[RFC,3/3] tcg/aarch64: Remove code buffer size limitation

Commit Message

Pranith Kumar June 7, 2017, 10:52 p.m. UTC

This enables indirect jump on aarch64 hosts. Tested by booting an x86 guest on aarch64 host.

Signed-off-by: Pranith Kumar <[e-mail]>
---
 include/exec/exec-all.h      |  6 +-----
 tcg/aarch64/tcg-target.inc.c | 25 ++++++-------------------
 translate-all.c              |  2 --
 3 files changed, 7 insertions(+), 26 deletions(-)

Comments

Richard Henderson June 8, 2017, 4:58 p.m. UTC | #1

On 06/07/2017 03:52 PM, Pranith Kumar wrote:
> This enables indirect jump on aarch64 hosts. Tested by booting an x86 guest on aarch64 host.

This patch is doing too many things.

One patch should be the switch to indirect chaining (which should also squash 
in the 2/3 patch, otherwise tcg_out_insn_3305 is unused and therefore provokes 
a Werror).

Second, you should consider merely softening the code buffer size limitation to 
2GB.  At that point, we still use just ADRP+ADD to form the return address for 
exit_tb.

Third, if you want to go beyond 2GB to no limit at all, then we need to 
consider reserving a register to contain the return address back to the epilogue.


r~

Patch

diff --git a/include/exec/exec-all.h b/include/exec/exec-all.h
index 724ec73dce..a6bd3c7d1e 100644
--- a/include/exec/exec-all.h
+++ b/include/exec/exec-all.h
@@ -301,9 +301,8 @@  static inline void tlb_flush_by_mmuidx_all_cpus_synced(CPUState *cpu,
 #define CODE_GEN_AVG_BLOCK_SIZE 150
 #endif
 
-#if defined(_ARCH_PPC) \
+#if defined(_ARCH_PPC) || defined(__sparc__)    \
     || defined(__x86_64__) || defined(__i386__) \
-    || defined(__sparc__) || defined(__aarch64__) \
     || defined(__s390x__) || defined(__mips__) \
     || defined(CONFIG_TCG_INTERPRETER)
 /* NOTE: Direct jump patching must be atomic to be thread-safe. */
@@ -398,9 +397,6 @@  static inline void tb_set_jmp_target1(uintptr_t jmp_addr, uintptr_t addr)
     atomic_set((int32_t *)jmp_addr, disp / 2);
     /* no need to flush icache explicitly */
 }
-#elif defined(__aarch64__)
-void aarch64_tb_set_jmp_target(uintptr_t jmp_addr, uintptr_t addr);
-#define tb_set_jmp_target1 aarch64_tb_set_jmp_target
 #elif defined(__sparc__) || defined(__mips__)
 void tb_set_jmp_target1(uintptr_t jmp_addr, uintptr_t addr);
 #else
diff --git a/tcg/aarch64/tcg-target.inc.c b/tcg/aarch64/tcg-target.inc.c
index e488aacadb..cc81a0d5ff 100644
--- a/tcg/aarch64/tcg-target.inc.c
+++ b/tcg/aarch64/tcg-target.inc.c
@@ -865,15 +865,6 @@  static inline void tcg_out_call(TCGContext *s, tcg_insn_unit *target)
     }
 }
 
-void aarch64_tb_set_jmp_target(uintptr_t jmp_addr, uintptr_t addr)
-{
-    tcg_insn_unit *code_ptr = (tcg_insn_unit *)jmp_addr;
-    tcg_insn_unit *target = (tcg_insn_unit *)addr;
-
-    reloc_pc26_atomic(code_ptr, target);
-    flush_icache_range(jmp_addr, jmp_addr + 4);
-}
-
 static inline void tcg_out_goto_label(TCGContext *s, TCGLabel *l)
 {
     if (!l->has_value) {
@@ -1385,16 +1376,12 @@  static void tcg_out_op(TCGContext *s, TCGOpcode opc,
         break;
 
     case INDEX_op_goto_tb:
-#ifndef USE_DIRECT_JUMP
-#error "USE_DIRECT_JUMP required for aarch64"
-#endif
-        /* consistency for USE_DIRECT_JUMP */
-        tcg_debug_assert(s->tb_jmp_insn_offset != NULL);
-        s->tb_jmp_insn_offset[a0] = tcg_current_code_size(s);
-        /* actual branch destination will be patched by
-           aarch64_tb_set_jmp_target later, beware retranslation. */
-        tcg_out_goto_noaddr(s);
-        s->tb_jmp_reset_offset[a0] = tcg_current_code_size(s);
+        {
+            intptr_t offset = tcg_pcrel_diff(s, (s->tb_jmp_target_addr + a0)) >> 2;
+            tcg_out_insn(s, 3305, LDR, offset, TCG_REG_TMP);
+            tcg_out_callr(s, TCG_REG_TMP);
+            s->tb_jmp_reset_offset[a0] = tcg_current_code_size(s);
+        }
         break;
 
     case INDEX_op_goto_ptr:
diff --git a/translate-all.c b/translate-all.c
index 966747ad60..e4cd849931 100644
--- a/translate-all.c
+++ b/translate-all.c
@@ -521,8 +521,6 @@  static inline PageDesc *page_find(tb_page_addr_t index)
 # define MAX_CODE_GEN_BUFFER_SIZE  (2ul * 1024 * 1024 * 1024)
 #elif defined(__powerpc__)
 # define MAX_CODE_GEN_BUFFER_SIZE  (32u * 1024 * 1024)
-#elif defined(__aarch64__)
-# define MAX_CODE_GEN_BUFFER_SIZE  (128ul * 1024 * 1024)
 #elif defined(__s390x__)
   /* We have a +- 4GB range on the branches; leave some slop.  */
 # define MAX_CODE_GEN_BUFFER_SIZE  (3ul * 1024 * 1024 * 1024)