|
 |
5544c1b |
From 8d5f3ca9ccace2374fd73d46fad56decc02e0a44 Mon Sep 17 00:00:00 2001
|
|
|
5544c1b |
From: Aurelien Jarno <aurelien@aurel32.net>
|
|
|
5544c1b |
Date: Tue, 18 Sep 2012 19:37:00 +0200
|
|
|
5544c1b |
Subject: [PATCH] tcg/optimize: further optimize brcond/movcond/setcond
|
|
|
5544c1b |
|
|
|
5544c1b |
When both argument of brcond/movcond/setcond are the same or when one
|
|
|
5544c1b |
of the two values is a constant equal to zero, it's possible to do
|
|
|
5544c1b |
further optimizations.
|
|
|
5544c1b |
|
|
|
5544c1b |
Reviewed-by: Richard Henderson <rth@twiddle.net>
|
|
|
5544c1b |
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
|
|
|
5544c1b |
Signed-off-by: Michael Roth <mdroth@linux.vnet.ibm.com>
|
|
|
5544c1b |
---
|
|
|
5544c1b |
tcg/optimize.c | 127 ++++++++++++++++++++++++++++++++++-----------------------
|
|
|
5544c1b |
1 file changed, 76 insertions(+), 51 deletions(-)
|
|
|
5544c1b |
|
|
|
5544c1b |
diff --git a/tcg/optimize.c b/tcg/optimize.c
|
|
|
5544c1b |
index ceea644..abe016a 100644
|
|
|
5544c1b |
--- a/tcg/optimize.c
|
|
|
5544c1b |
+++ b/tcg/optimize.c
|
|
|
5544c1b |
@@ -292,58 +292,88 @@ static TCGArg do_constant_folding(TCGOpcode op, TCGArg x, TCGArg y)
|
|
|
5544c1b |
return res;
|
|
|
5544c1b |
}
|
|
|
5544c1b |
|
|
|
5544c1b |
+/* Return 2 if the condition can't be simplified, and the result
|
|
|
5544c1b |
+ of the condition (0 or 1) if it can */
|
|
|
5544c1b |
static TCGArg do_constant_folding_cond(TCGOpcode op, TCGArg x,
|
|
|
5544c1b |
TCGArg y, TCGCond c)
|
|
|
5544c1b |
{
|
|
|
5544c1b |
- switch (op_bits(op)) {
|
|
|
5544c1b |
- case 32:
|
|
|
5544c1b |
+ if (temps[x].state == TCG_TEMP_CONST && temps[y].state == TCG_TEMP_CONST) {
|
|
|
5544c1b |
+ switch (op_bits(op)) {
|
|
|
5544c1b |
+ case 32:
|
|
|
5544c1b |
+ switch (c) {
|
|
|
5544c1b |
+ case TCG_COND_EQ:
|
|
|
5544c1b |
+ return (uint32_t)temps[x].val == (uint32_t)temps[y].val;
|
|
|
5544c1b |
+ case TCG_COND_NE:
|
|
|
5544c1b |
+ return (uint32_t)temps[x].val != (uint32_t)temps[y].val;
|
|
|
5544c1b |
+ case TCG_COND_LT:
|
|
|
5544c1b |
+ return (int32_t)temps[x].val < (int32_t)temps[y].val;
|
|
|
5544c1b |
+ case TCG_COND_GE:
|
|
|
5544c1b |
+ return (int32_t)temps[x].val >= (int32_t)temps[y].val;
|
|
|
5544c1b |
+ case TCG_COND_LE:
|
|
|
5544c1b |
+ return (int32_t)temps[x].val <= (int32_t)temps[y].val;
|
|
|
5544c1b |
+ case TCG_COND_GT:
|
|
|
5544c1b |
+ return (int32_t)temps[x].val > (int32_t)temps[y].val;
|
|
|
5544c1b |
+ case TCG_COND_LTU:
|
|
|
5544c1b |
+ return (uint32_t)temps[x].val < (uint32_t)temps[y].val;
|
|
|
5544c1b |
+ case TCG_COND_GEU:
|
|
|
5544c1b |
+ return (uint32_t)temps[x].val >= (uint32_t)temps[y].val;
|
|
|
5544c1b |
+ case TCG_COND_LEU:
|
|
|
5544c1b |
+ return (uint32_t)temps[x].val <= (uint32_t)temps[y].val;
|
|
|
5544c1b |
+ case TCG_COND_GTU:
|
|
|
5544c1b |
+ return (uint32_t)temps[x].val > (uint32_t)temps[y].val;
|
|
|
5544c1b |
+ }
|
|
|
5544c1b |
+ break;
|
|
|
5544c1b |
+ case 64:
|
|
