tcg-i386 cleanup and improvement, v2

PATCH Applied Richard Henderson 2 participants 48 messages Apr 28, 2010 – May 21, 2010

Share & tools

Richard Henderson

author

2010-04-28 18:30 UTC #0 lore raw

cover tcg-i386 cleanup and improvement, v2

Changes v1->v2:
* Dropped controversial bswap changes; bswap16 continues to use rolw.
* Tidy data16 as the last of the hard-coded constants.

Richard Henderson (22):
tcg-i386: Allocate call-saved registers first.
tcg-i386: Tidy initialization of tcg_target_call_clobber_regs.
tcg-i386: Tidy ext8u and ext16u operations.
tcg-i386: Tidy ext8s and ext16s operations.
tcg-i386: Tidy bswap operations.
tcg-i386: Tidy shift operations.
tcg-i386: Tidy move operations.
tcg-i386: Eliminate extra move from qemu_ld64.
tcg-i386: Tidy jumps.
tcg-i386: Tidy immediate arithmetic operations.
tcg-i386: Tidy non-immediate arithmetic operations.
tcg-i386: Tidy movi.
tcg-i386: Tidy push/pop.
tcg-i386: Tidy calls.
tcg-i386: Tidy ret.
tcg-i386: Tidy setcc.
tcg-i386: Tidy unary arithmetic.
tcg-i386: Tidy multiply.
tcg-i386: Tidy xchg.
tcg-i386: Tidy lea.
tcg-i386: Use lea for three-operand add.
tcg-i386: Tidy data16 prefixes.

 tcg/i386/tcg-target.c |  718 +++++++++++++++++++++++++++++--------------------
 1 files changed, 433 insertions(+), 285 deletions(-)

Richard Henderson

author

2010-04-28 18:30 UTC #1 lore raw

PATCH 1/22 tcg-i386: Allocate call-saved registers first.

Signed-off-by Richard Henderson <[email protected]>

---
 tcg/i386/tcg-target.c |   13 +++++++++----
 1 files changed, 9 insertions(+), 4 deletions(-)

tcg/i386/tcg-target.c +9 -4

index e684b33..f5c24f7 100644
@@ -36,16 +36,21 @@ static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
 #endif
 
 static const int tcg_target_reg_alloc_order[] = {
-    TCG_REG_EAX,
-    TCG_REG_EDX,
-    TCG_REG_ECX,
     TCG_REG_EBX,
     TCG_REG_ESI,
     TCG_REG_EDI,
     TCG_REG_EBP,
+    TCG_REG_ECX,
+    TCG_REG_EDX,
+    TCG_REG_EAX,
+};
+
+static const int tcg_target_call_iarg_regs[3] = {
+    TCG_REG_EAX,
+    TCG_REG_EDX,
+    TCG_REG_ECX
 };
 
-static const int tcg_target_call_iarg_regs[3] = { TCG_REG_EAX, TCG_REG_EDX, TCG_REG_ECX };
 static const int tcg_target_call_oarg_regs[2] = { TCG_REG_EAX, TCG_REG_EDX };
 
 static uint8_t *tb_ret_addr;

Signature · 1 line

1.6.6.1

Aurelien Jarno

Applied

2010-05-19 17:37 UTC #2 lore raw

Re: [Qemu-devel] [PATCH 01/22] tcg-i386: Allocate call-saved registers first.

Quoted 4 lines · Richard Henderson

Signed-off-by Richard Henderson <[email protected]>
---
 tcg/i386/tcg-target.c |   13 +++++++++----
 1 files changed, 9 insertions(+), 4 deletions(-)

Thanks, applied.

Quoted 36 lines

tcg/i386/tcg-target.c +9 -4

index e684b33..f5c24f7 100644
@@ -36,16 +36,21 @@ static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = {
 #endif
 
 static const int tcg_target_reg_alloc_order[] = {
-    TCG_REG_EAX,
-    TCG_REG_EDX,
-    TCG_REG_ECX,
     TCG_REG_EBX,
     TCG_REG_ESI,
     TCG_REG_EDI,
     TCG_REG_EBP,
+    TCG_REG_ECX,
+    TCG_REG_EDX,
+    TCG_REG_EAX,
+};
+
+static const int tcg_target_call_iarg_regs[3] = {
+    TCG_REG_EAX,
+    TCG_REG_EDX,
+    TCG_REG_ECX
 };
 
-static const int tcg_target_call_iarg_regs[3] = { TCG_REG_EAX, TCG_REG_EDX, TCG_REG_ECX };
 static const int tcg_target_call_oarg_regs[2] = { TCG_REG_EAX, TCG_REG_EDX };
 
 static uint8_t *tb_ret_addr;

Signature · 1 line

1.6.6.1

Signature · 2 lines

Aurelien Jarno                          GPG: 1024D/F1BCDB73
[email protected]                 http://www.aurel32.net

Richard Henderson

author

2010-04-28 18:30 UTC #3 lore raw

PATCH 2/22 tcg-i386: Tidy initialization of tcg_target_call_clobber_regs.

In reply to Richard Henderson

Setting the registers one by one is easier to read, and gets optimized by the compiler just the same.

Signed-off-by Richard Henderson <[email protected]>

---
 tcg/i386/tcg-target.c |   11 ++++++-----
 1 files changed, 6 insertions(+), 5 deletions(-)

tcg/i386/tcg-target.c +6 -5

index f5c24f7..359f81b 100644
@@ -1365,11 +1365,12 @@ void tcg_target_init(TCGContext *s)
 #endif
 
     tcg_regset_set32(tcg_target_available_regs[TCG_TYPE_I32], 0, 0xff);
-    tcg_regset_set32(tcg_target_call_clobber_regs, 0,
-                     (1 << TCG_REG_EAX) | 
-                     (1 << TCG_REG_EDX) | 
-                     (1 << TCG_REG_ECX));
-    
+
+    tcg_regset_clear(tcg_target_call_clobber_regs);
+    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EAX);
+    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EDX);
+    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_ECX);
+
     tcg_regset_clear(s->reserved_regs);
     tcg_regset_set_reg(s->reserved_regs, TCG_REG_ESP);

Signature · 1 line

1.6.6.1

Aurelien Jarno

Applied

2010-05-19 17:37 UTC #4 lore raw

Re: [Qemu-devel] [PATCH 02/22] tcg-i386: Tidy initialization of tcg_target_call_clobber_regs.

Quoted 7 lines · Richard Henderson

Setting the registers one by one is easier to read, and gets optimized by the compiler just the same.
Signed-off-by Richard Henderson <[email protected]>
---
 tcg/i386/tcg-target.c |   11 ++++++-----
 1 files changed, 6 insertions(+), 5 deletions(-)

Thanks, applied.

Quoted 28 lines

tcg/i386/tcg-target.c +6 -5

index f5c24f7..359f81b 100644
@@ -1365,11 +1365,12 @@ void tcg_target_init(TCGContext *s)
 #endif
 
     tcg_regset_set32(tcg_target_available_regs[TCG_TYPE_I32], 0, 0xff);
-    tcg_regset_set32(tcg_target_call_clobber_regs, 0,
-                     (1 << TCG_REG_EAX) | 
-                     (1 << TCG_REG_EDX) | 
-                     (1 << TCG_REG_ECX));
-    
+
+    tcg_regset_clear(tcg_target_call_clobber_regs);
+    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EAX);
+    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EDX);
+    tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_ECX);
+
     tcg_regset_clear(s->reserved_regs);
     tcg_regset_set_reg(s->reserved_regs, TCG_REG_ESP);

Signature · 1 line

1.6.6.1

Signature · 2 lines

Aurelien Jarno                          GPG: 1024D/F1BCDB73
[email protected]                 http://www.aurel32.net

Richard Henderson

author

2010-04-28 18:30 UTC #5 lore raw

PATCH 3/22 tcg-i386: Tidy ext8u and ext16u operations.

In reply to Richard Henderson

Define OPC_MOVZBL and OPC_MOVZWL. Factor opcode emission to separate functions. Don't restrict the input register to the low 4 "q" registers; emit an AND instead if needed.

Signed-off-by Richard Henderson <[email protected]>

---
 tcg/i386/tcg-target.c |   68 ++++++++++++++++++++++++++++++------------------
 1 files changed, 42 insertions(+), 26 deletions(-)

tcg/i386/tcg-target.c +42 ~~-26~~

index 359f81b..2cc1191 100644
@@ -161,6 +161,11 @@ static inline int tcg_target_const_match(tcg_target_long val,
         return 0;
 }
 
+#define P_EXT   0x100 /* 0x0f opcode prefix */
+
+#define OPC_MOVZBL	(0xb6 | P_EXT)
+#define OPC_MOVZWL	(0xb7 | P_EXT)
+
 #define ARITH_ADD 0
 #define ARITH_OR  1
 #define ARITH_ADC 2

@@ -194,8 +199,6 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define JCC_JLE 0xe
 #define JCC_JG  0xf
 
-#define P_EXT   0x100 /* 0x0f opcode prefix */
-
 static const uint8_t tcg_cond_to_jcc[10] = {
     [TCG_COND_EQ] = JCC_JE,
     [TCG_COND_NE] = JCC_JNE,

@@ -288,6 +291,27 @@ static inline void tcg_out_st(TCGContext *s, TCGType type, int arg,
     tcg_out_modrm_offset(s, 0x89, arg, arg1, arg2);
 }
 
+static void tcg_out_ext8u(TCGContext *s, int dest, int src)
+{
+    if (src >= 4) {
+        tcg_out_mov(s, dest, src);
+        if (dest >= 4) {
+            tcg_out_modrm(s, 0x81, ARITH_AND, dest);
+            tcg_out32(s, 0xff);
+            return;
+        }
+        src = dest;
+    }
+    /* movzbl */
+    tcg_out_modrm(s, OPC_MOVZBL, dest, src);
+}
+
+static inline void tcg_out_ext16u(TCGContext *s, int dest, int src)
+{
+    /* movzwl */
+    tcg_out_modrm(s, OPC_MOVZWL, dest, src);
+}
+
 static inline void tgen_arithi(TCGContext *s, int c, int r0, int32_t val, int cf)
 {
     if (!cf && ((c == ARITH_ADD && val == 1) || (c == ARITH_SUB && val == -1))) {

@@ -300,11 +324,9 @@ static inline void tgen_arithi(TCGContext *s, int c, int r0, int32_t val, int cf
         tcg_out_modrm(s, 0x83, c, r0);
         tcg_out8(s, val);
     } else if (c == ARITH_AND && val == 0xffu && r0 < 4) {
-        /* movzbl */
-        tcg_out_modrm(s, 0xb6 | P_EXT, r0, r0);
+        tcg_out_ext8u(s, r0, r0);
     } else if (c == ARITH_AND && val == 0xffffu) {
-        /* movzwl */
-        tcg_out_modrm(s, 0xb7 | P_EXT, r0, r0);
+        tcg_out_ext16u(s, r0, r0);
     } else {
         tcg_out_modrm(s, 0x81, c, r0);
         tcg_out32(s, val);

@@ -645,12 +667,10 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
         tcg_out_modrm(s, 0xbf | P_EXT, data_reg, TCG_REG_EAX);
         break;
     case 0:
-        /* movzbl */
-        tcg_out_modrm(s, 0xb6 | P_EXT, data_reg, TCG_REG_EAX);
+        tcg_out_ext8u(s, data_reg, TCG_REG_EAX);
         break;
     case 1:
-        /* movzwl */
-        tcg_out_modrm(s, 0xb7 | P_EXT, data_reg, TCG_REG_EAX);
+        tcg_out_ext16u(s, data_reg, TCG_REG_EAX);
         break;
     case 2:
     default:

@@ -690,7 +710,7 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
     switch(opc) {
     case 0:
         /* movzbl */
-        tcg_out_modrm_offset(s, 0xb6 | P_EXT, data_reg, r0, GUEST_BASE);
+        tcg_out_modrm_offset(s, OPC_MOVZBL, data_reg, r0, GUEST_BASE);
         break;
     case 0 | 4:
         /* movsbl */

@@ -698,7 +718,7 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
         break;
     case 1:
         /* movzwl */
-        tcg_out_modrm_offset(s, 0xb7 | P_EXT, data_reg, r0, GUEST_BASE);
+        tcg_out_modrm_offset(s, OPC_MOVZWL, data_reg, r0, GUEST_BASE);
         if (bswap) {
             /* rolw $8, data_reg */
             tcg_out8(s, 0x66);

@@ -850,12 +870,10 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
     } else {
         switch(opc) {
         case 0:
-            /* movzbl */
-            tcg_out_modrm(s, 0xb6 | P_EXT, TCG_REG_EDX, data_reg);
+            tcg_out_ext8u(s, TCG_REG_EDX, data_reg);
             break;
         case 1:
-            /* movzwl */
-            tcg_out_modrm(s, 0xb7 | P_EXT, TCG_REG_EDX, data_reg);
+            tcg_out_ext16u(s, TCG_REG_EDX, data_reg);
             break;
         case 2:
             tcg_out_mov(s, TCG_REG_EDX, data_reg);

@@ -881,12 +899,10 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
         tcg_out_mov(s, TCG_REG_EDX, addr_reg2);
         switch(opc) {
         case 0:
-            /* movzbl */
-            tcg_out_modrm(s, 0xb6 | P_EXT, TCG_REG_ECX, data_reg);
+            tcg_out_ext8u(s, TCG_REG_ECX, data_reg);
             break;
         case 1:
-            /* movzwl */
-            tcg_out_modrm(s, 0xb7 | P_EXT, TCG_REG_ECX, data_reg);
+            tcg_out_ext16u(s, TCG_REG_ECX, data_reg);
             break;
         case 2:
             tcg_out_mov(s, TCG_REG_ECX, data_reg);

@@ -1022,7 +1038,7 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
         break;
     case INDEX_op_ld8u_i32:
         /* movzbl */
-        tcg_out_modrm_offset(s, 0xb6 | P_EXT, args[0], args[1], args[2]);
+        tcg_out_modrm_offset(s, OPC_MOVZBL, args[0], args[1], args[2]);
         break;
     case INDEX_op_ld8s_i32:
         /* movsbl */

@@ -1030,7 +1046,7 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
         break;
     case INDEX_op_ld16u_i32:
         /* movzwl */
-        tcg_out_modrm_offset(s, 0xb7 | P_EXT, args[0], args[1], args[2]);
+        tcg_out_modrm_offset(s, OPC_MOVZWL, args[0], args[1], args[2]);
         break;
     case INDEX_op_ld16s_i32:
         /* movswl */

@@ -1177,10 +1193,10 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
         tcg_out_modrm(s, 0xbf | P_EXT, args[0], args[1]);
         break;
     case INDEX_op_ext8u_i32:
-        tcg_out_modrm(s, 0xb6 | P_EXT, args[0], args[1]);
+        tcg_out_ext8u(s, args[0], args[1]);
         break;
     case INDEX_op_ext16u_i32:
-        tcg_out_modrm(s, 0xb7 | P_EXT, args[0], args[1]);
+        tcg_out_ext16u(s, args[0], args[1]);
         break;
 
     case INDEX_op_setcond_i32:

@@ -1275,8 +1291,8 @@ static const TCGTargetOpDef x86_op_defs[] = {
 
     { INDEX_op_ext8s_i32, { "r", "q" } },
     { INDEX_op_ext16s_i32, { "r", "r" } },
-    { INDEX_op_ext8u_i32, { "r", "q"} },
-    { INDEX_op_ext16u_i32, { "r", "r"} },
+    { INDEX_op_ext8u_i32, { "r", "r" } },
+    { INDEX_op_ext16u_i32, { "r", "r" } },
 
     { INDEX_op_setcond_i32, { "q", "r", "ri" } },
     { INDEX_op_setcond2_i32, { "r", "r", "r", "ri", "ri" } },

Signature · 1 line

1.6.6.1

Aurelien Jarno

2010-05-19 17:37 UTC #6 lore raw

Re: [Qemu-devel] [PATCH 03/22] tcg-i386: Tidy ext8u and ext16u operations.

Hi,

Quoted 3 lines · Richard Henderson

Define OPC_MOVZBL and OPC_MOVZWL. Factor opcode emission to separate functions. Don't restrict the input register to the low 4 "q" registers; emit an AND instead if needed.

I am fine about the cleaning part, but I don't know what to think about the constraints change.

The reg allocator is able to issue move if needed, so the only improvement this patch is for doing a ext8u on both "q" registers.

OTOH the reg allocator knows this situation and will try to avoid this situation during the allocation. Cheating on the reg allocator might have some wrong effects, especially after your patch "Allocate call-saved registers first". I am thinking of the scenario where the value is in memory (which is likely to be the case given the limited number of registers), it will be likely loaded in a "r" register (they are now at the top priority), and then ext8u will be called, which will issue "mov" + "and" instructions instead of a "movzbl" instruction.

If there are still cases to optimize I think it should be done in the reg allocator instead, so it could benefit all ops and all targets. I have started to play on that with Laurent Desnogues, I have a few more ideas how to improve it, but unfortunately I don't have time to code them.

All of that is purely theoretical. Do you know how does it behave in practice?

Quoted 183 lines

Signed-off-by Richard Henderson <[email protected]>

---
 tcg/i386/tcg-target.c |   68 ++++++++++++++++++++++++++++++------------------
 1 files changed, 42 insertions(+), 26 deletions(-)

tcg/i386/tcg-target.c +42 ~~-26~~

index 359f81b..2cc1191 100644
@@ -161,6 +161,11 @@ static inline int tcg_target_const_match(tcg_target_long val,
         return 0;
 }
 
+#define P_EXT   0x100 /* 0x0f opcode prefix */
+
+#define OPC_MOVZBL	(0xb6 | P_EXT)
+#define OPC_MOVZWL	(0xb7 | P_EXT)
+
 #define ARITH_ADD 0
 #define ARITH_OR  1
 #define ARITH_ADC 2

@@ -194,8 +199,6 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define JCC_JLE 0xe
 #define JCC_JG  0xf
 
-#define P_EXT   0x100 /* 0x0f opcode prefix */
-
 static const uint8_t tcg_cond_to_jcc[10] = {
     [TCG_COND_EQ] = JCC_JE,
     [TCG_COND_NE] = JCC_JNE,

@@ -288,6 +291,27 @@ static inline void tcg_out_st(TCGContext *s, TCGType type, int arg,
     tcg_out_modrm_offset(s, 0x89, arg, arg1, arg2);
 }
 
+static void tcg_out_ext8u(TCGContext *s, int dest, int src)
+{
+    if (src >= 4) {
+        tcg_out_mov(s, dest, src);
+        if (dest >= 4) {
+            tcg_out_modrm(s, 0x81, ARITH_AND, dest);
+            tcg_out32(s, 0xff);
+            return;
+        }
+        src = dest;
+    }
+    /* movzbl */
+    tcg_out_modrm(s, OPC_MOVZBL, dest, src);
+}
+
+static inline void tcg_out_ext16u(TCGContext *s, int dest, int src)
+{
+    /* movzwl */
+    tcg_out_modrm(s, OPC_MOVZWL, dest, src);
+}
+
 static inline void tgen_arithi(TCGContext *s, int c, int r0, int32_t val, int cf)
 {
     if (!cf && ((c == ARITH_ADD && val == 1) || (c == ARITH_SUB && val == -1))) {

@@ -300,11 +324,9 @@ static inline void tgen_arithi(TCGContext *s, int c, int r0, int32_t val, int cf
         tcg_out_modrm(s, 0x83, c, r0);
         tcg_out8(s, val);
     } else if (c == ARITH_AND && val == 0xffu && r0 < 4) {
-        /* movzbl */
-        tcg_out_modrm(s, 0xb6 | P_EXT, r0, r0);
+        tcg_out_ext8u(s, r0, r0);
     } else if (c == ARITH_AND && val == 0xffffu) {
-        /* movzwl */
-        tcg_out_modrm(s, 0xb7 | P_EXT, r0, r0);
+        tcg_out_ext16u(s, r0, r0);
     } else {
         tcg_out_modrm(s, 0x81, c, r0);
         tcg_out32(s, val);

@@ -645,12 +667,10 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
         tcg_out_modrm(s, 0xbf | P_EXT, data_reg, TCG_REG_EAX);
         break;
     case 0:
-        /* movzbl */
-        tcg_out_modrm(s, 0xb6 | P_EXT, data_reg, TCG_REG_EAX);
+        tcg_out_ext8u(s, data_reg, TCG_REG_EAX);
         break;
     case 1:
-        /* movzwl */
-        tcg_out_modrm(s, 0xb7 | P_EXT, data_reg, TCG_REG_EAX);
+        tcg_out_ext16u(s, data_reg, TCG_REG_EAX);
         break;
     case 2:
     default:

@@ -690,7 +710,7 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
     switch(opc) {
     case 0:
         /* movzbl */
-        tcg_out_modrm_offset(s, 0xb6 | P_EXT, data_reg, r0, GUEST_BASE);
+        tcg_out_modrm_offset(s, OPC_MOVZBL, data_reg, r0, GUEST_BASE);
         break;
     case 0 | 4:
         /* movsbl */

@@ -698,7 +718,7 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
         break;
     case 1:
         /* movzwl */
-        tcg_out_modrm_offset(s, 0xb7 | P_EXT, data_reg, r0, GUEST_BASE);
+        tcg_out_modrm_offset(s, OPC_MOVZWL, data_reg, r0, GUEST_BASE);
         if (bswap) {
             /* rolw $8, data_reg */
             tcg_out8(s, 0x66);

@@ -850,12 +870,10 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
     } else {
         switch(opc) {
         case 0:
-            /* movzbl */
-            tcg_out_modrm(s, 0xb6 | P_EXT, TCG_REG_EDX, data_reg);
+            tcg_out_ext8u(s, TCG_REG_EDX, data_reg);
             break;
         case 1:
-            /* movzwl */
-            tcg_out_modrm(s, 0xb7 | P_EXT, TCG_REG_EDX, data_reg);
+            tcg_out_ext16u(s, TCG_REG_EDX, data_reg);
             break;
         case 2:
             tcg_out_mov(s, TCG_REG_EDX, data_reg);

@@ -881,12 +899,10 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
         tcg_out_mov(s, TCG_REG_EDX, addr_reg2);
         switch(opc) {
         case 0:
-            /* movzbl */
-            tcg_out_modrm(s, 0xb6 | P_EXT, TCG_REG_ECX, data_reg);
+            tcg_out_ext8u(s, TCG_REG_ECX, data_reg);
             break;
         case 1:
-            /* movzwl */
-            tcg_out_modrm(s, 0xb7 | P_EXT, TCG_REG_ECX, data_reg);
+            tcg_out_ext16u(s, TCG_REG_ECX, data_reg);
             break;
         case 2:
             tcg_out_mov(s, TCG_REG_ECX, data_reg);

@@ -1022,7 +1038,7 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
         break;
     case INDEX_op_ld8u_i32:
         /* movzbl */
-        tcg_out_modrm_offset(s, 0xb6 | P_EXT, args[0], args[1], args[2]);
+        tcg_out_modrm_offset(s, OPC_MOVZBL, args[0], args[1], args[2]);
         break;
     case INDEX_op_ld8s_i32:
         /* movsbl */

@@ -1030,7 +1046,7 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
         break;
     case INDEX_op_ld16u_i32:
         /* movzwl */
-        tcg_out_modrm_offset(s, 0xb7 | P_EXT, args[0], args[1], args[2]);
+        tcg_out_modrm_offset(s, OPC_MOVZWL, args[0], args[1], args[2]);
         break;
     case INDEX_op_ld16s_i32:
         /* movswl */

@@ -1177,10 +1193,10 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
         tcg_out_modrm(s, 0xbf | P_EXT, args[0], args[1]);
         break;
     case INDEX_op_ext8u_i32:
-        tcg_out_modrm(s, 0xb6 | P_EXT, args[0], args[1]);
+        tcg_out_ext8u(s, args[0], args[1]);
         break;
     case INDEX_op_ext16u_i32:
-        tcg_out_modrm(s, 0xb7 | P_EXT, args[0], args[1]);
+        tcg_out_ext16u(s, args[0], args[1]);
         break;
 
     case INDEX_op_setcond_i32:

@@ -1275,8 +1291,8 @@ static const TCGTargetOpDef x86_op_defs[] = {
 
     { INDEX_op_ext8s_i32, { "r", "q" } },
     { INDEX_op_ext16s_i32, { "r", "r" } },
-    { INDEX_op_ext8u_i32, { "r", "q"} },
-    { INDEX_op_ext16u_i32, { "r", "r"} },
+    { INDEX_op_ext8u_i32, { "r", "r" } },
+    { INDEX_op_ext16u_i32, { "r", "r" } },
 
     { INDEX_op_setcond_i32, { "q", "r", "ri" } },
     { INDEX_op_setcond2_i32, { "r", "r", "r", "ri", "ri" } },

Signature · 1 line

1.6.6.1

Signature · 2 lines

Aurelien Jarno                          GPG: 1024D/F1BCDB73
[email protected]                 http://www.aurel32.net

Richard Henderson

author

2010-05-19 18:31 UTC #7 lore raw

Re: [Qemu-devel] [PATCH 03/22] tcg-i386: Tidy ext8u and ext16u operations.

