Skip to contentSkip to conversation

RFC: TCG constant propagation.

PATCH In review Filip Navara 6 participants 10 messages Jul 21, 2009 – Aug 5, 2009

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 RFC: TCG constant propagation.

Add support for constant propagation to TCG. This has to be paired with the liveness analysis to remove the dead code. Not all possible operations are covered, but the most common ones are. This improves the code generation for several ARM instructions, like MVN (immediate), and it may help other targets as well.

---
 tcg/tcg.c |  164 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 1 files changed, 164 insertions(+), 0 deletions(-)
tcg/tcg.c +164 -0
index 4cb5934..10b6a99 100644
@@ -1025,7 +1025,169 @@ void tcg_add_target_add_op_defs(const TCGTargetOpDef *tdefs)
 #endif
         tdefs++;
     }
+}
+
+static void tcg_const_analysis(TCGContext *s)
+{
+    int nb_cargs, nb_iargs, nb_oargs, dest, src, src2, del_args, i;
+    TCGArg *args;
+    uint16_t op;
+    uint16_t *opc_ptr;
+    const TCGOpDef *def;
+    uint8_t *const_temps;
+    tcg_target_ulong *temp_values;
+    tcg_target_ulong val, mask;
+    tcg_target_ulong dest_val, src_val, src2_val;
+
+    const_temps = tcg_malloc(s->nb_temps);
+    memset(const_temps, 0, s->nb_temps);
+    temp_values = tcg_malloc(s->nb_temps * sizeof(uint32_t));
+
+    opc_ptr = gen_opc_buf;
+    args = gen_opparam_buf;
+    while (opc_ptr < gen_opc_ptr) {
+        op = *opc_ptr;
+        def = &tcg_op_defs[op];
+        nb_oargs = def->nb_oargs;
+        nb_iargs = def->nb_iargs;
+        nb_cargs = def->nb_cargs;
+        del_args = 0;
+        mask = ~((tcg_target_ulong)0);
 
+        switch(op) {
+        case INDEX_op_movi_i32:
+#if TCG_TARGET_REG_BITS == 64
+        case INDEX_op_movi_i64:
+#endif
+            dest = args[0];
+            val = args[1];
+            const_temps[dest] = 1;
+            temp_values[dest] = val;
+            break;
+        case INDEX_op_mov_i32:
+#if TCG_TARGET_REG_BITS == 64
+        case INDEX_op_mov_i64:
+#endif
+            dest = args[0];
+            src = args[1];
+            const_temps[dest] = const_temps[src];
+            temp_values[dest] = temp_values[src];
+            break;
+        case INDEX_op_not_i32:
+#if TCG_TARGET_REG_BITS == 64
+            mask = 0xffffffff;
+        case INDEX_op_not_i64:
+#endif
+            dest = args[0];
+            src = args[1];
+            if (const_temps[src]) {
+                const_temps[dest] = 1;
+                dest_val = ~temp_values[src];
+                *opc_ptr = INDEX_op_movi_i32;
+                args[1] = temp_values[dest] = dest_val & mask;
+            } else {
+                const_temps[dest] = 0;
+            }
+            break;
+        case INDEX_op_add_i32:
+        case INDEX_op_sub_i32:
+        case INDEX_op_mul_i32:
+        case INDEX_op_and_i32:
+        case INDEX_op_or_i32:
+        case INDEX_op_xor_i32:
+        case INDEX_op_shl_i32:
+        case INDEX_op_shr_i32:
+#if TCG_TARGET_REG_BITS == 64
+            mask = 0xffffffff;
+        case INDEX_op_add_i64:
+        case INDEX_op_sub_i64:
+        case INDEX_op_mul_i64:
+        case INDEX_op_and_i64:
+        case INDEX_op_or_i64:
+        case INDEX_op_xor_i64:
+        case INDEX_op_shl_i64:
+        case INDEX_op_shr_i64:
+#endif
+
+            dest = args[0];
+            src = args[1];
+            src2 = args[2];
+            if (const_temps[src] && const_temps[src2]) {
+                src_val = temp_values[src];
+                src2_val = temp_values[src2];
+                const_temps[dest] = 1;
+                switch (op) {
+                case INDEX_op_add_i32:
+                    dest_val = src_val + src2_val;
+                    break;
+                case INDEX_op_sub_i32:
+                    dest_val = src_val - src2_val;
+                    break;
+                case INDEX_op_mul_i32:
+                    dest_val = src_val * src2_val;
+                    break;
+                case INDEX_op_and_i32:
+                    dest_val = src_val & src2_val;
+                    break;
+                case INDEX_op_or_i32:
+                    dest_val = src_val | src2_val;
+                    break;
+                case INDEX_op_xor_i32:
+                    dest_val = src_val ^ src2_val;
+                    break;
+                case INDEX_op_shl_i32:
+                    dest_val = src_val << src2_val;
+                    break;
+                case INDEX_op_shr_i32:
+                    dest_val = src_val >> src2_val;
+                    break;
+                default:
+                    tcg_abort();
+                    return;
+                }
+                *opc_ptr = INDEX_op_movi_i32;                
+                args[1] = temp_values[dest] = dest_val & mask;
+                del_args = 1;
+            } else {
+                const_temps[dest] = 0;
+            }
+            break;
+        case INDEX_op_call:
+            nb_oargs = args[0] >> 16;
+            nb_iargs = args[0] & 0xffff;
+            nb_cargs = def->nb_cargs;
+            args++;
+            for (i = 0; i < nb_oargs; i++) {
+                const_temps[args[i]] = 0;
+            }
+            break;
+        case INDEX_op_nopn:
+            /* variable number of arguments */
+            nb_cargs = args[0];
+            break;
+        case INDEX_op_set_label:
+            memset(const_temps, 0, s->nb_temps);
+            break;
+        default:
+            if (def->flags & TCG_OPF_BB_END) {
+                memset(const_temps, 0, s->nb_temps);
+            } else {
+                for (i = 0; i < nb_oargs; i++) {
+                    const_temps[args[i]] = 0;
+                }
+            }
+            break;
+        }
+        opc_ptr++;
+        args += nb_iargs + nb_oargs + nb_cargs - del_args;
+        if (del_args > 0) {
+            gen_opparam_ptr -= del_args;
+            memmove(args, args + del_args, (gen_opparam_ptr - args) * sizeof(*args));
+        }
+    }
+
+    if (args != gen_opparam_ptr)
+        tcg_abort();
 }
 
