[v1,2/3] tcg-runtime: light re-factor of lookup_tb_ptr

Message ID 20170614140209.29847-3-alex.bennee@linaro.org
State New
Headers show
Series
  • Fixes for TCG hangs
Related show

Commit Message

Alex Bennée June 14, 2017, 2:02 p.m.
Just a little precursor re-factoring before I was going to add a trace
point:

  - single return point, defaulting to tcg_ctx.code_gen_epilogue
  - move cs_base, pc and flags inside the jump cache hit scope
  - calculate the tb_jmp_cache hash once

Signed-off-by: Alex Bennée <alex.bennee@linaro.org>

---
 tcg-runtime.c | 35 +++++++++++++++++++----------------
 1 file changed, 19 insertions(+), 16 deletions(-)

-- 
2.13.0

Comments

Richard Henderson June 14, 2017, 3:44 p.m. | #1
On 06/14/2017 07:02 AM, Alex Bennée wrote:
> Just a little precursor re-factoring before I was going to add a trace

> point:

> 

>    - single return point, defaulting to tcg_ctx.code_gen_epilogue


Why?  Any if you're going to do that, why not init ret = epilogue and avoid the 
null test at the end?

>    - move cs_base, pc and flags inside the jump cache hit scope


Funny story.  While looking at this again, I notice that there's no reason to 
avoid, and every reason not to avoid, calling tb_htable_lookup when 
tb_jmp_cache is empty.  So I'm now doing

     tb = atomic_rcu_read(...);
     cpu_get_tb_cpu_state(env, &pc, &cs_base, &flags);

     if (!(tb && tb->pc == addr && ...)) {
         tb = tb_htable_lookup(...);
         ...

>    - calculate the tb_jmp_cache hash once


I did look once and the compiler is doing the CSE.  But it does look cleaner.


r~
Alex Bennée June 14, 2017, 4:07 p.m. | #2
Richard Henderson <rth@twiddle.net> writes:

> On 06/14/2017 07:02 AM, Alex Bennée wrote:

>> Just a little precursor re-factoring before I was going to add a trace

>> point:

>>

>>    - single return point, defaulting to tcg_ctx.code_gen_epilogue

>

> Why?  Any if you're going to do that, why not init ret = epilogue and

> avoid the null test at the end?


That would be better. I think the NULL is left over from my originall
debugging when I had added a trace point and wanted to see when the
lookup had failed.

>

>>    - move cs_base, pc and flags inside the jump cache hit scope

>

> Funny story.  While looking at this again, I notice that there's no

> reason to avoid, and every reason not to avoid, calling

> tb_htable_lookup when tb_jmp_cache is empty.  So I'm now doing

>

>     tb = atomic_rcu_read(...);

>     cpu_get_tb_cpu_state(env, &pc, &cs_base, &flags);

>

>     if (!(tb && tb->pc == addr && ...)) {

>         tb = tb_htable_lookup(...);

>         ...


OK I can fix on the next iteration.

>>    - calculate the tb_jmp_cache hash once

>

> I did look once and the compiler is doing the CSE.  But it does look cleaner.

>

>

> r~



--
Alex Bennée
Pranith Kumar June 14, 2017, 5:10 p.m. | #3
Hi Alex,

On Wed, Jun 14, 2017 at 10:02 AM, Alex Bennée <alex.bennee@linaro.org> wrote:
> Just a little precursor re-factoring before I was going to add a trace

> point:

>

>   - single return point, defaulting to tcg_ctx.code_gen_epilogue

>   - move cs_base, pc and flags inside the jump cache hit scope

>   - calculate the tb_jmp_cache hash once

>

> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>

> ---

>  tcg-runtime.c | 35 +++++++++++++++++++----------------

>  1 file changed, 19 insertions(+), 16 deletions(-)


Phew, glad you got this figured out! I tested it on the images I have
and it works. Please add:

Tested-by: Pranith Kumar <bobby.prani@gmail.com>



>

> diff --git a/tcg-runtime.c b/tcg-runtime.c

> index 7fa90ce508..f4bfa9cea6 100644

> --- a/tcg-runtime.c

> +++ b/tcg-runtime.c

> @@ -147,30 +147,33 @@ uint64_t HELPER(ctpop_i64)(uint64_t arg)

>  void *HELPER(lookup_tb_ptr)(CPUArchState *env, target_ulong addr)

>  {

>      CPUState *cpu = ENV_GET_CPU(env);

> +    unsigned int addr_hash = tb_jmp_cache_hash_func(addr);

> +    void *code_ptr = NULL;

