/55235 > * expr.c (store_expr): Do not call emit_block_move for > non-BLKmode values. > > Patch was bootstrapped and regression tested on x86_64-pc-linux-gnu {,-m32}. > > OK for mainline? Oh, I didn't notice that Eric already approved the patch in the PR. Although he thinks the "feature" that the patch fixes is a bit controversial, it actually restores previous functionality. So, committed to mainline SVN. Uros.
Index: expr.c =================================================================== --- expr.c (revision 193296) +++ expr.c (working copy) @@ -5246,19 +5246,12 @@ store_expr (tree exp, rtx target, int call_param_p { if (GET_MODE (target) == BLKmode) { - if (REG_P (temp)) - { - if (TREE_CODE (exp) == CALL_EXPR) - copy_blkmode_from_reg (target, temp, TREE_TYPE (exp)); - else - store_bit_field (target, - INTVAL (expr_size (exp)) * BITS_PER_UNIT, - 0, 0, 0, GET_MODE (temp), temp); - } + if (REG_P (temp) && TREE_CODE (exp) == CALL_EXPR) + copy_blkmode_from_reg (target, temp, TREE_TYPE (exp)); else - emit_block_move (target, temp, expr_size (exp), - (call_param_p - ? BLOCK_OP_CALL_PARM : BLOCK_OP_NORMAL)); + store_bit_field (target, + INTVAL (expr_size (exp)) * BITS_PER_UNIT, + 0, 0, 0, GET_MODE (temp), temp); } else convert_move (target, temp, TYPE_UNSIGNED (TREE_TYPE (exp)));
Hello! Attached patch fixes an oversight, introduced in Revision 192641 [1] that caused following testsuite failure on i686: FAIL: gcc.target/i386/pr44948-2a.c (internal compiler error) FAIL: gcc.target/i386/pr44948-2a.c (test for excess errors) As shown in the PR [2], the referred patch activated the call to emit_block_move that was previously effectively dead code (see how modes of temp and target are checked). emit_block_move ICEs when constant is passed to it, so we got ICE for following arguments: (gdb) p debug_rtx (x) (mem/j/c:BLK (plus:SI (reg/f:SI 54 virtual-stack-vars) (const_int -16 [0xfffffffffffffff0])) [0 a.V4SF+0 S16 A128]) $1 = void (gdb) p debug_rtx (y) (const_vector:V4SF [ (const_double:SF 0.0 [0x0.0p+0]) (const_double:SF 1.0e+0 [0x0.8p+1]) (const_double:SF 2.0e+0 [0x0.8p+2]) (const_double:SF 3.0e+0 [0x0.cp+2]) ]) $2 = void The patch simply removes the call to emit_block_move, while still calling copy_blkmode_from_reg when appropriate. The patch fixes the testsuite failure and produces the same code as gcc-4.7. 2012-11-07 Uros Bizjak <ubizjak@gmail.com> PR middle-end/55235 * expr.c (store_expr): Do not call emit_block_move for non-BLKmode values. Patch was bootstrapped and regression tested on x86_64-pc-linux-gnu {,-m32}. OK for mainline? [1] http://gcc.gnu.org/ml/gcc-cvs/2012-10/msg00764.html [2] http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55235#c3 Uros.