 #ifdef USE_LIVENESS_ANALYSIS
@@ -1895,6 +2057,8 @@ static inline int tcg_gen_code_common(TCGContext *s, uint8_t *gen_code_buf,
     }
 #endif
 
+    tcg_const_analysis(s);
+
 #ifdef CONFIG_PROFILER
     s->la_time -= profile_getclock();
 #endif
Signature · 1 line
1.6.3.2.1299.gee46c
#1 lore raw
Quoted 7 lines · Filip Navara

Add support for constant propagation to TCG. This has to be paired with the liveness analysis to remove the dead code. Not all possible operations are covered, but the most common ones are. This improves the code generation for several ARM instructions, like MVN (immediate), and it may help other targets as well.
---
 tcg/tcg.c |  164 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++

 1 files changed, 164 insertions(+), 0 deletions(-)

Attached is updated version of the patch which adds profiling code for the TCG pass and hopefully fixes it for TCG 64-bit targets.

Also it looks that the actual code size is worse for TCG ARM target with this optimization. To mitigate it I propose the other attached patch which uses the barrel shifter for more optimal generation of the "movi" operation. I can't test it, so any help is welcome.

Best regards,
Filip Navara

From 403c7d19df34fb25a8b57d90000697a83488cd48 Mon Sep 17 00:00:00 2001
From: Filip Navara <[email protected]>
Date: Tue, 21 Jul 2009 16:31:40 +0200
Subject: [PATCH 1/2] TCG constant propagation.

Add support for constant propagation to TCG. This has to be paired with the liveness analysis to remove the dead code. Not all possible operations are covered, but the most common ones are. This improves the code generation for several ARM instructions, like MVN (immediate), and it may help other targets as well.

v1 -> v2:
Added profiling code and hopefully fixed for 64-bit TCG targets.

