1.1 --- a/src/sh4/sh4.c	Fri Nov 25 09:02:13 2011 +1000
     1.2 +++ b/src/sh4/sh4.c	Sun Nov 27 18:20:21 2011 +1000
     1.3 @@ -235,6 +235,9 @@
     1.4          /* If we were running with the translator, update new_pc and in_delay_slot */
     1.5          sh4r.new_pc = sh4r.pc+2;
     1.6          sh4r.in_delay_slot = FALSE;
     1.7 +        if( sh4_translate_get_profile_blocks() ) {
     1.8 +            xlat_dump_cache_by_activity(30);
     1.9 +        }
    1.10      }
    1.11  
    1.12  }

     2.1 --- a/src/sh4/sh4trans.h	Fri Nov 25 09:02:13 2011 +1000
     2.2 +++ b/src/sh4/sh4trans.h	Sun Nov 27 18:20:21 2011 +1000
     2.3 @@ -103,6 +103,16 @@
     2.4  void sh4_translate_set_fastmem( gboolean flag );
     2.5  
     2.6  /**
     2.7 + * Enable/disable basic block profiling
     2.8 + */
     2.9 +void sh4_translate_set_profile_blocks( gboolean flag );
    2.10 +
    2.11 +/**
    2.12 + * Get the boolean flag indicating whether block profiling is on.
    2.13 + */
    2.14 +gboolean sh4_translate_get_profile_blocks();
    2.15 +
    2.16 +/**
    2.17   * Set the address spaces for the translated code.
    2.18   */
    2.19  void sh4_translate_set_address_space( struct mem_region_fn **priv, struct mem_region_fn **user );

     3.1 --- a/src/sh4/sh4x86.in	Fri Nov 25 09:02:13 2011 +1000
     3.2 +++ b/src/sh4/sh4x86.in	Sun Nov 27 18:20:21 2011 +1000
     3.3 @@ -106,6 +106,7 @@
     3.4      xlat_block_begin_callback_t begin_callback;
     3.5      xlat_block_end_callback_t end_callback;
     3.6      gboolean fastmem;
     3.7 +    gboolean profile_blocks;
     3.8      
     3.9      /* Allocated memory for the (block-wide) back-patch list */
    3.10      struct backpatch_record *backpatch_list;
    3.11 @@ -166,6 +167,7 @@
    3.12      sh4_x86.end_callback = NULL;
    3.13      sh4_translate_set_address_space( sh4_address_space, sh4_user_address_space );
    3.14      sh4_x86.fastmem = TRUE;
    3.15 +    sh4_x86.profile_blocks = FALSE;
    3.16      sh4_x86.sse3_enabled = is_sse3_supported();
    3.17      x86_disasm_init();
    3.18      x86_set_symtab( x86_symbol_table, sizeof(x86_symbol_table)/sizeof(struct x86_symbol) );
    3.19 @@ -182,6 +184,16 @@
    3.20      sh4_x86.fastmem = flag;
    3.21  }
    3.22  
    3.23 +void sh4_translate_set_profile_blocks( gboolean flag )
    3.24 +{
    3.25 +    sh4_x86.profile_blocks = flag;
    3.26 +}
    3.27 +
    3.28 +gboolean sh4_translate_get_profile_blocks()
    3.29 +{
    3.30 +    return sh4_x86.profile_blocks;
    3.31 +}
    3.32 +
    3.33  /**
    3.34   * Disassemble the given translated code block, and it's source SH4 code block
    3.35   * side-by-side. The current native pc will be marked if non-null.
