A gcc build from today (gcc version 4.4.0 20080924) gets an ICE on PowerPC when building this (admittedly broken) code: # gcc -m64 -O2 -c test.c struct foo { int a; char lock; }; struct foo *foo; void testcase() { __sync_lock_test_and_set(&(foo->lock), 0); } It compiles if building 32bit. Even so, I wonder if gcc should warn when the variable is incompatible with __sync_* builtins.
How is this broken code?
After reading the gcc documentation I guess it is valid, and the 32bit lwarx/stwcx will overlap but not change surrounding memory.
Something like this fixes the issue: Index: config/rs6000/rs6000.c =================================================================== --- config/rs6000/rs6000.c (revision 140638) +++ config/rs6000/rs6000.c (working copy) @@ -14082,7 +14082,8 @@ rs6000_expand_compare_and_swapqhi (rtx d HOST_WIDE_INT imask = GET_MODE_MASK (mode); /* Shift amount for subword relative to aligned word. */ - addrSI = force_reg (SImode, gen_lowpart_common (SImode, XEXP (mem, 0))); + addrSI = force_reg (GET_MODE (XEXP (mem, 0)), XEXP (mem, 0)); + addrSI = force_reg (SImode, gen_lowpart_common (SImode, addrSI)); shift = gen_reg_rtx (SImode); emit_insn (gen_rlwinm (shift, addrSI, GEN_INT (3), GEN_INT (shift_mask))); --- CUT --- gen_lowpart_common does not work on non regs so force the addition to memory first and then get the lowerpart.
Mine, as far as I can tell this never worked.
Fixed on the trunk.
Subject: Bug 37640 Author: pinskia Date: Sun Sep 28 23:07:01 2008 New Revision: 140740 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=140740 Log: 2008-09-28 Andrew Pinski <andrew_pinski@playstation.sony.com> PR target/37640 * config/rs6000/rs6000.c (rs6000_expand_compare_and_swapqhi): Force address to a register before taking the lower part. 2008-09-28 Andrew Pinski <andrew_pinski@playstation.sony.com> PR target/37640 * gcc.c-torture/compile/sync-3.c: New testcase to check that addresses of non zero offset works. Added: trunk/gcc/testsuite/gcc.c-torture/compile/sync-3.c Modified: trunk/gcc/ChangeLog trunk/gcc/config/rs6000/rs6000.c trunk/gcc/testsuite/ChangeLog
Subject: Bug 37640 Author: bje Date: Tue Nov 18 22:07:58 2008 New Revision: 141980 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=141980 Log: Backport from mainline: PR target/37640 2008-09-28 Andrew Pinski <andrew_pinski@playstation.sony.com> * config/rs6000/rs6000.c (rs6000_expand_compare_and_swapqhi): Force address to a register before taking the lower part. testsuite/ * gcc.c-torture/compile/sync-3.c: New testcase to check that addresses of non zero offset works. Added: branches/gcc-4_3-branch/gcc/testsuite/gcc.c-torture/compile/sync-3.c Modified: branches/gcc-4_3-branch/gcc/ChangeLog branches/gcc-4_3-branch/gcc/config/rs6000/rs6000.c branches/gcc-4_3-branch/gcc/testsuite/ChangeLog