|
5544c1b |
+ switch (c) {
|
|
|
5544c1b |
+ case TCG_COND_EQ:
|
|
|
5544c1b |
+ return (uint64_t)temps[x].val == (uint64_t)temps[y].val;
|
|
|
5544c1b |
+ case TCG_COND_NE:
|
|
|
5544c1b |
+ return (uint64_t)temps[x].val != (uint64_t)temps[y].val;
|
|
|
5544c1b |
+ case TCG_COND_LT:
|
|
|
5544c1b |
+ return (int64_t)temps[x].val < (int64_t)temps[y].val;
|
|
|
5544c1b |
+ case TCG_COND_GE:
|
|
|
5544c1b |
+ return (int64_t)temps[x].val >= (int64_t)temps[y].val;
|
|
|
5544c1b |
+ case TCG_COND_LE:
|
|
|
5544c1b |
+ return (int64_t)temps[x].val <= (int64_t)temps[y].val;
|
|
|
5544c1b |
+ case TCG_COND_GT:
|
|
|
5544c1b |
+ return (int64_t)temps[x].val > (int64_t)temps[y].val;
|
|
|
5544c1b |
+ case TCG_COND_LTU:
|
|
|
5544c1b |
+ return (uint64_t)temps[x].val < (uint64_t)temps[y].val;
|
|
|
5544c1b |
+ case TCG_COND_GEU:
|
|
|
5544c1b |
+ return (uint64_t)temps[x].val >= (uint64_t)temps[y].val;
|
|
|
5544c1b |
+ case TCG_COND_LEU:
|
|
|
5544c1b |
+ return (uint64_t)temps[x].val <= (uint64_t)temps[y].val;
|
|
|
5544c1b |
+ case TCG_COND_GTU:
|
|
|
5544c1b |
+ return (uint64_t)temps[x].val > (uint64_t)temps[y].val;
|
|
|
5544c1b |
+ }
|
|
|
5544c1b |
+ break;
|
|
|
5544c1b |
+ }
|
|
|
5544c1b |
+ } else if (temps_are_copies(x, y)) {
|
|
|
5544c1b |
switch (c) {
|
|
|
5544c1b |
- case TCG_COND_EQ:
|
|
|
5544c1b |
- return (uint32_t)x == (uint32_t)y;
|
|
|
5544c1b |
- case TCG_COND_NE:
|
|
|
5544c1b |
- return (uint32_t)x != (uint32_t)y;
|
|
|
5544c1b |
- case TCG_COND_LT:
|
|
|
5544c1b |
- return (int32_t)x < (int32_t)y;
|
|
|
5544c1b |
- case TCG_COND_GE:
|
|
|
5544c1b |
- return (int32_t)x >= (int32_t)y;
|
|
|
5544c1b |
- case TCG_COND_LE:
|
|
|
5544c1b |
- return (int32_t)x <= (int32_t)y;
|
|
|
5544c1b |
case TCG_COND_GT:
|
|
|
5544c1b |
- return (int32_t)x > (int32_t)y;
|
|
|
5544c1b |
case TCG_COND_LTU:
|
|
|
5544c1b |
- return (uint32_t)x < (uint32_t)y;
|
|
|
5544c1b |
- case TCG_COND_GEU:
|
|
|
5544c1b |
- return (uint32_t)x >= (uint32_t)y;
|
|
|
5544c1b |
- case TCG_COND_LEU:
|
|
|
5544c1b |
- return (uint32_t)x <= (uint32_t)y;
|
|
|
5544c1b |
+ case TCG_COND_LT:
|
|
|
5544c1b |
case TCG_COND_GTU:
|
|
|
5544c1b |
- return (uint32_t)x > (uint32_t)y;
|
|
|
5544c1b |
- }
|
|
|
5544c1b |
- break;
|
|
|
5544c1b |
- case 64:
|
|
|
5544c1b |
- switch (c) {
|
|
|
5544c1b |
- case TCG_COND_EQ:
|
|
|
5544c1b |
- return (uint64_t)x == (uint64_t)y;
|
|
|
5544c1b |
case TCG_COND_NE:
|
|
|
5544c1b |
- return (uint64_t)x != (uint64_t)y;
|
|
|
5544c1b |
- case TCG_COND_LT:
|
|
|
5544c1b |
- return (int64_t)x < (int64_t)y;
|
|
|
5544c1b |
+ return 0;
|
|
|
5544c1b |
case TCG_COND_GE:
|
|
|
5544c1b |
- return (int64_t)x >= (int64_t)y;
|
|
|
5544c1b |
+ case TCG_COND_GEU:
|
|
|
5544c1b |
case TCG_COND_LE:
|
|
|
5544c1b |
- return (int64_t)x <= (int64_t)y;
|
|
|
5544c1b |
- case TCG_COND_GT:
|
|
|
5544c1b |
- return (int64_t)x > (int64_t)y;
|
|
|
5544c1b |
+ case TCG_COND_LEU:
|
|
|
5544c1b |
+ case TCG_COND_EQ:
|
|
|
5544c1b |
+ return 1;
|
|
|
5544c1b |