Quoted 11 lines · Aurelien Jarno

The reg allocator is able to issue move if needed, so the only improvement this patch is for doing a ext8u on both "q" registers.
OTOH the reg allocator knows this situation and will try to avoid this situation during the allocation. Cheating on the reg allocator might have some wrong effects, especially after your patch "Allocate call-saved registers first". I am thinking of the scenario where the value is in memory (which is likely to be the case given the limited number of registers), it will be likely loaded in a "r" register (they are now at the top priority), and then ext8u will be called, which will issue "mov" + "and" instructions instead of a "movzbl" instruction.

The case I was concerned with is the fact that if we have a value allocated to, say, %esi, and we need to to an ext8u, then the register allocator has been told that it must move the value to a "q" register in order to perform the movzbl. In this case, the new code will simply emit the andl.

I.e. the real problem is that we've told the register allocator one way that the extend can be implemented, but not every way.

Quoted 2 lines

All of that is purely theoretical. Do you know how does it behave in practice?

Picking the i386 target since it seems to use more extensions than any other target, from linux-user-test -d op_opt,out_asm i386/ls:

There are 176 instances of ext8u.
Of those, 83 instances are in-place, i.e. "ext8u_i32 tmp0,tmp0"

I examined the first 2 dozen appearances in the output assembly:

There are several instances of the value being in an "r" register:

shr_i32 tmp1,edx,tmp13
ext8u_i32 tmp1,tmp1
=>

0x601c5468:  shr    $0x8,%edi
0x601c546b:  and    $0xff,%edi

All of the instances that I looked at that were not in-place happened to already be using a "q" register -- usually %ebx. I assume that's because we place %ebx as the first allocation register and that's just how things happen to work out once we've flushed the registers before the qemu_ld.

qemu_ld8u tmp0,tmp2,$0xffffffff
ext8u_i32 tmp13,tmp0
=>
0x601c82f9: movzbl (%esi),%ebx
0x601c82fc: movzbl %bl,%ebx

Aurelien Jarno

2010-05-20 13:39 UTC #8 lore raw

Re: [Qemu-devel] [PATCH 03/22] tcg-i386: Tidy ext8u and ext16u operations.

Quoted 47 lines · Richard Henderson

Quoted 11 lines · Aurelien Jarno
The reg allocator is able to issue move if needed, so the only improvement this patch is for doing a ext8u on both "q" registers.
OTOH the reg allocator knows this situation and will try to avoid this situation during the allocation. Cheating on the reg allocator might have some wrong effects, especially after your patch "Allocate call-saved registers first". I am thinking of the scenario where the value is in memory (which is likely to be the case given the limited number of registers), it will be likely loaded in a "r" register (they are now at the top priority), and then ext8u will be called, which will issue "mov" + "and" instructions instead of a "movzbl" instruction.
The case I was concerned with is the fact that if we have a value allocated to, say, %esi, and we need to to an ext8u, then the register allocator has been told that it must move the value to a "q" register in order to perform the movzbl. In this case, the new code will simply emit the andl.
I.e. the real problem is that we've told the register allocator one way that the extend can be implemented, but not every way.
Quoted 2 lines
All of that is purely theoretical. Do you know how does it behave in practice?
Picking the i386 target since it seems to use more extensions than any other target, from linux-user-test -d op_opt,out_asm i386/ls:
There are 176 instances of ext8u.
Of those, 83 instances are in-place, i.e. "ext8u_i32 tmp0,tmp0"
I examined the first 2 dozen appearances in the output assembly:
There are several instances of the value being in an "r" register:
shr_i32 tmp1,edx,tmp13
ext8u_i32 tmp1,tmp1
=>
0x601c5468:  shr    $0x8,%edi
0x601c546b:  and    $0xff,%edi
All of the instances that I looked at that were not in-place happened to already be using a "q" register -- usually %ebx. I assume that's because we place %ebx as the first allocation register and that's just how things happen to work out once we've flushed the registers before the qemu_ld.

Quoted 6 lines

qemu_ld8u tmp0,tmp2,$0xffffffff
ext8u_i32 tmp13,tmp0
=>
0x601c82f9: movzbl (%esi),%ebx
0x601c82fc: movzbl %bl,%ebx

Do you have tried to compare the generated code before and after your patch? I expect a few cases where your patch has some drawbacks, so I don't know if there is a net gain on the size of the translated code.

Signature · 2 lines

Aurelien Jarno	                        GPG: 1024D/F1BCDB73
[email protected]                 http://www.aurel32.net

Aurelien Jarno

2010-05-20 14:04 UTC #9 lore raw

Re: [Qemu-devel] [PATCH 03/22] tcg-i386: Tidy ext8u and ext16u operations.

Quoted 55 lines · Aurelien Jarno

Quoted 47 lines · Richard Henderson
Quoted 11 lines · Aurelien Jarno
The reg allocator is able to issue move if needed, so the only improvement this patch is for doing a ext8u on both "q" registers.
OTOH the reg allocator knows this situation and will try to avoid this situation during the allocation. Cheating on the reg allocator might have some wrong effects, especially after your patch "Allocate call-saved registers first". I am thinking of the scenario where the value is in memory (which is likely to be the case given the limited number of registers), it will be likely loaded in a "r" register (they are now at the top priority), and then ext8u will be called, which will issue "mov" + "and" instructions instead of a "movzbl" instruction.
The case I was concerned with is the fact that if we have a value allocated to, say, %esi, and we need to to an ext8u, then the register allocator has been told that it must move the value to a "q" register in order to perform the movzbl. In this case, the new code will simply emit the andl.
I.e. the real problem is that we've told the register allocator one way that the extend can be implemented, but not every way.
Quoted 2 lines
All of that is purely theoretical. Do you know how does it behave in practice?
Picking the i386 target since it seems to use more extensions than any other target, from linux-user-test -d op_opt,out_asm i386/ls:
There are 176 instances of ext8u.
Of those, 83 instances are in-place, i.e. "ext8u_i32 tmp0,tmp0"
I examined the first 2 dozen appearances in the output assembly:
There are several instances of the value being in an "r" register:
shr_i32 tmp1,edx,tmp13
ext8u_i32 tmp1,tmp1
=>
0x601c5468:  shr    $0x8,%edi
0x601c546b:  and    $0xff,%edi
All of the instances that I looked at that were not in-place happened to already be using a "q" register -- usually %ebx. I assume that's because we place %ebx as the first allocation register and that's just how things happen to work out once we've flushed the registers before the qemu_ld.
Quoted 6 lines
qemu_ld8u tmp0,tmp2,$0xffffffff
ext8u_i32 tmp13,tmp0
=>
0x601c82f9: movzbl (%esi),%ebx
0x601c82fc: movzbl %bl,%ebx

Quoted 4 lines

Do you have tried to compare the generated code before and after your patch? I expect a few cases where your patch has some drawbacks, so I don't know if there is a net gain on the size of the translated code.

I have done a quick test on /bin/ls.

       |  instr |  size  |
       +--------+--------+
before | 101305 | 344770 |
after  | 101258 | 344829 |

In short a small gain in the number of instructions, and a small loss in the size of the translated code.

Signature · 2 lines

Aurelien Jarno                          GPG: 1024D/F1BCDB73
[email protected]                 http://www.aurel32.net

Richard Henderson

author

2010-05-20 14:41 UTC #10 lore raw

Re: [Qemu-devel] [PATCH 03/22] tcg-i386: Tidy ext8u and ext16u operations.

Quoted 5 lines · Aurelien Jarno

Quoted 4 lines
Do you have tried to compare the generated code before and after your patch? I expect a few cases where your patch has some drawbacks, so I don't know if there is a net gain on the size of the translated code.

Quoted 8 lines

I have done a quick test on /bin/ls.
       |  instr |  size  |
       +--------+--------+
before | 101305 | 344770 |
after  | 101258 | 344829 |
In short a small gain in the number of instructions, and a small loss in the size of the translated code.

That was pretty much the test I would have done.

So where are we? Is the patch acceptable as-is, or should I be re-writing it without the constraints change?

Aurelien Jarno

2010-05-20 18:50 UTC #11 lore raw

Re: [Qemu-devel] [PATCH 03/22] tcg-i386: Tidy ext8u and ext16u operations.

Quoted 12 lines · Richard Henderson

Quoted 11 lines · Aurelien Jarno
Quoted 4 lines
Do you have tried to compare the generated code before and after your patch? I expect a few cases where your patch has some drawbacks, so I don't know if there is a net gain on the size of the translated code.
I have done a quick test on /bin/ls.
       |  instr |  size  |
       +--------+--------+
before | 101305 | 344770 |
after  | 101258 | 344829 |

Quoted 8 lines

Quoted 2 lines
In short a small gain in the number of instructions, and a small loss in the size of the translated code.
That was pretty much the test I would have done.
So where are we? Is the patch acceptable as-is, or should I be re-writing it without the constraints change?

Given the tests do not show a real improvement and given that it complexify the code generation, I don't think we should have such a patch. Could you please rewrite it without the constraints change?

Signature · 2 lines

Aurelien Jarno	                        GPG: 1024D/F1BCDB73
[email protected]                 http://www.aurel32.net

Richard Henderson

author

2010-04-28 18:30 UTC #12 lore raw

PATCH 4/22 tcg-i386: Tidy ext8s and ext16s operations.

In reply to Richard Henderson

Define OPC_MOVSBL and OPC_MOVSWL. Factor opcode emission to separate functions. Don't restrict the input register to the low 4 "q" registers; emit shifts instead if needed.

Signed-off-by Richard Henderson <[email protected]>

---
 tcg/i386/tcg-target.c |   49 +++++++++++++++++++++++++++++++++++++------------
 1 files changed, 37 insertions(+), 12 deletions(-)

tcg/i386/tcg-target.c +37 ~~-12~~

index 2cc1191..75b9915 100644
@@ -165,6 +165,8 @@ static inline int tcg_target_const_match(tcg_target_long val,
 
 #define OPC_MOVZBL	(0xb6 | P_EXT)
 #define OPC_MOVZWL	(0xb7 | P_EXT)
+#define OPC_MOVSBL	(0xbe | P_EXT)
+#define OPC_MOVSWL	(0xbf | P_EXT)
 
 #define ARITH_ADD 0
 #define ARITH_OR  1

@@ -306,12 +308,37 @@ static void tcg_out_ext8u(TCGContext *s, int dest, int src)
     tcg_out_modrm(s, OPC_MOVZBL, dest, src);
 }
 
+static void tcg_out_ext8s(TCGContext *s, int dest, int src)
+{
+    if (src >= 4) {
+        tcg_out_mov(s, dest, src);
+        if (dest >= 4) {
+            /* shl $24, dest */
+            tcg_out_modrm(s, 0xc1, SHIFT_SHL, dest);
+            tcg_out8(s, 24);
+            /* sar $24, dest */
+            tcg_out_modrm(s, 0xc1, SHIFT_SAR, dest);
+            tcg_out8(s, 24);
+            return;
+        }
+        src = dest;
+    }
+    /* movsbl */
+    tcg_out_modrm(s, OPC_MOVSBL, dest, src);
+}
+
 static inline void tcg_out_ext16u(TCGContext *s, int dest, int src)
 {
     /* movzwl */
     tcg_out_modrm(s, OPC_MOVZWL, dest, src);
 }
 
+static inline void tcg_out_ext16s(TCGContext *s, int dest, int src)
+{
+    /* movswl */
+    tcg_out_modrm(s, OPC_MOVSWL, dest, src);
+}
+
 static inline void tgen_arithi(TCGContext *s, int c, int r0, int32_t val, int cf)
 {
     if (!cf && ((c == ARITH_ADD && val == 1) || (c == ARITH_SUB && val == -1))) {

@@ -659,12 +686,10 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
 
     switch(opc) {
     case 0 | 4:
-        /* movsbl */
-        tcg_out_modrm(s, 0xbe | P_EXT, data_reg, TCG_REG_EAX);
+        tcg_out_ext8s(s, data_reg, TCG_REG_EAX);
         break;
     case 1 | 4:
-        /* movswl */
-        tcg_out_modrm(s, 0xbf | P_EXT, data_reg, TCG_REG_EAX);
+        tcg_out_ext16s(s, data_reg, TCG_REG_EAX);
         break;
     case 0:
         tcg_out_ext8u(s, data_reg, TCG_REG_EAX);

@@ -714,7 +739,7 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
         break;
     case 0 | 4:
         /* movsbl */
-        tcg_out_modrm_offset(s, 0xbe | P_EXT, data_reg, r0, GUEST_BASE);
+        tcg_out_modrm_offset(s, OPC_MOVSBL, data_reg, r0, GUEST_BASE);
         break;
     case 1:
         /* movzwl */

@@ -728,7 +753,7 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
         break;
     case 1 | 4:
         /* movswl */
-        tcg_out_modrm_offset(s, 0xbf | P_EXT, data_reg, r0, GUEST_BASE);
+        tcg_out_modrm_offset(s, OPC_MOVSWL, data_reg, r0, GUEST_BASE);
         if (bswap) {
             /* rolw $8, data_reg */
             tcg_out8(s, 0x66);

@@ -736,7 +761,7 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
             tcg_out8(s, 8);
 
             /* movswl data_reg, data_reg */
-            tcg_out_modrm(s, 0xbf | P_EXT, data_reg, data_reg);
+            tcg_out_modrm(s, OPC_MOVSWL, data_reg, data_reg);
         }
         break;
     case 2:

@@ -1042,7 +1067,7 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
         break;
     case INDEX_op_ld8s_i32:
         /* movsbl */
-        tcg_out_modrm_offset(s, 0xbe | P_EXT, args[0], args[1], args[2]);
+        tcg_out_modrm_offset(s, OPC_MOVSBL, args[0], args[1], args[2]);
         break;
     case INDEX_op_ld16u_i32:
         /* movzwl */

@@ -1050,7 +1075,7 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
         break;
     case INDEX_op_ld16s_i32:
         /* movswl */
-        tcg_out_modrm_offset(s, 0xbf | P_EXT, args[0], args[1], args[2]);
+        tcg_out_modrm_offset(s, OPC_MOVSWL, args[0], args[1], args[2]);
         break;
     case INDEX_op_ld_i32:
         /* movl */

@@ -1187,10 +1212,10 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
         break;
 
     case INDEX_op_ext8s_i32:
-        tcg_out_modrm(s, 0xbe | P_EXT, args[0], args[1]);
+        tcg_out_ext8s(s, args[0], args[1]);
         break;
     case INDEX_op_ext16s_i32:
-        tcg_out_modrm(s, 0xbf | P_EXT, args[0], args[1]);
+        tcg_out_ext16s(s, args[0], args[1]);
         break;
     case INDEX_op_ext8u_i32:
         tcg_out_ext8u(s, args[0], args[1]);

@@ -1289,7 +1314,7 @@ static const TCGTargetOpDef x86_op_defs[] = {
 
     { INDEX_op_not_i32, { "r", "0" } },
 
-    { INDEX_op_ext8s_i32, { "r", "q" } },
+    { INDEX_op_ext8s_i32, { "r", "r" } },
     { INDEX_op_ext16s_i32, { "r", "r" } },
     { INDEX_op_ext8u_i32, { "r", "r" } },
     { INDEX_op_ext16u_i32, { "r", "r" } },

Signature · 1 line

1.6.6.1

Aurelien Jarno

2010-05-20 18:52 UTC #13 lore raw

Re: [Qemu-devel] [PATCH 04/22] tcg-i386: Tidy ext8s and ext16s operations.

Quoted 4 lines · Richard Henderson

Define OPC_MOVSBL and OPC_MOVSWL. Factor opcode emission to separate functions. Don't restrict the input register to the low 4 "q" registers; emit shifts instead if needed.

Given this patch is of the same type as the previous one, I have also benchmarked it, here are the results:

        |  instr |  size  |
        +--------+--------+
 before | 101258 | 344829 |
 after  | 101258 | 344833 |

This time the patch clearly doesn't bring an improvement, so I think it should also be rewritten without the constraints change.

Signature · 2 lines

Aurelien Jarno	                        GPG: 1024D/F1BCDB73
[email protected]                 http://www.aurel32.net

Richard Henderson

author

2010-04-28 18:30 UTC #14 lore raw

PATCH 5/22 tcg-i386: Tidy bswap operations.

In reply to Richard Henderson

Define OPC_BSWAP. Factor opcode emission to separate functions.

Signed-off-by Richard Henderson <[email protected]>

---
 tcg/i386/tcg-target.c |   52 ++++++++++++++++++++++--------------------------
 1 files changed, 24 insertions(+), 28 deletions(-)

tcg/i386/tcg-target.c +24 ~~-28~~

index 75b9915..094b2da 100644
@@ -163,6 +163,7 @@ static inline int tcg_target_const_match(tcg_target_long val,
 
 #define P_EXT   0x100 /* 0x0f opcode prefix */
 
+#define OPC_BSWAP	(0xc8 | P_EXT)
 #define OPC_MOVZBL	(0xb6 | P_EXT)
 #define OPC_MOVZWL	(0xb7 | P_EXT)
 #define OPC_MOVSBL	(0xbe | P_EXT)

@@ -339,6 +340,18 @@ static inline void tcg_out_ext16s(TCGContext *s, int dest, int src)
     tcg_out_modrm(s, OPC_MOVSWL, dest, src);
 }
 
+static inline void tcg_out_bswap32(TCGContext *s, int reg)
+{
+    tcg_out_opc(s, OPC_BSWAP + reg);
+}
+
+static inline void tcg_out_rolw_8(TCGContext *s, int reg)
+{
+    tcg_out8(s, 0x66); 
+    tcg_out_modrm(s, 0xc1, 0, reg);
+    tcg_out8(s, 8);
+}
+
 static inline void tgen_arithi(TCGContext *s, int c, int r0, int32_t val, int cf)
 {
     if (!cf && ((c == ARITH_ADD && val == 1) || (c == ARITH_SUB && val == -1))) {

@@ -745,22 +758,14 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
         /* movzwl */
         tcg_out_modrm_offset(s, OPC_MOVZWL, data_reg, r0, GUEST_BASE);
         if (bswap) {
-            /* rolw $8, data_reg */
-            tcg_out8(s, 0x66); 
-            tcg_out_modrm(s, 0xc1, 0, data_reg);
-            tcg_out8(s, 8);
+            tcg_out_rolw_8(s, data_reg);
         }
         break;
     case 1 | 4:
         /* movswl */
         tcg_out_modrm_offset(s, OPC_MOVSWL, data_reg, r0, GUEST_BASE);
         if (bswap) {
-            /* rolw $8, data_reg */
-            tcg_out8(s, 0x66); 
-            tcg_out_modrm(s, 0xc1, 0, data_reg);
-            tcg_out8(s, 8);
-
-            /* movswl data_reg, data_reg */
+            tcg_out_rolw_8(s, data_reg);
             tcg_out_modrm(s, OPC_MOVSWL, data_reg, data_reg);
         }
         break;

@@ -768,8 +773,7 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
         /* movl (r0), data_reg */
         tcg_out_modrm_offset(s, 0x8b, data_reg, r0, GUEST_BASE);
         if (bswap) {
-            /* bswap */
-            tcg_out_opc(s, (0xc8 + data_reg) | P_EXT);
+            tcg_out_bswap32(s, data_reg);
         }
         break;
     case 3:

@@ -786,11 +790,10 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
             tcg_out_modrm_offset(s, 0x8b, data_reg2, r0, GUEST_BASE + 4);
         } else {
             tcg_out_modrm_offset(s, 0x8b, data_reg, r0, GUEST_BASE + 4);
-            tcg_out_opc(s, (0xc8 + data_reg) | P_EXT);
+            tcg_out_bswap32(s, data_reg);
 
             tcg_out_modrm_offset(s, 0x8b, data_reg2, r0, GUEST_BASE);
-            /* bswap */
-            tcg_out_opc(s, (0xc8 + data_reg2) | P_EXT);
+            tcg_out_bswap32(s, data_reg2);
         }
         break;
     default:

@@ -970,9 +973,7 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
     case 1:
         if (bswap) {
             tcg_out_mov(s, r1, data_reg);
-            tcg_out8(s, 0x66); /* rolw $8, %ecx */
-            tcg_out_modrm(s, 0xc1, 0, r1);
-            tcg_out8(s, 8);
+            tcg_out_rolw_8(s, r1);
             data_reg = r1;
         }
         /* movw */

@@ -982,8 +983,7 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
     case 2:
         if (bswap) {
             tcg_out_mov(s, r1, data_reg);
-            /* bswap data_reg */
-            tcg_out_opc(s, (0xc8 + r1) | P_EXT);
+            tcg_out_bswap32(s, r1);
             data_reg = r1;
         }
         /* movl */

@@ -992,12 +992,10 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
     case 3:
         if (bswap) {
             tcg_out_mov(s, r1, data_reg2);
-            /* bswap data_reg */
-            tcg_out_opc(s, (0xc8 + r1) | P_EXT);
+            tcg_out_bswap32(s, r1);
             tcg_out_modrm_offset(s, 0x89, r1, r0, GUEST_BASE);
             tcg_out_mov(s, r1, data_reg);
-            /* bswap data_reg */
-            tcg_out_opc(s, (0xc8 + r1) | P_EXT);
+            tcg_out_bswap32(s, r1);
             tcg_out_modrm_offset(s, 0x89, r1, r0, GUEST_BASE + 4);
         } else {
             tcg_out_modrm_offset(s, 0x89, data_reg, r0, GUEST_BASE);

@@ -1195,12 +1193,10 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
         break;
 
     case INDEX_op_bswap16_i32:
-        tcg_out8(s, 0x66);
-        tcg_out_modrm(s, 0xc1, SHIFT_ROL, args[0]);
-        tcg_out8(s, 8);
+        tcg_out_rolw_8(s, args[0]);
         break;
     case INDEX_op_bswap32_i32:
-        tcg_out_opc(s, (0xc8 + args[0]) | P_EXT);
+        tcg_out_bswap32(s, args[0]);
         break;
 
     case INDEX_op_neg_i32:

Signature · 1 line

1.6.6.1

Richard Henderson

author

2010-04-28 18:30 UTC #15 lore raw

PATCH 6/22 tcg-i386: Tidy shift operations.

