From efe2f93bbdcdfd64e0bda0544e7f73405585cc1c Mon Sep 17 00:00:00 2001
From: Lucas Crowthers <lucasc@codeaurora.org>
Date: Thu, 25 Apr 2013 15:56:16 -0400
Subject: libm: optimistic pow calling

Call optimized pow optimistically and revert to full range
implementation if we detect an out-of-range input.

Change-Id: I02657c2a254e9796e795a8fcc6ae43e24c9d37b9
---
 libm/arm/e_pow.S        | 23 +++++++++++++++++++++--
 libm/src/e_pow.c        | 20 ++++----------------
 libm/src/math_private.h |  3 +++
 3 files changed, 28 insertions(+), 18 deletions(-)

diff --git a/libm/arm/e_pow.S b/libm/arm/e_pow.S
index 8ca9c79..b057308 100644
--- a/libm/arm/e_pow.S
+++ b/libm/arm/e_pow.S
@@ -34,6 +34,7 @@
 #define int_1               d29
 #define double_1            d28
 @ sign and 2^int_n fixup:
+#define maxrange            r12
 #define expadjustment       d7
 #define literals            r10
 @ Values which exist within both polynomial implementations:
@@ -100,13 +101,20 @@
 #define vmov_f32            fconsts
 #define vmovne_f64          fconstdne
 
-ENTRY(pow_neon)
+#if defined(KRAIT_NEON_OPTIMIZATION)
+#define KRAIT_NO_AAPCS_VFP_MODE
+#endif
+
+ENTRY(pow)
 #if defined(KRAIT_NO_AAPCS_VFP_MODE)
      @ ARM ABI has inputs coming in via r registers, lets move to a d register
     vmov            x, x_lw, x_hw
 #endif
     push            {r4, r5, r6, r7, r8, r9, r10, lr}
 
+    movw            maxrange, #0x0000
+    movt            maxrange, #0x4010
+
     @ pre-staged bp values
     vldr            bpa, .LbpA
     vldr            bpb, .LbpB
@@ -176,6 +184,11 @@ ENTRY(pow_neon)
 
 .Lxle2to3over5:
 
+    cmp             x_hw, maxrange
+    cmpls           y_hw, maxrange
+    movt            maxrange, #0x3f00
+    cmpls           maxrange, x_hw
+
     @ load log2 polynomial series constants
     vldm            literals!, {k4, k3, k2, k1}
 
@@ -187,6 +200,8 @@ ENTRY(pow_neon)
     vsub.f64        u, ss, bp
     vadd.f64        v, ss, bp
 
+    bhi             .LuseFullImpl
+
     @ s = (x-1)/(x+1)
     vdiv.f64        ss, u, v
 
@@ -393,6 +408,10 @@ ENTRY(pow_neon)
     @ leave directly returning whatever is in Return_lw and Return_hw
     pop             {r4, r5, r6, r7, r8, r9, r10, pc}
 
+.LuseFullImpl:
+    pop             {r4, r5, r6, r7, r8, r9, r10, lr}
+    b               __full_ieee754_pow
+
 .align 6
 .LliteralTable:
 @ Least-sqares tuned constants for 11th order (log2((1+s)/(1-s)):
@@ -440,4 +459,4 @@ ENTRY(pow_neon)
 
 .Ltwoto1o4: @ 2^1/4
     .long       0x0a31b715, 0x3ff306fe
-END(pow_neon)
+END(pow)
diff --git a/libm/src/e_pow.c b/libm/src/e_pow.c
index 67171e6..5bb7f7b 100644
--- a/libm/src/e_pow.c
+++ b/libm/src/e_pow.c
@@ -61,14 +61,6 @@ static char rcsid[] = "$FreeBSD: src/lib/msun/src/e_pow.c,v 1.11 2005/02/04 18:2
 #include "math.h"
 #include "math_private.h"
 
-#if defined(__ARM_NEON__)
-#if defined(KRAIT_NO_AAPCS_VFP_MODE)
-double pow_neon(double x, double y);
-#else
-double pow_neon(double x, double y, int32_t lx, int32_t hx) __attribute__((pcs("aapcs-vfp")));
-#endif
-#endif
-
 static const double
 bp[] = {1.0, 1.5,},
 dp_h[] = { 0.0, 5.84962487220764160156e-01,}, /* 0x3FE2B803, 0x40000000 */
@@ -103,7 +95,11 @@ ivln2_h  =  1.44269502162933349609e+00, /* 0x3FF71547, 0x60000000 =24b 1/ln2*/
 ivln2_l  =  1.92596299112661746887e-08; /* 0x3E54AE0B, 0xF85DDF44 =1/ln2 tail*/
 
 double
+#if defined(__ARM_NEON__)
+__full_ieee754_pow(double x, double y)
+#else
 __ieee754_pow(double x, double y)
+#endif
 {
 	double z,ax,z_h,z_l,p_h,p_l;
 	double y1,t1,t2,r,s,t,u,v,w;
@@ -229,14 +225,6 @@ __ieee754_pow(double x, double y)
 	    t1 = u+v;
 	    SET_LOW_WORD(t1,0);
 	    t2 = v-(t1-u);
-#if defined(__ARM_NEON__)
-	} else if (ix <= 0x40100000 && iy <= 0x40100000 && hy > 0 && hx > 0) {
-#if defined(KRAIT_NO_AAPCS_VFP_MODE)
-		return pow_neon(x,y);
-#else
-		return pow_neon(x,y,lx,hx);
-#endif
-#endif
 	} else {
 	    double ss,s2,s_h,s_l,t_h,t_l;
 	    n = 0;
diff --git a/libm/src/math_private.h b/libm/src/math_private.h
index 7cda2e9..888272d 100644
--- a/libm/src/math_private.h
+++ b/libm/src/math_private.h
@@ -257,6 +257,9 @@ cpackl(long double x, long double y)
 #define __ieee754_ldexpf ldexpf
 
 /* fdlibm kernel function */
+#if defined(__ARM_NEON__)
+double	__full_ieee754_pow(double,double);
+#endif
 #if defined(KRAIT_NEON_OPTIMIZATION)
 int	__ieee754_rem_pio2(double,double*) __attribute__((pcs("aapcs-vfp")));
 double	__kernel_sin(double,double,int) __attribute__((pcs("aapcs-vfp")));
-- 
cgit v1.1