From 4438b7e93059b4ea2c0c71009f84f26948289cd9 Mon Sep 17 00:00:00 2001
From: Saagar Jha <saagar@saagarjha.com>
Date: Sun, 14 Jun 2020 14:16:31 -0700
Subject: [PATCH] Put jit_frame directly in cpu_state, saving a copy

On CPU-bound code this should be a free 2x speedup
---
 emu/cpu.h   |  6 ++++++
 jit/frame.h |  6 ++++--
 jit/jit.c   | 18 ++++++++----------
 3 files changed, 18 insertions(+), 12 deletions(-)

diff --git a/emu/cpu.h b/emu/cpu.h
index 6915048dd1..67a293fd20 100644
--- a/emu/cpu.h
+++ b/emu/cpu.h
@@ -5,6 +5,7 @@
 #include "misc.h"
 #include "emu/float80.h"
 #include "emu/memory.h"
+#include "jit/frame.h"
 
 struct cpu_state;
 struct tlb;
@@ -26,8 +27,13 @@ static_assert(sizeof(union xmm_reg) == 16, "xmm_reg size");
 static_assert(sizeof(union mm_reg) == 8, "mm_reg size");
 
 struct cpu_state {
+    // DO NOT MOVE THIS!
+    // ...or do, i'm a comment, not a cop
+    struct jit_frame frame;
+
     struct mem *mem;
     struct jit *jit;
+
     long cycle;
 
     // general registers
diff --git a/jit/frame.h b/jit/frame.h
index f94e1e494d..73f270a733 100644
--- a/jit/frame.h
+++ b/jit/frame.h
@@ -1,14 +1,16 @@
-#include "emu/cpu.h"
+#ifndef JIT_FRAME_H
+#define JIT_FRAME_H
 
 // keep in sync with asm
 #define JIT_RETURN_CACHE_SIZE 4096
 #define JIT_RETURN_CACHE_HASH(x) ((x) & 0xFFF0) >> 4)
 
 struct jit_frame {
-    struct cpu_state cpu;
     void *bp;
     addr_t value_addr;
     uint64_t value[2]; // buffer for crosspage crap
     struct jit_block *last_block;
     long ret_cache[JIT_RETURN_CACHE_SIZE]; // a map of ip to pointer-to-call-gadget-arguments
 };
+
+#endif
diff --git a/jit/jit.c b/jit/jit.c
index 3fd7628fcd..422d547da6 100644
--- a/jit/jit.c
+++ b/jit/jit.c
@@ -158,7 +158,7 @@ static void jit_free_jetsam(struct jit *jit) {
     }
 }
 
-int jit_enter(struct jit_block *block, struct jit_frame *frame, struct tlb *tlb);
+int jit_enter(struct jit_block *block, struct cpu_state *cpu, struct tlb *tlb);
 
 static inline size_t jit_cache_hash(addr_t ip) {
     return (ip ^ (ip >> 12)) % JIT_CACHE_SIZE;
@@ -167,11 +167,11 @@ static inline size_t jit_cache_hash(addr_t ip) {
 static int cpu_step_to_interrupt(struct cpu_state *cpu, struct tlb *tlb) {
     struct jit *jit = cpu->mem->jit;
     struct jit_block *cache[JIT_CACHE_SIZE] = {};
-    struct jit_frame frame = {.cpu = *cpu, .ret_cache = {}};
+    cpu->frame = (struct jit_frame){};
 
     int interrupt = INT_NONE;
     while (interrupt == INT_NONE) {
-        addr_t ip = frame.cpu.eip;
+        addr_t ip = cpu->eip;
         size_t cache_index = jit_cache_hash(ip);
         struct jit_block *block = cache[cache_index];
         if (block == NULL || block->addr != ip) {
@@ -186,7 +186,7 @@ static int cpu_step_to_interrupt(struct cpu_state *cpu, struct tlb *tlb) {
             cache[cache_index] = block;
             unlock(&jit->lock);
         }
-        struct jit_block *last_block = frame.last_block;
+        struct jit_block *last_block = cpu->frame.last_block;
         if (last_block != NULL &&
                 (last_block->jump_ip[0] != NULL ||
                  last_block->jump_ip[1] != NULL)) {
@@ -205,15 +205,14 @@ static int cpu_step_to_interrupt(struct cpu_state *cpu, struct tlb *tlb) {
 
             unlock(&jit->lock);
         }
-        frame.last_block = block;
+        cpu->frame.last_block = block;
 
         // block may be jetsam, but that's ok, because it can't be freed until
         // every thread on this jit is not executing anything
 
         TRACE("%d %08x --- cycle %ld\n", current->pid, ip, cpu->cycle);
 
-        interrupt = jit_enter(block, &frame, tlb);
-        *cpu = frame.cpu;
+        interrupt = jit_enter(block, cpu, tlb);
         if (interrupt == INT_NONE && ++cpu->cycle % (1 << 10) == 0)
             interrupt = INT_TIMER;
     }
@@ -229,9 +228,8 @@ static int cpu_single_step(struct cpu_state *cpu, struct tlb *tlb) {
     gen_end(&state);
 
     struct jit_block *block = state.block;
-    struct jit_frame frame = {.cpu = *cpu};
-    int interrupt = jit_enter(block, &frame, tlb);
-    *cpu = frame.cpu;
+    cpu->frame = (struct jit_frame){};
+    int interrupt = jit_enter(block, cpu, tlb);
     jit_block_free(NULL, block);
     if (interrupt == INT_NONE)
         interrupt = INT_DEBUG;