In reply to Richard Henderson

Define OPC_SHIFT_{1,Ib,cl}. Factor opcode emission to a function.

Signed-off-by Richard Henderson <[email protected]>

---
 tcg/i386/tcg-target.c |   43 ++++++++++++++++++++++---------------------
 1 files changed, 22 insertions(+), 21 deletions(-)

tcg/i386/tcg-target.c +22 ~~-21~~

index 094b2da..c0133be 100644
@@ -168,6 +168,9 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define OPC_MOVZWL	(0xb7 | P_EXT)
 #define OPC_MOVSBL	(0xbe | P_EXT)
 #define OPC_MOVSWL	(0xbf | P_EXT)
+#define OPC_SHIFT_1	(0xd1)
+#define OPC_SHIFT_Ib	(0xc1)
+#define OPC_SHIFT_cl	(0xd3)
 
 #define ARITH_ADD 0
 #define ARITH_OR  1

@@ -294,6 +297,16 @@ static inline void tcg_out_st(TCGContext *s, TCGType type, int arg,
     tcg_out_modrm_offset(s, 0x89, arg, arg1, arg2);
 }
 
+static void tcg_out_shifti(TCGContext *s, int subopc, int reg, int count)
+{
+    if (count == 1) {
+        tcg_out_modrm(s, OPC_SHIFT_1, subopc, reg);
+    } else {
+        tcg_out_modrm(s, OPC_SHIFT_Ib, subopc, reg);
+        tcg_out8(s, count);
+    }
+}
+
 static void tcg_out_ext8u(TCGContext *s, int dest, int src)
 {
     if (src >= 4) {

@@ -314,12 +327,8 @@ static void tcg_out_ext8s(TCGContext *s, int dest, int src)
     if (src >= 4) {
         tcg_out_mov(s, dest, src);
         if (dest >= 4) {
-            /* shl $24, dest */
-            tcg_out_modrm(s, 0xc1, SHIFT_SHL, dest);
-            tcg_out8(s, 24);
-            /* sar $24, dest */
-            tcg_out_modrm(s, 0xc1, SHIFT_SAR, dest);
-            tcg_out8(s, 24);
+            tcg_out_shifti(s, SHIFT_SHL, dest, 24);
+            tcg_out_shifti(s, SHIFT_SAR, dest, 24);
             return;
         }
         src = dest;

@@ -348,8 +357,7 @@ static inline void tcg_out_bswap32(TCGContext *s, int reg)
 static inline void tcg_out_rolw_8(TCGContext *s, int reg)
 {
     tcg_out8(s, 0x66); 
-    tcg_out_modrm(s, 0xc1, 0, reg);
-    tcg_out8(s, 8);
+    tcg_out_shifti(s, SHIFT_ROL, reg, 8);
 }
 
 static inline void tgen_arithi(TCGContext *s, int c, int r0, int32_t val, int cf)

@@ -644,9 +652,8 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
 
     tcg_out_mov(s, r0, addr_reg); 
  
-    tcg_out_modrm(s, 0xc1, 5, r1); /* shr $x, r1 */
-    tcg_out8(s, TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS); 
-    
+    tcg_out_shifti(s, SHIFT_SHR, r1, TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS); 
+
     tcg_out_modrm(s, 0x81, 4, r0); /* andl $x, r0 */
     tcg_out32(s, TARGET_PAGE_MASK | ((1 << s_bits) - 1));

@@ -842,9 +849,8 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
 
     tcg_out_mov(s, r0, addr_reg); 
  
-    tcg_out_modrm(s, 0xc1, 5, r1); /* shr $x, r1 */
-    tcg_out8(s, TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS); 
-    
+    tcg_out_shifti(s, SHIFT_SHR, r1, TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS); 
+
     tcg_out_modrm(s, 0x81, 4, r0); /* andl $x, r0 */
     tcg_out32(s, TARGET_PAGE_MASK | ((1 << s_bits) - 1));

@@ -1141,14 +1147,9 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
         c = SHIFT_SHL;
     gen_shift32:
         if (const_args[2]) {
-            if (args[2] == 1) {
-                tcg_out_modrm(s, 0xd1, c, args[0]);
-            } else {
-                tcg_out_modrm(s, 0xc1, c, args[0]);
-                tcg_out8(s, args[2]);
-            }
+            tcg_out_shifti(s, c, args[0], args[2]);
         } else {
-            tcg_out_modrm(s, 0xd3, c, args[0]);
+            tcg_out_modrm(s, OPC_SHIFT_cl, c, args[0]);
         }
         break;
     case INDEX_op_shr_i32:

Signature · 1 line

1.6.6.1

Richard Henderson

author

2010-04-28 18:30 UTC #16 lore raw

PATCH 7/22 tcg-i386: Tidy move operations.

In reply to Richard Henderson

Define OPC_MOVB* and OPC_MOVL*; use them throughout. Use tcg_out_ld/st instead of bare tcg_out_modrm_offset when it makes sense.

Signed-off-by Richard Henderson <[email protected]>

---
 tcg/i386/tcg-target.c |   51 +++++++++++++++++++++++-------------------------
 1 files changed, 24 insertions(+), 27 deletions(-)

tcg/i386/tcg-target.c +24 ~~-27~~

index c0133be..a10f6ca 100644
@@ -164,6 +164,9 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define P_EXT   0x100 /* 0x0f opcode prefix */
 
 #define OPC_BSWAP	(0xc8 | P_EXT)
+#define OPC_MOVB_EvGv	(0x88)		/* stores, more or less */
+#define OPC_MOVL_EvGv	(0x89)		/* stores, more or less */
+#define OPC_MOVL_GvEv	(0x8b)		/* loads, more or less */
 #define OPC_MOVZBL	(0xb6 | P_EXT)
 #define OPC_MOVZWL	(0xb7 | P_EXT)
 #define OPC_MOVSBL	(0xbe | P_EXT)

@@ -267,8 +270,9 @@ static inline void tcg_out_modrm_offset(TCGContext *s, int opc, int r, int rm,
 
 static inline void tcg_out_mov(TCGContext *s, int ret, int arg)
 {
-    if (arg != ret)
-        tcg_out_modrm(s, 0x8b, ret, arg);
+    if (arg != ret) {
+        tcg_out_modrm(s, OPC_MOVL_GvEv, ret, arg);
+    }
 }
 
 static inline void tcg_out_movi(TCGContext *s, TCGType type,

@@ -286,15 +290,13 @@ static inline void tcg_out_movi(TCGContext *s, TCGType type,
 static inline void tcg_out_ld(TCGContext *s, TCGType type, int ret,
                               int arg1, tcg_target_long arg2)
 {
-    /* movl */
-    tcg_out_modrm_offset(s, 0x8b, ret, arg1, arg2);
+    tcg_out_modrm_offset(s, OPC_MOVL_GvEv, ret, arg1, arg2);
 }
 
 static inline void tcg_out_st(TCGContext *s, TCGType type, int arg,
                               int arg1, tcg_target_long arg2)
 {
-    /* movl */
-    tcg_out_modrm_offset(s, 0x89, arg, arg1, arg2);
+    tcg_out_modrm_offset(s, OPC_MOVL_EvGv, arg, arg1, arg2);
 }
 
 static void tcg_out_shifti(TCGContext *s, int subopc, int reg, int count)

@@ -777,8 +779,7 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
         }
         break;
     case 2:
-        /* movl (r0), data_reg */
-        tcg_out_modrm_offset(s, 0x8b, data_reg, r0, GUEST_BASE);
+        tcg_out_ld(s, TCG_TYPE_I32, data_reg, r0, GUEST_BASE);
         if (bswap) {
             tcg_out_bswap32(s, data_reg);
         }

@@ -793,13 +794,13 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
             r0 = r1;
         }
         if (!bswap) {
-            tcg_out_modrm_offset(s, 0x8b, data_reg, r0, GUEST_BASE);
-            tcg_out_modrm_offset(s, 0x8b, data_reg2, r0, GUEST_BASE + 4);
+            tcg_out_ld(s, TCG_TYPE_I32, data_reg, r0, GUEST_BASE);
+            tcg_out_ld(s, TCG_TYPE_I32, data_reg2, r0, GUEST_BASE + 4);
         } else {
-            tcg_out_modrm_offset(s, 0x8b, data_reg, r0, GUEST_BASE + 4);
+            tcg_out_ld(s, TCG_TYPE_I32, data_reg, r0, GUEST_BASE + 4);
             tcg_out_bswap32(s, data_reg);
 
-            tcg_out_modrm_offset(s, 0x8b, data_reg2, r0, GUEST_BASE);
+            tcg_out_ld(s, TCG_TYPE_I32, data_reg2, r0, GUEST_BASE);
             tcg_out_bswap32(s, data_reg2);
         }
         break;

@@ -973,8 +974,7 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
 #endif
     switch(opc) {
     case 0:
-        /* movb */
-        tcg_out_modrm_offset(s, 0x88, data_reg, r0, GUEST_BASE);
+        tcg_out_modrm_offset(s, OPC_MOVB_EvGv, data_reg, r0, GUEST_BASE);
         break;
     case 1:
         if (bswap) {

@@ -984,7 +984,7 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
         }
         /* movw */
         tcg_out8(s, 0x66);
-        tcg_out_modrm_offset(s, 0x89, data_reg, r0, GUEST_BASE);
+        tcg_out_modrm_offset(s, OPC_MOVL_EvGv, data_reg, r0, GUEST_BASE);
         break;
     case 2:
         if (bswap) {

@@ -992,20 +992,19 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
             tcg_out_bswap32(s, r1);
             data_reg = r1;
         }
-        /* movl */
-        tcg_out_modrm_offset(s, 0x89, data_reg, r0, GUEST_BASE);
+        tcg_out_st(s, TCG_TYPE_I32, data_reg, r0, GUEST_BASE);
         break;
     case 3:
         if (bswap) {
             tcg_out_mov(s, r1, data_reg2);
             tcg_out_bswap32(s, r1);
-            tcg_out_modrm_offset(s, 0x89, r1, r0, GUEST_BASE);
+            tcg_out_st(s, TCG_TYPE_I32, r1, r0, GUEST_BASE);
             tcg_out_mov(s, r1, data_reg);
             tcg_out_bswap32(s, r1);
-            tcg_out_modrm_offset(s, 0x89, r1, r0, GUEST_BASE + 4);
+            tcg_out_st(s, TCG_TYPE_I32, r1, r0, GUEST_BASE + 4);
         } else {
-            tcg_out_modrm_offset(s, 0x89, data_reg, r0, GUEST_BASE);
-            tcg_out_modrm_offset(s, 0x89, data_reg2, r0, GUEST_BASE + 4);
+            tcg_out_st(s, TCG_TYPE_I32, data_reg, r0, GUEST_BASE);
+            tcg_out_st(s, TCG_TYPE_I32, data_reg2, r0, GUEST_BASE + 4);
         }
         break;
     default:

@@ -1082,21 +1081,19 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
         tcg_out_modrm_offset(s, OPC_MOVSWL, args[0], args[1], args[2]);
         break;
     case INDEX_op_ld_i32:
-        /* movl */
-        tcg_out_modrm_offset(s, 0x8b, args[0], args[1], args[2]);
+        tcg_out_ld(s, TCG_TYPE_I32, args[0], args[1], args[2]);
         break;
     case INDEX_op_st8_i32:
         /* movb */
-        tcg_out_modrm_offset(s, 0x88, args[0], args[1], args[2]);
+        tcg_out_modrm_offset(s, OPC_MOVB_EvGv, args[0], args[1], args[2]);
         break;
     case INDEX_op_st16_i32:
         /* movw */
         tcg_out8(s, 0x66);
-        tcg_out_modrm_offset(s, 0x89, args[0], args[1], args[2]);
+        tcg_out_modrm_offset(s, OPC_MOVL_EvGv, args[0], args[1], args[2]);
         break;
     case INDEX_op_st_i32:
-        /* movl */
-        tcg_out_modrm_offset(s, 0x89, args[0], args[1], args[2]);
+        tcg_out_st(s, TCG_TYPE_I32, args[0], args[1], args[2]);
         break;
     case INDEX_op_sub_i32:
         c = ARITH_SUB;

Signature · 1 line

1.6.6.1

Richard Henderson

author

2010-04-28 18:30 UTC #17 lore raw

PATCH 8/22 tcg-i386: Eliminate extra move from qemu_ld64.

In reply to Richard Henderson

If the address register overlaps one of the output registers simply issue the clobbering load last, rather than emitting an extra move of the address register.

Signed-off-by Richard Henderson <[email protected]>

---
 tcg/i386/tcg-target.c |   20 +++++++++-----------
 1 files changed, 9 insertions(+), 11 deletions(-)

tcg/i386/tcg-target.c +9 ~~-11~~

index a10f6ca..489c2bf 100644
@@ -785,22 +785,20 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
         }
         break;
     case 3:
-        /* XXX: could be nicer */
-        if (r0 == data_reg) {
-            r1 = TCG_REG_EDX;
-            if (r1 == data_reg)
-                r1 = TCG_REG_EAX;
-            tcg_out_mov(s, r1, r0);
-            r0 = r1;
+        if (bswap) {
+            int t = data_reg;
+            data_reg = data_reg2;
+            data_reg2 = t;
         }
-        if (!bswap) {
+        if (r0 != data_reg) {
             tcg_out_ld(s, TCG_TYPE_I32, data_reg, r0, GUEST_BASE);
             tcg_out_ld(s, TCG_TYPE_I32, data_reg2, r0, GUEST_BASE + 4);
         } else {
-            tcg_out_ld(s, TCG_TYPE_I32, data_reg, r0, GUEST_BASE + 4);
+            tcg_out_ld(s, TCG_TYPE_I32, data_reg2, r0, GUEST_BASE + 4);
+            tcg_out_ld(s, TCG_TYPE_I32, data_reg, r0, GUEST_BASE);
+        }
+        if (bswap) {
             tcg_out_bswap32(s, data_reg);
-
-            tcg_out_ld(s, TCG_TYPE_I32, data_reg2, r0, GUEST_BASE);
             tcg_out_bswap32(s, data_reg2);
         }
         break;

Signature · 1 line

1.6.6.1

Richard Henderson

author

2010-04-28 18:30 UTC #18 lore raw

PATCH 9/22 tcg-i386: Tidy jumps.

In reply to Richard Henderson

Define OPC_JCC*, OC_JMP*, and EXT_JMPN_Ev. Use them throughout.

Signed-off-by Richard Henderson <[email protected]>

---
 tcg/i386/tcg-target.c |   58 +++++++++++++++++++++++++++---------------------
 1 files changed, 33 insertions(+), 25 deletions(-)

tcg/i386/tcg-target.c +33 ~~-25~~

index 489c2bf..aac2c5a 100644
@@ -164,6 +164,10 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define P_EXT   0x100 /* 0x0f opcode prefix */
 
 #define OPC_BSWAP	(0xc8 | P_EXT)
+#define OPC_JCC_long	(0x80 | P_EXT)	/* ... plus condition code */
+#define OPC_JCC_short	(0x70)		/* ... plus condition code */
+#define OPC_JMP_long	(0xe9)
+#define OPC_JMP_short	(0xeb)
 #define OPC_MOVB_EvGv	(0x88)		/* stores, more or less */
 #define OPC_MOVL_EvGv	(0x89)		/* stores, more or less */
 #define OPC_MOVL_GvEv	(0x8b)		/* loads, more or less */

@@ -175,6 +179,7 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define OPC_SHIFT_Ib	(0xc1)
 #define OPC_SHIFT_cl	(0xd3)
 
+/* Group 1 opcode extensions for 0x80-0x83.  */
 #define ARITH_ADD 0
 #define ARITH_OR  1
 #define ARITH_ADC 2

@@ -184,12 +189,17 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define ARITH_XOR 6
 #define ARITH_CMP 7
 
+/* Group 2 opcode extensions for 0xc0, 0xc1, 0xd0-0xd3.  */
 #define SHIFT_ROL 0
 #define SHIFT_ROR 1
 #define SHIFT_SHL 4
 #define SHIFT_SHR 5
 #define SHIFT_SAR 7
 
+/* Group 5 opcode extensions for 0xff.  */
+#define EXT_JMPN_Ev	4
+
+/* Condition codes to be added to OPC_JCC_{long,short}.  */
 #define JCC_JMP (-1)
 #define JCC_JO  0x0
 #define JCC_JNO 0x1

@@ -400,9 +410,9 @@ static void tcg_out_jxx(TCGContext *s, int opc, int label_index, int small)
         val1 = val - 2;
         if ((int8_t)val1 == val1) {
             if (opc == -1) {
-                tcg_out8(s, 0xeb);
+                tcg_out8(s, OPC_JMP_short);
             } else {
-                tcg_out8(s, 0x70 + opc);
+                tcg_out8(s, OPC_JCC_short + opc);
             }
             tcg_out8(s, val1);
         } else {

@@ -410,28 +420,26 @@ static void tcg_out_jxx(TCGContext *s, int opc, int label_index, int small)
                 tcg_abort();
             }
             if (opc == -1) {
-                tcg_out8(s, 0xe9);
+                tcg_out8(s, OPC_JMP_long);
                 tcg_out32(s, val - 5);
             } else {
-                tcg_out8(s, 0x0f);
-                tcg_out8(s, 0x80 + opc);
+                tcg_out_opc(s, OPC_JCC_long + opc);
                 tcg_out32(s, val - 6);
             }
         }
     } else if (small) {
         if (opc == -1) {
-            tcg_out8(s, 0xeb);
+            tcg_out8(s, OPC_JMP_short);
         } else {
-            tcg_out8(s, 0x70 + opc);
+            tcg_out8(s, OPC_JCC_short + opc);
         }
         tcg_out_reloc(s, s->code_ptr, R_386_PC8, label_index, -1);
         s->code_ptr += 1;
     } else {
         if (opc == -1) {
-            tcg_out8(s, 0xe9);
+            tcg_out8(s, OPC_JMP_long);
         } else {
-            tcg_out8(s, 0x0f);
-            tcg_out8(s, 0x80 + opc);
+            tcg_out_opc(s, OPC_JCC_long + opc);
         }
         tcg_out_reloc(s, s->code_ptr, R_386_PC32, label_index, -4);
         s->code_ptr += 4;

@@ -674,12 +682,12 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
     
 #if TARGET_LONG_BITS == 32
     /* je label1 */
-    tcg_out8(s, 0x70 + JCC_JE);
+    tcg_out8(s, OPC_JCC_short + JCC_JE);
     label1_ptr = s->code_ptr;
     s->code_ptr++;
 #else
     /* jne label3 */
-    tcg_out8(s, 0x70 + JCC_JNE);
+    tcg_out8(s, OPC_JCC_short + JCC_JNE);
     label3_ptr = s->code_ptr;
     s->code_ptr++;

@@ -687,7 +695,7 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
     tcg_out_modrm_offset(s, 0x3b, addr_reg2, r1, 4);
 
     /* je label1 */
-    tcg_out8(s, 0x70 + JCC_JE);
+    tcg_out8(s, OPC_JCC_short + JCC_JE);
     label1_ptr = s->code_ptr;
     s->code_ptr++;

@@ -735,7 +743,7 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
     }
 
     /* jmp label2 */
-    tcg_out8(s, 0xeb);
+    tcg_out8(s, OPC_JMP_short);
     label2_ptr = s->code_ptr;
     s->code_ptr++;

@@ -868,12 +876,12 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
     
 #if TARGET_LONG_BITS == 32
     /* je label1 */
-    tcg_out8(s, 0x70 + JCC_JE);
+    tcg_out8(s, OPC_JCC_short + JCC_JE);
     label1_ptr = s->code_ptr;
     s->code_ptr++;
 #else
     /* jne label3 */
-    tcg_out8(s, 0x70 + JCC_JNE);
+    tcg_out8(s, OPC_JCC_short + JCC_JNE);
     label3_ptr = s->code_ptr;
     s->code_ptr++;

@@ -881,7 +889,7 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
     tcg_out_modrm_offset(s, 0x3b, addr_reg2, r1, 4);
 
     /* je label1 */
-    tcg_out8(s, 0x70 + JCC_JE);
+    tcg_out8(s, OPC_JCC_short + JCC_JE);
     label1_ptr = s->code_ptr;
     s->code_ptr++;

@@ -951,7 +959,7 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
 #endif
     
     /* jmp label2 */
-    tcg_out8(s, 0xeb);
+    tcg_out8(s, OPC_JMP_short);
     label2_ptr = s->code_ptr;
     s->code_ptr++;

@@ -1023,19 +1031,18 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
     switch(opc) {
     case INDEX_op_exit_tb:
         tcg_out_movi(s, TCG_TYPE_I32, TCG_REG_EAX, args[0]);
-        tcg_out8(s, 0xe9); /* jmp tb_ret_addr */
+        tcg_out8(s, OPC_JMP_long); /* jmp tb_ret_addr */
         tcg_out32(s, tb_ret_addr - s->code_ptr - 4);
         break;
     case INDEX_op_goto_tb:
         if (s->tb_jmp_offset) {
             /* direct jump method */
-            tcg_out8(s, 0xe9); /* jmp im */
+            tcg_out8(s, OPC_JMP_long); /* jmp im */
             s->tb_jmp_offset[args[0]] = s->code_ptr - s->code_buf;
             tcg_out32(s, 0);
         } else {
             /* indirect jump method */
-            /* jmp Ev */
-            tcg_out_modrm_offset(s, 0xff, 4, -1, 
+            tcg_out_modrm_offset(s, 0xff, EXT_JMPN_Ev, -1, 
                                  (tcg_target_long)(s->tb_next + args[0]));
         }
         s->tb_next_offset[args[0]] = s->code_ptr - s->code_buf;

@@ -1050,10 +1057,11 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
         break;
     case INDEX_op_jmp:
         if (const_args[0]) {
-            tcg_out8(s, 0xe9);
+            tcg_out8(s, OPC_JMP_long);
             tcg_out32(s, args[0] - (tcg_target_long)s->code_ptr - 4);
         } else {
-            tcg_out_modrm(s, 0xff, 4, args[0]);
+            /* jmp *reg */
+            tcg_out_modrm(s, 0xff, EXT_JMPN_Ev, args[0]);
         }
         break;
     case INDEX_op_br:

@@ -1378,7 +1386,7 @@ void tcg_target_qemu_prologue(TCGContext *s)
     stack_addend = frame_size - push_size;
     tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
 
-    tcg_out_modrm(s, 0xff, 4, TCG_REG_EAX); /* jmp *%eax */
+    tcg_out_modrm(s, 0xff, EXT_JMPN_Ev, TCG_REG_EAX); /* jmp *%eax */
     
     /* TB epilogue */
     tb_ret_addr = s->code_ptr;

Signature · 1 line

1.6.6.1

Richard Henderson

author

2010-04-28 18:30 UTC #19 lore raw

PATCH 10/22 tcg-i386: Tidy immediate arithmetic operations.

