Author: leo
Date: Sun Feb 19 08:56:30 2006
New Revision: 11666
Modified:
trunk/src/jit/i386/core.jit
trunk/src/jit/i386/jit_emit.h
trunk/src/ops/experimental.ops
Log:
JIT/x86 and core - improve pow_n_n_i
* speedup core and JITted pow_n_n_i by doing more work in the loop
* enable JIT version
Modified: trunk/src/jit/i386/core.jit
==============================================================================
--- trunk/src/jit/i386/core.jit (original)
+++ trunk/src/jit/i386/core.jit Sun Feb 19 08:56:30 2006
@@ -1682,20 +1682,13 @@ Parrot_isle_i_nc_n {
}
;
-; this code works, but it's slower than the builtin pow (disabled XXX)
-; TODO check alignment
+; TODO check alignment, and maybe move 1 msk into reg
;
-Parrot_XXX_pow_n_n_i {
+Parrot_pow_n_n_i {
int e_reg, saved = 0;
char *L1, *L2, *L3, *L4, *L5, *L6;
- if (MAP[2]) {
- emitm_fld(NATIVECODE, MAP[2]);
- }
- else {
- jit_emit_fload_mb_n(NATIVECODE, emit_EBX, ROFFS_NUM(2));
- } /* n2 = $2 ST(1) */
- emitm_fld1(NATIVECODE); /* res = 1.0 ST(0) */
+ emitm_fld1(NATIVECODE); /* res = 1.0 ST(1) */
if (MAP[3]) {
jit_emit_mov_rr_i(NATIVECODE, emit_EAX, MAP[3]);
}
@@ -1705,50 +1698,59 @@ Parrot_XXX_pow_n_n_i {
jit_emit_test_r_i(NATIVECODE, emit_EAX); /* e == 0? */
L1 = NATIVECODE;
emitm_jxs(NATIVECODE, emitm_jz, 0); /* jz L1 */
+ if (MAP[2]) {
+ emitm_fld(NATIVECODE, (MAP[2] + 1));
+ }
+ else {
+ jit_emit_fload_mb_n(NATIVECODE, emit_EBX, ROFFS_NUM(2));
+ } /* n2 = $2 ST(0) */
if (intreg_is_used(jit_info, emit_ECX)) {
emitm_pushl_r(NATIVECODE, emit_ECX);
saved = 1;
}
- jit_emit_mov_ri_i(NATIVECODE, emit_ECX, 1); /* s = 1 */
+ jit_emit_mov_ri_i(NATIVECODE, emit_ECX, 1); /* s = 1 */
L2 = NATIVECODE;
emitm_jxs(NATIVECODE, emitm_jg, 0); /* jg L2 */
- jit_emit_neg_r_i(NATIVECODE, emit_ECX); /* s = -1 */
- jit_emit_neg_r_i(NATIVECODE, emit_EAX); /* e = -e */
- L2[1] = NATIVECODE - L2 - 2; /* L2: */
- /* L3: while (e) */
- jit_emit_test_r_i(NATIVECODE, emit_EAX); /* e == 0? */
- L3 = NATIVECODE;
+ jit_emit_neg_r_i(NATIVECODE, emit_ECX); /* s = -1 */
+ jit_emit_neg_r_i(NATIVECODE, emit_EAX); /* e = -e */
+ L2[1] = NATIVECODE - L2 - 2;
+ /* L2: */
+ /* while (e) */
+ jit_emit_test_r_i(NATIVECODE, emit_EAX); /* e == 0? */
L4 = NATIVECODE;
- emitm_jxs(NATIVECODE, emitm_jz, 0); /* jz L4 */
- jit_emit_test_ri_i(NATIVECODE, emit_EAX, 1); /* e & 1 ? */
+ emitm_jxs(NATIVECODE, emitm_jz, 0); /* jz L4 */
+ L3 = NATIVECODE;
+ /* L3: */
+ jit_emit_test_ri_i(NATIVECODE, emit_EAX, 1); /* e & 1 ? */
L5 = NATIVECODE;
- emitm_jxs(NATIVECODE, emitm_jz, 0); /* jz L5 */
- emitm_fmul(NATIVECODE, 1); /* res *= n2 */
