| Message ID | 1485451129-16439-2-git-send-email-lgustavo@codesourcery.com |
|---|---|
| State | New |
| Headers | show |
> From: Luis Machado <lgustavo@codesourcery.com> > Date: Thu, 26 Jan 2017 11:18:49 -0600 > > The following patch addresses BZ 21005, which is gdb failing to > recognize an rdrand instruction. > > It enables support for both rdrand and rdseed and handles extended register > addressing (R8~R15) for 16-bit, 32-bit and 64-bit. > > I tested this by hand with quite a few lines of inline asm. I thought we > had arch-specific tests in testsuite/, but it looks like we don't exercise > every little instruction with record/replay, so no testcase. Let me know > if one is required. > > Regression-tested on Ubuntu 16.04 x86-64. > > Testsuite results are as follows: > > === gdb Summary === > > # of expected passes 159 > > OK? > > gdb/ChangeLog > > 2017-01-26 Luis Machado <lgustavo@codesourcery.com> > > * NEWS: Mention support for record/replay of Intel 64 rdrand and > rdseed instructions. > i386-tdep.c (i386_process_record): Handle Intel 64 rdrand and rseed. OK for the NEWS part.
On 01/26/2017 05:18 PM, Luis Machado wrote: > The following patch addresses BZ 21005, which is gdb failing to > recognize an rdrand instruction. Suggest saying "this patch". "following" won't make sense when this sentence is part of the git log history. > [....] I thought we > had arch-specific tests in testsuite/, but it looks like we don't exercise > every little instruction with record/replay, so no testcase. Let me know > if one is required. This remark looks out of date? > > OK? Please remember to remove this, as it doesn't make sense to have it in a commit log. > > gdb/ChangeLog > > 2017-01-26 Luis Machado <lgustavo@codesourcery.com> > > * NEWS: Mention support for record/replay of Intel 64 rdrand and > rdseed instructions. > i386-tdep.c (i386_process_record): Handle Intel 64 rdrand and rseed. OK. Thanks, Pedro Alves
diff --git a/gdb/NEWS b/gdb/NEWS index 21e8cd3..0f63d9f 100644 --- a/gdb/NEWS +++ b/gdb/NEWS @@ -3,6 +3,9 @@ *** Changes since GDB 7.12 +* GDB now supports recording and replaying rdrand and rdseed Intel 64 + instructions. + * Building GDB and GDBserver now requires a C++11 compiler. For example, GCC 4.8 or later. diff --git a/gdb/i386-tdep.c b/gdb/i386-tdep.c index 8a4d59f..50327bb 100644 --- a/gdb/i386-tdep.c +++ b/gdb/i386-tdep.c @@ -5501,14 +5501,36 @@ i386_process_record (struct gdbarch *gdbarch, struct regcache *regcache, I386_RECORD_FULL_ARCH_LIST_ADD_REG (X86_RECORD_EFLAGS_REGNUM); break; - case 0x0fc7: /* cmpxchg8b */ + case 0x0fc7: /* cmpxchg8b / rdrand / rdseed */ if (i386_record_modrm (&ir)) return -1; if (ir.mod == 3) { - ir.addr -= 2; - opcode = opcode << 8 | ir.modrm; - goto no_support; + /* rdrand and rdseed use the 3 bits of the REG field of ModR/M as + an extended opcode. rdrand has bits 110 (/6) and rdseed + has bits 111 (/7). */ + if (ir.reg == 6 || ir.reg == 7) + { + /* The storage register is described by the 3 R/M bits, but the + REX.B prefix may be used to give access to registers + R8~R15. In this case ir.rex_b + R/M will give us the register + in the range R8~R15. + + REX.W may also be used to access 64-bit registers, but we + already record entire registers and not just partial bits + of them. */ + I386_RECORD_FULL_ARCH_LIST_ADD_REG (ir.rex_b + ir.rm); + /* These instructions also set conditional bits. */ + I386_RECORD_FULL_ARCH_LIST_ADD_REG (X86_RECORD_EFLAGS_REGNUM); + break; + } + else + { + /* We don't handle this particular instruction yet. */ + ir.addr -= 2; + opcode = opcode << 8 | ir.modrm; + goto no_support; + } } I386_RECORD_FULL_ARCH_LIST_ADD_REG (X86_RECORD_REAX_REGNUM); I386_RECORD_FULL_ARCH_LIST_ADD_REG (X86_RECORD_REDX_REGNUM);
The following patch addresses BZ 21005, which is gdb failing to recognize an rdrand instruction. It enables support for both rdrand and rdseed and handles extended register addressing (R8~R15) for 16-bit, 32-bit and 64-bit. I tested this by hand with quite a few lines of inline asm. I thought we had arch-specific tests in testsuite/, but it looks like we don't exercise every little instruction with record/replay, so no testcase. Let me know if one is required. Regression-tested on Ubuntu 16.04 x86-64. Testsuite results are as follows: === gdb Summary === # of expected passes 159 OK? gdb/ChangeLog 2017-01-26 Luis Machado <lgustavo@codesourcery.com> * NEWS: Mention support for record/replay of Intel 64 rdrand and rdseed instructions. i386-tdep.c (i386_process_record): Handle Intel 64 rdrand and rseed. --- gdb/NEWS | 3 +++ gdb/i386-tdep.c | 30 ++++++++++++++++++++++++++---- 2 files changed, 29 insertions(+), 4 deletions(-)