In reply to Richard Henderson

Define OPC_ARITH_EvI[bz]; use throughout. Use tcg_out_ext8u directly in setcond. Use tgen_arithi in qemu_ld/st.

Signed-off-by Richard Henderson <[email protected]>

---
 tcg/i386/tcg-target.c |   28 +++++++++++-----------------
 1 files changed, 11 insertions(+), 17 deletions(-)

tcg/i386/tcg-target.c +11 ~~-17~~

index aac2c5a..df1bdfc 100644
@@ -163,6 +163,8 @@ static inline int tcg_target_const_match(tcg_target_long val,
 
 #define P_EXT   0x100 /* 0x0f opcode prefix */
 
+#define OPC_ARITH_EvIz	(0x81)
+#define OPC_ARITH_EvIb	(0x83)
 #define OPC_BSWAP	(0xc8 | P_EXT)
 #define OPC_JCC_long	(0x80 | P_EXT)	/* ... plus condition code */
 #define OPC_JCC_short	(0x70)		/* ... plus condition code */

@@ -324,7 +326,7 @@ static void tcg_out_ext8u(TCGContext *s, int dest, int src)
     if (src >= 4) {
         tcg_out_mov(s, dest, src);
         if (dest >= 4) {
-            tcg_out_modrm(s, 0x81, ARITH_AND, dest);
+            tcg_out_modrm(s, OPC_ARITH_EvIz, ARITH_AND, dest);
             tcg_out32(s, 0xff);
             return;
         }

@@ -381,14 +383,14 @@ static inline void tgen_arithi(TCGContext *s, int c, int r0, int32_t val, int cf
         /* dec */
         tcg_out_opc(s, 0x48 + r0);
     } else if (val == (int8_t)val) {
-        tcg_out_modrm(s, 0x83, c, r0);
+        tcg_out_modrm(s, OPC_ARITH_EvIb, c, r0);
         tcg_out8(s, val);
     } else if (c == ARITH_AND && val == 0xffu && r0 < 4) {
         tcg_out_ext8u(s, r0, r0);
     } else if (c == ARITH_AND && val == 0xffffu) {
         tcg_out_ext16u(s, r0, r0);
     } else {
-        tcg_out_modrm(s, 0x81, c, r0);
+        tcg_out_modrm(s, OPC_ARITH_EvIz, c, r0);
         tcg_out32(s, val);
     }
 }

@@ -557,7 +559,7 @@ static void tcg_out_setcond(TCGContext *s, TCGCond cond, TCGArg dest,
     tcg_out_cmp(s, arg1, arg2, const_arg2);
     /* setcc */
     tcg_out_modrm(s, 0x90 | tcg_cond_to_jcc[cond] | P_EXT, 0, dest);
-    tgen_arithi(s, ARITH_AND, dest, 0xff, 0);
+    tcg_out_ext8u(s, dest, dest);
 }
 
 static void tcg_out_setcond2(TCGContext *s, const TCGArg *args,

@@ -659,16 +661,12 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
 
 #if defined(CONFIG_SOFTMMU)
     tcg_out_mov(s, r1, addr_reg); 
-
     tcg_out_mov(s, r0, addr_reg); 
- 
+
     tcg_out_shifti(s, SHIFT_SHR, r1, TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS); 
 
-    tcg_out_modrm(s, 0x81, 4, r0); /* andl $x, r0 */
-    tcg_out32(s, TARGET_PAGE_MASK | ((1 << s_bits) - 1));
-    
-    tcg_out_modrm(s, 0x81, 4, r1); /* andl $x, r1 */
-    tcg_out32(s, (CPU_TLB_SIZE - 1) << CPU_TLB_ENTRY_BITS);
+    tgen_arithi(s, ARITH_AND, r0, TARGET_PAGE_MASK | ((1 << s_bits) - 1), 0);
+    tgen_arithi(s, ARITH_AND, r1, (CPU_TLB_SIZE - 1) << CPU_TLB_ENTRY_BITS, 0);
 
     tcg_out_opc(s, 0x8d); /* lea offset(r1, %ebp), r1 */
     tcg_out8(s, 0x80 | (r1 << 3) | 0x04);

@@ -853,16 +851,12 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
 
 #if defined(CONFIG_SOFTMMU)
     tcg_out_mov(s, r1, addr_reg); 
-
     tcg_out_mov(s, r0, addr_reg); 
  
     tcg_out_shifti(s, SHIFT_SHR, r1, TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS); 
 
-    tcg_out_modrm(s, 0x81, 4, r0); /* andl $x, r0 */
-    tcg_out32(s, TARGET_PAGE_MASK | ((1 << s_bits) - 1));
-    
-    tcg_out_modrm(s, 0x81, 4, r1); /* andl $x, r1 */
-    tcg_out32(s, (CPU_TLB_SIZE - 1) << CPU_TLB_ENTRY_BITS);
+    tgen_arithi(s, ARITH_AND, r0, TARGET_PAGE_MASK | ((1 << s_bits) - 1), 0);
+    tgen_arithi(s, ARITH_AND, r1, (CPU_TLB_SIZE - 1) << CPU_TLB_ENTRY_BITS, 0);
 
     tcg_out_opc(s, 0x8d); /* lea offset(r1, %ebp), r1 */
     tcg_out8(s, 0x80 | (r1 << 3) | 0x04);

Signature · 1 line

1.6.6.1

Aurelien Jarno

Acked-by on v1 10/22

2010-05-21 09:38 UTC #20 lore raw

Re: [Qemu-devel] [PATCH 10/22] tcg-i386: Tidy immediate arithmetic operations.

Quoted 4 lines · Richard Henderson

Define OPC_ARITH_EvI[bz]; use throughout. Use tcg_out_ext8u directly in setcond. Use tgen_arithi in qemu_ld/st.
Signed-off-by Richard Henderson <[email protected]>

Acked-by Aurelien Jarno <[email protected]>

Quoted 97 lines

---
 tcg/i386/tcg-target.c |   28 +++++++++++-----------------
 1 files changed, 11 insertions(+), 17 deletions(-)

tcg/i386/tcg-target.c +11 ~~-17~~

index aac2c5a..df1bdfc 100644
@@ -163,6 +163,8 @@ static inline int tcg_target_const_match(tcg_target_long val,
 
 #define P_EXT   0x100 /* 0x0f opcode prefix */
 
+#define OPC_ARITH_EvIz	(0x81)
+#define OPC_ARITH_EvIb	(0x83)
 #define OPC_BSWAP	(0xc8 | P_EXT)
 #define OPC_JCC_long	(0x80 | P_EXT)	/* ... plus condition code */
 #define OPC_JCC_short	(0x70)		/* ... plus condition code */

@@ -324,7 +326,7 @@ static void tcg_out_ext8u(TCGContext *s, int dest, int src)
     if (src >= 4) {
         tcg_out_mov(s, dest, src);
         if (dest >= 4) {
-            tcg_out_modrm(s, 0x81, ARITH_AND, dest);
+            tcg_out_modrm(s, OPC_ARITH_EvIz, ARITH_AND, dest);
             tcg_out32(s, 0xff);
             return;
         }

@@ -381,14 +383,14 @@ static inline void tgen_arithi(TCGContext *s, int c, int r0, int32_t val, int cf
         /* dec */
         tcg_out_opc(s, 0x48 + r0);
     } else if (val == (int8_t)val) {
-        tcg_out_modrm(s, 0x83, c, r0);
+        tcg_out_modrm(s, OPC_ARITH_EvIb, c, r0);
         tcg_out8(s, val);
     } else if (c == ARITH_AND && val == 0xffu && r0 < 4) {
         tcg_out_ext8u(s, r0, r0);
     } else if (c == ARITH_AND && val == 0xffffu) {
         tcg_out_ext16u(s, r0, r0);
     } else {
-        tcg_out_modrm(s, 0x81, c, r0);
+        tcg_out_modrm(s, OPC_ARITH_EvIz, c, r0);
         tcg_out32(s, val);
     }
 }

@@ -557,7 +559,7 @@ static void tcg_out_setcond(TCGContext *s, TCGCond cond, TCGArg dest,
     tcg_out_cmp(s, arg1, arg2, const_arg2);
     /* setcc */
     tcg_out_modrm(s, 0x90 | tcg_cond_to_jcc[cond] | P_EXT, 0, dest);
-    tgen_arithi(s, ARITH_AND, dest, 0xff, 0);
+    tcg_out_ext8u(s, dest, dest);
 }
 
 static void tcg_out_setcond2(TCGContext *s, const TCGArg *args,

@@ -659,16 +661,12 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
 
 #if defined(CONFIG_SOFTMMU)
     tcg_out_mov(s, r1, addr_reg); 
-
     tcg_out_mov(s, r0, addr_reg); 
- 
+
     tcg_out_shifti(s, SHIFT_SHR, r1, TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS); 
 
-    tcg_out_modrm(s, 0x81, 4, r0); /* andl $x, r0 */
-    tcg_out32(s, TARGET_PAGE_MASK | ((1 << s_bits) - 1));
-    
-    tcg_out_modrm(s, 0x81, 4, r1); /* andl $x, r1 */
-    tcg_out32(s, (CPU_TLB_SIZE - 1) << CPU_TLB_ENTRY_BITS);
+    tgen_arithi(s, ARITH_AND, r0, TARGET_PAGE_MASK | ((1 << s_bits) - 1), 0);
+    tgen_arithi(s, ARITH_AND, r1, (CPU_TLB_SIZE - 1) << CPU_TLB_ENTRY_BITS, 0);
 
     tcg_out_opc(s, 0x8d); /* lea offset(r1, %ebp), r1 */
     tcg_out8(s, 0x80 | (r1 << 3) | 0x04);

@@ -853,16 +851,12 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
 
 #if defined(CONFIG_SOFTMMU)
     tcg_out_mov(s, r1, addr_reg); 
-
     tcg_out_mov(s, r0, addr_reg); 
 
     tcg_out_shifti(s, SHIFT_SHR, r1, TARGET_PAGE_BITS - CPU_TLB_ENTRY_BITS); 
 
-    tcg_out_modrm(s, 0x81, 4, r0); /* andl $x, r0 */
-    tcg_out32(s, TARGET_PAGE_MASK | ((1 << s_bits) - 1));
-    
-    tcg_out_modrm(s, 0x81, 4, r1); /* andl $x, r1 */
-    tcg_out32(s, (CPU_TLB_SIZE - 1) << CPU_TLB_ENTRY_BITS);
+    tgen_arithi(s, ARITH_AND, r0, TARGET_PAGE_MASK | ((1 << s_bits) - 1), 0);
+    tgen_arithi(s, ARITH_AND, r1, (CPU_TLB_SIZE - 1) << CPU_TLB_ENTRY_BITS, 0);
 
     tcg_out_opc(s, 0x8d); /* lea offset(r1, %ebp), r1 */
     tcg_out8(s, 0x80 | (r1 << 3) | 0x04);

Signature · 1 line

1.6.6.1

Signature · 2 lines

Aurelien Jarno                          GPG: 1024D/F1BCDB73
[email protected]                 http://www.aurel32.net

Richard Henderson

author

2010-04-28 18:30 UTC #21 lore raw

PATCH 11/22 tcg-i386: Tidy non-immediate arithmetic operations.

In reply to Richard Henderson

Add more OPC values, and tgen_arithr. Use the later throughout.

Note that normal reg/reg arithmetic now uses the Gv,Ev opcode form instead of the Ev,Gv opcode form used previously. Both forms disassemble properly, and so there's no visible change when diffing log files before and after the change. This change makes the operand ordering within the output routines more natural, and avoids the need to define an OPC_ARITH_EvGv since a read-modify-write with memory is not needed within TCG.

Signed-off-by Richard Henderson <[email protected]>

---
 tcg/i386/tcg-target.c |   78 ++++++++++++++++++++++++++++++-------------------
 1 files changed, 48 insertions(+), 30 deletions(-)

tcg/i386/tcg-target.c +48 ~~-30~~

index df1bdfc..b4e8e74 100644
@@ -165,7 +165,12 @@ static inline int tcg_target_const_match(tcg_target_long val,
 
 #define OPC_ARITH_EvIz	(0x81)
 #define OPC_ARITH_EvIb	(0x83)
+#define OPC_ARITH_GvEv	(0x03)		/* ... plus (ARITH_FOO << 3) */
+#define OPC_ADD_GvEv	(OPC_ARITH_GvEv | (ARITH_ADD << 3))
+#define OPC_CMP_GvEv	(OPC_ARITH_GvEv | (ARITH_CMP << 3))
+#define OPC_DEC_r32	(0x48)
 #define OPC_BSWAP	(0xc8 | P_EXT)
+#define OPC_INC_r32	(0x40)
 #define OPC_JCC_long	(0x80 | P_EXT)	/* ... plus condition code */
 #define OPC_JCC_short	(0x70)		/* ... plus condition code */
 #define OPC_JMP_long	(0xe9)

@@ -180,6 +185,7 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define OPC_SHIFT_1	(0xd1)
 #define OPC_SHIFT_Ib	(0xc1)
 #define OPC_SHIFT_cl	(0xd3)
+#define OPC_TESTL	(0x85)
 
 /* Group 1 opcode extensions for 0x80-0x83.  */
 #define ARITH_ADD 0

@@ -280,6 +286,12 @@ static inline void tcg_out_modrm_offset(TCGContext *s, int opc, int r, int rm,
     }
 }
 
+/* Generate dest op= src.  Uses the same ARITH_* codes as tgen_arithi.  */
+static inline void tgen_arithr(TCGContext *s, int subop, int dest, int src)
+{
+    tcg_out_modrm(s, OPC_ARITH_GvEv + (subop << 3), dest, src);
+}
+
 static inline void tcg_out_mov(TCGContext *s, int ret, int arg)
 {
     if (arg != ret) {

@@ -291,8 +303,7 @@ static inline void tcg_out_movi(TCGContext *s, TCGType type,
                                 int ret, int32_t arg)
 {
     if (arg == 0) {
-        /* xor r0,r0 */
-        tcg_out_modrm(s, 0x01 | (ARITH_XOR << 3), ret, ret);
+        tgen_arithr(s, ARITH_XOR, ret, ret);
     } else {
         tcg_out8(s, 0xb8 + ret);
         tcg_out32(s, arg);

@@ -374,14 +385,15 @@ static inline void tcg_out_rolw_8(TCGContext *s, int reg)
     tcg_out_shifti(s, SHIFT_ROL, reg, 8);
 }
 
-static inline void tgen_arithi(TCGContext *s, int c, int r0, int32_t val, int cf)
+static inline void tgen_arithi(TCGContext *s, int c, int r0,
+                               int32_t val, int cf)
 {
-    if (!cf && ((c == ARITH_ADD && val == 1) || (c == ARITH_SUB && val == -1))) {
-        /* inc */
-        tcg_out_opc(s, 0x40 + r0);
-    } else if (!cf && ((c == ARITH_ADD && val == -1) || (c == ARITH_SUB && val == 1))) {
-        /* dec */
-        tcg_out_opc(s, 0x48 + r0);
+    /* ??? While INC is 2 bytes shorter than ADDL $1, they also induce
+       partial flags update stalls on Pentium4 and are not recommended
+       by current Intel optimization manuals.  */
+    if (!cf && (c == ARITH_ADD || c == ARITH_SUB) && (val == 1 || val == -1)) {
+        int opc = ((c == ARITH_ADD) ^ (val < 0) ? OPC_INC_r32 : OPC_DEC_r32);
+        tcg_out_opc(s, opc + r0);
     } else if (val == (int8_t)val) {
         tcg_out_modrm(s, OPC_ARITH_EvIb, c, r0);
         tcg_out8(s, val);

@@ -454,12 +466,12 @@ static void tcg_out_cmp(TCGContext *s, TCGArg arg1, TCGArg arg2,
     if (const_arg2) {
         if (arg2 == 0) {
             /* test r, r */
-            tcg_out_modrm(s, 0x85, arg1, arg1);
+            tcg_out_modrm(s, OPC_TESTL, arg1, arg1);
         } else {
             tgen_arithi(s, ARITH_CMP, arg1, arg2, 0);
         }
     } else {
-        tcg_out_modrm(s, 0x01 | (ARITH_CMP << 3), arg2, arg1);
+        tgen_arithr(s, ARITH_CMP, arg1, arg2);
     }
 }

@@ -674,7 +686,7 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
     tcg_out32(s, offsetof(CPUState, tlb_table[mem_index][0].addr_read));
 
     /* cmp 0(r1), r0 */
-    tcg_out_modrm_offset(s, 0x3b, r0, r1, 0);
+    tcg_out_modrm_offset(s, OPC_CMP_GvEv, r0, r1, 0);
     
     tcg_out_mov(s, r0, addr_reg);

@@ -690,7 +702,7 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
     s->code_ptr++;
     
     /* cmp 4(r1), addr_reg2 */
-    tcg_out_modrm_offset(s, 0x3b, addr_reg2, r1, 4);
+    tcg_out_modrm_offset(s, OPC_CMP_GvEv, addr_reg2, r1, 4);
 
     /* je label1 */
     tcg_out8(s, OPC_JCC_short + JCC_JE);

@@ -749,7 +761,8 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
     *label1_ptr = s->code_ptr - label1_ptr - 1;
 
     /* add x(r1), r0 */
-    tcg_out_modrm_offset(s, 0x03, r0, r1, offsetof(CPUTLBEntry, addend) - 
+    tcg_out_modrm_offset(s, OPC_ADD_GvEv, r0, r1,
+                         offsetof(CPUTLBEntry, addend) - 
                          offsetof(CPUTLBEntry, addr_read));
 #else
     r0 = addr_reg;

@@ -864,7 +877,7 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
     tcg_out32(s, offsetof(CPUState, tlb_table[mem_index][0].addr_write));
 
     /* cmp 0(r1), r0 */
-    tcg_out_modrm_offset(s, 0x3b, r0, r1, 0);
+    tcg_out_modrm_offset(s, OPC_CMP_GvEv, r0, r1, 0);
     
     tcg_out_mov(s, r0, addr_reg);

@@ -880,7 +893,7 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
     s->code_ptr++;
     
     /* cmp 4(r1), addr_reg2 */
-    tcg_out_modrm_offset(s, 0x3b, addr_reg2, r1, 4);
+    tcg_out_modrm_offset(s, OPC_CMP_GvEv, addr_reg2, r1, 4);
 
     /* je label1 */
     tcg_out8(s, OPC_JCC_short + JCC_JE);

@@ -961,7 +974,8 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
     *label1_ptr = s->code_ptr - label1_ptr - 1;
 
     /* add x(r1), r0 */
-    tcg_out_modrm_offset(s, 0x03, r0, r1, offsetof(CPUTLBEntry, addend) - 
+    tcg_out_modrm_offset(s, OPC_ADD_GvEv, r0, r1,
+                         offsetof(CPUTLBEntry, addend) - 
                          offsetof(CPUTLBEntry, addr_write));
 #else
     r0 = addr_reg;

@@ -1113,7 +1127,7 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
         if (const_args[2]) {
             tgen_arithi(s, c, args[0], args[2], 0);
         } else {
-            tcg_out_modrm(s, 0x01 | (c << 3), args[2], args[0]);
+            tgen_arithr(s, c, args[0], args[2]);
         }
         break;
     case INDEX_op_mul_i32:

@@ -1163,24 +1177,28 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
         goto gen_shift32;
 
     case INDEX_op_add2_i32:
-        if (const_args[4]) 
+        if (const_args[4]) {
             tgen_arithi(s, ARITH_ADD, args[0], args[4], 1);
-        else
-            tcg_out_modrm(s, 0x01 | (ARITH_ADD << 3), args[4], args[0]);
-        if (const_args[5]) 
+        } else {
+            tgen_arithr(s, ARITH_ADD, args[0], args[4]);
+        }
+        if (const_args[5]) {
             tgen_arithi(s, ARITH_ADC, args[1], args[5], 1);
-        else
-            tcg_out_modrm(s, 0x01 | (ARITH_ADC << 3), args[5], args[1]);
+        } else {
+            tgen_arithr(s, ARITH_ADC, args[0], args[5]);
+        }
         break;
     case INDEX_op_sub2_i32:
-        if (const_args[4]) 
+        if (const_args[4]) {
             tgen_arithi(s, ARITH_SUB, args[0], args[4], 1);
-        else
-            tcg_out_modrm(s, 0x01 | (ARITH_SUB << 3), args[4], args[0]);
-        if (const_args[5]) 
+        } else {
+            tgen_arithr(s, ARITH_SUB, args[0], args[4]);
+        }
+        if (const_args[5]) {
             tgen_arithi(s, ARITH_SBB, args[1], args[5], 1);
-        else
-            tcg_out_modrm(s, 0x01 | (ARITH_SBB << 3), args[5], args[1]);
+        } else {
+            tgen_arithr(s, ARITH_SBB, args[1], args[5]);
+        }
         break;
     case INDEX_op_brcond_i32:
         tcg_out_brcond(s, args[2], args[0], args[1], const_args[1],

Signature · 1 line

1.6.6.1

Aurelien Jarno

Acked-by on v1 11/22

2010-05-21 09:39 UTC #22 lore raw

Re: [Qemu-devel] [PATCH 11/22] tcg-i386: Tidy non-immediate arithmetic operations.

