AArch64: Improve codegen in SVE expm1f and users

Use unpredicated muls, use absolute compare and improve memory access.
Expm1f, sinhf and tanhf show 7%, 5% and 1% improvement in throughput
microbenchmark on Neoverse V1.

4 files changed, 44 insertions, 45 deletions

diff --git a/sysdeps/aarch64/fpu/expm1f_sve.c b/sysdeps/aarch64/fpu/expm1f_sve.c
index 742ebbe..21cb2e1 100644
--- a/sysdeps/aarch64/fpu/expm1f_sve.c
+++ b/sysdeps/aarch64/fpu/expm1f_sve.c
@@ -18,7 +18,6 @@
    <https://www.gnu.org/licenses/>.  */
 
 #include "sv_math.h"
-#include "poly_sve_f32.h"
 
 /* Largest value of x for which expm1(x) should round to -1.  */
 #define SpecialBound 0x1.5ebc4p+6f
@@ -28,20 +27,17 @@ static const struct data
   /* These 4 are grouped together so they can be loaded as one quadword, then
      used with _lane forms of svmla/svmls.  */
   float c2, c4, ln2_hi, ln2_lo;
-  float c0, c1, c3, inv_ln2, special_bound, shift;
+  float c0, inv_ln2, c1, c3, special_bound;
 } data = {
   /* Generated using fpminimax.  */
   .c0 = 0x1.fffffep-2,		 .c1 = 0x1.5554aep-3,
   .c2 = 0x1.555736p-5,		 .c3 = 0x1.12287cp-7,
-  .c4 = 0x1.6b55a2p-10,
+  .c4 = 0x1.6b55a2p-10,		 .inv_ln2 = 0x1.715476p+0f,
+  .special_bound = SpecialBound, .ln2_lo = 0x1.7f7d1cp-20f,
+  .ln2_hi = 0x1.62e4p-1f,
 
-  .special_bound = SpecialBound, .shift = 0x1.8p23f,
-  .inv_ln2 = 0x1.715476p+0f,	 .ln2_hi = 0x1.62e4p-1f,
-  .ln2_lo = 0x1.7f7d1cp-20f,
 };
 
-#define C(i) sv_f32 (d->c##i)
-
 static svfloat32_t NOINLINE
 special_case (svfloat32_t x, svbool_t pg)
 {
@@ -71,9 +67,8 @@ svfloat32_t SV_NAME_F1 (expm1) (svfloat32_t x, svbool_t pg)
      and f = x - i * ln2, then f is in [-ln2/2, ln2/2].
      exp(x) - 1 = 2^i * (expm1(f) + 1) - 1
      where 2^i is exact because i is an integer.  */
-  svfloat32_t j = svmla_x (pg, sv_f32 (d->shift), x, d->inv_ln2);
-  j = svsub_x (pg, j, d->shift);
-  svint32_t i = svcvt_s32_x (pg, j);
+  svfloat32_t j = svmul_x (svptrue_b32 (), x, d->inv_ln2);
+  j = svrinta_x (pg, j);
 
   svfloat32_t f = svmls_lane (x, j, lane_constants, 2);
   f = svmls_lane (f, j, lane_constants, 3);
@@ -83,17 +78,17 @@ svfloat32_t SV_NAME_F1 (expm1) (svfloat32_t x, svbool_t pg)
 	 x + ax^2 + bx^3 + cx^4 ....
      So we calculate the polynomial P(f) = a + bf + cf^2 + ...
      and assemble the approximation expm1(f) ~= f + f^2 * P(f).  */
-  svfloat32_t p12 = svmla_lane (C (1), f, lane_constants, 0);
-  svfloat32_t p34 = svmla_lane (C (3), f, lane_constants, 1);
-  svfloat32_t f2 = svmul_x (pg, f, f);
+  svfloat32_t p12 = svmla_lane (sv_f32 (d->c1), f, lane_constants, 0);
+  svfloat32_t p34 = svmla_lane (sv_f32 (d->c3), f, lane_constants, 1);
+  svfloat32_t f2 = svmul_x (svptrue_b32 (), f, f);
   svfloat32_t p = svmla_x (pg, p12, f2, p34);
-  p = svmla_x (pg, C (0), f, p);
+
+  p = svmla_x (pg, sv_f32 (d->c0), f, p);
   p = svmla_x (pg, f, f2, p);
 
