Il 17/09/2014 08:22, Paolo Bonzini ha scritto:
>
>> What if instead of having a "mmu_index" for the mmu arrays, we have a pointer
>> to the "mmu context". This does imply an extra memory load on the fast path,
>> but probably not an extra instruction.
>>
>> With this, we can suddenly afford to have a relatively large number of mmu
>> contexts, with which we could implement address space numbers for relevant
>> targets.
>>
>> It is, of course, a much larger change, but perhaps it's of larger benefit.
>
> Sounds good. I can give it a shot---in the meanwhile, since I forgot to
> Cc qemu-ppc, Alex can you review/apply patch 1?
Much simpler: let's cut the size of the TLB in half on affected targets.
This does sacrifice some speed, but you still get about two thirds of
the improvement (boot speed of a Debian installation ISO: 30s without
patches, 24s with small TLB, 22s with large TLB) compared to the current
TCG target.
For 32-bit target and 32-bit host we can still use the full TLB size.
The following can be easily squashed in patch 2:
diff --git a/include/exec/cpu-defs.h b/include/exec/cpu-defs.h
index 0ca6f0b..ed78884 100644
--- a/include/exec/cpu-defs.h
+++ b/include/exec/cpu-defs.h
@@ -69,8 +69,6 @@ typedef uint64_t target_ulong;
#define TB_JMP_PAGE_MASK (TB_JMP_CACHE_SIZE - TB_JMP_PAGE_SIZE)
#if !defined(CONFIG_USER_ONLY)
-#define CPU_TLB_BITS 8
-#define CPU_TLB_SIZE (1 << CPU_TLB_BITS)
/* use a fully associative victim tlb of 8 entries */
#define CPU_VTLB_SIZE 8
@@ -80,6 +78,16 @@ typedef uint64_t target_ulong;
#define CPU_TLB_ENTRY_BITS 5
#endif
+/* All the TLBs together must be smaller than 64k on RISC machines */
+#if !defined(__i386__) && !defined(__x86_64__) && !defined(__aarch64__) \
+ && !defined(__sparc__) && !defined(CONFIG_TCG_INTERPRETER)
+#define CPU_TLB_BITS (NB_MMU_MODES < 8 ? 8 : 12 - CPU_TLB_ENTRY_BITS)
+#else
+#define CPU_TLB_BITS 8
+#endif
+
+#define CPU_TLB_SIZE (1 << CPU_TLB_BITS)
+
typedef struct CPUTLBEntry {
/* bit TARGET_LONG_BITS to TARGET_PAGE_BITS : virtual address
bit TARGET_PAGE_BITS-1..4 : Nonzero for accesses that should not
Tom, can you test this on PPC?
Paolo