Quoted 11 lines · Richard Henderson

Add more OPC values, and tgen_arithr. Use the later throughout.
Note that normal reg/reg arithmetic now uses the Gv,Ev opcode form instead of the Ev,Gv opcode form used previously. Both forms disassemble properly, and so there's no visible change when diffing log files before and after the change. This change makes the operand ordering within the output routines more natural, and avoids the need to define an OPC_ARITH_EvGv since a read-modify-write with memory is not needed within TCG.
Signed-off-by Richard Henderson <[email protected]>

Acked-by Aurelien Jarno <[email protected]>

Quoted 202 lines

---
 tcg/i386/tcg-target.c |   78 ++++++++++++++++++++++++++++++-------------------
 1 files changed, 48 insertions(+), 30 deletions(-)

tcg/i386/tcg-target.c +48 ~~-30~~

index df1bdfc..b4e8e74 100644
@@ -165,7 +165,12 @@ static inline int tcg_target_const_match(tcg_target_long val,
 
 #define OPC_ARITH_EvIz	(0x81)
 #define OPC_ARITH_EvIb	(0x83)
+#define OPC_ARITH_GvEv	(0x03)		/* ... plus (ARITH_FOO << 3) */
+#define OPC_ADD_GvEv	(OPC_ARITH_GvEv | (ARITH_ADD << 3))
+#define OPC_CMP_GvEv	(OPC_ARITH_GvEv | (ARITH_CMP << 3))
+#define OPC_DEC_r32	(0x48)
 #define OPC_BSWAP	(0xc8 | P_EXT)
+#define OPC_INC_r32	(0x40)
 #define OPC_JCC_long	(0x80 | P_EXT)	/* ... plus condition code */
 #define OPC_JCC_short	(0x70)		/* ... plus condition code */
 #define OPC_JMP_long	(0xe9)

@@ -180,6 +185,7 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define OPC_SHIFT_1	(0xd1)
 #define OPC_SHIFT_Ib	(0xc1)
 #define OPC_SHIFT_cl	(0xd3)
+#define OPC_TESTL	(0x85)
 
 /* Group 1 opcode extensions for 0x80-0x83.  */
 #define ARITH_ADD 0

@@ -280,6 +286,12 @@ static inline void tcg_out_modrm_offset(TCGContext *s, int opc, int r, int rm,
     }
 }
 
+/* Generate dest op= src.  Uses the same ARITH_* codes as tgen_arithi.  */
+static inline void tgen_arithr(TCGContext *s, int subop, int dest, int src)
+{
+    tcg_out_modrm(s, OPC_ARITH_GvEv + (subop << 3), dest, src);
+}
+
 static inline void tcg_out_mov(TCGContext *s, int ret, int arg)
 {
     if (arg != ret) {

@@ -291,8 +303,7 @@ static inline void tcg_out_movi(TCGContext *s, TCGType type,
                                 int ret, int32_t arg)
 {
     if (arg == 0) {
-        /* xor r0,r0 */
-        tcg_out_modrm(s, 0x01 | (ARITH_XOR << 3), ret, ret);
+        tgen_arithr(s, ARITH_XOR, ret, ret);
     } else {
         tcg_out8(s, 0xb8 + ret);
         tcg_out32(s, arg);

@@ -374,14 +385,15 @@ static inline void tcg_out_rolw_8(TCGContext *s, int reg)
     tcg_out_shifti(s, SHIFT_ROL, reg, 8);
 }
 
-static inline void tgen_arithi(TCGContext *s, int c, int r0, int32_t val, int cf)
+static inline void tgen_arithi(TCGContext *s, int c, int r0,
+                               int32_t val, int cf)
 {
-    if (!cf && ((c == ARITH_ADD && val == 1) || (c == ARITH_SUB && val == -1))) {
-        /* inc */
-        tcg_out_opc(s, 0x40 + r0);
-    } else if (!cf && ((c == ARITH_ADD && val == -1) || (c == ARITH_SUB && val == 1))) {
-        /* dec */
-        tcg_out_opc(s, 0x48 + r0);
+    /* ??? While INC is 2 bytes shorter than ADDL $1, they also induce
+       partial flags update stalls on Pentium4 and are not recommended
+       by current Intel optimization manuals.  */
+    if (!cf && (c == ARITH_ADD || c == ARITH_SUB) && (val == 1 || val == -1)) {
+        int opc = ((c == ARITH_ADD) ^ (val < 0) ? OPC_INC_r32 : OPC_DEC_r32);
+        tcg_out_opc(s, opc + r0);
     } else if (val == (int8_t)val) {
         tcg_out_modrm(s, OPC_ARITH_EvIb, c, r0);
         tcg_out8(s, val);

@@ -454,12 +466,12 @@ static void tcg_out_cmp(TCGContext *s, TCGArg arg1, TCGArg arg2,
     if (const_arg2) {
         if (arg2 == 0) {
             /* test r, r */
-            tcg_out_modrm(s, 0x85, arg1, arg1);
+            tcg_out_modrm(s, OPC_TESTL, arg1, arg1);
         } else {
             tgen_arithi(s, ARITH_CMP, arg1, arg2, 0);
         }
     } else {
-        tcg_out_modrm(s, 0x01 | (ARITH_CMP << 3), arg2, arg1);
+        tgen_arithr(s, ARITH_CMP, arg1, arg2);
     }
 }

@@ -674,7 +686,7 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
     tcg_out32(s, offsetof(CPUState, tlb_table[mem_index][0].addr_read));
 
     /* cmp 0(r1), r0 */
-    tcg_out_modrm_offset(s, 0x3b, r0, r1, 0);
+    tcg_out_modrm_offset(s, OPC_CMP_GvEv, r0, r1, 0);
 
     tcg_out_mov(s, r0, addr_reg);

@@ -690,7 +702,7 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
     s->code_ptr++;
 
     /* cmp 4(r1), addr_reg2 */
-    tcg_out_modrm_offset(s, 0x3b, addr_reg2, r1, 4);
+    tcg_out_modrm_offset(s, OPC_CMP_GvEv, addr_reg2, r1, 4);
 
     /* je label1 */
     tcg_out8(s, OPC_JCC_short + JCC_JE);

@@ -749,7 +761,8 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
     *label1_ptr = s->code_ptr - label1_ptr - 1;
 
     /* add x(r1), r0 */
-    tcg_out_modrm_offset(s, 0x03, r0, r1, offsetof(CPUTLBEntry, addend) - 
+    tcg_out_modrm_offset(s, OPC_ADD_GvEv, r0, r1,
+                         offsetof(CPUTLBEntry, addend) - 
                          offsetof(CPUTLBEntry, addr_read));
 #else
     r0 = addr_reg;

@@ -864,7 +877,7 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
     tcg_out32(s, offsetof(CPUState, tlb_table[mem_index][0].addr_write));
 
     /* cmp 0(r1), r0 */
-    tcg_out_modrm_offset(s, 0x3b, r0, r1, 0);
+    tcg_out_modrm_offset(s, OPC_CMP_GvEv, r0, r1, 0);
 
     tcg_out_mov(s, r0, addr_reg);

@@ -880,7 +893,7 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
     s->code_ptr++;
 
     /* cmp 4(r1), addr_reg2 */
-    tcg_out_modrm_offset(s, 0x3b, addr_reg2, r1, 4);
+    tcg_out_modrm_offset(s, OPC_CMP_GvEv, addr_reg2, r1, 4);
 
     /* je label1 */
     tcg_out8(s, OPC_JCC_short + JCC_JE);

@@ -961,7 +974,8 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
     *label1_ptr = s->code_ptr - label1_ptr - 1;
 
     /* add x(r1), r0 */
-    tcg_out_modrm_offset(s, 0x03, r0, r1, offsetof(CPUTLBEntry, addend) - 
+    tcg_out_modrm_offset(s, OPC_ADD_GvEv, r0, r1,
+                         offsetof(CPUTLBEntry, addend) - 
                          offsetof(CPUTLBEntry, addr_write));
 #else
     r0 = addr_reg;

@@ -1113,7 +1127,7 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
         if (const_args[2]) {
             tgen_arithi(s, c, args[0], args[2], 0);
         } else {
-            tcg_out_modrm(s, 0x01 | (c << 3), args[2], args[0]);
+            tgen_arithr(s, c, args[0], args[2]);
         }
         break;
     case INDEX_op_mul_i32:

@@ -1163,24 +1177,28 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
         goto gen_shift32;
 
     case INDEX_op_add2_i32:
-        if (const_args[4]) 
+        if (const_args[4]) {
             tgen_arithi(s, ARITH_ADD, args[0], args[4], 1);
-        else
-            tcg_out_modrm(s, 0x01 | (ARITH_ADD << 3), args[4], args[0]);
-        if (const_args[5]) 
+        } else {
+            tgen_arithr(s, ARITH_ADD, args[0], args[4]);
+        }
+        if (const_args[5]) {
             tgen_arithi(s, ARITH_ADC, args[1], args[5], 1);
-        else
-            tcg_out_modrm(s, 0x01 | (ARITH_ADC << 3), args[5], args[1]);
+        } else {
+            tgen_arithr(s, ARITH_ADC, args[0], args[5]);
+        }
         break;
     case INDEX_op_sub2_i32:
-        if (const_args[4]) 
+        if (const_args[4]) {
             tgen_arithi(s, ARITH_SUB, args[0], args[4], 1);
-        else
-            tcg_out_modrm(s, 0x01 | (ARITH_SUB << 3), args[4], args[0]);
-        if (const_args[5]) 
+        } else {
+            tgen_arithr(s, ARITH_SUB, args[0], args[4]);
+        }
+        if (const_args[5]) {
             tgen_arithi(s, ARITH_SBB, args[1], args[5], 1);
-        else
-            tcg_out_modrm(s, 0x01 | (ARITH_SBB << 3), args[5], args[1]);
+        } else {
+            tgen_arithr(s, ARITH_SBB, args[1], args[5]);
+        }
         break;
     case INDEX_op_brcond_i32:
         tcg_out_brcond(s, args[2], args[0], args[1], const_args[1],

Signature · 1 line

1.6.6.1

Signature · 2 lines

Aurelien Jarno                          GPG: 1024D/F1BCDB73
[email protected]                 http://www.aurel32.net

Richard Henderson

author

2010-04-28 18:30 UTC #23 lore raw

PATCH 12/22 tcg-i386: Tidy movi.

In reply to Richard Henderson

Define and use OPC_MOVL_Iv.

Signed-off-by Richard Henderson <[email protected]>

---
 tcg/i386/tcg-target.c |    3 ++-
 1 files changed, 2 insertions(+), 1 deletions(-)

tcg/i386/tcg-target.c +2 -1

index b4e8e74..06946e5 100644
@@ -178,6 +178,7 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define OPC_MOVB_EvGv	(0x88)		/* stores, more or less */
 #define OPC_MOVL_EvGv	(0x89)		/* stores, more or less */
 #define OPC_MOVL_GvEv	(0x8b)		/* loads, more or less */
+#define OPC_MOVL_Iv     (0xb8)
 #define OPC_MOVZBL	(0xb6 | P_EXT)
 #define OPC_MOVZWL	(0xb7 | P_EXT)
 #define OPC_MOVSBL	(0xbe | P_EXT)

@@ -305,7 +306,7 @@ static inline void tcg_out_movi(TCGContext *s, TCGType type,
     if (arg == 0) {
         tgen_arithr(s, ARITH_XOR, ret, ret);
     } else {
-        tcg_out8(s, 0xb8 + ret);
+        tcg_out8(s, OPC_MOVL_Iv + ret);
         tcg_out32(s, arg);
     }
 }

Signature · 1 line

1.6.6.1

Aurelien Jarno

Acked-by on v1 12/22

2010-05-21 09:39 UTC #24 lore raw

Re: [Qemu-devel] [PATCH 12/22] tcg-i386: Tidy movi.

Quoted 3 lines · Richard Henderson

Define and use OPC_MOVL_Iv.
Signed-off-by Richard Henderson <[email protected]>

Acked-by Aurelien Jarno <[email protected]>

Quoted 31 lines

---
 tcg/i386/tcg-target.c |    3 ++-
 1 files changed, 2 insertions(+), 1 deletions(-)

tcg/i386/tcg-target.c +2 -1

index b4e8e74..06946e5 100644
@@ -178,6 +178,7 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define OPC_MOVB_EvGv	(0x88)		/* stores, more or less */
 #define OPC_MOVL_EvGv	(0x89)		/* stores, more or less */
 #define OPC_MOVL_GvEv	(0x8b)		/* loads, more or less */
+#define OPC_MOVL_Iv     (0xb8)
 #define OPC_MOVZBL	(0xb6 | P_EXT)
 #define OPC_MOVZWL	(0xb7 | P_EXT)
 #define OPC_MOVSBL	(0xbe | P_EXT)

@@ -305,7 +306,7 @@ static inline void tcg_out_movi(TCGContext *s, TCGType type,
     if (arg == 0) {
         tgen_arithr(s, ARITH_XOR, ret, ret);
     } else {
-        tcg_out8(s, 0xb8 + ret);
+        tcg_out8(s, OPC_MOVL_Iv + ret);
         tcg_out32(s, arg);
     }
 }

Signature · 1 line

1.6.6.1

Signature · 2 lines

Aurelien Jarno                          GPG: 1024D/F1BCDB73
[email protected]                 http://www.aurel32.net

Richard Henderson

author

2010-04-28 18:30 UTC #25 lore raw

PATCH 13/22 tcg-i386: Tidy push/pop.

In reply to Richard Henderson

Move tcg_out_push/pop up in the file so that they can be used by qemu_ld/st. Define a tcg_out_pushi to be used as well.

Signed-off-by Richard Henderson <[email protected]>

---
 tcg/i386/tcg-target.c |   50 +++++++++++++++++++++++++++++++-----------------
 1 files changed, 32 insertions(+), 18 deletions(-)

tcg/i386/tcg-target.c +32 ~~-18~~

index 06946e5..58fcd23 100644
@@ -183,6 +183,10 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define OPC_MOVZWL	(0xb7 | P_EXT)
 #define OPC_MOVSBL	(0xbe | P_EXT)
 #define OPC_MOVSWL	(0xbf | P_EXT)
+#define OPC_POP_r32	(0x58)
+#define OPC_PUSH_r32	(0x50)
+#define OPC_PUSH_Iv	(0x68)
+#define OPC_PUSH_Ib	(0x6a)
 #define OPC_SHIFT_1	(0xd1)
 #define OPC_SHIFT_Ib	(0xc1)
 #define OPC_SHIFT_cl	(0xd3)

@@ -311,6 +315,29 @@ static inline void tcg_out_movi(TCGContext *s, TCGType type,
     }
 }
 
+#if defined(CONFIG_SOFTMMU)
+static void tcg_out_pushi(TCGContext *s, tcg_target_long val)
+{
+    if (val == (int8_t)val) {
+        tcg_out_opc(s, OPC_PUSH_Ib);
+        tcg_out8(s, val);
+    } else {
+        tcg_out_opc(s, OPC_PUSH_Iv);
+        tcg_out32(s, val);
+    }
+}
+#endif
+
+static inline void tcg_out_push(TCGContext *s, int reg)
+{
+    tcg_out_opc(s, OPC_PUSH_r32 + reg);
+}
+
+static inline void tcg_out_pop(TCGContext *s, int reg)
+{
+    tcg_out_opc(s, OPC_POP_r32 + reg);
+}
+
 static inline void tcg_out_ld(TCGContext *s, TCGType type, int ret,
                               int arg1, tcg_target_long arg2)
 {

@@ -910,8 +937,7 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
     if (opc == 3) {
         tcg_out_mov(s, TCG_REG_EDX, data_reg);
         tcg_out_mov(s, TCG_REG_ECX, data_reg2);
-        tcg_out8(s, 0x6a); /* push Ib */
-        tcg_out8(s, mem_index);
+        tcg_out_pushi(s, mem_index);
         tcg_out8(s, 0xe8);
         tcg_out32(s, (tcg_target_long)qemu_st_helpers[s_bits] - 
                   (tcg_target_long)s->code_ptr - 4);

@@ -936,10 +962,9 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
 #else
     if (opc == 3) {
         tcg_out_mov(s, TCG_REG_EDX, addr_reg2);
-        tcg_out8(s, 0x6a); /* push Ib */
-        tcg_out8(s, mem_index);
-        tcg_out_opc(s, 0x50 + data_reg2); /* push */
-        tcg_out_opc(s, 0x50 + data_reg); /* push */
+        tcg_out_pushi(s, mem_index);
+        tcg_out_push(s, data_reg2);
+        tcg_out_push(s, data_reg);
         tcg_out8(s, 0xe8);
         tcg_out32(s, (tcg_target_long)qemu_st_helpers[s_bits] - 
                   (tcg_target_long)s->code_ptr - 4);

@@ -957,8 +982,7 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
             tcg_out_mov(s, TCG_REG_ECX, data_reg);
             break;
         }
-        tcg_out8(s, 0x6a); /* push Ib */
-        tcg_out8(s, mem_index);
+        tcg_out_pushi(s, mem_index);
         tcg_out8(s, 0xe8);
         tcg_out32(s, (tcg_target_long)qemu_st_helpers[s_bits] - 
                   (tcg_target_long)s->code_ptr - 4);

@@ -1371,16 +1395,6 @@ static int tcg_target_callee_save_regs[] = {
     TCG_REG_EDI,
 };
 
-static inline void tcg_out_push(TCGContext *s, int reg)
-{
-    tcg_out_opc(s, 0x50 + reg);
-}
-
-static inline void tcg_out_pop(TCGContext *s, int reg)
-{
-    tcg_out_opc(s, 0x58 + reg);
-}
-
 /* Generate global QEMU prologue and epilogue code */
 void tcg_target_qemu_prologue(TCGContext *s)
 {

Signature · 1 line

1.6.6.1

Aurelien Jarno

2010-05-21 09:38 UTC #26 lore raw

Re: [Qemu-devel] [PATCH 13/22] tcg-i386: Tidy push/pop.

Quoted 29 lines · Richard Henderson

Move tcg_out_push/pop up in the file so that they can be used by qemu_ld/st. Define a tcg_out_pushi to be used as well.

Signed-off-by Richard Henderson <[email protected]>

---
 tcg/i386/tcg-target.c |   50 +++++++++++++++++++++++++++++++-----------------
 1 files changed, 32 insertions(+), 18 deletions(-)

tcg/i386/tcg-target.c +6 -0

index 06946e5..58fcd23 100644
@@ -183,6 +183,10 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define OPC_MOVZWL	(0xb7 | P_EXT)
 #define OPC_MOVSBL	(0xbe | P_EXT)
 #define OPC_MOVSWL	(0xbf | P_EXT)
+#define OPC_POP_r32	(0x58)
+#define OPC_PUSH_r32	(0x50)
+#define OPC_PUSH_Iv	(0x68)
+#define OPC_PUSH_Ib	(0x6a)
 #define OPC_SHIFT_1	(0xd1)
 #define OPC_SHIFT_Ib	(0xc1)
 #define OPC_SHIFT_cl	(0xd3)

@@ -311,6 +315,29 @@ static inline void tcg_out_movi(TCGContext *s, TCGType type,
     }
 }
 
+#if defined(CONFIG_SOFTMMU)
+static void tcg_out_pushi(TCGContext *s, tcg_target_long val)

tcg/i386/tcg-target.c static inline void tcg_out_movi(TCGContext *s, TCGType type, · line 329

Quoted 10 lines

tcg/i386/tcg-target.c +10 -0

+{
+    if (val == (int8_t)val) {
+        tcg_out_opc(s, OPC_PUSH_Ib);
+        tcg_out8(s, val);
+    } else {
+        tcg_out_opc(s, OPC_PUSH_Iv);
+        tcg_out32(s, val);
+    }
+}
+#endif

It might be better to declare this function as inline, so that you can remove the #if defined(CONFIG_SOFTMMU).

