An initial cut at conditional moves for the i386 backend.
Signed-off-by: Richard Henderson <r...@twiddle.net>
---
elf.h | 2 +
tcg/i386/tcg-target.c | 280 ++++++++++++++++++++++++++++++++++++++++---------
2 files changed, 233 insertions(+), 49 deletions(-)
diff --git a/elf.h b/elf.h
index 11674d7..c84c8ab 100644
--- a/elf.h
+++ b/elf.h
@@ -243,6 +243,8 @@ typedef struct {
#define R_386_GOTOFF 9
#define R_386_GOTPC 10
#define R_386_NUM 11
+/* Not a dynamic reloc, so not included in R_386_NUM. Used in TCG. */
+#define R_386_PC8 23
#define R_MIPS_NONE 0
#define R_MIPS_16 1
diff --git a/tcg/i386/tcg-target.c b/tcg/i386/tcg-target.c
index 972b102..90dbbe9 100644
--- a/tcg/i386/tcg-target.c
+++ b/tcg/i386/tcg-target.c
@@ -61,6 +61,9 @@ static void patch_reloc(uint8_t *code_ptr, int type,
case R_386_PC32:
*(uint32_t *)code_ptr = value - (long)code_ptr;
break;
+ case R_386_PC8:
+ *(uint8_t *)code_ptr = value - (long)code_ptr;
+ break;
default:
tcg_abort();
}
@@ -305,7 +308,8 @@ static void tcg_out_addi(TCGContext *s, int reg,
tcg_target_long val)
tgen_arithi(s, ARITH_ADD, reg, val, 0);
}
-static void tcg_out_jxx(TCGContext *s, int opc, int label_index)
+/* Use SMALL != 0 to force a short forward branch. */
+static void tcg_out_jxx(TCGContext *s, int opc, int label_index, int small)
{
int32_t val, val1;
TCGLabel *l = &s->labels[label_index];
@@ -320,6 +324,7 @@ static void tcg_out_jxx(TCGContext *s, int opc, int
label_index)
tcg_out8(s, 0x70 + opc);
tcg_out8(s, val1);
} else {
+ assert (!small);
if (opc == -1) {
tcg_out8(s, 0xe9);
tcg_out32(s, val - 5);
@@ -329,6 +334,15 @@ static void tcg_out_jxx(TCGContext *s, int opc, int
label_index)
tcg_out32(s, val - 6);
}
}
+ } else if (small) {
+ if (opc == -1) {
+ tcg_out8(s, 0xeb);
+ } else {
+ tcg_out8(s, 0x0f);
+ tcg_out8(s, 0x70 + 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);
@@ -341,9 +355,8 @@ static void tcg_out_jxx(TCGContext *s, int opc, int
label_index)
}
}
-static void tcg_out_brcond(TCGContext *s, int cond,
- TCGArg arg1, TCGArg arg2, int const_arg2,
- int label_index)
+static void tcg_out_cond(TCGContext *s, int cond,
+ TCGArg arg1, TCGArg arg2, int const_arg2)
{
if (const_arg2) {
if (arg2 == 0) {
@@ -355,71 +368,225 @@ static void tcg_out_brcond(TCGContext *s, int cond,
} else {
tcg_out_modrm(s, 0x01 | (ARITH_CMP << 3), arg2, arg1);
}
- tcg_out_jxx(s, tcg_cond_to_jcc[cond], label_index);
+}
+
+static void tcg_out_brcond(TCGContext *s, int cond,
+ TCGArg arg1, TCGArg arg2, int const_arg2,
+ int label_index, int small)
+{
+ tcg_out_cond(s, cond, arg1, arg2, const_arg2);
+ tcg_out_jxx(s, tcg_cond_to_jcc[cond], label_index, small);
}
/* XXX: we implement it at the target level to avoid having to
handle cross basic blocks temporaries */
-static void tcg_out_brcond2(TCGContext *s,