---
 tcg/tcg.c |  202 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 tcg/tcg.h |    1 +
 2 files changed, 203 insertions(+), 0 deletions(-)
tcg/tcg.c +202 -0
index 4cb5934..e23e303 100644
@@ -1025,7 +1025,199 @@ void tcg_add_target_add_op_defs(const TCGTargetOpDef *tdefs)
 #endif
         tdefs++;
     }
+}
+
+static void tcg_const_analysis(TCGContext *s)
+{
+    int nb_cargs, nb_iargs, nb_oargs, dest, src, src2, del_args, i;
+    TCGArg *args;
+    uint16_t op;
+    uint16_t *opc_ptr;
+    const TCGOpDef *def;
+    uint8_t *const_temps;
+    tcg_target_ulong *temp_values;
+    tcg_target_ulong val, mask;
+    tcg_target_ulong dest_val, src_val, src2_val;
+
+    const_temps = tcg_malloc(s->nb_temps);
+    memset(const_temps, 0, s->nb_temps);
+    temp_values = tcg_malloc(s->nb_temps * sizeof(uint32_t));
+
+    opc_ptr = gen_opc_buf;
+    args = gen_opparam_buf;
+    while (opc_ptr < gen_opc_ptr) {
+        op = *opc_ptr;
+        def = &tcg_op_defs[op];
+        nb_oargs = def->nb_oargs;
+        nb_iargs = def->nb_iargs;
+        nb_cargs = def->nb_cargs;
+        del_args = 0;
+        mask = ~((tcg_target_ulong)0);
+
+        switch(op) {
+        case INDEX_op_movi_i32:
+#if TCG_TARGET_REG_BITS == 64
+        case INDEX_op_movi_i64:
+#endif
+            dest = args[0];
+            val = args[1];
+            const_temps[dest] = 1;
+            temp_values[dest] = val;
+            break;
+        case INDEX_op_mov_i32:
+#if TCG_TARGET_REG_BITS == 64
+        case INDEX_op_mov_i64:
+#endif
+            dest = args[0];
+            src = args[1];
+            const_temps[dest] = const_temps[src];
+            temp_values[dest] = temp_values[src];
+            break;
+#if defined(TCG_TARGET_HAS_not_i32) || defined(TCG_TARGET_HAS_not_i64)
+#ifdef TCG_TARGET_HAS_not_i32
+        case INDEX_op_not_i32:
+#endif
+#if TCG_TARGET_REG_BITS == 64
+            mask = 0xffffffff;
+#ifdef TCG_TARGET_HAS_not_i64
+        case INDEX_op_not_i64:
+#endif
+#endif
+            dest = args[0];
+            src = args[1];
+            if (const_temps[src]) {
+                const_temps[dest] = 1;
+                dest_val = ~temp_values[src];
+                *opc_ptr = INDEX_op_movi_i32;
+                args[1] = temp_values[dest] = dest_val & mask;
+            } else {
+                const_temps[dest] = 0;
+            }
+            break;
+#endif
+        case INDEX_op_add_i32:
+        case INDEX_op_sub_i32:
+        case INDEX_op_mul_i32:
+        case INDEX_op_and_i32:
+        case INDEX_op_or_i32:
+        case INDEX_op_xor_i32:
+        case INDEX_op_shl_i32:
+        case INDEX_op_shr_i32:
+#if TCG_TARGET_REG_BITS == 64
+            mask = 0xffffffff;
+        case INDEX_op_add_i64:
+        case INDEX_op_sub_i64:
+        case INDEX_op_mul_i64:
+        case INDEX_op_and_i64:
+        case INDEX_op_or_i64:
+        case INDEX_op_xor_i64:
+        case INDEX_op_shl_i64:
+        case INDEX_op_shr_i64:
+#endif
+
+            dest = args[0];
+            src = args[1];
+            src2 = args[2];
+            if (const_temps[src] && const_temps[src2]) {
+                src_val = temp_values[src];
+                src2_val = temp_values[src2];
+                const_temps[dest] = 1;
+                switch (op) {
+                case INDEX_op_add_i32:
+#if TCG_TARGET_REG_BITS == 64
+                case INDEX_op_add_i64:
+#endif
+                    dest_val = src_val + src2_val;
+                    break;
+                case INDEX_op_sub_i32:
+#if TCG_TARGET_REG_BITS == 64
+                case INDEX_op_sub_i64:
+#endif
+                    dest_val = src_val - src2_val;
+                    break;
+                case INDEX_op_mul_i32:
+#if TCG_TARGET_REG_BITS == 64
+                case INDEX_op_mul_i64:
+#endif
+                    dest_val = src_val * src2_val;
+                    break;
+                case INDEX_op_and_i32:
+#if TCG_TARGET_REG_BITS == 64
+                case INDEX_op_and_i64:
+#endif
+                    dest_val = src_val & src2_val;
+                    break;
+                case INDEX_op_or_i32:
+#if TCG_TARGET_REG_BITS == 64
+                case INDEX_op_or_i64:
+#endif
+                    dest_val = src_val | src2_val;
+                    break;
+                case INDEX_op_xor_i32:
+#if TCG_TARGET_REG_BITS == 64
+                case INDEX_op_xor_i64:
+#endif
+                    dest_val = src_val ^ src2_val;
+                    break;
+                case INDEX_op_shl_i32:
+#if TCG_TARGET_REG_BITS == 64
+                case INDEX_op_shl_i64:
+#endif
+                    dest_val = src_val << src2_val;
+                    break;
+                case INDEX_op_shr_i32:
+#if TCG_TARGET_REG_BITS == 64
+                case INDEX_op_shr_i64:
+#endif
+                    dest_val = src_val >> src2_val;
+                    break;
+                default:
+                    tcg_abort();
+                    return;
+                }
+                *opc_ptr = INDEX_op_movi_i32;                
+                args[1] = temp_values[dest] = dest_val & mask;
+                del_args = 1;
+            } else {
+                const_temps[dest] = 0;
+            }
+            break;
+        case INDEX_op_call:
+            nb_oargs = args[0] >> 16;
+            nb_iargs = args[0] & 0xffff;
+            nb_cargs = def->nb_cargs;
+            args++;
+            for (i = 0; i < nb_oargs; i++) {
+                const_temps[args[i]] = 0;
+            }
+            break;
+        case INDEX_op_nopn:
+            /* variable number of arguments */
+            nb_cargs = args[0];
+            break;
+        case INDEX_op_set_label:
+            memset(const_temps, 0, s->nb_temps);
+            break;
+        default:
+            if (def->flags & TCG_OPF_BB_END) {
+                memset(const_temps, 0, s->nb_temps);
+            } else {
+                for (i = 0; i < nb_oargs; i++) {
+                    const_temps[args[i]] = 0;
+                }
+            }
+            break;
+        }
+        opc_ptr++;
+        args += nb_iargs + nb_oargs + nb_cargs - del_args;
+        if (del_args > 0) {
+            gen_opparam_ptr -= del_args;
+            memmove(args, args + del_args, (gen_opparam_ptr - args) * sizeof(*args));
+        }
+    }
 