   /* Assemble the result.
      expm1(x) ~= 2^i * (p + 1) - 1
      Let t = 2^i.  */
-  svfloat32_t t = svreinterpret_f32 (
-      svadd_x (pg, svreinterpret_u32 (svlsl_x (pg, i, 23)), 0x3f800000));
-  return svmla_x (pg, svsub_x (pg, t, 1), p, t);
+  svfloat32_t t = svscale_x (pg, sv_f32 (1.0f), svcvt_s32_x (pg, j));
+  return svmla_x (pg, svsub_x (pg, t, 1.0f), p, t);
 }
diff --git a/sysdeps/aarch64/fpu/sinhf_sve.c b/sysdeps/aarch64/fpu/sinhf_sve.c
index a6dc0cd..90692ac 100644
--- a/sysdeps/aarch64/fpu/sinhf_sve.c
+++ b/sysdeps/aarch64/fpu/sinhf_sve.c
@@ -63,5 +63,5 @@ svfloat32_t SV_NAME_F1 (sinh) (svfloat32_t x, const svbool_t pg)
   if (__glibc_unlikely (svptest_any (pg, special)))
     return special_case (x, svmul_x (pg, t, halfsign), special);
 
-  return svmul_x (pg, t, halfsign);
+  return svmul_x (svptrue_b32 (), t, halfsign);
 }
diff --git a/sysdeps/aarch64/fpu/sv_expm1f_inline.h b/sysdeps/aarch64/fpu/sv_expm1f_inline.h
index c5af5fe..ef2b8ea 100644
--- a/sysdeps/aarch64/fpu/sv_expm1f_inline.h
+++ b/sysdeps/aarch64/fpu/sv_expm1f_inline.h
@@ -27,21 +27,18 @@ struct sv_expm1f_data
   /* These 4 are grouped together so they can be loaded as one quadword, then
    used with _lane forms of svmla/svmls.  */
   float32_t c2, c4, ln2_hi, ln2_lo;
-  float32_t c0, c1, c3, inv_ln2, shift;
+  float c0, inv_ln2, c1, c3, special_bound;
 };
 
 /* Coefficients generated using fpminimax.  */
 #define SV_EXPM1F_DATA                                                        \
   {                                                                           \
-    .c0 = 0x1.fffffep-2, .c1 = 0x1.5554aep-3, .c2 = 0x1.555736p-5,            \
-    .c3 = 0x1.12287cp-7, .c4 = 0x1.6b55a2p-10,                                \
+    .c0 = 0x1.fffffep-2, .c1 = 0x1.5554aep-3, .inv_ln2 = 0x1.715476p+0f,      \
+    .c2 = 0x1.555736p-5, .c3 = 0x1.12287cp-7,                                 \
                                                                               \
-    .shift = 0x1.8p23f, .inv_ln2 = 0x1.715476p+0f, .ln2_hi = 0x1.62e4p-1f,    \
-    .ln2_lo = 0x1.7f7d1cp-20f,                                                \
+    .c4 = 0x1.6b55a2p-10, .ln2_lo = 0x1.7f7d1cp-20f, .ln2_hi = 0x1.62e4p-1f,  \
   }
 
-#define C(i) sv_f32 (d->c##i)
-
 static inline svfloat32_t
 expm1f_inline (svfloat32_t x, svbool_t pg, const struct sv_expm1f_data *d)
 {
@@ -55,9 +52,8 @@ expm1f_inline (svfloat32_t x, svbool_t pg, const struct sv_expm1f_data *d)
      and f = x - i * ln2, then f is in [-ln2/2, ln2/2].
      exp(x) - 1 = 2^i * (expm1(f) + 1) - 1
      where 2^i is exact because i is an integer.  */
-  svfloat32_t j = svmla_x (pg, sv_f32 (d->shift), x, d->inv_ln2);
-  j = svsub_x (pg, j, d->shift);
-  svint32_t i = svcvt_s32_x (pg, j);
+  svfloat32_t j = svmul_x (svptrue_b32 (), x, d->inv_ln2);
+  j = svrinta_x (pg, j);
 