>      TranslationBlock *tb;

> -    target_ulong cs_base, pc;

> -    uint32_t flags;

>

> -    tb = atomic_rcu_read(&cpu->tb_jmp_cache[tb_jmp_cache_hash_func(addr)]);

> +    tb = atomic_rcu_read(&cpu->tb_jmp_cache[addr_hash]);

>      if (likely(tb)) {

> +        target_ulong cs_base, pc;

> +        uint32_t flags;

> +

>          cpu_get_tb_cpu_state(env, &pc, &cs_base, &flags);

> +

>          if (likely(tb->pc == addr && tb->cs_base == cs_base &&

>                     tb->flags == flags)) {

> -            goto found;

> -        }

> -        tb = tb_htable_lookup(cpu, addr, cs_base, flags);

> -        if (likely(tb)) {

> -            atomic_set(&cpu->tb_jmp_cache[tb_jmp_cache_hash_func(addr)], tb);

> -            goto found;

> +            code_ptr = tb->tc_ptr;

> +        } else {

> +            /* If we didn't find it in the jmp_cache we still might

> +             * find it in the global tb_htable

> +             */

> +            tb = tb_htable_lookup(cpu, addr, cs_base, flags);

> +            if (likely(tb)) {

> +                atomic_set(&cpu->tb_jmp_cache[addr_hash], tb);

> +                code_ptr = tb->tc_ptr;

> +            }

>          }

>      }

> -    return tcg_ctx.code_gen_epilogue;

> - found:

> -    qemu_log_mask_and_addr(CPU_LOG_EXEC, addr,

> -                           "Chain %p [%d: " TARGET_FMT_lx "] %s\n",

> -                           tb->tc_ptr, cpu->cpu_index, addr,

> -                           lookup_symbol(addr));

> -    return tb->tc_ptr;

> +

> +    return code_ptr ? code_ptr : tcg_ctx.code_gen_epilogue;

>  }

>

>  void HELPER(exit_atomic)(CPUArchState *env)

> --

> 2.13.0

>

>

Patch hide | download patch | download mbox

diff --git a/tcg-runtime.c b/tcg-runtime.c
index 7fa90ce508..f4bfa9cea6 100644
--- a/tcg-runtime.c
+++ b/tcg-runtime.c
@@ -147,30 +147,33 @@  uint64_t HELPER(ctpop_i64)(uint64_t arg)
 void *HELPER(lookup_tb_ptr)(CPUArchState *env, target_ulong addr)
 {
     CPUState *cpu = ENV_GET_CPU(env);
+    unsigned int addr_hash = tb_jmp_cache_hash_func(addr);
+    void *code_ptr = NULL;
     TranslationBlock *tb;
-    target_ulong cs_base, pc;
-    uint32_t flags;
 
-    tb = atomic_rcu_read(&cpu->tb_jmp_cache[tb_jmp_cache_hash_func(addr)]);
+    tb = atomic_rcu_read(&cpu->tb_jmp_cache[addr_hash]);
     if (likely(tb)) {
+        target_ulong cs_base, pc;
+        uint32_t flags;
+
         cpu_get_tb_cpu_state(env, &pc, &cs_base, &flags);
+
         if (likely(tb->pc == addr && tb->cs_base == cs_base &&
                    tb->flags == flags)) {
-            goto found;
-        }
-        tb = tb_htable_lookup(cpu, addr, cs_base, flags);
-        if (likely(tb)) {
-            atomic_set(&cpu->tb_jmp_cache[tb_jmp_cache_hash_func(addr)], tb);
-            goto found;
+            code_ptr = tb->tc_ptr;
+        } else {
+            /* If we didn't find it in the jmp_cache we still might
+             * find it in the global tb_htable
+             */
+            tb = tb_htable_lookup(cpu, addr, cs_base, flags);
+            if (likely(tb)) {
+                atomic_set(&cpu->tb_jmp_cache[addr_hash], tb);
+                code_ptr = tb->tc_ptr;
+            }
         }
     }
-    return tcg_ctx.code_gen_epilogue;
- found:
-    qemu_log_mask_and_addr(CPU_LOG_EXEC, addr,
-                           "Chain %p [%d: " TARGET_FMT_lx "] %s\n",
-                           tb->tc_ptr, cpu->cpu_index, addr,
-                           lookup_symbol(addr));
-    return tb->tc_ptr;
+
+    return code_ptr ? code_ptr : tcg_ctx.code_gen_epilogue;
 }
 
 void HELPER(exit_atomic)(CPUArchState *env)