+    if (args != gen_opparam_ptr)
+        tcg_abort();
 }
 
 #ifdef USE_LIVENESS_ANALYSIS
@@ -1896,6 +2088,14 @@ static inline int tcg_gen_code_common(TCGContext *s, uint8_t *gen_code_buf,
 #endif
 
 #ifdef CONFIG_PROFILER
+    s->const_time -= profile_getclock();
+#endif
+    tcg_const_analysis(s);
+#ifdef CONFIG_PROFILER
+    s->const_time += profile_getclock();
+#endif
+
+#ifdef CONFIG_PROFILER
     s->la_time -= profile_getclock();
 #endif
     tcg_liveness_analysis(s);
@@ -2068,6 +2268,8 @@ void tcg_dump_info(FILE *f,
                 (double)s->interm_time / tot * 100.0);
     cpu_fprintf(f, "  gen_code time     %0.1f%%\n", 
                 (double)s->code_time / tot * 100.0);
+    cpu_fprintf(f, "const/code time     %0.1f%%\n", 
+                (double)s->const_time / (s->code_time ? s->code_time : 1) * 100.0);
     cpu_fprintf(f, "liveness/code time  %0.1f%%\n", 
                 (double)s->la_time / (s->code_time ? s->code_time : 1) * 100.0);
     cpu_fprintf(f, "cpu_restore count   %" PRId64 "\n",
tcg/tcg.h +1 -0
index ad0bd14..f8322f2 100644
@@ -288,6 +288,7 @@ struct TCGContext {
     int64_t code_out_len;
     int64_t interm_time;
     int64_t code_time;
+    int64_t const_time;
     int64_t la_time;
     int64_t restore_count;
     int64_t restore_time;
Signature · 1 line
1.6.3.2.1299.gee46c

