Between 20120302 and 20120307, many 64-bit execution tests all over the testsuite started to fail on Solaris 10 and 11/x86 when using Sun as. With gas, all is fine. E.g. FAIL: gcc.c-torture/execute/20021120-1.c execution, -O3 -fomit-frame-pointer -funroll-loops The test aborts, and can be reproduced with -m64 -O1 -funroll-loops. If I add printf's to investigate the failure (gdb cannot print optimized-out values here), the failure vanishes. Between .s files for as and gas, there are no codegen differences when using -fno-dwarf2-cfi-asm. When I compare the 4.7 and mainline .s files with as, I find lots of changes like --- 20021120-1.s 2012-03-08 13:00:25.054484596 +0100 +++ /var/gcc/regression/trunk/10-gcc/build/gcc/testsuite/gcc/20021120-1.s 2012-03-08 12:48:21.890112842 +0100 @@ -534,37 +534,37 @@ movsd %xmm2, gd(,%rcx,8) cvtsi2ss %eax, %xmm3 movss %xmm3, gf(,%rcx,4) - leal 1(%rax), %edi + leal 1(%eax), %edi cvtsi2sd %edi, %xmm4 movslq %edi, %r8 movsd %xmm4, gd(,%r8,8) A reghunt revealed that this was caused by this patch: 2012-03-04 H.J. Lu <hongjiu.lu@intel.com> * config/i386/i386.c (ix86_print_operand_address): Only handle zero-extended DImode addresses. I'm now trying another bootstrap with this change reverted to see if this fixes all the failures. Rainer
Ouch. Before the change, we always used "q" modifier for addresses, only in two special cases we emitted "l". This "q" modifier forced DImode address even for SImode operands, avoiding addr32 prefixes. The change - while correct - cancelled this optimization. So, please revert this patch ASAP. Also, Sun is buggy in the way addr32 prefix is handled.
(In reply to comment #1) > Ouch. > > Before the change, we always used "q" modifier for addresses, only in two > special cases we emitted "l". This "q" modifier forced DImode address even for > SImode operands, avoiding addr32 prefixes. > > The change - while correct - cancelled this optimization. > > So, please revert this patch ASAP. > Uros, can you revert it for me? Thanks.
Author: uros Date: Thu Mar 8 15:19:32 2012 New Revision: 185103 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=185103 Log: PR target/52530 Revert: 2012-03-04 H.J. Lu <hongjiu.lu@intel.com> * config/i386/i386.c (ix86_print_operand_address): Only handle zero-extended DImode addresses. Modified: trunk/gcc/ChangeLog trunk/gcc/config/i386/i386.c
Fixed by revert.
However, there is a problem with Pmode != DImode. Consider following test: struct foo { int *f; int i; }; void __attribute__ ((noinline)) bar (struct foo x) { *(x.f) = 1; } This will compile with -mx32 to: bar: movl $1, (%rdi) ret which is wrong. The move is: #(insn:TI 6 3 13 2 (set (mem:SI (reg:SI 5 di [orig:60 x ] [60]) [4 *D.1704_1+0 S4 A32]) # (const_int 1 [0x1])) ptr.c:11 64 {*movsi_internal} # (expr_list:REG_DEAD (reg:SI 5 di [orig:60 x ] [60]) # (nil))) movl $1, (%rdi) # 6 *movsi_internal/2 [length = 6] So we want to output address in SImode, while avoiding addr32 prefixes for LEAs. The patch from H.J. should be applied and LEAs should be fixed.
