Hi all,
This patch adds the following NEON intrinsics to the ARM Aarch64 GCC:
* vmaxnm_f64
* vminnm_f64
* vmax_f64
* vmin_f64
Some refactoring was done as well to remove some superfluous UNSPECS
and iterators that could be combined into one.
<fmaxmin><mode>3 was removed from aarch64-simd.md because after the
combining of the unspec FMAXMIN with FMAXMIN_UNS the pattern
immediately above it now generates the same cases.
Added new tests for these and ran regression tests on aarch64-none-elf.
Ok for trunk?
Thanks,
Tamar
gcc/
2016-07-08 Tamar Christina <tamar.christ...@arm.com>
* config/aarch64/aarch64-simd-builtins.def
(__builtin_aarch64_fmindf): Change BUILTIN_VDQF to BUILTIN_VDQF_DF.
(__builtin_aarch64_fmaxdf): Likewise.
(__builtin_aarch64_smin_nandf): Likewise.
(__builtin_aarch64_smax_nandf): Likewise.
* config/aarch64/aarch64-simd.md (<fmaxmin><mode>3): Remove.
* config/aarch64/aarch64.md (<fmaxmin><mode>3): Rename to...
(<fmaxmin><mode>3): ...this.
* config/aarch64/arm_neon.h (vmaxnm_f64): New.
(vminnm_f64): Likewise.
(vmin_f64): Likewise.
(vmax_f64): Likewise.
* config/aarch64/iterators.md (FMAXMIN): Merge with...
(FMAXMIN_UNS): ...this.
(fmaxmin): Merged with
(fmaxmin_op): ...this...
(maxmin_uns_op): ...in to this.
gcc/testsuite/
2016-07-11 Tamar Christina <tamar.christ...@arm.com>
* gcc.target/aarch64/vminmaxnm.c: New.
* gcc.target/aarch64/simd/vminmaxnm_1.c (main): Added float64x1_t
tests.
:100644 100644 c7fe08b... e1154b4... M gcc/config/aarch64/aarch64-simd-builtins.def
:100644 100644 0bf3ac8... f2575a0... M gcc/config/aarch64/aarch64-simd.md
:100644 100644 7d8b394... 08ce915... M gcc/config/aarch64/aarch64.md
:100644 100644 ab3a00c... 07c04a3... M gcc/config/aarch64/arm_neon.h
:100644 100644 91e2e64... f35d8d7... M gcc/config/aarch64/iterators.md
:100644 100644 96608eb... 192bad9... M gcc/testsuite/gcc.target/aarch64/simd/vminmaxnm_1.c
:000000 100644 0000000... bdaa564... A gcc/testsuite/gcc.target/aarch64/vminmaxnm.c
diff --git a/gcc/config/aarch64/aarch64-simd-builtins.def b/gcc/config/aarch64/aarch64-simd-builtins.def
index c7fe08b..e1154b4 100644
--- a/gcc/config/aarch64/aarch64-simd-builtins.def
+++ b/gcc/config/aarch64/aarch64-simd-builtins.def
@@ -241,19 +241,19 @@
BUILTIN_VHSDF (UNOP, reduc_smax_nan_scal_, 10)
BUILTIN_VHSDF (UNOP, reduc_smin_nan_scal_, 10)
- /* Implemented by <maxmin><mode>3.
+ /* Implemented by <maxmin_uns><mode>3.