From 203b7284477109415989d44b71dbd149d05caadf Mon Sep 17 00:00:00 2001
From: Filip Navara <[email protected]>
Date: Thu, 23 Jul 2009 11:01:32 +0200
Subject: [PATCH 2/2] Use ARM barrel shifter for more optimal generation of MVN, MOV instructions for the MOVI TCG operation.

---
 tcg/arm/tcg-target.c |   12 ++++++++++++
 1 files changed, 12 insertions(+), 0 deletions(-)
tcg/arm/tcg-target.c +12 -0
index 7ef2b89..0c59984 100644
@@ -334,6 +334,7 @@ static inline void tcg_out_movi32(TCGContext *s,
                 int cond, int rd, int32_t arg)
 {
     int offset = (uint32_t) arg - ((uint32_t) s->code_ptr + 8);
+    int shift;
 
     /* TODO: This is very suboptimal, we can easily have a constant
      * pool somewhere after all the instructions.  */
@@ -346,6 +347,17 @@ static inline void tcg_out_movi32(TCGContext *s,
                 tcg_out_dat_imm(s, cond, ARITH_ADD, rd, 15, offset) :
                 tcg_out_dat_imm(s, cond, ARITH_SUB, rd, 15, -offset);
 
+    if (arg < 0x100)
+        tcg_out_dat_imm(s, cond, ARITH_MOV, rd, 0, arg & 0xff);
+
+    shift = ctz32(arg) & ~1;
+    if (!((arg >> shift) & ~0xff))
+        return tcg_out_dat_imm(s, cond, ARITH_MOV, rd, 0, (arg >> shift) + (shift << 7));
+
+    shift = ctz32(~arg) & ~1;
+    if (!((~arg >> shift) & ~0xff))
+        return tcg_out_dat_imm(s, cond, ARITH_MVN, rd, 0, (~arg >> shift) + (shift << 7));
+
 #ifdef __ARM_ARCH_7A__
     /* use movw/movt */
     /* movw */
Signature · 1 line
1.6.3.2.1299.gee46c
#2 lore raw
Quoted 4 lines · Filip Navara

Also it looks that the actual code size is worse for TCG ARM target with this optimization. To mitigate it I propose the other attached patch which uses the barrel shifter for more optimal generation of the "movi" operation. I can't test it, so any help is welcome.

This is wrong for two reasons:
- this is a rotation, you only check for shifts
- the rotation amount is multiplied by 2 before being applied to the 8-bit immediate field.

Laurent

#3 lore raw
Quoted 1 line · Laurent Desnogues

On Thu, Jul 23, 2009 at 11:08 AM, Filip Navara<[email protected]>

Quoted 7 lines · wrote:
Quoted 4 lines

Also it looks that the actual code size is worse for TCG ARM target with this optimization. To mitigate it I propose the other attached patch which uses the barrel shifter for more optimal generation of the "movi" operation. I can't test it, so any help is welcome.

- the rotation amount is multiplied by 2 before being applied to the 8-bit immediate field.

This is compensated for by the later << 7

Paul

#4 lore raw
In reply to Laurent Desnogues
Quoted 8 lines · Laurent Desnogues
Quoted 4 lines · Filip Navara

Also it looks that the actual code size is worse for TCG ARM target with this optimization. To mitigate it I propose the other attached patch which uses the barrel shifter for more optimal generation of the "movi" operation. I can't test it, so any help is welcome.

This is wrong for two reasons:
  - this is a rotation, you only check for shifts

I deliberately didn't do it, but it's surely an opportunity for further optimization.

Quoted 2 lines

  - the rotation amount is multiplied by 2 before being applied to the 8-bit immediate field.

As Paul already pointed out, that's compensated by the shift (<< 7 instead of << 8) and the "& ~1" after ctz32 call.

Attached is fixed version of the patch, I forgot one "return" there.

