From 152f702439a34bf20b74f472695490edb42ad56f Mon Sep 17 00:00:00 2001 From: Cary Coutant Date: Sun, 22 Mar 2015 11:59:21 -0700 Subject: Fix support for i386 TLS GD-to-IE optimization. There are two cases to support, one with an SIB-form (6-byte) LEA, the other with a 5-byte LEA and a NOP after the call __tls_get_addr. Gold did not yet support the second case. This patch adds that support. gold/ PR gold/18106 * i386.cc (Target_i386::Relocate::tls_gd_to_ie): Fix support for non-SIB form of lea, with nop after the call. --- gold/ChangeLog | 6 ++++++ gold/i386.cc | 27 ++++++++------------------- 2 files changed, 14 insertions(+), 19 deletions(-) diff --git a/gold/ChangeLog b/gold/ChangeLog index 7d281a9f23f..71887d3300b 100644 --- a/gold/ChangeLog +++ b/gold/ChangeLog @@ -1,3 +1,9 @@ +2015-03-22 Cary Coutant + + PR gold/18106 + * i386.cc (Target_i386::Relocate::tls_gd_to_ie): Fix support for + non-SIB form of lea, with nop after the call. + 2015-03-21 Cary Coutant PR gold/14217 diff --git a/gold/i386.cc b/gold/i386.cc index 24f41036f54..3bc3d9461ca 100644 --- a/gold/i386.cc +++ b/gold/i386.cc @@ -3253,6 +3253,8 @@ Target_i386::Relocate::tls_gd_to_ie(const Relocate_info<32, false>* relinfo, { // leal foo(,%ebx,1),%eax; call ___tls_get_addr // ==> movl %gs:0,%eax; addl foo@gotntpoff(%ebx),%eax + // leal foo(%ebx),%eax; call ___tls_get_addr; nop + // ==> movl %gs:0,%eax; addl foo@gotntpoff(%ebx),%eax tls::check_range(relinfo, relnum, rel.get_r_offset(), view_size, -2); tls::check_range(relinfo, relnum, rel.get_r_offset(), view_size, 9); @@ -3264,10 +3266,7 @@ Target_i386::Relocate::tls_gd_to_ie(const Relocate_info<32, false>* relinfo, op2 == 0x8d || op2 == 0x04); tls::check_tls(relinfo, relnum, rel.get_r_offset(), view[4] == 0xe8); - int roff = 5; - - // FIXME: For now, support only the first (SIB) form. - tls::check_tls(relinfo, relnum, rel.get_r_offset(), op2 == 0x04); + int roff; if (op2 == 0x04) { @@ -3275,28 +3274,18 @@ Target_i386::Relocate::tls_gd_to_ie(const Relocate_info<32, false>* relinfo, tls::check_tls(relinfo, relnum, rel.get_r_offset(), view[-3] == 0x8d); tls::check_tls(relinfo, relnum, rel.get_r_offset(), ((op1 & 0xc7) == 0x05 && op1 != (4 << 3))); - memcpy(view - 3, "\x65\xa1\0\0\0\0\x03\x83\0\0\0", 12); + roff = 5; } else { + tls::check_range(relinfo, relnum, rel.get_r_offset(), view_size, 10); tls::check_tls(relinfo, relnum, rel.get_r_offset(), (op1 & 0xf8) == 0x80 && (op1 & 7) != 4); - if (rel.get_r_offset() + 9 < view_size - && view[9] == 0x90) - { - // FIXME: This is not the right instruction sequence. - // There is a trailing nop. Use the size byte subl. - memcpy(view - 2, "\x65\xa1\0\0\0\0\x81\xe8\0\0\0", 12); - roff = 6; - } - else - { - // FIXME: This is not the right instruction sequence. - // Use the five byte subl. - memcpy(view - 2, "\x65\xa1\0\0\0\0\x2d\0\0\0", 11); - } + tls::check_tls(relinfo, relnum, rel.get_r_offset(), view[9] == 0x90); + roff = 6; } + memcpy(view + roff - 8, "\x65\xa1\0\0\0\0\x03\x83\0\0\0", 12); Relocate_functions<32, false>::rel32(view + roff, value); // The next reloc should be a PLT32 reloc against __tls_get_addr. -- cgit v1.2.1