- jit_emit_dec_r_i(NATIVECODE, emit_EAX); /* --e */
- /* dec and lsr are setting flags - branch past test at L3 */
- emitm_jumps(NATIVECODE, L3 - NATIVECODE - 1); /* jmp L3 */
+ emitm_jxs(NATIVECODE, emitm_jz, 0); /* jz L5 */
+ emitm_fmulr(NATIVECODE, 1); /* res *= n2 */
/* L5: */
- L5[1] = NATIVECODE - L5 - 2; /* L5: */
- jit_emit_mul_rr_n(NATIVECODE, 1, 1); /* n2 *= n2 */
- jit_emit_lsr_ri_i(NATIVECODE, emit_EAX, 1); /* e >>= 1 */
- emitm_jumps(NATIVECODE, L3 - NATIVECODE - 1); /* jmp L3 */
+ L5[1] = NATIVECODE - L5 - 2;
+ jit_emit_lsr_ri_i(NATIVECODE, emit_EAX, 1); /* e >>= 1 */
+ emitm_fmul(NATIVECODE, 0); /* n2 *= n2 */
+ /* lsr is setting flags - branch past test at L3 */
+ emitm_jxs(NATIVECODE, emitm_jnz, (L3 - NATIVECODE - 1)); /* jmp L3 */
/* endwhile */
- L1[1] = NATIVECODE - L1 - 2; /* L2: */
- L4[1] = NATIVECODE - L4 - 2; /* L4: */
+ L4[1] = NATIVECODE - L4 - 2;
+ /* L4: */
jit_emit_test_r_i(NATIVECODE, emit_ECX); /* s ? */
L6 = NATIVECODE;
emitm_jxs(NATIVECODE, emitm_jg, 0); /* jg L6 */
- emitm_fld1(NATIVECODE); /* push 1.0 */
- emitm_fxch(NATIVECODE, 1);
- emitm_fdivp(NATIVECODE, 1); /* res = 1.0/res */
- L6[1] = NATIVECODE - L6 - 2; /* L6: */
+ emitm_fld1(NATIVECODE); /* push 1.0 */
+ emitm_fxch(NATIVECODE, 2);
+ emitm_fdivp(NATIVECODE, 2); /* res = 1.0/res */
+ L6[1] = NATIVECODE - L6 - 2;
+ /* L6: */
+ emitm_fstp(NATIVECODE, 0); /* pop n2 */
+ L1[1] = NATIVECODE - L1 - 2;
+ /* L1: */
if (MAP[1]) {
- emitm_fstp(NATIVECODE, (MAP[1] + 2));
+ emitm_fstp(NATIVECODE, (MAP[1] + 1)); /* store res */
}
else {
jit_emit_fstore_m_n(NATIVECODE, ROFFS_NUM(1));
}
- emitm_fstp(NATIVECODE, 0); /* pop 0 */
if (saved)
emitm_popl_r(NATIVECODE, emit_ECX);
}
Modified: trunk/src/jit/i386/jit_emit.h
==============================================================================
--- trunk/src/jit/i386/jit_emit.h (original)
+++ trunk/src/jit/i386/jit_emit.h Sun Feb 19 08:56:30 2006
@@ -1076,6 +1076,7 @@ static unsigned char *lastpc;
/* FCMOV*, FCOMI PPRO */
/* 0xDC like 0xD8 with reversed operands */
+# define emitm_fmulr(pc, sti) emitm_fl_3(pc, emit_b100, emit_b001, sti)
/* 0xDD ops */
/* FFree ST(i) */
Modified: trunk/src/ops/experimental.ops
==============================================================================
--- trunk/src/ops/experimental.ops (original)
+++ trunk/src/ops/experimental.ops Sun Feb 19 08:56:30 2006
@@ -302,12 +302,9 @@ inline op pow(out NUM, in NUM, in INT) :
while (e) {
if (e & 1) {
res *= n2;
- --e;
- }
- else {
- n2 *= n2;
- e >>= 1;
}
+ n2 *= n2;
+ e >>= 1;
}
if (s < 0) {
res = 1.0/res;