    3.36 @@ -476,6 +488,11 @@
    3.37  
    3.38  #define SLOTILLEGAL() exit_block_exc(EXC_SLOT_ILLEGAL, pc-2); sh4_x86.in_delay_slot = DELAY_NONE; return 2;
    3.39  
    3.40 +/** Offset of xlat_sh4_mode field relative to the code pointer */ 
    3.41 +#define XLAT_SH4_MODE_CODE_OFFSET  (uint32_t)(offsetof(struct xlat_cache_block, xlat_sh4_mode) - offsetof(struct xlat_cache_block,code) )
    3.42 +#define XLAT_CHAIN_CODE_OFFSET (uint32_t)(offsetof(struct xlat_cache_block, chain) - offsetof(struct xlat_cache_block,code) )
    3.43 +#define XLAT_ACTIVE_CODE_OFFSET (uint32_t)(offsetof(struct xlat_cache_block, active) - offsetof(struct xlat_cache_block,code) )
    3.44 +
    3.45  void sh4_translate_begin_block( sh4addr_t pc ) 
    3.46  {
    3.47  	sh4_x86.code = xlat_output;
    3.48 @@ -493,6 +510,10 @@
    3.49      if( sh4_x86.begin_callback ) {
    3.50          CALL_ptr( sh4_x86.begin_callback );
    3.51      }
    3.52 +    if( sh4_x86.profile_blocks ) {
    3.53 +    	MOVP_immptr_rptr( ((uintptr_t)sh4_x86.code) + XLAT_ACTIVE_CODE_OFFSET, REG_EAX );
    3.54 +    	ADDL_imms_r32disp( 1, REG_EAX, 0 );
    3.55 +    }  
    3.56  }
    3.57  
    3.58  
    3.59 @@ -519,10 +540,6 @@
    3.60  
    3.61  #define UNTRANSLATABLE(pc) !IS_IN_ICACHE(pc)
    3.62  
    3.63 -/** Offset of xlat_sh4_mode field relative to the code pointer */ 
    3.64 -#define XLAT_SH4_MODE_CODE_OFFSET  (uint32_t)(offsetof(struct xlat_cache_block, xlat_sh4_mode) - offsetof(struct xlat_cache_block,code) )
    3.65 -#define XLAT_CHAIN_CODE_OFFSET (uint32_t)(offsetof(struct xlat_cache_block, chain) - offsetof(struct xlat_cache_block,code) )
    3.66 -
    3.67  /**
    3.68   * Test if the loaded target code pointer in %eax is valid, and if so jump
    3.69   * directly into it, bypassing the normal exit.

     4.1 --- a/src/xlat/xltcache.c	Fri Nov 25 09:02:13 2011 +1000
     4.2 +++ b/src/xlat/xltcache.c	Sun Nov 27 18:20:21 2011 +1000
     4.3 @@ -37,6 +37,8 @@
     4.4  #define XLAT_LUT_ENTRY_EMPTY (void *)0
     4.5  #define XLAT_LUT_ENTRY_USED  (void *)1
     4.6  
     4.7 +#define XLAT_ADDR_FROM_ENTRY(pagenum,entrynum) ((((pagenum)&0xFFFF)<<13)|(((entrynum)<<1)&0x1FFE))
     4.8 +
     4.9  #define NEXT(block) ( (xlat_cache_block_t)&((block)->code[(block)->size]))
    4.10  #define IS_ENTRY_POINT(ent) (ent > XLAT_LUT_ENTRY_USED)
    4.11  #define IS_ENTRY_USED(ent) (ent != XLAT_LUT_ENTRY_EMPTY)
    4.12 @@ -575,3 +577,92 @@
    4.13  #endif
    4.14  }
    4.15  
    4.16 +typedef struct {
    4.17 +    xlat_cache_block_t block;
    4.18 +    sh4addr_t sh4_pc;
    4.19 +} block_sh4_entry;
    4.20 +
    4.21 +unsigned int xlat_get_active_block_count()
    4.22 +{
    4.23 +    unsigned int count = 0;
    4.24 +    xlat_cache_block_t ptr = xlat_new_cache;
    4.25 +    while( ptr->size != 0 ) {
    4.26 +        if( ptr->active != 0 ) {
    4.27 +            count++;
    4.28 +        }
    4.29 +        ptr = NEXT(ptr);
    4.30 +    }
    4.31 +    return count;
    4.32 +}
    4.33 +
    4.34 +unsigned int xlat_get_active_blocks( block_sh4_entry *blocks, unsigned int size )
    4.35 +{
    4.36 +    unsigned int count = 0;
    4.37 +    xlat_cache_block_t ptr = xlat_new_cache;
    4.38 +    while( ptr->size != 0 ) {
    4.39 +        if( ptr->active != 0 ) {
    4.40 +            blocks[count].block = ptr;
    4.41 +            blocks[count].sh4_pc = 0;