smax variants map to fmaxnm,
smax_nan variants map to fmax. */
BUILTIN_VDQ_BHSI (BINOP, smax, 3)
BUILTIN_VDQ_BHSI (BINOP, smin, 3)
BUILTIN_VDQ_BHSI (BINOP, umax, 3)
BUILTIN_VDQ_BHSI (BINOP, umin, 3)
- BUILTIN_VHSDF (BINOP, smax_nan, 3)
- BUILTIN_VHSDF (BINOP, smin_nan, 3)
+ BUILTIN_VHSDF_DF (BINOP, smax_nan, 3)
+ BUILTIN_VHSDF_DF (BINOP, smin_nan, 3)
- /* Implemented by <fmaxmin><mode>3. */
- BUILTIN_VHSDF (BINOP, fmax, 3)
- BUILTIN_VHSDF (BINOP, fmin, 3)
+ /* Implemented by <maxmin_uns><mode>3. */
+ BUILTIN_VHSDF_HSDF (BINOP, fmax, 3)
+ BUILTIN_VHSDF_HSDF (BINOP, fmin, 3)
/* Implemented by aarch64_<maxmin_uns>p<mode>. */
BUILTIN_VDQ_BHSI (BINOP, smaxp, 0)
@@ -549,8 +549,4 @@
BUILTIN_GPI (UNOP, fix_truncdf, 2)
BUILTIN_GPI_I16 (UNOPUS, fixuns_trunchf, 2)
BUILTIN_GPI (UNOPUS, fixuns_truncsf, 2)
- BUILTIN_GPI (UNOPUS, fixuns_truncdf, 2)
-
- /* Implemented by <fmaxmin><mode>3. */
- VAR1 (BINOP, fmax, 3, hf)
- VAR1 (BINOP, fmin, 3, hf)
+ BUILTIN_GPI (UNOPUS, fixuns_truncdf, 2)
\ No newline at end of file
diff --git a/gcc/config/aarch64/aarch64-simd.md b/gcc/config/aarch64/aarch64-simd.md
index 0bf3ac8..f2575a0 100644
--- a/gcc/config/aarch64/aarch64-simd.md
+++ b/gcc/config/aarch64/aarch64-simd.md
@@ -2038,6 +2038,9 @@
[(set_attr "type" "neon_fp_minmax_<stype><q>")]
)
+;; Vector forms for fmax, fmin, fmaxnm, fminnm.
+;; fmaxnm and fminnm are used for the fmax<mode>3 standard pattern names,
+;; which implement the IEEE fmax ()/fmin () functions.
(define_insn "<maxmin_uns><mode>3"
[(set (match_operand:VHSDF 0 "register_operand" "=w")
(unspec:VHSDF [(match_operand:VHSDF 1 "register_operand" "w")
@@ -2048,17 +2051,6 @@
[(set_attr "type" "neon_fp_minmax_<stype><q>")]
)
-;; Auto-vectorized forms for the IEEE-754 fmax()/fmin() functions
-(define_insn "<fmaxmin><mode>3"
- [(set (match_operand:VHSDF 0 "register_operand" "=w")
- (unspec:VHSDF [(match_operand:VHSDF 1 "register_operand" "w")
- (match_operand:VHSDF 2 "register_operand" "w")]
- FMAXMIN))]
- "TARGET_SIMD"
- "<fmaxmin_op>\\t%0.<Vtype>, %1.<Vtype>, %2.<Vtype>"
- [(set_attr "type" "neon_fp_minmax_<stype><q>")]
-)
-
;; 'across lanes' add.
(define_expand "reduc_plus_scal_<mode>"
diff --git a/gcc/config/aarch64/aarch64.md b/gcc/config/aarch64/aarch64.md
index 7d8b394..08ce915 100644
--- a/gcc/config/aarch64/aarch64.md
+++ b/gcc/config/aarch64/aarch64.md
@@ -4831,14 +4831,16 @@
[(set_attr "type" "f_minmax<s>")]
)
-;; Scalar forms for the IEEE-754 fmax()/fmin() functions
-(define_insn "<fmaxmin><mode>3"
+;; Scalar forms for fmax, fmin, fmaxnm, fminnm.
+;; fmaxnm and fminnm are used for the fmax<mode>3 standard pattern names,
+;; which implement the IEEE fmax ()/fmin () functions.
