From 8afa65cc961177caf6a8b40eb81e680fa4d3a44a Mon Sep 17 00:00:00 2001
From: gbeauche <>
Date: Tue, 27 Jan 2004 13:54:51 +0000
Subject: [PATCH] Inline fast basic block lookups. Only check top tag as it is
 a hit more than 95% of the time. Overall, this improves performance by more
 than 2x on a P4.

---
 SheepShaver/src/kpx_cpu/src/cpu/block-cache.hpp | 12 ++++++++++++
 SheepShaver/src/kpx_cpu/src/cpu/ppc/ppc-cpu.cpp |  5 ++++-
 .../src/kpx_cpu/src/cpu/ppc/ppc-dyngen-ops.cpp  | 17 +++++++++++++++++
 .../src/kpx_cpu/src/cpu/ppc/ppc-dyngen.hpp      |  1 +
 .../src/kpx_cpu/src/cpu/ppc/ppc-translate.cpp   |  3 +++
 5 files changed, 37 insertions(+), 1 deletion(-)

diff --git a/SheepShaver/src/kpx_cpu/src/cpu/block-cache.hpp b/SheepShaver/src/kpx_cpu/src/cpu/block-cache.hpp
index 313c3fab..ef36c1f0 100644
--- a/SheepShaver/src/kpx_cpu/src/cpu/block-cache.hpp
+++ b/SheepShaver/src/kpx_cpu/src/cpu/block-cache.hpp
@@ -60,6 +60,7 @@ public:
 	void initialize();
 	void clear();
 	void clear_range(uintptr start, uintptr end);
+	block_info *fast_find(uintptr pc);
 	block_info *find(uintptr pc);
 
 	void remove_from_cl_list(block_info *bi);
@@ -148,6 +149,17 @@ inline void block_cache< block_info, block_allocator >::delete_blockinfo(block_i
 	allocator.release(bce);
 }
 
+template< class block_info, template<class T> class block_allocator >
+inline block_info *block_cache< block_info, block_allocator >::fast_find(uintptr pc)
+{
+	// Hit: return immediately (that covers more than 95% of the cases)
+	entry * bce = cache_tags[cacheline(pc)];
+	if (bce && bce->pc == pc)
+		return bce;
+
+	return NULL;
+}
+
 template< class block_info, template<class T> class block_allocator >
 block_info *block_cache< block_info, block_allocator >::find(uintptr pc)
 {
diff --git a/SheepShaver/src/kpx_cpu/src/cpu/ppc/ppc-cpu.cpp b/SheepShaver/src/kpx_cpu/src/cpu/ppc/ppc-cpu.cpp
index 80d13e1b..0a06bf7e 100644
--- a/SheepShaver/src/kpx_cpu/src/cpu/ppc/ppc-cpu.cpp
+++ b/SheepShaver/src/kpx_cpu/src/cpu/ppc/ppc-cpu.cpp
@@ -461,7 +461,10 @@ void powerpc_cpu::execute(uint32 entry)
 						}
 					}
 
-					if ((bi->pc != pc()) && ((bi = block_cache.find(pc())) == NULL))
+					// Don't check for backward branches here as this
+					// is now done by generated code. Besides, we will
+					// get here if the fast cache lookup failed too.
+					if ((bi = block_cache.find(pc())) == NULL)
 						break;
 				}
 			}
diff --git a/SheepShaver/src/kpx_cpu/src/cpu/ppc/ppc-dyngen-ops.cpp b/SheepShaver/src/kpx_cpu/src/cpu/ppc/ppc-dyngen-ops.cpp
index b3afb370..b80f7a80 100644
--- a/SheepShaver/src/kpx_cpu/src/cpu/ppc/ppc-dyngen-ops.cpp
+++ b/SheepShaver/src/kpx_cpu/src/cpu/ppc/ppc-dyngen-ops.cpp
@@ -86,6 +86,8 @@ struct powerpc_dyngen_helper {
 	static double & fp_result()					{ return CPU->fp_result(); }
 	static uint64 & fp_result_dw()				{ return CPU->fp_result_dw(); }
 	static inline void set_cr(int crfd, int v)	{ CPU->cr().set(crfd, v); }
+
+	static inline powerpc_block_info *find_block(uint32 pc) { return CPU->block_cache.fast_find(pc); }
 };
 
 
@@ -1206,3 +1208,18 @@ void OPPROTO op_nego_T0(void)
 	T0 = -T0;
 }
 
+/**
+ *		Generate possible call to next basic block without going
+ *		through register state restore & full cache lookup
+ **/
+
+void OPPROTO op_jump_next_A0(void)
+{
+	// Make sure there is no pending interrupt request
+	if (powerpc_dyngen_helper::spcflags().empty()) {
+		powerpc_block_info *bi = (powerpc_block_info *)reg_A0;
+		uint32 pc = powerpc_dyngen_helper::get_pc();
+		if (bi->pc == pc || (bi = powerpc_dyngen_helper::find_block(pc)) != NULL)
+			goto *(bi->entry_point);
+	}
+}
diff --git a/SheepShaver/src/kpx_cpu/src/cpu/ppc/ppc-dyngen.hpp b/SheepShaver/src/kpx_cpu/src/cpu/ppc/ppc-dyngen.hpp
index 7810dcfb..9d5d8652 100644
--- a/SheepShaver/src/kpx_cpu/src/cpu/ppc/ppc-dyngen.hpp
+++ b/SheepShaver/src/kpx_cpu/src/cpu/ppc/ppc-dyngen.hpp
@@ -119,6 +119,7 @@ public:
 	DEFINE_ALIAS(decrement_ctr_T0,0);
 	DEFINE_ALIAS(branch_A0_if_T0,1);
 	DEFINE_ALIAS(branch_A0_if_not_T0,1);
+	DEFINE_ALIAS(jump_next_A0,0);
 
 	// Compare & Record instructions
 	DEFINE_ALIAS(record_cr0_T0,0);
diff --git a/SheepShaver/src/kpx_cpu/src/cpu/ppc/ppc-translate.cpp b/SheepShaver/src/kpx_cpu/src/cpu/ppc/ppc-translate.cpp
index 38d35f91..76689db1 100644
--- a/SheepShaver/src/kpx_cpu/src/cpu/ppc/ppc-translate.cpp
+++ b/SheepShaver/src/kpx_cpu/src/cpu/ppc/ppc-translate.cpp
@@ -1303,6 +1303,9 @@ powerpc_cpu::compile_block(uint32 entry_point)
 			goto again;
 		}
 	}
+	// TODO: optimize this to a direct jump to pregenerated code?
+	dg.gen_mov_ad_A0_im((uintptr)bi);
+	dg.gen_jump_next_A0();
 	dg.gen_exec_return();
 	dg.gen_end();
 	bi->end_pc = dpc;