Quoted 71 lines

tcg/i386/tcg-target.c +16 ~~-18~~

+
+static inline void tcg_out_push(TCGContext *s, int reg)
+{
+    tcg_out_opc(s, OPC_PUSH_r32 + reg);
+}
+
+static inline void tcg_out_pop(TCGContext *s, int reg)
+{
+    tcg_out_opc(s, OPC_POP_r32 + reg);
+}
+
 static inline void tcg_out_ld(TCGContext *s, TCGType type, int ret,
                               int arg1, tcg_target_long arg2)
 {

@@ -910,8 +937,7 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
     if (opc == 3) {
         tcg_out_mov(s, TCG_REG_EDX, data_reg);
         tcg_out_mov(s, TCG_REG_ECX, data_reg2);
-        tcg_out8(s, 0x6a); /* push Ib */
-        tcg_out8(s, mem_index);
+        tcg_out_pushi(s, mem_index);
         tcg_out8(s, 0xe8);
         tcg_out32(s, (tcg_target_long)qemu_st_helpers[s_bits] - 
                   (tcg_target_long)s->code_ptr - 4);

@@ -936,10 +962,9 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
 #else
     if (opc == 3) {
         tcg_out_mov(s, TCG_REG_EDX, addr_reg2);
-        tcg_out8(s, 0x6a); /* push Ib */
-        tcg_out8(s, mem_index);
-        tcg_out_opc(s, 0x50 + data_reg2); /* push */
-        tcg_out_opc(s, 0x50 + data_reg); /* push */
+        tcg_out_pushi(s, mem_index);
+        tcg_out_push(s, data_reg2);
+        tcg_out_push(s, data_reg);
         tcg_out8(s, 0xe8);
         tcg_out32(s, (tcg_target_long)qemu_st_helpers[s_bits] - 
                   (tcg_target_long)s->code_ptr - 4);

@@ -957,8 +982,7 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
             tcg_out_mov(s, TCG_REG_ECX, data_reg);
             break;
         }
-        tcg_out8(s, 0x6a); /* push Ib */
-        tcg_out8(s, mem_index);
+        tcg_out_pushi(s, mem_index);
         tcg_out8(s, 0xe8);
         tcg_out32(s, (tcg_target_long)qemu_st_helpers[s_bits] - 
                   (tcg_target_long)s->code_ptr - 4);

@@ -1371,16 +1395,6 @@ static int tcg_target_callee_save_regs[] = {
     TCG_REG_EDI,
 };
 
-static inline void tcg_out_push(TCGContext *s, int reg)
-{
-    tcg_out_opc(s, 0x50 + reg);
-}
-
-static inline void tcg_out_pop(TCGContext *s, int reg)
-{
-    tcg_out_opc(s, 0x58 + reg);
-}
-
 /* Generate global QEMU prologue and epilogue code */
 void tcg_target_qemu_prologue(TCGContext *s)
 {

Signature · 1 line

1.6.6.1

Signature · 2 lines

Aurelien Jarno                          GPG: 1024D/F1BCDB73
[email protected]                 http://www.aurel32.net

Richard Henderson

author

2010-04-28 18:30 UTC #27 lore raw

PATCH 14/22 tcg-i386: Tidy calls.

In reply to Richard Henderson

Define OPC_CALL_Jz, generated by tcg_out_calli; use the later throughout. Unify the calls within qemu_st; adjust the stack with a single pop if applicable.

Define and use EXT_CALLN_Ev for indirect calls.

Signed-off-by Richard Henderson <[email protected]>

---
 tcg/i386/tcg-target.c |   49 +++++++++++++++++++++++++++----------------------
 1 files changed, 27 insertions(+), 22 deletions(-)

tcg/i386/tcg-target.c +27 ~~-22~~

index 58fcd23..e82788d 100644
@@ -167,6 +167,7 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define OPC_ARITH_EvIb	(0x83)
 #define OPC_ARITH_GvEv	(0x03)		/* ... plus (ARITH_FOO << 3) */
 #define OPC_ADD_GvEv	(OPC_ARITH_GvEv | (ARITH_ADD << 3))
+#define OPC_CALL_Jz	(0xe8)
 #define OPC_CMP_GvEv	(OPC_ARITH_GvEv | (ARITH_CMP << 3))
 #define OPC_DEC_r32	(0x48)
 #define OPC_BSWAP	(0xc8 | P_EXT)

@@ -210,6 +211,7 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define SHIFT_SAR 7
 
 /* Group 5 opcode extensions for 0xff.  */
+#define EXT_CALLN_Ev	2
 #define EXT_JMPN_Ev	4
 
 /* Condition codes to be added to OPC_JCC_{long,short}.  */

@@ -644,6 +646,12 @@ static void tcg_out_setcond2(TCGContext *s, const TCGArg *args,
     }
 }
 
+static void tcg_out_calli(TCGContext *s, tcg_target_long dest)
+{
+    tcg_out_opc(s, OPC_CALL_Jz);
+    tcg_out32(s, dest - (tcg_target_long)s->code_ptr - 4);
+}
+
 #if defined(CONFIG_SOFTMMU)
 
 #include "../../softmmu_defs.h"

@@ -748,9 +756,7 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
     tcg_out_mov(s, TCG_REG_EDX, addr_reg2);
     tcg_out_movi(s, TCG_TYPE_I32, TCG_REG_ECX, mem_index);
 #endif
-    tcg_out8(s, 0xe8);
-    tcg_out32(s, (tcg_target_long)qemu_ld_helpers[s_bits] - 
-              (tcg_target_long)s->code_ptr - 4);
+    tcg_out_calli(s, (tcg_target_long)qemu_ld_helpers[s_bits]);
 
     switch(opc) {
     case 0 | 4:

@@ -865,6 +871,7 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
 {
     int addr_reg, data_reg, data_reg2, r0, r1, mem_index, s_bits, bswap;
 #if defined(CONFIG_SOFTMMU)
+    int stack_adjust;
     uint8_t *label1_ptr, *label2_ptr;
 #endif
 #if TARGET_LONG_BITS == 64

@@ -938,10 +945,7 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
         tcg_out_mov(s, TCG_REG_EDX, data_reg);
         tcg_out_mov(s, TCG_REG_ECX, data_reg2);
         tcg_out_pushi(s, mem_index);
-        tcg_out8(s, 0xe8);
-        tcg_out32(s, (tcg_target_long)qemu_st_helpers[s_bits] - 
-                  (tcg_target_long)s->code_ptr - 4);
-        tcg_out_addi(s, TCG_REG_ESP, 4);
+        stack_adjust = 4;
     } else {
         switch(opc) {
         case 0:

@@ -955,9 +959,7 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
             break;
         }
         tcg_out_movi(s, TCG_TYPE_I32, TCG_REG_ECX, mem_index);
-        tcg_out8(s, 0xe8);
-        tcg_out32(s, (tcg_target_long)qemu_st_helpers[s_bits] - 
-                  (tcg_target_long)s->code_ptr - 4);
+        stack_adjust = 0;
     }
 #else
     if (opc == 3) {

@@ -965,10 +967,7 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
         tcg_out_pushi(s, mem_index);
         tcg_out_push(s, data_reg2);
         tcg_out_push(s, data_reg);
-        tcg_out8(s, 0xe8);
-        tcg_out32(s, (tcg_target_long)qemu_st_helpers[s_bits] - 
-                  (tcg_target_long)s->code_ptr - 4);
-        tcg_out_addi(s, TCG_REG_ESP, 12);
+        stack_adjust = 12;
     } else {
         tcg_out_mov(s, TCG_REG_EDX, addr_reg2);
         switch(opc) {

@@ -983,13 +982,19 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
             break;
         }
         tcg_out_pushi(s, mem_index);
-        tcg_out8(s, 0xe8);
-        tcg_out32(s, (tcg_target_long)qemu_st_helpers[s_bits] - 
-                  (tcg_target_long)s->code_ptr - 4);
-        tcg_out_addi(s, TCG_REG_ESP, 4);
+        stack_adjust = 4;
     }
 #endif
-    
+
+    tcg_out_calli(s, (tcg_target_long)qemu_st_helpers[s_bits]);
+
+    if (stack_adjust == 4) {
+        /* Pop and discard.  This is 2 bytes smaller than the add.  */
+        tcg_out_pop(s, TCG_REG_ECX);
+    } else if (stack_adjust != 0) {
+        tcg_out_addi(s, TCG_REG_ESP, stack_adjust);
+    }
+
     /* jmp label2 */
     tcg_out8(s, OPC_JMP_short);
     label2_ptr = s->code_ptr;

@@ -1082,10 +1087,10 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
         break;
     case INDEX_op_call:
         if (const_args[0]) {
-            tcg_out8(s, 0xe8);
-            tcg_out32(s, args[0] - (tcg_target_long)s->code_ptr - 4);
+            tcg_out_calli(s, args[0]);
         } else {
-            tcg_out_modrm(s, 0xff, 2, args[0]);
+            /* call *reg */
+            tcg_out_modrm(s, 0xff, EXT_CALLN_Ev, args[0]);
         }
         break;
     case INDEX_op_jmp:

Signature · 1 line

1.6.6.1

Aurelien Jarno

Acked-by on v1 14/22

2010-05-21 09:40 UTC #28 lore raw

Re: [Qemu-devel] [PATCH 14/22] tcg-i386: Tidy calls.

Quoted 7 lines · Richard Henderson

Define OPC_CALL_Jz, generated by tcg_out_calli; use the later throughout. Unify the calls within qemu_st; adjust the stack with a single pop if applicable.
Define and use EXT_CALLN_Ev for indirect calls.
Signed-off-by Richard Henderson <[email protected]>

Acked-by Aurelien Jarno <[email protected]>

Quoted 136 lines

---
 tcg/i386/tcg-target.c |   49 +++++++++++++++++++++++++++----------------------
 1 files changed, 27 insertions(+), 22 deletions(-)

tcg/i386/tcg-target.c +27 ~~-22~~

index 58fcd23..e82788d 100644
@@ -167,6 +167,7 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define OPC_ARITH_EvIb	(0x83)
 #define OPC_ARITH_GvEv	(0x03)		/* ... plus (ARITH_FOO << 3) */
 #define OPC_ADD_GvEv	(OPC_ARITH_GvEv | (ARITH_ADD << 3))
+#define OPC_CALL_Jz	(0xe8)
 #define OPC_CMP_GvEv	(OPC_ARITH_GvEv | (ARITH_CMP << 3))
 #define OPC_DEC_r32	(0x48)
 #define OPC_BSWAP	(0xc8 | P_EXT)

@@ -210,6 +211,7 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define SHIFT_SAR 7
 
 /* Group 5 opcode extensions for 0xff.  */
+#define EXT_CALLN_Ev	2
 #define EXT_JMPN_Ev	4
 
 /* Condition codes to be added to OPC_JCC_{long,short}.  */

@@ -644,6 +646,12 @@ static void tcg_out_setcond2(TCGContext *s, const TCGArg *args,
     }
 }
 
+static void tcg_out_calli(TCGContext *s, tcg_target_long dest)
+{
+    tcg_out_opc(s, OPC_CALL_Jz);
+    tcg_out32(s, dest - (tcg_target_long)s->code_ptr - 4);
+}
+
 #if defined(CONFIG_SOFTMMU)
 
 #include "../../softmmu_defs.h"

@@ -748,9 +756,7 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
     tcg_out_mov(s, TCG_REG_EDX, addr_reg2);
     tcg_out_movi(s, TCG_TYPE_I32, TCG_REG_ECX, mem_index);
 #endif
-    tcg_out8(s, 0xe8);
-    tcg_out32(s, (tcg_target_long)qemu_ld_helpers[s_bits] - 
-              (tcg_target_long)s->code_ptr - 4);
+    tcg_out_calli(s, (tcg_target_long)qemu_ld_helpers[s_bits]);
 
     switch(opc) {
     case 0 | 4:

@@ -865,6 +871,7 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
 {
     int addr_reg, data_reg, data_reg2, r0, r1, mem_index, s_bits, bswap;
 #if defined(CONFIG_SOFTMMU)
+    int stack_adjust;
     uint8_t *label1_ptr, *label2_ptr;
 #endif
 #if TARGET_LONG_BITS == 64

@@ -938,10 +945,7 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
         tcg_out_mov(s, TCG_REG_EDX, data_reg);
         tcg_out_mov(s, TCG_REG_ECX, data_reg2);
         tcg_out_pushi(s, mem_index);
-        tcg_out8(s, 0xe8);
-        tcg_out32(s, (tcg_target_long)qemu_st_helpers[s_bits] - 
-                  (tcg_target_long)s->code_ptr - 4);
-        tcg_out_addi(s, TCG_REG_ESP, 4);
+        stack_adjust = 4;
     } else {
         switch(opc) {
         case 0:

@@ -955,9 +959,7 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
             break;
         }
         tcg_out_movi(s, TCG_TYPE_I32, TCG_REG_ECX, mem_index);
-        tcg_out8(s, 0xe8);
-        tcg_out32(s, (tcg_target_long)qemu_st_helpers[s_bits] - 
-                  (tcg_target_long)s->code_ptr - 4);
+        stack_adjust = 0;
     }
 #else
     if (opc == 3) {

@@ -965,10 +967,7 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
         tcg_out_pushi(s, mem_index);
         tcg_out_push(s, data_reg2);
         tcg_out_push(s, data_reg);
-        tcg_out8(s, 0xe8);
-        tcg_out32(s, (tcg_target_long)qemu_st_helpers[s_bits] - 
-                  (tcg_target_long)s->code_ptr - 4);
-        tcg_out_addi(s, TCG_REG_ESP, 12);
+        stack_adjust = 12;
     } else {
         tcg_out_mov(s, TCG_REG_EDX, addr_reg2);
         switch(opc) {

@@ -983,13 +982,19 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
             break;
         }
         tcg_out_pushi(s, mem_index);
-        tcg_out8(s, 0xe8);
-        tcg_out32(s, (tcg_target_long)qemu_st_helpers[s_bits] - 
-                  (tcg_target_long)s->code_ptr - 4);
-        tcg_out_addi(s, TCG_REG_ESP, 4);
+        stack_adjust = 4;
     }
 #endif
-    
+
+    tcg_out_calli(s, (tcg_target_long)qemu_st_helpers[s_bits]);
+
+    if (stack_adjust == 4) {
+        /* Pop and discard.  This is 2 bytes smaller than the add.  */
+        tcg_out_pop(s, TCG_REG_ECX);
+    } else if (stack_adjust != 0) {
+        tcg_out_addi(s, TCG_REG_ESP, stack_adjust);
+    }
+
     /* jmp label2 */
     tcg_out8(s, OPC_JMP_short);
     label2_ptr = s->code_ptr;

@@ -1082,10 +1087,10 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
         break;
     case INDEX_op_call:
         if (const_args[0]) {
-            tcg_out8(s, 0xe8);
-            tcg_out32(s, args[0] - (tcg_target_long)s->code_ptr - 4);
+            tcg_out_calli(s, args[0]);
         } else {
-            tcg_out_modrm(s, 0xff, 2, args[0]);
+            /* call *reg */
+            tcg_out_modrm(s, 0xff, EXT_CALLN_Ev, args[0]);
         }
         break;
     case INDEX_op_jmp:

Signature · 1 line

1.6.6.1

Signature · 2 lines

Aurelien Jarno                          GPG: 1024D/F1BCDB73
[email protected]                 http://www.aurel32.net

Richard Henderson

author

2010-04-28 18:30 UTC #29 lore raw

PATCH 15/22 tcg-i386: Tidy ret.

In reply to Richard Henderson

Define and use OPC_RET.

Signed-off-by Richard Henderson <[email protected]>

---
 tcg/i386/tcg-target.c |    3 ++-
 1 files changed, 2 insertions(+), 1 deletions(-)

tcg/i386/tcg-target.c +2 -1

index e82788d..27e9e9e 100644
@@ -188,6 +188,7 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define OPC_PUSH_r32	(0x50)
 #define OPC_PUSH_Iv	(0x68)
 #define OPC_PUSH_Ib	(0x6a)
+#define OPC_RET		(0xc3)
 #define OPC_SHIFT_1	(0xd1)
 #define OPC_SHIFT_Ib	(0xc1)
 #define OPC_SHIFT_cl	(0xd3)

@@ -1426,7 +1427,7 @@ void tcg_target_qemu_prologue(TCGContext *s)
     for(i = ARRAY_SIZE(tcg_target_callee_save_regs) - 1; i >= 0; i--) {
         tcg_out_pop(s, tcg_target_callee_save_regs[i]);
     }
-    tcg_out8(s, 0xc3); /* ret */
+    tcg_out_opc(s, OPC_RET);
 }
 
 void tcg_target_init(TCGContext *s)

Signature · 1 line

1.6.6.1

Aurelien Jarno

Acked-by on v1 15/22

2010-05-21 09:41 UTC #30 lore raw

Re: [Qemu-devel] [PATCH 15/22] tcg-i386: Tidy ret.

Quoted 3 lines · Richard Henderson

Define and use OPC_RET.
Signed-off-by Richard Henderson <[email protected]>

Acked-by Aurelien Jarno <[email protected]>

Quoted 31 lines

---
 tcg/i386/tcg-target.c |    3 ++-
 1 files changed, 2 insertions(+), 1 deletions(-)

tcg/i386/tcg-target.c +2 -1

index e82788d..27e9e9e 100644
@@ -188,6 +188,7 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define OPC_PUSH_r32	(0x50)
 #define OPC_PUSH_Iv	(0x68)
 #define OPC_PUSH_Ib	(0x6a)
+#define OPC_RET		(0xc3)
 #define OPC_SHIFT_1	(0xd1)
 #define OPC_SHIFT_Ib	(0xc1)
 #define OPC_SHIFT_cl	(0xd3)

@@ -1426,7 +1427,7 @@ void tcg_target_qemu_prologue(TCGContext *s)
     for(i = ARRAY_SIZE(tcg_target_callee_save_regs) - 1; i >= 0; i--) {
         tcg_out_pop(s, tcg_target_callee_save_regs[i]);
     }
-    tcg_out8(s, 0xc3); /* ret */
+    tcg_out_opc(s, OPC_RET);
 }
 
 void tcg_target_init(TCGContext *s)

Signature · 1 line

1.6.6.1

Signature · 2 lines

Aurelien Jarno                          GPG: 1024D/F1BCDB73
[email protected]                 http://www.aurel32.net

Richard Henderson

author

2010-04-28 18:30 UTC #31 lore raw

PATCH 16/22 tcg-i386: Tidy setcc.

In reply to Richard Henderson

Define and use OPC_SETCC.

Signed-off-by Richard Henderson <[email protected]>

---
 tcg/i386/tcg-target.c |    4 ++--
 1 files changed, 2 insertions(+), 2 deletions(-)

tcg/i386/tcg-target.c +2 -2

index 27e9e9e..0c1a53a 100644
@@ -189,6 +189,7 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define OPC_PUSH_Iv	(0x68)
 #define OPC_PUSH_Ib	(0x6a)
 #define OPC_RET		(0xc3)
+#define OPC_SETCC	(0x90 | P_EXT)	/* ... plus condition code */
 #define OPC_SHIFT_1	(0xd1)
 #define OPC_SHIFT_Ib	(0xc1)
 #define OPC_SHIFT_cl	(0xd3)

@@ -600,8 +601,7 @@ static void tcg_out_setcond(TCGContext *s, TCGCond cond, TCGArg dest,
                             TCGArg arg1, TCGArg arg2, int const_arg2)
 {
     tcg_out_cmp(s, arg1, arg2, const_arg2);
-    /* setcc */
-    tcg_out_modrm(s, 0x90 | tcg_cond_to_jcc[cond] | P_EXT, 0, dest);
+    tcg_out_modrm(s, OPC_SETCC | tcg_cond_to_jcc[cond], 0, dest);
     tcg_out_ext8u(s, dest, dest);
 }

Signature · 1 line

1.6.6.1

Aurelien Jarno

Acked-by on v1 16/22

2010-05-21 09:41 UTC #32 lore raw

Re: [Qemu-devel] [PATCH 16/22] tcg-i386: Tidy setcc.

Quoted 3 lines · Richard Henderson

Define and use OPC_SETCC.
Signed-off-by Richard Henderson <[email protected]>

Acked-by Aurelien Jarno <[email protected]>

Quoted 32 lines

---
 tcg/i386/tcg-target.c |    4 ++--
 1 files changed, 2 insertions(+), 2 deletions(-)

tcg/i386/tcg-target.c +2 -2

index 27e9e9e..0c1a53a 100644
@@ -189,6 +189,7 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define OPC_PUSH_Iv	(0x68)
 #define OPC_PUSH_Ib	(0x6a)
 #define OPC_RET		(0xc3)
+#define OPC_SETCC	(0x90 | P_EXT)	/* ... plus condition code */
 #define OPC_SHIFT_1	(0xd1)
 #define OPC_SHIFT_Ib	(0xc1)
 #define OPC_SHIFT_cl	(0xd3)

@@ -600,8 +601,7 @@ static void tcg_out_setcond(TCGContext *s, TCGCond cond, TCGArg dest,
                             TCGArg arg1, TCGArg arg2, int const_arg2)
 {
     tcg_out_cmp(s, arg1, arg2, const_arg2);
-    /* setcc */
-    tcg_out_modrm(s, 0x90 | tcg_cond_to_jcc[cond] | P_EXT, 0, dest);
+    tcg_out_modrm(s, OPC_SETCC | tcg_cond_to_jcc[cond], 0, dest);
     tcg_out_ext8u(s, dest, dest);
 }

Signature · 1 line

1.6.6.1

Signature · 2 lines

Aurelien Jarno                          GPG: 1024D/F1BCDB73
[email protected]                 http://www.aurel32.net

Richard Henderson

author

2010-04-28 18:30 UTC #33 lore raw

PATCH 17/22 tcg-i386: Tidy unary arithmetic.

In reply to Richard Henderson

Define OPC_GRP3 and EXT3_FOO to match. Use them instead of bare constants.

Define OPC_GRP5 and rename the existing EXT_BAR to EXT5_BAR to make it clear which extension should be used with which opcode.

