Author: leo
Date: Sun Feb 19 11:07:16 2006
New Revision: 11669
Modified:
trunk/src/jit/i386/core.jit
trunk/src/jit/i386/jit_emit.h
Log:
JIT/x86 - really implement JITted sin, cos, sqrt
* providing JIT code for non-existing ops doesn't help at all
(the JIT compiler should warn here)
* <op>_n_n is the thing we need
* partialsums.pir is now ~30% faster than gcc (sse2) and ~70% faster than
gcc w shootout optimizations
Modified: trunk/src/jit/i386/core.jit
==============================================================================
--- trunk/src/jit/i386/core.jit (original)
+++ trunk/src/jit/i386/core.jit Sun Feb 19 11:07:16 2006
@@ -852,17 +852,6 @@ Parrot_neg_n {
unary_x s/<_N>/_n/ s/INT/NUM/ s/<op>/neg/
}
-Parrot_sin {
- unary_x s/<_N>/_n/ s/INT/NUM/ s/<op>/sin/
-}
-
-Parrot_cos_n {
- unary_x s/<_N>/_n/ s/INT/NUM/ s/<op>/cos/
-}
-
-Parrot_sqrt_n {
- unary_x s/<_N>/_n/ s/INT/NUM/ s/<op>/sqrt/
-}
Parrot_bnot_i {
unary_x s/<_N>/_i/ s/<op>/not/
@@ -916,6 +905,18 @@ Parrot_abs_n_n {
unary_x_x s/<_N>/_n/ s/ISR/FSR/ s/INT/NUM/ s/<op>/abs/
}
+Parrot_sin_n_n {
+ unary_x_x s/<_N>/_n/ s/ISR/FSR/ s/INT/NUM/ s/<op>/sin/
+}
+
+Parrot_cos_n_n {
+ unary_x_x s/<_N>/_n/ s/ISR/FSR/ s/INT/NUM/ s/<op>/cos/
+}
+
+Parrot_sqrt_n_n {
+ unary_x_x s/<_N>/_n/ s/ISR/FSR/ s/INT/NUM/ s/<op>/sqrt/
+}
+
TEMPLATE Parrot_ifunless_x_ic {
/*
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 11:07:16 2006
@@ -1259,16 +1259,16 @@ static unsigned char *lastpc;
}
# endif
-# define jit_emit_sin_n(pc, r) \
+# define jit_emit_sin_r_n(pc, r) \
if (r) { \
emitm_fld(pc, r); \
} \
- emitm_fsqrt(pc); \
+ emitm_fsin(pc); \
if (r) { \
emitm_fstp(pc, (r+1)); \
}
-# define jit_emit_cos_n(pc, r) \
+# define jit_emit_cos_r_n(pc, r) \
if (r) { \
emitm_fld(pc, r); \
} \
@@ -1277,7 +1277,7 @@ static unsigned char *lastpc;
emitm_fstp(pc, (r+1)); \
}
-# define jit_emit_sqrt_n(pc, r) \
+# define jit_emit_sqrt_r_n(pc, r) \
if (r) { \
emitm_fld(pc, r); \
} \