Best regards,
Filip Navara

From d326aabe85412a91c17e0be2ebbfd078bec16713 Mon Sep 17 00:00:00 2001
From: Filip Navara <[email protected]>
Date: Thu, 23 Jul 2009 11:01:32 +0200
Subject: [PATCH 2/2] Use ARM barrel shifter for more optimal generation of MVN, MOV instructions for the MOVI TCG operation.

---
 tcg/arm/tcg-target.c |   12 ++++++++++++
 1 files changed, 12 insertions(+), 0 deletions(-)
tcg/arm/tcg-target.c +12 -0
index 7ef2b89..9748ce2 100644
@@ -334,6 +334,7 @@ static inline void tcg_out_movi32(TCGContext *s,
                 int cond, int rd, int32_t arg)
 {
     int offset = (uint32_t) arg - ((uint32_t) s->code_ptr + 8);
+    int shift;
 
     /* TODO: This is very suboptimal, we can easily have a constant
      * pool somewhere after all the instructions.  */
@@ -346,6 +347,17 @@ static inline void tcg_out_movi32(TCGContext *s,
                 tcg_out_dat_imm(s, cond, ARITH_ADD, rd, 15, offset) :
                 tcg_out_dat_imm(s, cond, ARITH_SUB, rd, 15, -offset);
 
+    if (arg < 0x100)
+        return tcg_out_dat_imm(s, cond, ARITH_MOV, rd, 0, arg & 0xff);
+
+    shift = ctz32(arg) & ~1;
+    if (!((arg >> shift) & ~0xff))
+        return tcg_out_dat_imm(s, cond, ARITH_MOV, rd, 0, (arg >> shift) + (shift << 7));
+
+    shift = ctz32(~arg) & ~1;
+    if (!((~arg >> shift) & ~0xff))
+        return tcg_out_dat_imm(s, cond, ARITH_MVN, rd, 0, (~arg >> shift) + (shift << 7));
+
 #ifdef __ARM_ARCH_7A__
     /* use movw/movt */
     /* movw */
Signature · 1 line
1.6.3.2.1299.gee46c
#5 lore raw
In reply to Filip Navara
+            dest = args[0];
+            src = args[1];
+            if (const_temps[src]) {
+                const_temps[dest] = 1;
+                dest_val = ~temp_values[src];
+                *opc_ptr = INDEX_op_movi_i32;

Hrm... is it really right to be setting *opc_ptr = INDEX_op_movi_i32 even for 64-bit ops? Applies to both 'not'[0] and the binary ops...

Cheers,

Signature · 4 lines
Stuart Brady

[0] BTW, the #ifdefs will need shuffling if other unary ops (e.g. neg)
are wanted... it's trivial, but I can't see a reason not to do it. :-)
#6 lore raw
Quoted 9 lines · Stuart Brady

+            dest = args[0];
+            src = args[1];
+            if (const_temps[src]) {
+                const_temps[dest] = 1;
+                dest_val = ~temp_values[src];
+                *opc_ptr = INDEX_op_movi_i32;

Hrm... is it really right to be setting *opc_ptr = INDEX_op_movi_i32 even for 64-bit ops?  Applies to both 'not'[0] and the binary ops...

You are right, another bug :) I wish I could test it on some 64-bit target, but QEMU is not working on MinGW64 (unsigned long != intptr_t).

Thanks,
Filip Navara

#7 lore raw
In reply to Filip Navara

tcg/arm/tcg-target.c static inline void tcg_out_movi32(TCGContext *s, · line 351

Quoted 2 lines · Filip Navara
tcg/arm/tcg-target.c +2 -0
+    if (arg < 0x100)
+        tcg_out_dat_imm(s, cond, ARITH_MOV, rd, 0, arg & 0xff);

return?

Signature · 2 lines
Daniel Jacobowitz
CodeSourcery
#8 lore raw
In reply to Filip Navara

Filip Navara kirjoitti:

Quoted 4 lines

Add support for constant propagation to TCG. This has to be paired with the liveness analysis to remove the dead code. Not all possible operations are covered, but the most common ones are. This improves the code generation for several ARM instructions, like MVN (immediate), and it may help other targets as well.