- const TCGArg *args, const int *const_args)
+static void tcg_out_brcond2(TCGContext *s, const TCGArg *args,
+ const int *const_args, int small)
{
- int label_next;
- label_next = gen_new_label();
- switch(args[4]) {
+ int label_next = gen_new_label();
+ int label_dest = args[5];
+ int cond = args[4], c1, c2, c3;
+
+ switch (cond) {
case TCG_COND_EQ:
- tcg_out_brcond(s, TCG_COND_NE, args[0], args[2], const_args[2],
label_next);
- tcg_out_brcond(s, TCG_COND_EQ, args[1], args[3], const_args[3],
args[5]);
+ c1 = -1, c2 = TCG_COND_NE, c3 = TCG_COND_EQ;
break;
case TCG_COND_NE:
- tcg_out_brcond(s, TCG_COND_NE, args[0], args[2], const_args[2],
args[5]);
- tcg_out_brcond(s, TCG_COND_NE, args[1], args[3], const_args[3],
args[5]);
+ c1 = TCG_COND_NE, c2 = -1, c3 = TCG_COND_NE;
break;
case TCG_COND_LT:
- tcg_out_brcond(s, TCG_COND_LT, args[1], args[3], const_args[3],
args[5]);
- tcg_out_jxx(s, JCC_JNE, label_next);
- tcg_out_brcond(s, TCG_COND_LTU, args[0], args[2], const_args[2],
args[5]);
- break;
- case TCG_COND_LE:
- tcg_out_brcond(s, TCG_COND_LT, args[1], args[3], const_args[3],
args[5]);
- tcg_out_jxx(s, JCC_JNE, label_next);
- tcg_out_brcond(s, TCG_COND_LEU, args[0], args[2], const_args[2],
args[5]);
- break;
- case TCG_COND_GT:
- tcg_out_brcond(s, TCG_COND_GT, args[1], args[3], const_args[3],
args[5]);
- tcg_out_jxx(s, JCC_JNE, label_next);
- tcg_out_brcond(s, TCG_COND_GTU, args[0], args[2], const_args[2],
args[5]);
- break;
- case TCG_COND_GE:
- tcg_out_brcond(s, TCG_COND_GT, args[1], args[3], const_args[3],
args[5]);
- tcg_out_jxx(s, JCC_JNE, label_next);
- tcg_out_brcond(s, TCG_COND_GEU, args[0], args[2], const_args[2],
args[5]);
- break;
case TCG_COND_LTU:
- tcg_out_brcond(s, TCG_COND_LTU, args[1], args[3], const_args[3],
args[5]);
- tcg_out_jxx(s, JCC_JNE, label_next);
- tcg_out_brcond(s, TCG_COND_LTU, args[0], args[2], const_args[2],
args[5]);
+ c1 = cond, c2 = TCG_COND_NE, c3 = TCG_COND_LTU;
break;
+ case TCG_COND_LE:
case TCG_COND_LEU:
- tcg_out_brcond(s, TCG_COND_LTU, args[1], args[3], const_args[3],
args[5]);
- tcg_out_jxx(s, JCC_JNE, label_next);
- tcg_out_brcond(s, TCG_COND_LEU, args[0], args[2], const_args[2],
args[5]);
+ c1 = cond, c2 = TCG_COND_NE, c3 = TCG_COND_LEU;
break;
+ case TCG_COND_GT:
case TCG_COND_GTU:
- tcg_out_brcond(s, TCG_COND_GTU, args[1], args[3], const_args[3],
args[5]);
- tcg_out_jxx(s, JCC_JNE, label_next);
- tcg_out_brcond(s, TCG_COND_GTU, args[0], args[2], const_args[2],
args[5]);
+ c1 = cond, c2 = TCG_COND_NE, c3 = TCG_COND_GTU;
break;
+ case TCG_COND_GE:
case TCG_COND_GEU:
- tcg_out_brcond(s, TCG_COND_GTU, args[1], args[3], const_args[3],
args[5]);
- tcg_out_jxx(s, JCC_JNE, label_next);
- tcg_out_brcond(s, TCG_COND_GEU, args[0], args[2], const_args[2],
args[5]);
+ c1 = cond, c2 = TCG_COND_NE, c3 = TCG_COND_GEU;
break;
default:
- tcg_abort();