+(define_insn "<maxmin_uns><mode>3"
[(set (match_operand:GPF_F16 0 "register_operand" "=w")
(unspec:GPF_F16 [(match_operand:GPF_F16 1 "register_operand" "w")
(match_operand:GPF_F16 2 "register_operand" "w")]
- FMAXMIN))]
+ FMAXMIN_UNS))]
"TARGET_FLOAT"
- "<fmaxmin_op>\\t%<s>0, %<s>1, %<s>2"
+ "<maxmin_uns_op>\\t%<s>0, %<s>1, %<s>2"
[(set_attr "type" "f_minmax<stype>")]
)
diff --git a/gcc/config/aarch64/arm_neon.h b/gcc/config/aarch64/arm_neon.h
index ab3a00c..07c04a3 100644
--- a/gcc/config/aarch64/arm_neon.h
+++ b/gcc/config/aarch64/arm_neon.h
@@ -17201,6 +17201,14 @@ vmax_f32 (float32x2_t __a, float32x2_t __b)
return __builtin_aarch64_smax_nanv2sf (__a, __b);
}
+__extension__ static __inline float64x1_t __attribute__ ((__always_inline__))
+vmax_f64 (float64x1_t __a, float64x1_t __b)
+{
+ return (float64x1_t)
+ { __builtin_aarch64_smax_nandf (vget_lane_f64 (__a, 0),
+ vget_lane_f64 (__b, 0)) };
+}
+
__extension__ static __inline int8x8_t __attribute__ ((__always_inline__))
vmax_s8 (int8x8_t __a, int8x8_t __b)
{
@@ -17692,6 +17700,14 @@ vmaxnm_f32 (float32x2_t __a, float32x2_t __b)
return __builtin_aarch64_fmaxv2sf (__a, __b);
}
+__extension__ static __inline float64x1_t __attribute__ ((__always_inline__))
+vmaxnm_f64 (float64x1_t __a, float64x1_t __b)
+{
+ return (float64x1_t)
+ { __builtin_aarch64_fmaxdf (vget_lane_f64 (__a, 0),
+ vget_lane_f64 (__b, 0)) };
+}
+
__extension__ static __inline float32x4_t __attribute__ ((__always_inline__))
vmaxnmq_f32 (float32x4_t __a, float32x4_t __b)
{
@@ -17824,6 +17840,14 @@ vmin_f32 (float32x2_t __a, float32x2_t __b)
return __builtin_aarch64_smin_nanv2sf (__a, __b);
}
+__extension__ static __inline float64x1_t __attribute__ ((__always_inline__))
+vmin_f64 (float64x1_t __a, float64x1_t __b)
+{
+ return (float64x1_t)
+ { __builtin_aarch64_smin_nandf (vget_lane_f64 (__a, 0),
+ vget_lane_f64 (__b, 0)) };
+}
+
__extension__ static __inline int8x8_t __attribute__ ((__always_inline__))
vmin_s8 (int8x8_t __a, int8x8_t __b)
{
@@ -17922,6 +17946,14 @@ vminnm_f32 (float32x2_t __a, float32x2_t __b)
return __builtin_aarch64_fminv2sf (__a, __b);
}
+__extension__ static __inline float64x1_t __attribute__ ((__always_inline__))
+vminnm_f64 (float64x1_t __a, float64x1_t __b)
+{
+ return (float64x1_t)
+ { __builtin_aarch64_fmindf (vget_lane_f64 (__a, 0),
+ vget_lane_f64 (__b, 0)) };
+}
+
__extension__ static __inline float32x4_t __attribute__ ((__always_inline__))
vminnmq_f32 (float32x4_t __a, float32x4_t __b)
{
diff --git a/gcc/config/aarch64/iterators.md b/gcc/config/aarch64/iterators.md
index 91e2e64..f35d8d7 100644
--- a/gcc/config/aarch64/iterators.md
+++ b/gcc/config/aarch64/iterators.md
@@ -1019,9 +1019,8 @@
(define_int_iterator ADDSUBHN2 [UNSPEC_ADDHN2 UNSPEC_RADDHN2
UNSPEC_SUBHN2 UNSPEC_RSUBHN2])
-(define_int_iterator FMAXMIN_UNS [UNSPEC_FMAX UNSPEC_FMIN])
-