This patch works for me: --- diff --git a/gcc/ChangeLog.addr32 b/gcc/ChangeLog.addr32 index 066f1ec..a191e47 100644 --- a/gcc/ChangeLog.addr32 +++ b/gcc/ChangeLog.addr32 @@ -1,3 +1,8 @@ +2012-03-08 H.J. Lu <hongjiu.lu@intel.com> + + * config/i386/i386.c (ix86_print_operand_address): Only handle + zero-extended DImode addresses if Pmode == DImode. + 2012-03-06 Uros Bizjak <ubizjak@gmail.com> * config/i386/i386.md (*zero_extendsidi2_rex64): Allow loading diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c index 69cb6ae..c2cad5a 100644 --- a/gcc/config/i386/i386.c +++ b/gcc/config/i386/i386.c @@ -14548,7 +14548,7 @@ ix86_print_operand_address (FILE *file, rtx addr) /* Print SImode registers for zero-extended addresses to force addr32 prefix. Otherwise print DImode registers to avoid it. */ - if (TARGET_64BIT) + if (Pmode == DImode) code = ((GET_CODE (addr) == ZERO_EXTEND || GET_CODE (addr) == AND) ? 'l' --
(In reply to comment #6) > This patch works for me: > > --- > diff --git a/gcc/ChangeLog.addr32 b/gcc/ChangeLog.addr32 > index 066f1ec..a191e47 100644 > --- a/gcc/ChangeLog.addr32 > +++ b/gcc/ChangeLog.addr32 > @@ -1,3 +1,8 @@ > +2012-03-08 H.J. Lu <hongjiu.lu@intel.com> > + > + * config/i386/i386.c (ix86_print_operand_address): Only handle > + zero-extended DImode addresses if Pmode == DImode. > + > 2012-03-06 Uros Bizjak <ubizjak@gmail.com> > > * config/i386/i386.md (*zero_extendsidi2_rex64): Allow loading > diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c > index 69cb6ae..c2cad5a 100644 > --- a/gcc/config/i386/i386.c > +++ b/gcc/config/i386/i386.c > @@ -14548,7 +14548,7 @@ ix86_print_operand_address (FILE *file, rtx addr) > > /* Print SImode registers for zero-extended addresses to force > addr32 prefix. Otherwise print DImode registers to avoid it. */ > - if (TARGET_64BIT) > + if (Pmode == DImode) > code = ((GET_CODE (addr) == ZERO_EXTEND > || GET_CODE (addr) == AND) > ? 'l' > -- You will have addr32 added to all lea(%SImode),%SImode instructions. I have a patch in testing that emits addresses in their natural mode (SImode or DImode), forces 'l' override for the above RTXes and 'q' for all LEA insns. For LEAs, we introduce %E modifier that enables special handling: + case 'E': + /* Wrap address in an UNSPEC to declare special handling. */ + if (TARGET_64BIT) + x = gen_rtx_UNSPEC (DImode, gen_rtvec (1, x), UNSPEC_LEA_ADDR); + + output_address (x); + return; And in ix86_print_operand_address: + else if (GET_CODE (addr) == UNSPEC && XINT (addr, 1) == UNSPEC_LEA_ADDR) + { + gcc_assert (TARGET_64BIT); + ok = ix86_decompose_address (XVECEXP (addr, 0, 0), &parts); + code = 'q'; + }
Created attachment 26865 [details] Patch that introduces %E modifier Patch in testing.
Author: uros Date: Fri Mar 9 18:01:47 2012 New Revision: 185148 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=185148 Log: PR target/52530 * config/i386/i386.c (ix86_print_operand): Handle 'E' operand modifier. (ix86_print_operand_address): Handle UNSPEC_LEA_ADDR. Do not fallback to set code to 'q'. * config/i386/i386.md (UNSPEC_LEA_ADDR): New unspec. (*movdi_internal_rex64): Use %E operand modifier for lea. (*movsi_internal): Ditto. (*lea_1): Ditto. (*lea<mode>_2): Ditto. (*lea_{3,4,5,6}_zext): Ditto. (*tls_global_dynamic_32_gnu): Ditto. (*tls_global_dynamic_64): Ditto. (*tls_dynamic_gnu2_lea_32): Ditto. (*tls_dynamic_gnu2_lea_64): Ditto. (pro_epilogue_adjust_stack_<mode>_add): Ditto. Modified: trunk/gcc/ChangeLog trunk/gcc/config/i386/i386.c trunk/gcc/config/i386/i386.md
Author: uros Date: Sun Mar 11 19:47:58 2012 New Revision: 185201 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=185201 Log: PR target/52530 * gcc.dg/torture/pr52530.c: New test. Added: trunk/gcc/testsuite/gcc.dg/torture/pr52530.c Modified: trunk/gcc/testsuite/ChangeLog
Fixed.
Author: uros Date: Wed Aug 1 19:14:59 2012 New Revision: 190048 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=190048 Log: Backport from mainline 2012-03-09 Uros Bizjak <ubizjak@gmail.com> PR target/52530 * config/i386/i386.c (ix86_print_operand): Handle 'E' operand modifier. (ix86_print_operand_address): Handle UNSPEC_LEA_ADDR. Do not fallback to set code to 'q'. * config/i386/i386.md (UNSPEC_LEA_ADDR): New unspec. (*movdi_internal_rex64): Use %E operand modifier for lea. (*movsi_internal): Ditto. (*lea_1): Ditto. (*lea<mode>_2): Ditto. (*lea_{3,4,5,6}_zext): Ditto. (*tls_global_dynamic_32_gnu): Ditto. (*tls_global_dynamic_64): Ditto. (*tls_dynamic_gnu2_lea_32): Ditto. (*tls_dynamic_gnu2_lea_64): Ditto. (pro_epilogue_adjust_stack_<mode>_add): Ditto. testsuite/ChangeLog: Backport from mainline 2012-03-11 Uros Bizjak <ubizjak@gmail.com> PR target/52530 * gcc.dg/torture/pr52530.c: New test. Added: branches/gcc-4_7-branch/gcc/testsuite/gcc.dg/pr52530.c Modified: branches/gcc-4_7-branch/gcc/ChangeLog branches/gcc-4_7-branch/gcc/config/i386/i386.c branches/gcc-4_7-branch/gcc/config/i386/i386.md branches/gcc-4_7-branch/gcc/testsuite/ChangeLog