On my small benchmark, qemu-system-sh4 was about 3% slower on Intel Xeon [email protected]. I'm running 64-bit mode. My mini benchmark is to build zlib 1.2.3, so it's 'real' world work load. Ran the benchmark several times and results seems to be pretty constant.

ps. I added INDEX_op_*_i64 cases to the evaluation part. I'm not completly sure if those &mask should be there.

Pablo Virolainen

tcg/tcg.c +189 -0
@@ -1021,7 +1021,194 @@
 #endif
         tdefs++;
     }
+}
 
+static void tcg_const_analysis(TCGContext *s)
+{
+    int nb_cargs, nb_iargs, nb_oargs, dest, src, src2, del_args, i;
+    TCGArg *args;
+    uint16_t op;
+    uint16_t *opc_ptr;
+    const TCGOpDef *def;
+    uint8_t *const_temps;
+    tcg_target_ulong *temp_values;
+    tcg_target_ulong val, mask;
+    tcg_target_ulong dest_val, src_val, src2_val;
+
+    const_temps = tcg_malloc(s->nb_temps);
+    memset(const_temps, 0, s->nb_temps);
+    temp_values = tcg_malloc(s->nb_temps * sizeof(uint32_t));
+
+    opc_ptr = gen_opc_buf;
+    args = gen_opparam_buf;
+    while (opc_ptr < gen_opc_ptr) {
+        op = *opc_ptr;
+        def = &tcg_op_defs[op];
+        nb_oargs = def->nb_oargs;
+        nb_iargs = def->nb_iargs;
+        nb_cargs = def->nb_cargs;
+        del_args = 0;
+        mask = ~((tcg_target_ulong)0);
+
+        switch(op) {
+        case INDEX_op_movi_i32:
+#if TCG_TARGET_REG_BITS == 64
+        case INDEX_op_movi_i64:
+#endif
+            dest = args[0];
+            val = args[1];
+            const_temps[dest] = 1;
+            temp_values[dest] = val;
+            break;
+        case INDEX_op_mov_i32:
+#if TCG_TARGET_REG_BITS == 64
+        case INDEX_op_mov_i64:
+#endif
+            dest = args[0];
+            src = args[1];
+            const_temps[dest] = const_temps[src];
+            temp_values[dest] = temp_values[src];
+            break;
+        case INDEX_op_not_i32:
+#if TCG_TARGET_REG_BITS == 64
+            mask = 0xffffffff;
+        case INDEX_op_not_i64:
+#endif
+            dest = args[0];
+            src = args[1];
+            if (const_temps[src]) {
+                const_temps[dest] = 1;
+                dest_val = ~temp_values[src];
+                *opc_ptr = INDEX_op_movi_i32;
+                args[1] = temp_values[dest] = dest_val & mask;
+            } else {
+                const_temps[dest] = 0;
+            }
+            break;
+        case INDEX_op_add_i32:
+        case INDEX_op_sub_i32:
+        case INDEX_op_mul_i32:
+        case INDEX_op_and_i32:
+        case INDEX_op_or_i32:
+        case INDEX_op_xor_i32:
+        case INDEX_op_shl_i32:
+        case INDEX_op_shr_i32:
+#if TCG_TARGET_REG_BITS == 64
+            mask = 0xffffffff;
+        case INDEX_op_add_i64:
+        case INDEX_op_sub_i64:
+        case INDEX_op_mul_i64:
+        case INDEX_op_and_i64:
+        case INDEX_op_or_i64:
+        case INDEX_op_xor_i64:
+        case INDEX_op_shl_i64:
+        case INDEX_op_shr_i64:
+#endif
+
+            dest = args[0];
+            src = args[1];
+            src2 = args[2];
+            if (const_temps[src] && const_temps[src2]) {
+                src_val = temp_values[src];
+                src2_val = temp_values[src2];
+                const_temps[dest] = 1;
+                switch (op) {
+                case INDEX_op_add_i32:
+                    dest_val = src_val + src2_val;
+                    break;
+                case INDEX_op_add_i64:
+		    dest_val = (src_val + src2_val) & mask;
+		    break;
+                case INDEX_op_sub_i32:
+                    dest_val = src_val - src2_val;
+                    break;
+                case INDEX_op_sub_i64:
+		    dest_val = (src_val - src2_val) & mask;
+		    break;
+                case INDEX_op_mul_i32:
+                    dest_val = src_val * src2_val;
+                    break;
+                case INDEX_op_mul_i64:
+		    dest_val = (src_val * src2_val) & mask;
+		    break;
+                case INDEX_op_and_i32:
+                    dest_val = src_val & src2_val;
+                    break;
+                case INDEX_op_and_i64:
+		    dest_val = src_val & src2_val & mask;
+		    break;
+                case INDEX_op_or_i32:
+                    dest_val = src_val | src2_val;
+                    break;
+                case INDEX_op_or_i64:
+		    dest_val = (src_val | src2_val) & mask;
+		    break;
+                case INDEX_op_xor_i32:
+                    dest_val = src_val ^ src2_val;
+                    break;
+                case INDEX_op_xor_i64:
+		    dest_val = (src_val ^ src2_val) & mask;
+		    break;
+                case INDEX_op_shl_i32:
+                    dest_val = src_val << src2_val;
+                    break;
+                case INDEX_op_shl_i64:
+		    dest_val = (src_val << src2_val) & mask;
+		    break;
+                case INDEX_op_shr_i32:
+                    dest_val = src_val >> src2_val;
+                    break;
+                case INDEX_op_shr_i64:
+		    dest_val = (src_val >> src2_val) & mask;
+		    break;
+                default:
+		  fprintf(stderr,"index op %i\n",op);
+                    tcg_abort();
+                    return;
+                }
+                *opc_ptr = INDEX_op_movi_i32;                
+                args[1] = temp_values[dest] = dest_val & mask;
+                del_args = 1;
+            } else {
+                const_temps[dest] = 0;
+            }
+            break;
+        case INDEX_op_call:
+            nb_oargs = args[0] >> 16;
+            nb_iargs = args[0] & 0xffff;
+            nb_cargs = def->nb_cargs;
+            args++;
+            for (i = 0; i < nb_oargs; i++) {
+                const_temps[args[i]] = 0;
+            }
+            break;
+        case INDEX_op_nopn:
+            /* variable number of arguments */
+            nb_cargs = args[0];
+            break;
+        case INDEX_op_set_label:
+            memset(const_temps, 0, s->nb_temps);
+            break;
+        default:
+            if (def->flags & TCG_OPF_BB_END) {
+                memset(const_temps, 0, s->nb_temps);
+            } else {
+                for (i = 0; i < nb_oargs; i++) {
+                    const_temps[args[i]] = 0;
+                }
+            }
+            break;
+        }
+        opc_ptr++;
+        args += nb_iargs + nb_oargs + nb_cargs - del_args;
+        if (del_args > 0) {
+            gen_opparam_ptr -= del_args;
+            memmove(args, args + del_args, (gen_opparam_ptr - args) * sizeof(*args));
+        }
+    }
+
+    if (args != gen_opparam_ptr)
+        tcg_abort();
 }
 