-(define_int_iterator FMAXMIN [UNSPEC_FMAXNM UNSPEC_FMINNM])
+(define_int_iterator FMAXMIN_UNS [UNSPEC_FMAX UNSPEC_FMIN
+ UNSPEC_FMAXNM UNSPEC_FMINNM])
(define_int_iterator VQDMULH [UNSPEC_SQDMULH UNSPEC_SQRDMULH])
@@ -1105,7 +1104,9 @@
(UNSPEC_FMAXV "smax_nan")
(UNSPEC_FMIN "smin_nan")
(UNSPEC_FMINNMV "smin")
- (UNSPEC_FMINV "smin_nan")])
+ (UNSPEC_FMINV "smin_nan")
+ (UNSPEC_FMAXNM "fmax")
+ (UNSPEC_FMINNM "fmin")])
(define_int_attr maxmin_uns_op [(UNSPEC_UMAXV "umax")
(UNSPEC_UMINV "umin")
@@ -1116,13 +1117,9 @@
(UNSPEC_FMAXV "fmax")
(UNSPEC_FMIN "fmin")
(UNSPEC_FMINNMV "fminnm")
- (UNSPEC_FMINV "fmin")])
-
-(define_int_attr fmaxmin [(UNSPEC_FMAXNM "fmax")
- (UNSPEC_FMINNM "fmin")])
-
-(define_int_attr fmaxmin_op [(UNSPEC_FMAXNM "fmaxnm")
- (UNSPEC_FMINNM "fminnm")])
+ (UNSPEC_FMINV "fmin")
+ (UNSPEC_FMAXNM "fmaxnm")
+ (UNSPEC_FMINNM "fminnm")])
(define_int_attr sur [(UNSPEC_SHADD "s") (UNSPEC_UHADD "u")
(UNSPEC_SRHADD "sr") (UNSPEC_URHADD "ur")
diff --git a/gcc/testsuite/gcc.target/aarch64/simd/vminmaxnm_1.c b/gcc/testsuite/gcc.target/aarch64/simd/vminmaxnm_1.c
index 96608eb..192bad9 100644
--- a/gcc/testsuite/gcc.target/aarch64/simd/vminmaxnm_1.c
+++ b/gcc/testsuite/gcc.target/aarch64/simd/vminmaxnm_1.c
@@ -1,4 +1,4 @@
-/* Test the `v[min|max]nm{q}_f*' AArch64 SIMD intrinsic. */
+/* Test the `v[min|max]{nm}{q}_f*' AArch64 SIMD intrinsic. */
/* { dg-do run } */
/* { dg-options "-O2" } */
@@ -18,6 +18,7 @@ extern void abort ();
int
main (int argc, char **argv)
{
+ /* v{min|max}nm_f32 normal. */
float32x2_t f32x2_input1 = vdup_n_f32 (-1.0);
float32x2_t f32x2_input2 = vdup_n_f32 (0.0);
float32x2_t f32x2_exp_minnm = vdup_n_f32 (-1.0);
@@ -28,6 +29,7 @@ main (int argc, char **argv)
CHECK (uint32_t, 2, f32x2_ret_minnm, f32x2_exp_minnm);
CHECK (uint32_t, 2, f32x2_ret_maxnm, f32x2_exp_maxnm);
+ /* v{min|max}nm_f32 NaN. */
f32x2_input1 = vdup_n_f32 (__builtin_nanf (""));
f32x2_input2 = vdup_n_f32 (1.0);
f32x2_exp_minnm = vdup_n_f32 (1.0);
@@ -38,6 +40,7 @@ main (int argc, char **argv)
CHECK (uint32_t, 2, f32x2_ret_minnm, f32x2_exp_minnm);
CHECK (uint32_t, 2, f32x2_ret_maxnm, f32x2_exp_maxnm);
+ /* v{min|max}nmq_f32 normal. */
float32x4_t f32x4_input1 = vdupq_n_f32 (-1024.0);
float32x4_t f32x4_input2 = vdupq_n_f32 (77.0);
float32x4_t f32x4_exp_minnm = vdupq_n_f32 (-1024.0);
@@ -48,6 +51,7 @@ main (int argc, char **argv)
CHECK (uint32_t, 4, f32x4_ret_minnm, f32x4_exp_minnm);
CHECK (uint32_t, 4, f32x4_ret_maxnm, f32x4_exp_maxnm);
+ /* v{min|max}nmq_f32 NaN. */
f32x4_input1 = vdupq_n_f32 (-__builtin_nanf (""));
f32x4_input2 = vdupq_n_f32 (-1.0);
f32x4_exp_minnm = vdupq_n_f32 (-1.0);
@@ -58,16 +62,57 @@ main (int argc, char **argv)
CHECK (uint32_t, 4, f32x4_ret_minnm, f32x4_exp_minnm);
CHECK (uint32_t, 4, f32x4_ret_maxnm, f32x4_exp_maxnm);
+ /* v{min|max}nm_f64 normal. */
+ float64x1_t f64x1_input1 = vdup_n_f64 (1.23);
