[Qemu-devel] [PATCH 2/3] Get rid of temporary variable cache.

[Qemu-devel] [PATCH 2/3] Get rid of temporary variable cache.

[Filip Navara]

The temporary variable cache actually hurts TCG more than it helps. Empirical results show that when running the Dhrystone benchmark with this patch improves the speed by about 2%.
---
 target-arm/translate.c |   31 +++----------------------------
 1 files changed, 3 insertions(+), 28 deletions(-)

diff --git a/target-arm/translate.c b/target-arm/translate.c
index 9a39536..2b8b770 100644
--- a/target-arm/translate.c
+++ b/target-arm/translate.c
@@ -111,44 +111,20 @@ void arm_translate_init(void)
 #include "helpers.h"
 }
 
-/* The code generator doesn't like lots of temporaries, so maintain our own
-   cache for reuse within a function.  */
-#define MAX_TEMPS 8
 static int num_temps;
-static TCGv temps[MAX_TEMPS];
 
 /* Allocate a temporary variable.  */
 static TCGv_i32 new_tmp(void)
 {
-    TCGv tmp;
-    if (num_temps == MAX_TEMPS)
-        abort();
-
-    if (GET_TCGV_I32(temps[num_temps]))
-      return temps[num_temps++];
-
-    tmp = tcg_temp_new_i32();
-    temps[num_temps++] = tmp;
-    return tmp;
+    num_temps++;
+    return tcg_temp_new_i32();
 }
 
 /* Release a temporary variable.  */
 static void dead_tmp(TCGv tmp)
 {
-    int i;
+    tcg_temp_free(tmp);
     num_temps--;
-    i = num_temps;
-    if (TCGV_EQUAL(temps[i], tmp))
-        return;
-
-    /* Shuffle this temp to the last slot.  */
-    while (!TCGV_EQUAL(temps[i], tmp))
-        i--;
-    while (i < num_temps) {
-        temps[i] = temps[i + 1];
-        i++;
-    }
-    temps[i] = tmp;
 }
 
 static inline TCGv load_cpu_offset(int offset)
@@ -8761,7 +8737,6 @@ static inline void gen_intermediate_code_internal(CPUState *env,
 
     /* generate intermediate code */
     num_temps = 0;
-    memset(temps, 0, sizeof(temps));
 
     pc_start = tb->pc;
 
-- 
1.6.3.msysgit.0

Re: [Qemu-devel] [PATCH 2/3] Get rid of temporary variable cache.

[Paul Brook]

On Monday 29 June 2009, Filip Navara wrote:
> The temporary variable cache actually hurts TCG more than it helps.
> Empirical results show that when running the Dhrystone benchmark with this
> patch improves the speed by about 2%.

I find it hard to believe this patch has any significant effect on runtime, 
other than random variation. My guess is it spends almost all of its time in a 
few guest loops, and no time doing translation. I'd expect the generated code 
should be the same before and after.

I'm not saying this patch is a bad idea (new_tmp/dead_tmp predate 
tcg_temp_free and are now redundant), just that I don't believe your benchmark 
results.

Paul

Re: [Qemu-devel] [PATCH 2/3] Get rid of temporary variable cache.

[Filip Navara]

On Mon, Jun 29, 2009 at 8:34 PM, Paul Brook<paul@codesourcery.com> wrote:
> On Monday 29 June 2009, Filip Navara wrote:
>> The temporary variable cache actually hurts TCG more than it helps.
>> Empirical results show that when running the Dhrystone benchmark with this
>> patch improves the speed by about 2%.
>
> I find it hard to believe this patch has any significant effect on runtime,
> other than random variation. My guess is it spends almost all of its time in a
> few guest loops, and no time doing translation. I'd expect the generated code
> should be the same before and after.

You are right. That will teach me not to trust benchmarks and validate
the results by other means also.

> I'm not saying this patch is a bad idea (new_tmp/dead_tmp predate
> tcg_temp_free and are now redundant), just that I don't believe your benchmark
> results.

And you were right not to believe them. Thanks for looking at the patch.

Best regards,
Filip Navara