 #ifdef USE_LIVENESS_ANALYSIS
@@ -1891,6 +2078,8 @@
     }
 #endif
 
+    tcg_const_analysis(s);
+
 #ifdef CONFIG_PROFILER
     s->la_time -= profile_getclock();
 #endif
#9 lore raw
Quoted 10 lines · Pablo Virolainen

Filip Navara kirjoitti:

Quoted 8 lines

Add support for constant propagation to TCG. This has to be paired with the liveness
analysis to remove the dead code. Not all possible operations are covered, but the
most common ones are. This improves the code generation for several ARM instructions,
like MVN (immediate), and it may help other targets as well.

Quoted 4 lines

On my small benchmark, qemu-system-sh4 was about 3% slower on Intel Xeon [email protected]. I'm running 64-bit mode. My mini benchmark is to build zlib 1.2.3, so it's 'real' world work load. Ran the benchmark several times and results seems to be pretty constant.

Thanks for testing and reporting the results. I'll see if I can reduce the overhead or if I should ditch the patch.

Quoted 2 lines

ps. I added INDEX_op_*_i64 cases to the evaluation part. I'm not completly sure if those &mask should be there.

I've rewritten the patch and fixed the 64-bit mess that was there. It is available at
http://repo.or.cz/w/qemu/navara.git?a=commit;h=5df3a524fc0b923cf2e5e1883ff550d055d36eb5

Best regards,
Filip Navara