From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from eggs.gnu.org ([2001:4830:134:3::10]:52771) by lists.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1XgZ6U-0006PX-HR for qemu-devel@nongnu.org; Tue, 21 Oct 2014 09:04:04 -0400 Received: from Debian-exim by eggs.gnu.org with spam-scanned (Exim 4.71) (envelope-from ) id 1XgZ6O-0004vq-El for qemu-devel@nongnu.org; Tue, 21 Oct 2014 09:03:58 -0400 Received: from mail.ispras.ru ([83.149.199.45]:48503) by eggs.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1XgZ6O-0004vi-1k for qemu-devel@nongnu.org; Tue, 21 Oct 2014 09:03:52 -0400 From: Pavel Dovgalyuk Date: Tue, 21 Oct 2014 17:03:51 +0400 Message-ID: <20141021130351.3780.41554.stgit@PASHA-ISP> MIME-Version: 1.0 Content-Type: text/plain; charset="utf-8" Content-Transfer-Encoding: 7bit Subject: [Qemu-devel] [PATCH] i386: fix icount processing for repz instructions List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , To: qemu-devel@nongnu.org Cc: mark.burton@greensocs.com, batuzovk@ispras.ru, maria.klimushenkova@ispras.ru, pavel.dovgaluk@ispras.ru, pbonzini@redhat.com, zealot351@gmail.com, fred.konrad@greensocs.com TCG generates optimized code for i386 repz instructions. It means that when ecx becomes 0, execution of the string instruction breaks immediately without an additional iteration for ecx==0 (which will only check ecx and set the flags). Omitting this iteration leads to different instructions counting in singlestep mode and in normal execution. This patch disables optimization of this last iteration for icount mode. Signed-off-by: Pavel Dovgalyuk --- target-i386/translate.c | 17 +++++++++++++++-- 1 files changed, 15 insertions(+), 2 deletions(-) diff --git a/target-i386/translate.c b/target-i386/translate.c index 418173e..1284173 100644 --- a/target-i386/translate.c +++ b/target-i386/translate.c @@ -115,6 +115,7 @@ typedef struct DisasContext { int tf; /* TF cpu flag */ int singlestep_enabled; /* "hardware" single step enabled */ int jmp_opt; /* use direct block chaining for direct jumps */ + int repz_opt; /* optimize jumps within repz instructions */ int mem_index; /* select memory access functions */ uint64_t flags; /* all execution flags */ struct TranslationBlock *tb; @@ -1215,7 +1216,7 @@ static inline void gen_repz_ ## op(DisasContext *s, TCGMemOp ot, \ gen_op_add_reg_im(s->aflag, R_ECX, -1); \ /* a loop would cause two single step exceptions if ECX = 1 \ before rep string_insn */ \ - if (!s->jmp_opt) \ + if (!s->repz_opt) \ gen_op_jz_ecx(s->aflag, l2); \ gen_jmp(s, cur_eip); \ } @@ -1233,7 +1234,7 @@ static inline void gen_repz_ ## op(DisasContext *s, TCGMemOp ot, \ gen_op_add_reg_im(s->aflag, R_ECX, -1); \ gen_update_cc_op(s); \ gen_jcc1(s, (JCC_Z << 1) | (nz ^ 1), l2); \ - if (!s->jmp_opt) \ + if (!s->repz_opt) \ gen_op_jz_ecx(s->aflag, l2); \ gen_jmp(s, cur_eip); \ } @@ -7951,6 +7952,18 @@ static inline void gen_intermediate_code_internal(X86CPU *cpu, || (flags & HF_SOFTMMU_MASK) #endif ); + dc->repz_opt = dc->jmp_opt + /* Do not optimize repz jumps at all in icount mode, because + rep movsS instructions are execured with different paths + in repz_opt and !repz_opt modes. The first one was used + always except single step mode. And this setting + disables jumps optimization and control paths become + equivalent in run and single step modes. + Now there will be no jump optimization for repz in + trace and replay modes and there will always be an + additional step for ecx=0. + */ + || use_icount; #if 0 /* check addseg logic */ if (!dc->addseg && (dc->vm86 || !dc->pe || !dc->code32))