   svfloat32_t f = svmls_lane (x, j, lane_constants, 2);
   f = svmls_lane (f, j, lane_constants, 3);
@@ -67,18 +63,18 @@ expm1f_inline (svfloat32_t x, svbool_t pg, const struct sv_expm1f_data *d)
 	 x + ax^2 + bx^3 + cx^4 ....
      So we calculate the polynomial P(f) = a + bf + cf^2 + ...
      and assemble the approximation expm1(f) ~= f + f^2 * P(f).  */
-  svfloat32_t p12 = svmla_lane (C (1), f, lane_constants, 0);
-  svfloat32_t p34 = svmla_lane (C (3), f, lane_constants, 1);
-  svfloat32_t f2 = svmul_x (pg, f, f);
+  svfloat32_t p12 = svmla_lane (sv_f32 (d->c1), f, lane_constants, 0);
+  svfloat32_t p34 = svmla_lane (sv_f32 (d->c3), f, lane_constants, 1);
+  svfloat32_t f2 = svmul_x (svptrue_b32 (), f, f);
   svfloat32_t p = svmla_x (pg, p12, f2, p34);
-  p = svmla_x (pg, C (0), f, p);
+  p = svmla_x (pg, sv_f32 (d->c0), f, p);
   p = svmla_x (pg, f, f2, p);
 
   /* Assemble the result.
      expm1(x) ~= 2^i * (p + 1) - 1
      Let t = 2^i.  */
-  svfloat32_t t = svscale_x (pg, sv_f32 (1), i);
-  return svmla_x (pg, svsub_x (pg, t, 1), p, t);
+  svfloat32_t t = svscale_x (pg, sv_f32 (1.0f), svcvt_s32_x (pg, j));
+  return svmla_x (pg, svsub_x (pg, t, 1.0f), p, t);
 }
 
 #endif
diff --git a/sysdeps/aarch64/fpu/tanhf_sve.c b/sysdeps/aarch64/fpu/tanhf_sve.c
index c33bf09..e12f86d 100644
--- a/sysdeps/aarch64/fpu/tanhf_sve.c
+++ b/sysdeps/aarch64/fpu/tanhf_sve.c
@@ -19,20 +19,27 @@
 
 #include "sv_expm1f_inline.h"
 
+/* Largest value of x for which tanhf(x) rounds to 1 (or -1 for negative).  */
+#define BoringBound 0x1.205966p+3f
+
 static const struct data
 {
   struct sv_expm1f_data expm1f_consts;
-  uint32_t boring_bound, onef;
+  uint32_t onef, special_bound;
+  float boring_bound;
 } data = {
   .expm1f_consts = SV_EXPM1F_DATA,
-  /* 0x1.205966p+3, above which tanhf rounds to 1 (or -1 for negative).  */
-  .boring_bound = 0x41102cb3,
   .onef = 0x3f800000,
+  .special_bound = 0x7f800000,
+  .boring_bound = BoringBound,
 };
 
 static svfloat32_t NOINLINE
-special_case (svfloat32_t x, svfloat32_t y, svbool_t special)
+special_case (svfloat32_t x, svbool_t pg, svbool_t is_boring,
+	      svfloat32_t boring, svfloat32_t q, svbool_t special)
 {
+  svfloat32_t y
+      = svsel_f32 (is_boring, boring, svdiv_x (pg, q, svadd_x (pg, q, 2.0)));
   return sv_call_f32 (tanhf, x, y, special);
 }
 
@@ -47,15 +54,16 @@ svfloat32_t SV_NAME_F1 (tanh) (svfloat32_t x, const svbool_t pg)
   svfloat32_t ax = svabs_x (pg, x);
   svuint32_t iax = svreinterpret_u32 (ax);
   svuint32_t sign = sveor_x (pg, svreinterpret_u32 (x), iax);
-  svbool_t is_boring = svcmpgt (pg, iax, d->boring_bound);
   svfloat32_t boring = svreinterpret_f32 (svorr_x (pg, sign, d->onef));
-
-  svbool_t special = svcmpgt (pg, iax, 0x7f800000);
+  svbool_t special = svcmpgt (pg, iax, d->special_bound);
+  svbool_t is_boring = svacgt (pg, x, d->boring_bound);
 
   /* tanh(x) = (e^2x - 1) / (e^2x + 1).  */
-  svfloat32_t q = expm1f_inline (svmul_x (pg, x, 2.0), pg, &d->expm1f_consts);
-  svfloat32_t y = svdiv_x (pg, q, svadd_x (pg, q, 2.0));
+  svfloat32_t q = expm1f_inline (svmul_x (svptrue_b32 (), x, 2.0), pg,
+				 &d->expm1f_consts);
+
   if (__glibc_unlikely (svptest_any (pg, special)))
-    return special_case (x, svsel_f32 (is_boring, boring, y), special);
+    return special_case (x, pg, is_boring, boring, q, special);
+  svfloat32_t y = svdiv_x (pg, q, svadd_x (pg, q, 2.0));
   return svsel_f32 (is_boring, boring, y);
 }