+ tcg_abort ();
+ }
+
+ tcg_out_cond(s, cond, args[1], args[3], const_args[3]);
+ if (c1 != -1) {
+ tcg_out_jxx(s, tcg_cond_to_jcc[c1], label_dest, small);
+ }
+ if (c2 != -1) {
+ tcg_out_jxx(s, tcg_cond_to_jcc[c2], label_next, 1);
}
+ tcg_out_brcond(s, c3, args[0], args[2], const_args[2], label_dest, small);
+
tcg_out_label(s, label_next, (tcg_target_long)s->code_ptr);
}
+static void tcg_out_setcond(TCGContext *s, int cond, TCGArg arg0,
+ TCGArg arg1, TCGArg arg2, int const_arg2)
+{
+ int use_xor = (arg0 != arg1 && (const_arg2 || arg0 != arg2));
+
+ if (use_xor)
+ tcg_out_movi(s, TCG_TYPE_I32, arg0, 0);
+ tcg_out_cond(s, cond, arg1, arg2, const_arg2);
+ tcg_out_modrm(s, 0x90 | tcg_cond_to_jcc[cond] | P_EXT, 0, arg0);
+ if (!use_xor)
+ tgen_arithi(s, ARITH_AND, arg0, 0xff, 0);
+}
+
+static void tcg_out_setcond2(TCGContext *s, const TCGArg *args,
+ const int *const_args)
+{
+ int overlapl, overlaph;
+ TCGArg new_args[6];
+ int label_true, label_over;
+
+ overlapl = (args[0] == args[1] || (!const_args[3] && args[0] == args[3]));
+ overlaph = (args[0] == args[2] || (!const_args[4] && args[0] == args[4]));
+ memcpy(new_args, args+1, 5*sizeof(TCGArg));
+
+ if (!overlapl && !overlaph) {
+ /* ??? For EQ and NE, and output register in 'q', we could
+ implement this as cmp lows; setb %al; cmp highs; setb %ah;
+ andb %ah, %al; movzbl %al, %eax it's not clear it's worth
+ it though. */
+
+ /* When possible, clear the destination first and increment in
+ the true case. This results in smaller code than the
+ general case below. */
+ tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
+
+ label_over = gen_new_label();
+ new_args[5] = label_over;
+ tcg_out_brcond2(s, new_args, const_args+1, 1);
+
+ tgen_arithi(s, ARITH_ADD, args[0], 1, 0);
+ tcg_out_label(s, label_over, (tcg_target_long)s->code_ptr);
+ } else {
+ label_true = gen_new_label();
+ label_over = gen_new_label();
+
+ new_args[5] = label_true;
+ tcg_out_brcond2(s, new_args, const_args+1, 1);
+
+ tcg_out_movi(s, TCG_TYPE_I32, args[0], 0);
+ tcg_out_jxx(s, JCC_JMP, label_over, 1);
+ tcg_out_label(s, label_true, (tcg_target_long)s->code_ptr);
+
+ tcg_out_movi(s, TCG_TYPE_I32, args[0], 1);
+ tcg_out_label(s, label_over, (tcg_target_long)s->code_ptr);
+ }
+}
+
+static inline int have_cmov(void)
+{
+#ifdef __i686__
+ /* Compiler options say that cmov is available. */
+ return 1;
+#else
+ /* ??? Use cpuid or something and figure out what's running. */
+ return 0;
+#endif
+}
+
+static void tcg_out_movcond(TCGContext *s, const TCGArg *args,
+ const int *const_args)
+{
+ int vtc, vfc, cond, use_cmov = 0, do_swap = 0;
+ TCGArg d, vt, vf;
+
+ d = args[0];
+ vt = args[3];
+ vf = args[4];
+ vtc = const_args[3];
+ vfc = const_args[4];
+
+ /* ??? The jcc code path below assumes that one mov insn must be skipped.