Signed-off-by Richard Henderson <[email protected]>

---
 tcg/i386/tcg-target.c |   38 +++++++++++++++++++++++++-------------
 1 files changed, 25 insertions(+), 13 deletions(-)

tcg/i386/tcg-target.c +25 ~~-13~~

index 0c1a53a..c3f3e4d 100644
@@ -195,7 +195,11 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define OPC_SHIFT_cl	(0xd3)
 #define OPC_TESTL	(0x85)
 
-/* Group 1 opcode extensions for 0x80-0x83.  */
+#define OPC_GRP3_Ev	(0xf7)
+#define OPC_GRP5	(0xff)
+
+/* Group 1 opcode extensions for 0x80-0x83.
+   These are also used as modifiers for OPC_ARITH.  */
 #define ARITH_ADD 0
 #define ARITH_OR  1
 #define ARITH_ADC 2

@@ -212,9 +216,17 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define SHIFT_SHR 5
 #define SHIFT_SAR 7
 
-/* Group 5 opcode extensions for 0xff.  */
-#define EXT_CALLN_Ev	2
-#define EXT_JMPN_Ev	4
+/* Group 3 opcode extensions for 0xf6, 0xf7.  To be used with OPC_GRP3.  */
+#define EXT3_NOT   2
+#define EXT3_NEG   3
+#define EXT3_MUL   4
+#define EXT3_IMUL  5
+#define EXT3_DIV   6
+#define EXT3_IDIV  7
+
+/* Group 5 opcode extensions for 0xff.  To be used with OPC_GRP5.  */
+#define EXT5_CALLN_Ev	2
+#define EXT5_JMPN_Ev	4
 
 /* Condition codes to be added to OPC_JCC_{long,short}.  */
 #define JCC_JMP (-1)

@@ -1081,7 +1093,7 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
             tcg_out32(s, 0);
         } else {
             /* indirect jump method */
-            tcg_out_modrm_offset(s, 0xff, EXT_JMPN_Ev, -1, 
+            tcg_out_modrm_offset(s, OPC_GRP5, EXT5_JMPN_Ev, -1, 
                                  (tcg_target_long)(s->tb_next + args[0]));
         }
         s->tb_next_offset[args[0]] = s->code_ptr - s->code_buf;

@@ -1091,7 +1103,7 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
             tcg_out_calli(s, args[0]);
         } else {
             /* call *reg */
-            tcg_out_modrm(s, 0xff, EXT_CALLN_Ev, args[0]);
+            tcg_out_modrm(s, OPC_GRP5, EXT5_CALLN_Ev, args[0]);
         }
         break;
     case INDEX_op_jmp:

@@ -1100,7 +1112,7 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
             tcg_out32(s, args[0] - (tcg_target_long)s->code_ptr - 4);
         } else {
             /* jmp *reg */
-            tcg_out_modrm(s, 0xff, EXT_JMPN_Ev, args[0]);
+            tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, args[0]);
         }
         break;
     case INDEX_op_br:

@@ -1177,13 +1189,13 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
         }
         break;
     case INDEX_op_mulu2_i32:
-        tcg_out_modrm(s, 0xf7, 4, args[3]);
+        tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_MUL, args[3]);
         break;
     case INDEX_op_div2_i32:
-        tcg_out_modrm(s, 0xf7, 7, args[4]);
+        tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_IDIV, args[4]);
         break;
     case INDEX_op_divu2_i32:
-        tcg_out_modrm(s, 0xf7, 6, args[4]);
+        tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_DIV, args[4]);
         break;
     case INDEX_op_shl_i32:
         c = SHIFT_SHL;

@@ -1247,11 +1259,11 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
         break;
 
     case INDEX_op_neg_i32:
-        tcg_out_modrm(s, 0xf7, 3, args[0]);
+        tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_NEG, args[0]);
         break;
 
     case INDEX_op_not_i32:
-        tcg_out_modrm(s, 0xf7, 2, args[0]);
+        tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_NOT, args[0]);
         break;
 
     case INDEX_op_ext8s_i32:

@@ -1419,7 +1431,7 @@ void tcg_target_qemu_prologue(TCGContext *s)
     stack_addend = frame_size - push_size;
     tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
 
-    tcg_out_modrm(s, 0xff, EXT_JMPN_Ev, TCG_REG_EAX); /* jmp *%eax */
+    tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, TCG_REG_EAX); /* jmp *%eax */
     
     /* TB epilogue */
     tb_ret_addr = s->code_ptr;

Signature · 1 line

1.6.6.1

Aurelien Jarno

Acked-by on v1 17/22

2010-05-21 09:43 UTC #34 lore raw

Re: [Qemu-devel] [PATCH 17/22] tcg-i386: Tidy unary arithmetic.

Quoted 7 lines · Richard Henderson

Define OPC_GRP3 and EXT3_FOO to match. Use them instead of bare constants.
Define OPC_GRP5 and rename the existing EXT_BAR to EXT5_BAR to make it clear which extension should be used with which opcode.
Signed-off-by Richard Henderson <[email protected]>

Acked-by Aurelien Jarno <[email protected]>

Quoted 115 lines

---
 tcg/i386/tcg-target.c |   38 +++++++++++++++++++++++++-------------
 1 files changed, 25 insertions(+), 13 deletions(-)

tcg/i386/tcg-target.c +25 ~~-13~~

index 0c1a53a..c3f3e4d 100644
@@ -195,7 +195,11 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define OPC_SHIFT_cl	(0xd3)
 #define OPC_TESTL	(0x85)
 
-/* Group 1 opcode extensions for 0x80-0x83.  */
+#define OPC_GRP3_Ev	(0xf7)
+#define OPC_GRP5	(0xff)
+
+/* Group 1 opcode extensions for 0x80-0x83.
+   These are also used as modifiers for OPC_ARITH.  */
 #define ARITH_ADD 0
 #define ARITH_OR  1
 #define ARITH_ADC 2

@@ -212,9 +216,17 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define SHIFT_SHR 5
 #define SHIFT_SAR 7
 
-/* Group 5 opcode extensions for 0xff.  */
-#define EXT_CALLN_Ev	2
-#define EXT_JMPN_Ev	4
+/* Group 3 opcode extensions for 0xf6, 0xf7.  To be used with OPC_GRP3.  */
+#define EXT3_NOT   2
+#define EXT3_NEG   3
+#define EXT3_MUL   4
+#define EXT3_IMUL  5
+#define EXT3_DIV   6
+#define EXT3_IDIV  7
+
+/* Group 5 opcode extensions for 0xff.  To be used with OPC_GRP5.  */
+#define EXT5_CALLN_Ev	2
+#define EXT5_JMPN_Ev	4
 
 /* Condition codes to be added to OPC_JCC_{long,short}.  */
 #define JCC_JMP (-1)

@@ -1081,7 +1093,7 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
             tcg_out32(s, 0);
         } else {
             /* indirect jump method */
-            tcg_out_modrm_offset(s, 0xff, EXT_JMPN_Ev, -1, 
+            tcg_out_modrm_offset(s, OPC_GRP5, EXT5_JMPN_Ev, -1, 
                                  (tcg_target_long)(s->tb_next + args[0]));
         }
         s->tb_next_offset[args[0]] = s->code_ptr - s->code_buf;

@@ -1091,7 +1103,7 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
             tcg_out_calli(s, args[0]);
         } else {
             /* call *reg */
-            tcg_out_modrm(s, 0xff, EXT_CALLN_Ev, args[0]);
+            tcg_out_modrm(s, OPC_GRP5, EXT5_CALLN_Ev, args[0]);
         }
         break;
     case INDEX_op_jmp:

@@ -1100,7 +1112,7 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
             tcg_out32(s, args[0] - (tcg_target_long)s->code_ptr - 4);
         } else {
             /* jmp *reg */
-            tcg_out_modrm(s, 0xff, EXT_JMPN_Ev, args[0]);
+            tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, args[0]);
         }
         break;
     case INDEX_op_br:

@@ -1177,13 +1189,13 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
         }
         break;
     case INDEX_op_mulu2_i32:
-        tcg_out_modrm(s, 0xf7, 4, args[3]);
+        tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_MUL, args[3]);
         break;
     case INDEX_op_div2_i32:
-        tcg_out_modrm(s, 0xf7, 7, args[4]);
+        tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_IDIV, args[4]);
         break;
     case INDEX_op_divu2_i32:
-        tcg_out_modrm(s, 0xf7, 6, args[4]);
+        tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_DIV, args[4]);
         break;
     case INDEX_op_shl_i32:
         c = SHIFT_SHL;

@@ -1247,11 +1259,11 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
         break;
 
     case INDEX_op_neg_i32:
-        tcg_out_modrm(s, 0xf7, 3, args[0]);
+        tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_NEG, args[0]);
         break;
 
     case INDEX_op_not_i32:
-        tcg_out_modrm(s, 0xf7, 2, args[0]);
+        tcg_out_modrm(s, OPC_GRP3_Ev, EXT3_NOT, args[0]);
         break;
 
     case INDEX_op_ext8s_i32:

@@ -1419,7 +1431,7 @@ void tcg_target_qemu_prologue(TCGContext *s)
     stack_addend = frame_size - push_size;
     tcg_out_addi(s, TCG_REG_ESP, -stack_addend);
 
-    tcg_out_modrm(s, 0xff, EXT_JMPN_Ev, TCG_REG_EAX); /* jmp *%eax */
+    tcg_out_modrm(s, OPC_GRP5, EXT5_JMPN_Ev, TCG_REG_EAX); /* jmp *%eax */
 
     /* TB epilogue */
     tb_ret_addr = s->code_ptr;

Signature · 1 line

1.6.6.1

Signature · 2 lines

Aurelien Jarno                          GPG: 1024D/F1BCDB73
[email protected]                 http://www.aurel32.net

Richard Henderson

author

2010-04-28 18:30 UTC #35 lore raw

PATCH 18/22 tcg-i386: Tidy multiply.

In reply to Richard Henderson

Define and use OPC_IMUL_GvEv{,Ib,Iz}.

Signed-off-by Richard Henderson <[email protected]>

---
 tcg/i386/tcg-target.c |    9 ++++++---
 1 files changed, 6 insertions(+), 3 deletions(-)

tcg/i386/tcg-target.c +6 -3

index c3f3e4d..020faf0 100644
@@ -184,6 +184,9 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define OPC_MOVZWL	(0xb7 | P_EXT)
 #define OPC_MOVSBL	(0xbe | P_EXT)
 #define OPC_MOVSWL	(0xbf | P_EXT)
+#define OPC_IMUL_GvEv	(0xaf | P_EXT)
+#define OPC_IMUL_GvEvIb	(0x6b)
+#define OPC_IMUL_GvEvIz	(0x69)
 #define OPC_POP_r32	(0x58)
 #define OPC_PUSH_r32	(0x50)
 #define OPC_PUSH_Iv	(0x68)

@@ -1178,14 +1181,14 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
             int32_t val;
             val = args[2];
             if (val == (int8_t)val) {
-                tcg_out_modrm(s, 0x6b, args[0], args[0]);
+                tcg_out_modrm(s, OPC_IMUL_GvEvIb, args[0], args[0]);
                 tcg_out8(s, val);
             } else {
-                tcg_out_modrm(s, 0x69, args[0], args[0]);
+                tcg_out_modrm(s, OPC_IMUL_GvEvIz, args[0], args[0]);
                 tcg_out32(s, val);
             }
         } else {
-            tcg_out_modrm(s, 0xaf | P_EXT, args[0], args[2]);
+            tcg_out_modrm(s, OPC_IMUL_GvEv, args[0], args[2]);
         }
         break;
     case INDEX_op_mulu2_i32:

Signature · 1 line

1.6.6.1

Aurelien Jarno

Acked-by on v1 18/22

2010-05-21 09:43 UTC #36 lore raw

Re: [Qemu-devel] [PATCH 18/22] tcg-i386: Tidy multiply.

Quoted 3 lines · Richard Henderson

Define and use OPC_IMUL_GvEv{,Ib,Iz}.
Signed-off-by Richard Henderson <[email protected]>

Acked-by Aurelien Jarno <[email protected]>

Quoted 42 lines

---
 tcg/i386/tcg-target.c |    9 ++++++---
 1 files changed, 6 insertions(+), 3 deletions(-)

tcg/i386/tcg-target.c +6 -3

index c3f3e4d..020faf0 100644
@@ -184,6 +184,9 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define OPC_MOVZWL	(0xb7 | P_EXT)
 #define OPC_MOVSBL	(0xbe | P_EXT)
 #define OPC_MOVSWL	(0xbf | P_EXT)
+#define OPC_IMUL_GvEv	(0xaf | P_EXT)
+#define OPC_IMUL_GvEvIb	(0x6b)
+#define OPC_IMUL_GvEvIz	(0x69)
 #define OPC_POP_r32	(0x58)
 #define OPC_PUSH_r32	(0x50)
 #define OPC_PUSH_Iv	(0x68)

@@ -1178,14 +1181,14 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
             int32_t val;
             val = args[2];
             if (val == (int8_t)val) {
-                tcg_out_modrm(s, 0x6b, args[0], args[0]);
+                tcg_out_modrm(s, OPC_IMUL_GvEvIb, args[0], args[0]);
                 tcg_out8(s, val);
             } else {
-                tcg_out_modrm(s, 0x69, args[0], args[0]);
+                tcg_out_modrm(s, OPC_IMUL_GvEvIz, args[0], args[0]);
                 tcg_out32(s, val);
             }
         } else {
-            tcg_out_modrm(s, 0xaf | P_EXT, args[0], args[2]);
+            tcg_out_modrm(s, OPC_IMUL_GvEv, args[0], args[2]);
         }
         break;
     case INDEX_op_mulu2_i32:

Signature · 1 line

1.6.6.1

Signature · 2 lines

Aurelien Jarno                          GPG: 1024D/F1BCDB73
[email protected]                 http://www.aurel32.net

Richard Henderson

author

2010-04-28 18:30 UTC #37 lore raw

PATCH 19/22 tcg-i386: Tidy xchg.

In reply to Richard Henderson

Define and use OPC_XCHG_ax_r32.

Signed-off-by Richard Henderson <[email protected]>

---
 tcg/i386/tcg-target.c |    4 +++-
 1 files changed, 3 insertions(+), 1 deletions(-)

tcg/i386/tcg-target.c +3 -1

index 020faf0..bf3c0d6 100644
@@ -197,6 +197,7 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define OPC_SHIFT_Ib	(0xc1)
 #define OPC_SHIFT_cl	(0xd3)
 #define OPC_TESTL	(0x85)
+#define OPC_XCHG_ax_r32	(0x90)
 
 #define OPC_GRP3_Ev	(0xf7)
 #define OPC_GRP5	(0xff)

@@ -793,7 +794,8 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
         break;
     case 3:
         if (data_reg == TCG_REG_EDX) {
-            tcg_out_opc(s, 0x90 + TCG_REG_EDX); /* xchg %edx, %eax */
+            /* xchg %edx, %eax */
+            tcg_out_opc(s, OPC_XCHG_ax_r32 + TCG_REG_EDX);
             tcg_out_mov(s, data_reg2, TCG_REG_EAX);
         } else {
             tcg_out_mov(s, data_reg, TCG_REG_EAX);

Signature · 1 line

1.6.6.1

Aurelien Jarno

Acked-by on v1 19/22

2010-05-21 09:43 UTC #38 lore raw

Re: [Qemu-devel] [PATCH 19/22] tcg-i386: Tidy xchg.

Quoted 3 lines · Richard Henderson

Define and use OPC_XCHG_ax_r32.
Signed-off-by Richard Henderson <[email protected]>

Acked-by Aurelien Jarno <[email protected]>

Quoted 32 lines

---
 tcg/i386/tcg-target.c |    4 +++-
 1 files changed, 3 insertions(+), 1 deletions(-)

tcg/i386/tcg-target.c +3 -1

index 020faf0..bf3c0d6 100644
@@ -197,6 +197,7 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define OPC_SHIFT_Ib	(0xc1)
 #define OPC_SHIFT_cl	(0xd3)
 #define OPC_TESTL	(0x85)
+#define OPC_XCHG_ax_r32	(0x90)
 
 #define OPC_GRP3_Ev	(0xf7)
 #define OPC_GRP5	(0xff)

@@ -793,7 +794,8 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
         break;
     case 3:
         if (data_reg == TCG_REG_EDX) {
-            tcg_out_opc(s, 0x90 + TCG_REG_EDX); /* xchg %edx, %eax */
+            /* xchg %edx, %eax */
+            tcg_out_opc(s, OPC_XCHG_ax_r32 + TCG_REG_EDX);
             tcg_out_mov(s, data_reg2, TCG_REG_EAX);
         } else {
             tcg_out_mov(s, data_reg, TCG_REG_EAX);

Signature · 1 line

1.6.6.1

Signature · 2 lines

Aurelien Jarno                          GPG: 1024D/F1BCDB73
[email protected]                 http://www.aurel32.net

Richard Henderson

author

2010-04-28 18:30 UTC #39 lore raw

PATCH 20/22 tcg-i386: Tidy lea.

In reply to Richard Henderson

Implement full modrm+sib addressing mode processing. Use that in qemu_ld/st to output the LEA.

Signed-off-by Richard Henderson <[email protected]>

---
 tcg/i386/tcg-target.c |   91 ++++++++++++++++++++++++++++++++-----------------
 1 files changed, 60 insertions(+), 31 deletions(-)

tcg/i386/tcg-target.c +60 ~~-31~~

index bf3c0d6..755d46d 100644
@@ -176,6 +176,7 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define OPC_JCC_short	(0x70)		/* ... plus condition code */
 #define OPC_JMP_long	(0xe9)
 #define OPC_JMP_short	(0xeb)
+#define OPC_LEA         (0x8d)
 #define OPC_MOVB_EvGv	(0x88)		/* stores, more or less */
 #define OPC_MOVL_EvGv	(0x89)		/* stores, more or less */
 #define OPC_MOVL_GvEv	(0x8b)		/* loads, more or less */

@@ -277,40 +278,70 @@ static inline void tcg_out_modrm(TCGContext *s, int opc, int r, int rm)
     tcg_out8(s, 0xc0 | (r << 3) | rm);
 }
 
-/* rm == -1 means no register index */
-static inline void tcg_out_modrm_offset(TCGContext *s, int opc, int r, int rm, 
-                                        int32_t offset)
+/* Output an opcode with a full "rm + (index<<shift) + offset" address mode.  
+   We handle either RM and INDEX missing with a -1 value.  */
+
+static void tcg_out_modrm_sib_offset(TCGContext *s, int opc, int r, int rm,
+                                     int index, int shift, int32_t offset)
 {
+    int mod, len;
+
+    if (index == -1 && rm == -1) {
+        /* Absolute address.  */
+        tcg_out_opc(s, opc);
+        tcg_out8(s, (r << 3) | 5);
+        tcg_out32(s, offset);
+        return;
+    }
+
     tcg_out_opc(s, opc);
+
+    /* Find the length of the immediate addend.  Note that the encoding
+       that would be used for (%ebp) indicates absolute addressing.  */
     if (rm == -1) {
-        tcg_out8(s, 0x05 | (r << 3));
-        tcg_out32(s, offset);
+        mod = 0, len = 4, rm = 5;
     } else if (offset == 0 && rm != TCG_REG_EBP) {
-        if (rm == TCG_REG_ESP) {
-            tcg_out8(s, 0x04 | (r << 3));
-            tcg_out8(s, 0x24);
-        } else {
-            tcg_out8(s, 0x00 | (r << 3) | rm);
-        }
-    } else if ((int8_t)offset == offset) {
-        if (rm == TCG_REG_ESP) {
-            tcg_out8(s, 0x44 | (r << 3));
-            tcg_out8(s, 0x24);
-        } else {
-            tcg_out8(s, 0x40 | (r << 3) | rm);
-        }
-        tcg_out8(s, offset);
+        mod = 0, len = 0;
+    } else if (offset == (int8_t)offset) {
+        mod = 0x40, len = 1;
     } else {
-        if (rm == TCG_REG_ESP) {
-            tcg_out8(s, 0x84 | (r << 3));
-            tcg_out8(s, 0x24);
+        mod = 0x80, len = 4;
+    }
+
+    /* Use a single byte MODRM format if possible.  Note that the encoding
+       that would be used for %esp is the escape to the two byte form.  */
+    if (index == -1 && rm != TCG_REG_ESP) {
+        /* Single byte MODRM format.  */
+        tcg_out8(s, mod | (r << 3) | rm);
+    } else {
+        /* Two byte MODRM+SIB format.  */
+
+        /* Note that the encoding that would place %esp into the index
+           field indicates no index register.  */
+        if (index == -1) {
+            index = 4;
         } else {
-            tcg_out8(s, 0x80 | (r << 3) | rm);
+            assert(index != TCG_REG_ESP);
         }
+
+        tcg_out8(s, mod | (r << 3) | 4);
+        tcg_out8(s, (shift << 6) | (index << 3) | rm);
+    }
+
+    if (len == 1) {
+        tcg_out8(s, offset);
+    } else if (len == 4) {
         tcg_out32(s, offset);
     }
 }
 
+/* rm == -1 means no register index */
+static inline void tcg_out_modrm_offset(TCGContext *s, int opc, int r, int rm, 
+                                        int32_t offset)
+{
+    tcg_out_modrm_sib_offset(s, opc, r, rm, -1, 0, offset);
+}
+
 /* Generate dest op= src.  Uses the same ARITH_* codes as tgen_arithi.  */
 static inline void tgen_arithr(TCGContext *s, int subop, int dest, int src)
 {

@@ -733,10 +764,9 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
     tgen_arithi(s, ARITH_AND, r0, TARGET_PAGE_MASK | ((1 << s_bits) - 1), 0);
     tgen_arithi(s, ARITH_AND, r1, (CPU_TLB_SIZE - 1) << CPU_TLB_ENTRY_BITS, 0);
 
-    tcg_out_opc(s, 0x8d); /* lea offset(r1, %ebp), r1 */
-    tcg_out8(s, 0x80 | (r1 << 3) | 0x04);
-    tcg_out8(s, (5 << 3) | r1);
-    tcg_out32(s, offsetof(CPUState, tlb_table[mem_index][0].addr_read));
+    tcg_out_modrm_sib_offset(s, OPC_LEA, r1, TCG_AREG0, r1, 0,
+                             offsetof(CPUState,
+                                      tlb_table[mem_index][0].addr_read));
 
     /* cmp 0(r1), r0 */
     tcg_out_modrm_offset(s, OPC_CMP_GvEv, r0, r1, 0);

@@ -924,10 +954,9 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
     tgen_arithi(s, ARITH_AND, r0, TARGET_PAGE_MASK | ((1 << s_bits) - 1), 0);
     tgen_arithi(s, ARITH_AND, r1, (CPU_TLB_SIZE - 1) << CPU_TLB_ENTRY_BITS, 0);
 
-    tcg_out_opc(s, 0x8d); /* lea offset(r1, %ebp), r1 */
-    tcg_out8(s, 0x80 | (r1 << 3) | 0x04);
-    tcg_out8(s, (5 << 3) | r1);
-    tcg_out32(s, offsetof(CPUState, tlb_table[mem_index][0].addr_write));
+    tcg_out_modrm_sib_offset(s, OPC_LEA, r1, TCG_AREG0, r1, 0,
+                             offsetof(CPUState,
+                                      tlb_table[mem_index][0].addr_write));
 
     /* cmp 0(r1), r0 */
     tcg_out_modrm_offset(s, OPC_CMP_GvEv, r0, r1, 0);

Signature · 1 line

1.6.6.1

Aurelien Jarno

Acked-by on v1 20/22

2010-05-21 09:45 UTC #40 lore raw

Re: [Qemu-devel] [PATCH 20/22] tcg-i386: Tidy lea.

