[PATCH, RTX]: Additional fix for PR 57003

Uros Bizjak

2014-10-03 17:54:01 UTC

Hello!

My r215428 change exposed another PR 57003 problem on x86_64. When
compiling gcc.target/i386/pr57003.c we refer to clobbered %rdi
register after the call to memcpy:

--- pr57003.s 2014-10-03 15:08:24.000000000 +0200
+++ pr57003_.s 2014-10-03 15:08:19.000000000 +0200
@@ -78,7 +78,7 @@
leaq -20(%rbx), %rdx
movq %rax, %rdi
call memcpy
- movq %rdi, c(%rip)
+ movq %rax, c(%rip)
.L8:
movaps (%rsp), %xmm6
movaps 16(%rsp), %xmm7
@@ -321,5 +321,5 @@
.byte 0xb
.align 8
.LEFDE7:
- .ident "GCC: (GNU) 5.0.0 20141002 (experimental) [trunk
revision 215797]"
+ .ident "GCC: (GNU) 4.9.2 20141001 (prerelease)
[gcc-4_9-branch revision 215749]"
.section .note.GNU-stack,"",@progbits

The runtime failure happens only on CentOS5 (and not in Fedora20),
which supports findings in Comment #17 of the PR.

The difference is, that now we emit memcpy for MS_ABI->ELF_ABI
cross-ABI call as:

#(call_insn:TI 24 23 27 3 (set (reg:DI 0 ax)
# (call (mem:QI (symbol_ref:DI ("memcpy") [flags 0x41]
<function_decl 0x7fce6f586438 memcpy>) [0 memcpy S1 A8])
# (const_int 0 [0]))) pr57003.c:32 661 {*call_value}
# (expr_list:REG_DEAD (reg:DI 5 di)
# (expr_list:REG_DEAD (reg:DI 4 si)
# (expr_list:REG_DEAD (reg:DI 1 dx)
# (expr_list:REG_UNUSED (reg:DI 0 ax)
# (expr_list:REG_RETURNED (reg/v/f:DI 2 cx [orig:87 e ] [87])
# (expr_list:REG_CALL_DECL (symbol_ref:DI
("memcpy") [flags 0x41] <function_decl 0x7fce6f586438 memcpy>)
# (expr_list:REG_EH_REGION (const_int 0 [0])
# (nil))))))))
# (expr_list (clobber (reg:TI 52 xmm15))
# (expr_list (clobber (reg:TI 51 xmm14))
# (expr_list (clobber (reg:TI 50 xmm13))
# (expr_list (clobber (reg:TI 49 xmm12))
# (expr_list (clobber (reg:TI 48 xmm11))
# (expr_list (clobber (reg:TI 47 xmm10))
# (expr_list (clobber (reg:TI 46 xmm9))
# (expr_list (clobber (reg:TI 45 xmm8))
# (expr_list (clobber (reg:TI 28 xmm7))
# (expr_list (clobber (reg:TI 27 xmm6))
# (expr_list (clobber (reg:DI 5 di))
# (expr_list (clobber
(reg:DI 4 si))
# (expr_list:DI
(set (reg:DI 0 ax)
# (reg:DI 5 di))
# (expr_list:DI
(use (reg:DI 5 di))
#
(expr_list:DI (use (reg:DI 4 si))
#
(expr_list:DI (use (reg:DI 1 dx))
#
(nil))))))))))))))))))

which is alternate, but equivalent form of what was generated previously:

#(call_insn:TI 24 23 27 3 (parallel [
# (set (reg:DI 0 ax)
# (call (mem:QI (symbol_ref:DI ("memcpy") [flags 0x41]
<function_decl 0x7fd91824a800 memcpy>) [0 memcpy S1 A8])
# (const_int 0 [0])))
# (unspec [
# (const_int 0 [0])
# ] UNSPEC_MS_TO_SYSV_CALL)
# (clobber (reg:DI 4 si))
# (clobber (reg:DI 5 di))
# (clobber (reg:TI 27 xmm6))
# (clobber (reg:TI 28 xmm7))
# (clobber (reg:TI 45 xmm8))
# (clobber (reg:TI 46 xmm9))
# (clobber (reg:TI 47 xmm10))
# (clobber (reg:TI 48 xmm11))
# (clobber (reg:TI 49 xmm12))
# (clobber (reg:TI 50 xmm13))
# (clobber (reg:TI 51 xmm14))
# (clobber (reg:TI 52 xmm15))
# ]) pr57003.c:32 652 {*call_value_rex64_ms_sysv}
# (expr_list:REG_DEAD (reg:DI 5 di)
# (expr_list:REG_DEAD (reg:DI 4 si)
# (expr_list:REG_DEAD (reg:DI 1 dx)
# (expr_list:REG_RETURNED (reg/v/f:DI 2 cx [orig:87 e ] [87])
# (expr_list:REG_EH_REGION (const_int 0 [0])
# (nil))))))
# (expr_list:DI (set (reg:DI 0 ax)
# (reg:DI 5 di))
# (expr_list:DI (use (reg:DI 5 di))
# (expr_list:DI (use (reg:DI 4 si))
# (expr_list:DI (use (reg:DI 1 dx))
# (nil))))))

It looks that Jakub's patch, proposed in Comment #21 doesn't cover
alternative form, so it doesn't record clobbers properly.

Attached patch fixes this omission.

2014-10-03 Uros Bizjak <***@gmail.com>

PR rtl-optimization/57003
* regcprop.c (copyprop_hardreg_forward_1): If ksvd.ignore_set_reg,
also check CALL_INSN_FUNCTION_USAGE for clobbers again after
killing regs_invalidated_by_call.

Tested on x86_64-linux-gnu {,-m32}.

OK for mainline and release branches?

Uros.