+ Rather than complicate the code below, make sure to simplify the
+ conditional move here. */
+ if (vtc == vfc && vt == vf) {
+ if (vtc)
+ tcg_out_movi(s, TCG_TYPE_I32, d, vt);
+ else
+ tcg_out_mov(s, d, vt);
+ return;
+ }
+
+ cond = args[5];
+
+ /* If both arguments are constants, we *could* do all the funny bits that
+ gcc does with sbc, masks, etc. There's likely no point. Just use the
+ jcc version in this case. We also have to be careful about clobbering
+ inputs when trying to move constants into position. */
+
+ if (have_cmov()) {
+ use_cmov = 1;
+ if (vtc) {
+ if (vfc || d == vf)
+ use_cmov = 0;
+ else
+ do_swap = 1;
+ } else if (d == vt) {
+ if (vfc)
+ use_cmov = 0;
+ else
+ do_swap = 1;
+ }
+ }
+
+ if (!use_cmov) {
+ /* We're going to follow the lead of cmov and set D=VF first,
+ which means inverting the condition upon which we jump. */
+ cond = tcg_invert_cond(cond);
+
+ /* Don't allow the move we jump over to be a nop. */
+ do_swap = (!vtc && d == vt);
+ }
+
+ if (do_swap) {
+ TCGArg t;
+ cond = tcg_invert_cond(cond);
+ t = vf, vf = vt, vt = t;
+ t = vfc, vfc = vtc, vtc = t;
+ }
+
+ /* If possible, set D=0 before the compare, so that we can use XOR. */
+ if (vfc && vf == 0 && d != args[1] && (const_args[2] || d != args[2])) {
+ tcg_out_movi(s, TCG_TYPE_I32, d, vf);
+ vf = d, vfc = 0;
+ }
+
+ tcg_out_cond(s, cond, args[1], args[2], const_args[2]);
+
+ if (vfc) {
+ /* Force the use of "mov $0, d" to avoid clobbering flags. */
+ tcg_out8(s, 0xb8 + d);
+ tcg_out32(s, vf);
+ } else {
+ tcg_out_mov(s, d, vf);
+ }
+
+ if (use_cmov) {
+ assert (!vtc);
+ tcg_out_modrm(s, 0x40 | tcg_cond_to_jcc[cond] | P_EXT, d, vt);
+ } else {
+ int label_next = gen_new_label();
+
+ tcg_out_jxx(s, tcg_cond_to_jcc[cond], label_next, 1);
+ if (vtc)
+ tcg_out_movi(s, TCG_TYPE_I32, d, vt);
+ else
+ tcg_out_mov(s, d, vt);
+
+ tcg_out_label(s, label_next, (tcg_target_long)s->code_ptr);
+ }
+}
+
#if defined(CONFIG_SOFTMMU)
#include "../../softmmu_defs.h"
@@ -913,7 +1080,7 @@ static inline void tcg_out_op(TCGContext *s, int opc,
}
break;
case INDEX_op_br:
- tcg_out_jxx(s, JCC_JMP, args[0]);
+ tcg_out_jxx(s, JCC_JMP, args[0], 0);
break;
case INDEX_op_movi_i32:
tcg_out_movi(s, TCG_TYPE_I32, args[0], args[1]);
@@ -1044,10 +1211,11 @@ static inline void tcg_out_op(TCGContext *s, int opc,
tcg_out_modrm(s, 0x01 | (ARITH_SBB << 3), args[5], args[1]);
break;
case INDEX_op_brcond_i32:
- tcg_out_brcond(s, args[2], args[0], args[1], const_args[1], args[3]);
+ tcg_out_brcond(s, args[2], args[0], args[1], const_args[1],
+ args[3], 0);
break;
case INDEX_op_brcond2_i32:
- tcg_out_brcond2(s, args, const_args);
+ tcg_out_brcond2(s, args, const_args, 0);
break;
case INDEX_op_bswap16_i32:
@@ -1080,6 +1248,16 @@ static inline void tcg_out_op(TCGContext *s, int opc,
tcg_out_modrm(s, 0xb7 | P_EXT, args[0], args[1]);
break;
+ case INDEX_op_setcond_i32:
+ tcg_out_setcond(s, args[3], args[0], args[1], args[2], const_args[2]);
+ break;
+ case INDEX_op_movcond_i32:
+ tcg_out_movcond(s, args, const_args);
+ break;
+ case INDEX_op_setcond2_i32:
+ tcg_out_setcond2(s, args, const_args);
+ break;
+
case INDEX_op_qemu_ld8u:
tcg_out_qemu_ld(s, args, 0);
break;
@@ -1168,6 +1346,10 @@ static const TCGTargetOpDef x86_op_defs[] = {
{ INDEX_op_ext8u_i32, { "r", "q"} },
{ INDEX_op_ext16u_i32, { "r", "r"} },
+ { INDEX_op_setcond_i32, { "q", "r", "ri" } },
+ { INDEX_op_movcond_i32, { "r", "r", "ri", "ri", "ri" } },
+ { INDEX_op_setcond2_i32, { "r", "r", "r", "ri", "ri" } },
+
#if TARGET_LONG_BITS == 32
{ INDEX_op_qemu_ld8u, { "r", "L" } },
{ INDEX_op_qemu_ld8s, { "r", "L" } },
--
1.6.2.5