Quoted 4 lines · Richard Henderson

Implement full modrm+sib addressing mode processing. Use that in qemu_ld/st to output the LEA.
Signed-off-by Richard Henderson <[email protected]>

Acked-by Aurelien Jarno <[email protected]>

Quoted 144 lines

---
 tcg/i386/tcg-target.c |   91 ++++++++++++++++++++++++++++++++-----------------
 1 files changed, 60 insertions(+), 31 deletions(-)

tcg/i386/tcg-target.c +60 ~~-31~~

index bf3c0d6..755d46d 100644
@@ -176,6 +176,7 @@ static inline int tcg_target_const_match(tcg_target_long val,
 #define OPC_JCC_short	(0x70)		/* ... plus condition code */
 #define OPC_JMP_long	(0xe9)
 #define OPC_JMP_short	(0xeb)
+#define OPC_LEA         (0x8d)
 #define OPC_MOVB_EvGv	(0x88)		/* stores, more or less */
 #define OPC_MOVL_EvGv	(0x89)		/* stores, more or less */
 #define OPC_MOVL_GvEv	(0x8b)		/* loads, more or less */

@@ -277,40 +278,70 @@ static inline void tcg_out_modrm(TCGContext *s, int opc, int r, int rm)
     tcg_out8(s, 0xc0 | (r << 3) | rm);
 }
 
-/* rm == -1 means no register index */
-static inline void tcg_out_modrm_offset(TCGContext *s, int opc, int r, int rm, 
-                                        int32_t offset)
+/* Output an opcode with a full "rm + (index<<shift) + offset" address mode.  
+   We handle either RM and INDEX missing with a -1 value.  */
+
+static void tcg_out_modrm_sib_offset(TCGContext *s, int opc, int r, int rm,
+                                     int index, int shift, int32_t offset)
 {
+    int mod, len;
+
+    if (index == -1 && rm == -1) {
+        /* Absolute address.  */
+        tcg_out_opc(s, opc);
+        tcg_out8(s, (r << 3) | 5);
+        tcg_out32(s, offset);
+        return;
+    }
+
     tcg_out_opc(s, opc);
+
+    /* Find the length of the immediate addend.  Note that the encoding
+       that would be used for (%ebp) indicates absolute addressing.  */
     if (rm == -1) {
-        tcg_out8(s, 0x05 | (r << 3));
-        tcg_out32(s, offset);
+        mod = 0, len = 4, rm = 5;
     } else if (offset == 0 && rm != TCG_REG_EBP) {
-        if (rm == TCG_REG_ESP) {
-            tcg_out8(s, 0x04 | (r << 3));
-            tcg_out8(s, 0x24);
-        } else {
-            tcg_out8(s, 0x00 | (r << 3) | rm);
-        }
-    } else if ((int8_t)offset == offset) {
-        if (rm == TCG_REG_ESP) {
-            tcg_out8(s, 0x44 | (r << 3));
-            tcg_out8(s, 0x24);
-        } else {
-            tcg_out8(s, 0x40 | (r << 3) | rm);
-        }
-        tcg_out8(s, offset);
+        mod = 0, len = 0;
+    } else if (offset == (int8_t)offset) {
+        mod = 0x40, len = 1;
     } else {
-        if (rm == TCG_REG_ESP) {
-            tcg_out8(s, 0x84 | (r << 3));
-            tcg_out8(s, 0x24);
+        mod = 0x80, len = 4;
+    }
+
+    /* Use a single byte MODRM format if possible.  Note that the encoding
+       that would be used for %esp is the escape to the two byte form.  */
+    if (index == -1 && rm != TCG_REG_ESP) {
+        /* Single byte MODRM format.  */
+        tcg_out8(s, mod | (r << 3) | rm);
+    } else {
+        /* Two byte MODRM+SIB format.  */
+
+        /* Note that the encoding that would place %esp into the index
+           field indicates no index register.  */
+        if (index == -1) {
+            index = 4;
         } else {
-            tcg_out8(s, 0x80 | (r << 3) | rm);
+            assert(index != TCG_REG_ESP);
         }
+
+        tcg_out8(s, mod | (r << 3) | 4);
+        tcg_out8(s, (shift << 6) | (index << 3) | rm);
+    }
+
+    if (len == 1) {
+        tcg_out8(s, offset);
+    } else if (len == 4) {
         tcg_out32(s, offset);
     }
 }
 
+/* rm == -1 means no register index */
+static inline void tcg_out_modrm_offset(TCGContext *s, int opc, int r, int rm, 
+                                        int32_t offset)
+{
+    tcg_out_modrm_sib_offset(s, opc, r, rm, -1, 0, offset);
+}
+
 /* Generate dest op= src.  Uses the same ARITH_* codes as tgen_arithi.  */
 static inline void tgen_arithr(TCGContext *s, int subop, int dest, int src)
 {

@@ -733,10 +764,9 @@ static void tcg_out_qemu_ld(TCGContext *s, const TCGArg *args,
     tgen_arithi(s, ARITH_AND, r0, TARGET_PAGE_MASK | ((1 << s_bits) - 1), 0);
     tgen_arithi(s, ARITH_AND, r1, (CPU_TLB_SIZE - 1) << CPU_TLB_ENTRY_BITS, 0);
 
-    tcg_out_opc(s, 0x8d); /* lea offset(r1, %ebp), r1 */
-    tcg_out8(s, 0x80 | (r1 << 3) | 0x04);
-    tcg_out8(s, (5 << 3) | r1);
-    tcg_out32(s, offsetof(CPUState, tlb_table[mem_index][0].addr_read));
+    tcg_out_modrm_sib_offset(s, OPC_LEA, r1, TCG_AREG0, r1, 0,
+                             offsetof(CPUState,
+                                      tlb_table[mem_index][0].addr_read));
 
     /* cmp 0(r1), r0 */
     tcg_out_modrm_offset(s, OPC_CMP_GvEv, r0, r1, 0);

@@ -924,10 +954,9 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
     tgen_arithi(s, ARITH_AND, r0, TARGET_PAGE_MASK | ((1 << s_bits) - 1), 0);
     tgen_arithi(s, ARITH_AND, r1, (CPU_TLB_SIZE - 1) << CPU_TLB_ENTRY_BITS, 0);
 
-    tcg_out_opc(s, 0x8d); /* lea offset(r1, %ebp), r1 */
-    tcg_out8(s, 0x80 | (r1 << 3) | 0x04);
-    tcg_out8(s, (5 << 3) | r1);
-    tcg_out32(s, offsetof(CPUState, tlb_table[mem_index][0].addr_write));
+    tcg_out_modrm_sib_offset(s, OPC_LEA, r1, TCG_AREG0, r1, 0,
+                             offsetof(CPUState,
+                                      tlb_table[mem_index][0].addr_write));
 
     /* cmp 0(r1), r0 */
     tcg_out_modrm_offset(s, OPC_CMP_GvEv, r0, r1, 0);

Signature · 1 line

1.6.6.1

Signature · 2 lines

Aurelien Jarno                          GPG: 1024D/F1BCDB73
[email protected]                 http://www.aurel32.net

Richard Henderson

author

2010-04-28 18:30 UTC #41 lore raw

PATCH 21/22 tcg-i386: Use lea for three-operand add.

In reply to Richard Henderson

The result is shorter than the mov+add that TCG would otherwise generate for us.

Signed-off-by Richard Henderson <[email protected]>

---
 tcg/i386/tcg-target.c |   23 ++++++++++++++++++++---
 1 files changed, 20 insertions(+), 3 deletions(-)

tcg/i386/tcg-target.c +20 -3

index 755d46d..646a7b6 100644
@@ -1186,6 +1186,25 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
     case INDEX_op_st_i32:
         tcg_out_st(s, TCG_TYPE_I32, args[0], args[1], args[2]);
         break;
+    case INDEX_op_add_i32:
+        /* For 3-operand addition, use LEA.  */
+        if (args[0] != args[1]) {
+            TCGArg a0 = args[0], a1 = args[1], a2 = args[2], c3 = 0;
+
+            if (const_args[2]) {
+                c3 = a2, a2 = -1;
+            } else if (a0 == a2) {
+                /* Watch out for dest = src + dest, since we've removed
+                   the matching constraint on the add.  */
+                tgen_arithr(s, ARITH_ADD, a0, a1);
+                break;
+            }
+
+            tcg_out_modrm_sib_offset(s, OPC_LEA, a0, a1, a2, 0, c3);
+            break;
+        }
+        c = ARITH_ADD;
+        goto gen_arith;
     case INDEX_op_sub_i32:
         c = ARITH_SUB;
         goto gen_arith;

@@ -1198,8 +1217,6 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
     case INDEX_op_xor_i32:
         c = ARITH_XOR;
         goto gen_arith;
-    case INDEX_op_add_i32:
-        c = ARITH_ADD;
     gen_arith:
         if (const_args[2]) {
             tgen_arithi(s, c, args[0], args[2], 0);

@@ -1374,7 +1391,7 @@ static const TCGTargetOpDef x86_op_defs[] = {
     { INDEX_op_st16_i32, { "r", "r" } },
     { INDEX_op_st_i32, { "r", "r" } },
 
-    { INDEX_op_add_i32, { "r", "0", "ri" } },
+    { INDEX_op_add_i32, { "r", "r", "ri" } },
     { INDEX_op_sub_i32, { "r", "0", "ri" } },
     { INDEX_op_mul_i32, { "r", "0", "ri" } },
     { INDEX_op_mulu2_i32, { "a", "d", "a", "r" } },

Signature · 1 line

1.6.6.1

Aurelien Jarno

Acked-by on v1 21/22

2010-05-21 09:47 UTC #42 lore raw

Re: [Qemu-devel] [PATCH 21/22] tcg-i386: Use lea for three-operand add.

Quoted 4 lines · Richard Henderson

The result is shorter than the mov+add that TCG would otherwise generate for us.
Signed-off-by Richard Henderson <[email protected]>

Acked-by Aurelien Jarno <[email protected]>

Quoted 58 lines

---
 tcg/i386/tcg-target.c |   23 ++++++++++++++++++++---
 1 files changed, 20 insertions(+), 3 deletions(-)

tcg/i386/tcg-target.c +20 -3

index 755d46d..646a7b6 100644
@@ -1186,6 +1186,25 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
     case INDEX_op_st_i32:
         tcg_out_st(s, TCG_TYPE_I32, args[0], args[1], args[2]);
         break;
+    case INDEX_op_add_i32:
+        /* For 3-operand addition, use LEA.  */
+        if (args[0] != args[1]) {
+            TCGArg a0 = args[0], a1 = args[1], a2 = args[2], c3 = 0;
+
+            if (const_args[2]) {
+                c3 = a2, a2 = -1;
+            } else if (a0 == a2) {
+                /* Watch out for dest = src + dest, since we've removed
+                   the matching constraint on the add.  */
+                tgen_arithr(s, ARITH_ADD, a0, a1);
+                break;
+            }
+
+            tcg_out_modrm_sib_offset(s, OPC_LEA, a0, a1, a2, 0, c3);
+            break;
+        }
+        c = ARITH_ADD;
+        goto gen_arith;
     case INDEX_op_sub_i32:
         c = ARITH_SUB;
         goto gen_arith;

@@ -1198,8 +1217,6 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
     case INDEX_op_xor_i32:
         c = ARITH_XOR;
         goto gen_arith;
-    case INDEX_op_add_i32:
-        c = ARITH_ADD;
     gen_arith:
         if (const_args[2]) {
             tgen_arithi(s, c, args[0], args[2], 0);

@@ -1374,7 +1391,7 @@ static const TCGTargetOpDef x86_op_defs[] = {
     { INDEX_op_st16_i32, { "r", "r" } },
     { INDEX_op_st_i32, { "r", "r" } },
 
-    { INDEX_op_add_i32, { "r", "0", "ri" } },
+    { INDEX_op_add_i32, { "r", "r", "ri" } },
     { INDEX_op_sub_i32, { "r", "0", "ri" } },
     { INDEX_op_mul_i32, { "r", "0", "ri" } },
     { INDEX_op_mulu2_i32, { "a", "d", "a", "r" } },

Signature · 1 line

1.6.6.1

Signature · 2 lines

Aurelien Jarno                          GPG: 1024D/F1BCDB73
[email protected]                 http://www.aurel32.net

Richard Henderson

author

2010-04-28 18:30 UTC #43 lore raw

PATCH 22/22 tcg-i386: Tidy data16 prefixes.

In reply to Richard Henderson

Include it in the opcode as an extension, as with P_EXT or the REX bits in the x86-64 port.

Signed-off-by Richard Henderson <[email protected]>

---
 tcg/i386/tcg-target.c |   21 +++++++++++++--------
 1 files changed, 13 insertions(+), 8 deletions(-)

tcg/i386/tcg-target.c +13 -8

index 646a7b6..2f096fa 100644
@@ -161,7 +161,8 @@ static inline int tcg_target_const_match(tcg_target_long val,
         return 0;
 }
 
-#define P_EXT   0x100 /* 0x0f opcode prefix */
+#define P_EXT		0x100		/* 0x0f opcode prefix */
+#define P_DATA16	0x200		/* 0x66 opcode prefix */
 
 #define OPC_ARITH_EvIz	(0x81)
 #define OPC_ARITH_EvIb	(0x83)

@@ -267,8 +268,12 @@ static const uint8_t tcg_cond_to_jcc[10] = {
 
 static inline void tcg_out_opc(TCGContext *s, int opc)
 {
-    if (opc & P_EXT)
+    if (opc & P_DATA16) {
+        tcg_out8(s, 0x66);
+    }
+    if (opc & P_EXT) {
         tcg_out8(s, 0x0f);
+    }
     tcg_out8(s, opc);
 }

@@ -460,8 +465,8 @@ static inline void tcg_out_bswap32(TCGContext *s, int reg)
 
 static inline void tcg_out_rolw_8(TCGContext *s, int reg)
 {
-    tcg_out8(s, 0x66); 
-    tcg_out_shifti(s, SHIFT_ROL, reg, 8);
+    tcg_out_modrm(s, OPC_SHIFT_Ib | P_DATA16, SHIFT_ROL, reg);
+    tcg_out8(s, 8); 
 }
 
 static inline void tgen_arithi(TCGContext *s, int c, int r0,

@@ -1074,8 +1079,8 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
             data_reg = r1;
         }
         /* movw */
-        tcg_out8(s, 0x66);
-        tcg_out_modrm_offset(s, OPC_MOVL_EvGv, data_reg, r0, GUEST_BASE);
+        tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_DATA16,
+                             data_reg, r0, GUEST_BASE);
         break;
     case 2:
         if (bswap) {

@@ -1180,8 +1185,8 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
         break;
     case INDEX_op_st16_i32:
         /* movw */
-        tcg_out8(s, 0x66);
-        tcg_out_modrm_offset(s, OPC_MOVL_EvGv, args[0], args[1], args[2]);
+        tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_DATA16,
+                             args[0], args[1], args[2]);
         break;
     case INDEX_op_st_i32:
         tcg_out_st(s, TCG_TYPE_I32, args[0], args[1], args[2]);

Signature · 1 line

1.6.6.1

Aurelien Jarno

Acked-by on v1 22/22

2010-05-21 09:47 UTC #44 lore raw

Re: [Qemu-devel] [PATCH 22/22] tcg-i386: Tidy data16 prefixes.

Quoted 4 lines · Richard Henderson

Include it in the opcode as an extension, as with P_EXT or the REX bits in the x86-64 port.
Signed-off-by Richard Henderson <[email protected]>

Acked-by Aurelien Jarno <[email protected]>

Quoted 71 lines

---
 tcg/i386/tcg-target.c |   21 +++++++++++++--------
 1 files changed, 13 insertions(+), 8 deletions(-)

tcg/i386/tcg-target.c +13 -8

index 646a7b6..2f096fa 100644
@@ -161,7 +161,8 @@ static inline int tcg_target_const_match(tcg_target_long val,
         return 0;
 }
 
-#define P_EXT   0x100 /* 0x0f opcode prefix */
+#define P_EXT		0x100		/* 0x0f opcode prefix */
+#define P_DATA16	0x200		/* 0x66 opcode prefix */
 
 #define OPC_ARITH_EvIz	(0x81)
 #define OPC_ARITH_EvIb	(0x83)

@@ -267,8 +268,12 @@ static const uint8_t tcg_cond_to_jcc[10] = {
 
 static inline void tcg_out_opc(TCGContext *s, int opc)
 {
-    if (opc & P_EXT)
+    if (opc & P_DATA16) {
+        tcg_out8(s, 0x66);
+    }
+    if (opc & P_EXT) {
         tcg_out8(s, 0x0f);
+    }
     tcg_out8(s, opc);
 }

@@ -460,8 +465,8 @@ static inline void tcg_out_bswap32(TCGContext *s, int reg)
 
 static inline void tcg_out_rolw_8(TCGContext *s, int reg)
 {
-    tcg_out8(s, 0x66); 
-    tcg_out_shifti(s, SHIFT_ROL, reg, 8);
+    tcg_out_modrm(s, OPC_SHIFT_Ib | P_DATA16, SHIFT_ROL, reg);
+    tcg_out8(s, 8); 
 }
 
 static inline void tgen_arithi(TCGContext *s, int c, int r0,

@@ -1074,8 +1079,8 @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args,
             data_reg = r1;
         }
         /* movw */
-        tcg_out8(s, 0x66);
-        tcg_out_modrm_offset(s, OPC_MOVL_EvGv, data_reg, r0, GUEST_BASE);
+        tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_DATA16,
+                             data_reg, r0, GUEST_BASE);
         break;
     case 2:
         if (bswap) {

@@ -1180,8 +1185,8 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc,
         break;
     case INDEX_op_st16_i32:
         /* movw */
-        tcg_out8(s, 0x66);
-        tcg_out_modrm_offset(s, OPC_MOVL_EvGv, args[0], args[1], args[2]);
+        tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_DATA16,
+                             args[0], args[1], args[2]);
         break;
     case INDEX_op_st_i32:
         tcg_out_st(s, TCG_TYPE_I32, args[0], args[1], args[2]);

Signature · 1 line

1.6.6.1

Signature · 2 lines

Aurelien Jarno                          GPG: 1024D/F1BCDB73
[email protected]                 http://www.aurel32.net

Richard Henderson

author

2010-05-17 18:26 UTC #45 lore raw

In reply to Richard Henderson

Ping?

Quoted 37 lines · Richard Henderson

Changes v1->v2:
* Dropped controversial bswap changes; bswap16 continues to use rolw.
* Tidy data16 as the last of the hard-coded constants.
r~
Richard Henderson (22):
tcg-i386: Allocate call-saved registers first.
tcg-i386: Tidy initialization of tcg_target_call_clobber_regs.
tcg-i386: Tidy ext8u and ext16u operations.
tcg-i386: Tidy ext8s and ext16s operations.
tcg-i386: Tidy bswap operations.
tcg-i386: Tidy shift operations.
tcg-i386: Tidy move operations.
tcg-i386: Eliminate extra move from qemu_ld64.
tcg-i386: Tidy jumps.
tcg-i386: Tidy immediate arithmetic operations.
tcg-i386: Tidy non-immediate arithmetic operations.
tcg-i386: Tidy movi.
tcg-i386: Tidy push/pop.
tcg-i386: Tidy calls.
tcg-i386: Tidy ret.
tcg-i386: Tidy setcc.
tcg-i386: Tidy unary arithmetic.
tcg-i386: Tidy multiply.
tcg-i386: Tidy xchg.
tcg-i386: Tidy lea.
tcg-i386: Use lea for three-operand add.
tcg-i386: Tidy data16 prefixes.
 tcg/i386/tcg-target.c |  718 +++++++++++++++++++++++++++++--------------------
 1 files changed, 433 insertions(+), 285 deletions(-)

Aurelien Jarno

2010-05-17 19:54 UTC #46 lore raw

Pong.

Sorry -ENOTIME for QEMU

Quoted 42 lines · Richard Henderson

Ping?
Quoted 37 lines · Richard Henderson
Changes v1->v2:
* Dropped controversial bswap changes; bswap16 continues to use rolw.
* Tidy data16 as the last of the hard-coded constants.
r~
Richard Henderson (22):
tcg-i386: Allocate call-saved registers first.
tcg-i386: Tidy initialization of tcg_target_call_clobber_regs.
tcg-i386: Tidy ext8u and ext16u operations.
tcg-i386: Tidy ext8s and ext16s operations.
tcg-i386: Tidy bswap operations.
tcg-i386: Tidy shift operations.
tcg-i386: Tidy move operations.
tcg-i386: Eliminate extra move from qemu_ld64.
tcg-i386: Tidy jumps.
tcg-i386: Tidy immediate arithmetic operations.
tcg-i386: Tidy non-immediate arithmetic operations.
tcg-i386: Tidy movi.
tcg-i386: Tidy push/pop.
tcg-i386: Tidy calls.
tcg-i386: Tidy ret.
tcg-i386: Tidy setcc.
tcg-i386: Tidy unary arithmetic.
tcg-i386: Tidy multiply.
tcg-i386: Tidy xchg.
tcg-i386: Tidy lea.
tcg-i386: Use lea for three-operand add.
tcg-i386: Tidy data16 prefixes.
 tcg/i386/tcg-target.c |  718 +++++++++++++++++++++++++++++--------------------
 1 files changed, 433 insertions(+), 285 deletions(-)

Signature · 2 lines

Aurelien Jarno	                        GPG: 1024D/F1BCDB73
[email protected]                 http://www.aurel32.net

Aurelien Jarno

2010-05-21 09:48 UTC #47 lore raw

In reply to Richard Henderson

Quoted 4 lines · Richard Henderson

Changes v1->v2:
* Dropped controversial bswap changes; bswap16 continues to use rolw.
* Tidy data16 as the last of the hard-coded constants.

I have applied 1-2 and 5-9, but I can't apply the other without patches 3 and 4. The others are fine except a minor issue on patch 13. Could you please resend a series with the changes?

Quoted 31 lines

Richard Henderson (22):
tcg-i386: Allocate call-saved registers first.
tcg-i386: Tidy initialization of tcg_target_call_clobber_regs.
tcg-i386: Tidy ext8u and ext16u operations.
tcg-i386: Tidy ext8s and ext16s operations.
tcg-i386: Tidy bswap operations.
tcg-i386: Tidy shift operations.
tcg-i386: Tidy move operations.
tcg-i386: Eliminate extra move from qemu_ld64.
tcg-i386: Tidy jumps.
tcg-i386: Tidy immediate arithmetic operations.
tcg-i386: Tidy non-immediate arithmetic operations.
tcg-i386: Tidy movi.
tcg-i386: Tidy push/pop.
tcg-i386: Tidy calls.
tcg-i386: Tidy ret.
tcg-i386: Tidy setcc.
tcg-i386: Tidy unary arithmetic.
tcg-i386: Tidy multiply.
tcg-i386: Tidy xchg.
tcg-i386: Tidy lea.
tcg-i386: Use lea for three-operand add.
tcg-i386: Tidy data16 prefixes.
 tcg/i386/tcg-target.c |  718 +++++++++++++++++++++++++++++--------------------
 1 files changed, 433 insertions(+), 285 deletions(-)

Signature · 2 lines

Aurelien Jarno                          GPG: 1024D/F1BCDB73
[email protected]                 http://www.aurel32.net