    4.42 +            count++;
    4.43 +        }
    4.44 +        if( count >= size )
    4.45 +            break;
    4.46 +        ptr = NEXT(ptr);
    4.47 +    }
    4.48 +    return count;
    4.49 +}
    4.50 +
    4.51 +void xlat_get_block_sh4addrs( block_sh4_entry *blocks, unsigned int size )
    4.52 +{
    4.53 +    unsigned i;
    4.54 +    for( i=0; i<XLAT_LUT_PAGES;i ++ ) {
    4.55 +        void **page = xlat_lut[i];
    4.56 +        if( page != NULL ) {
    4.57 +            for( unsigned j=0; j < XLAT_LUT_PAGE_ENTRIES; j++ ) {
    4.58 +                void *code = (void *)(((uintptr_t)(page[j])) & (~((uintptr_t)0x03)));
    4.59 +                if( code != NULL ) {
    4.60 +                    xlat_cache_block_t ptr = XLAT_BLOCK_FOR_CODE(code);
    4.61 +                    sh4addr_t pc = XLAT_ADDR_FROM_ENTRY(i,j);
    4.62 +                    for( unsigned k=0; k<size; k++ ) {
    4.63 +                        if( blocks[k].block == ptr ) {
    4.64 +                            blocks[k].sh4_pc = pc;
    4.65 +                            ptr = ptr->chain;
    4.66 +                            if( ptr == NULL )
    4.67 +                                break;
    4.68 +                            else {
    4.69 +                                ptr = XLAT_BLOCK_FOR_CODE(ptr);
    4.70 +                                k = 0;
    4.71 +                            }
    4.72 +                        }
    4.73 +                    }
    4.74 +                }
    4.75 +            }
    4.76 +        }
    4.77 +    }
    4.78 +}
    4.79 +
    4.80 +static int xlat_compare_active_field( const void *a, const void *b )
    4.81 +{
    4.82 +    const block_sh4_entry *ptra = (const block_sh4_entry *)a;
    4.83 +    const block_sh4_entry *ptrb = (const block_sh4_entry *)b;
    4.84 +    return ptrb->block->active - ptra->block->active;
    4.85 +}
    4.86 +
    4.87 +void xlat_dump_cache_by_activity( unsigned int topN )
    4.88 +{
    4.89 +    int i=0;
    4.90 +    int count = xlat_get_active_block_count();
    4.91 +
    4.92 +    block_sh4_entry blocks[count];
    4.93 +    xlat_get_active_blocks(blocks, count);
    4.94 +    xlat_get_block_sh4addrs(blocks,count);
    4.95 +    qsort(blocks, count, sizeof(block_sh4_entry), xlat_compare_active_field);
    4.96 +
    4.97 +    if( topN == 0 || topN > count )
    4.98 +        topN = count;
    4.99 +    for( unsigned int i=0; i<topN; i++ ) {
   4.100 +        fprintf(stderr, "0x%08X (%p): %d\n", blocks[i].sh4_pc, blocks[i].block->code, blocks[i].block->active);
   4.101 +        sh4_translate_disasm_block( stderr, blocks[i].block->code, blocks[i].sh4_pc, NULL );
   4.102 +        fprintf(stderr, "\n");
   4.103 +    }
   4.104 +}

     5.1 --- a/src/xlat/xltcache.h	Fri Nov 25 09:02:13 2011 +1000
     5.2 +++ b/src/xlat/xltcache.h	Sun Nov 27 18:20:21 2011 +1000
     5.3 @@ -181,4 +181,11 @@
     5.4   */
     5.5  void xlat_check_integrity();
     5.6  
     5.7 +/**
     5.8 + * Dump out the top N translated blocks by number of executions (requires block
     5.9 + * profiling to be turned on in order to give meaningful results).
    5.10 + * @param topN Number of blocks to print. If 0, print all blocks in the cache
    5.11 + */
    5.12 +void xlat_dump_cache_by_activity( unsigned int topN );
    5.13 +
    5.14  #endif /* lxdream_xltcache_H */