+ }
|
|
|
5544c1b |
+ } else if (temps[y].state == TCG_TEMP_CONST && temps[y].val == 0) {
|
|
|
5544c1b |
+ switch (c) {
|
|
|
5544c1b |
case TCG_COND_LTU:
|
|
|
5544c1b |
- return (uint64_t)x < (uint64_t)y;
|
|
|
5544c1b |
+ return 0;
|
|
|
5544c1b |
case TCG_COND_GEU:
|
|
|
5544c1b |
- return (uint64_t)x >= (uint64_t)y;
|
|
|
5544c1b |
- case TCG_COND_LEU:
|
|
|
5544c1b |
- return (uint64_t)x <= (uint64_t)y;
|
|
|
5544c1b |
- case TCG_COND_GTU:
|
|
|
5544c1b |
- return (uint64_t)x > (uint64_t)y;
|
|
|
5544c1b |
+ return 1;
|
|
|
5544c1b |
+ default:
|
|
|
5544c1b |
+ return 2;
|
|
|
5544c1b |
}
|
|
|
5544c1b |
- break;
|
|
|
5544c1b |
+ } else {
|
|
|
5544c1b |
+ return 2;
|
|
|
5544c1b |
}
|
|
|
5544c1b |
|
|
|
5544c1b |
fprintf(stderr,
|
|
|
5544c1b |
@@ -636,11 +666,9 @@ static TCGArg *tcg_constant_folding(TCGContext *s, uint16_t *tcg_opc_ptr,
|
|
|
5544c1b |
args += 3;
|
|
|
5544c1b |
break;
|
|
|
5544c1b |
CASE_OP_32_64(setcond):
|
|
|
5544c1b |
- if (temps[args[1]].state == TCG_TEMP_CONST
|
|
|
5544c1b |
- && temps[args[2]].state == TCG_TEMP_CONST) {
|
|
|
5544c1b |
+ tmp = do_constant_folding_cond(op, args[1], args[2], args[3]);
|
|
|
5544c1b |
+ if (tmp != 2) {
|
|
|
5544c1b |
gen_opc_buf[op_index] = op_to_movi(op);
|
|
|
5544c1b |
- tmp = do_constant_folding_cond(op, temps[args[1]].val,
|
|
|
5544c1b |
- temps[args[2]].val, args[3]);
|
|
|
5544c1b |
tcg_opt_gen_movi(gen_args, args[0], tmp);
|
|
|
5544c1b |
gen_args += 2;
|
|
|
5544c1b |
} else {
|
|
|
5544c1b |
@@ -654,10 +682,9 @@ static TCGArg *tcg_constant_folding(TCGContext *s, uint16_t *tcg_opc_ptr,
|
|
|
5544c1b |
args += 4;
|
|
|
5544c1b |
break;
|
|
|
5544c1b |
CASE_OP_32_64(brcond):
|
|
|
5544c1b |
- if (temps[args[0]].state == TCG_TEMP_CONST
|
|
|
5544c1b |
- && temps[args[1]].state == TCG_TEMP_CONST) {
|
|
|
5544c1b |
- if (do_constant_folding_cond(op, temps[args[0]].val,
|
|
|
5544c1b |
- temps[args[1]].val, args[2])) {
|
|
|
5544c1b |
+ tmp = do_constant_folding_cond(op, args[0], args[1], args[2]);
|
|
|
5544c1b |
+ if (tmp != 2) {
|
|
|
5544c1b |
+ if (tmp) {
|
|
|
5544c1b |
memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info));
|
|
|
5544c1b |
gen_opc_buf[op_index] = INDEX_op_br;
|
|
|
5544c1b |
gen_args[0] = args[3];
|
|
|
5544c1b |
@@ -677,10 +704,8 @@ static TCGArg *tcg_constant_folding(TCGContext *s, uint16_t *tcg_opc_ptr,
|
|
|
5544c1b |
args += 4;
|
|
|
5544c1b |
break;
|
|
|
5544c1b |
CASE_OP_32_64(movcond):
|
|
|
5544c1b |
- if (temps[args[1]].state == TCG_TEMP_CONST
|
|
|
5544c1b |
- && temps[args[2]].state == TCG_TEMP_CONST) {
|
|
|
5544c1b |
- tmp = do_constant_folding_cond(op, temps[args[1]].val,
|
|
|
5544c1b |
- temps[args[2]].val, args[5]);
|
|
|
5544c1b |
+ tmp = do_constant_folding_cond(op, args[1], args[2], args[5]);
|
|
|
5544c1b |
+ if (tmp != 2) {
|
|
|
5544c1b |
if (temps_are_copies(args[0], args[4-tmp])) {
|
|
|
5544c1b |
gen_opc_buf[op_index] = INDEX_op_nop;
|
|
|
5544c1b |
} else if (temps[args[4-tmp]].state == TCG_TEMP_CONST) {
|