+ float64x1_t f64x1_input2 = vdup_n_f64 (4.56);
+ float64x1_t f64x1_exp_minnm = vdup_n_f64 (1.23);
+ float64x1_t f64x1_exp_maxnm = vdup_n_f64 (4.56);
+ float64x1_t f64x1_ret_minnm = vminnm_f64 (f64x1_input1, f64x1_input2);
+ float64x1_t f64x1_ret_maxnm = vmaxnm_f64 (f64x1_input1, f64x1_input2);
+ CHECK (uint64_t, 1, f64x1_ret_minnm, f64x1_exp_minnm);
+ CHECK (uint64_t, 1, f64x1_ret_maxnm, f64x1_exp_maxnm);
+
+ /* v{min|max}_f64 normal. */
+ float64x1_t f64x1_exp_min = vdup_n_f64 (1.23);
+ float64x1_t f64x1_exp_max = vdup_n_f64 (4.56);
+ float64x1_t f64x1_ret_min = vmin_f64 (f64x1_input1, f64x1_input2);
+ float64x1_t f64x1_ret_max = vmax_f64 (f64x1_input1, f64x1_input2);
+ CHECK (uint64_t, 1, f64x1_ret_min, f64x1_exp_min);
+ CHECK (uint64_t, 1, f64x1_ret_max, f64x1_exp_max);
+
+ /* v{min|max}nmq_f64 normal. */
float64x2_t f64x2_input1 = vdupq_n_f64 (1.23);
float64x2_t f64x2_input2 = vdupq_n_f64 (4.56);
float64x2_t f64x2_exp_minnm = vdupq_n_f64 (1.23);
float64x2_t f64x2_exp_maxnm = vdupq_n_f64 (4.56);
float64x2_t f64x2_ret_minnm = vminnmq_f64 (f64x2_input1, f64x2_input2);
float64x2_t f64x2_ret_maxnm = vmaxnmq_f64 (f64x2_input1, f64x2_input2);
-
CHECK (uint64_t, 2, f64x2_ret_minnm, f64x2_exp_minnm);
CHECK (uint64_t, 2, f64x2_ret_maxnm, f64x2_exp_maxnm);
+ /* v{min|max}nm_f64 NaN. */
+ f64x1_input1 = vdup_n_f64 (-__builtin_nanf (""));
+ f64x1_input2 = vdup_n_f64 (1.0);
+ f64x1_exp_minnm = vdup_n_f64 (1.0);
+ f64x1_exp_maxnm = vdup_n_f64 (1.0);
+ f64x1_ret_minnm = vminnm_f64 (f64x1_input1, f64x1_input2);
+ f64x1_ret_maxnm = vmaxnm_f64 (f64x1_input1, f64x1_input2);
+
+ CHECK (uint64_t, 1, f64x1_ret_minnm, f64x1_exp_minnm);
+ CHECK (uint64_t, 1, f64x1_ret_maxnm, f64x1_exp_maxnm);
+
+ /* v{min|max}_f64 NaN. */
+ f64x1_input1 = vdup_n_f64 (-__builtin_nanf (""));
+ f64x1_input2 = vdup_n_f64 (1.0);
+ f64x1_exp_minnm = vdup_n_f64 (-__builtin_nanf (""));
+ f64x1_exp_maxnm = vdup_n_f64 (-__builtin_nanf (""));
+ f64x1_ret_minnm = vmin_f64 (f64x1_input1, f64x1_input2);
+ f64x1_ret_maxnm = vmax_f64 (f64x1_input1, f64x1_input2);
+
+ CHECK (uint64_t, 1, f64x1_ret_minnm, f64x1_exp_minnm);
+ CHECK (uint64_t, 1, f64x1_ret_maxnm, f64x1_exp_maxnm);
+
+ /* v{min|max}nmq_f64 NaN. */
f64x2_input1 = vdupq_n_f64 (-__builtin_nan (""));
f64x2_input2 = vdupq_n_f64 (1.0);
f64x2_exp_minnm = vdupq_n_f64 (1.0);
diff --git a/gcc/testsuite/gcc.target/aarch64/vminmaxnm.c b/gcc/testsuite/gcc.target/aarch64/vminmaxnm.c
new file mode 100644
index 0000000..bdaa564
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/vminmaxnm.c
@@ -0,0 +1,37 @@
+/* { dg-do compile } */
+/* { dg-options "-O2" } */
+
+#include "arm_neon.h"
+
+/* For each of these intrinsics, we map directly to an unspec in RTL.
+ We're just using the argument directly and returning the result, so we
+ can precisely specify the exact instruction pattern and register
+ allocations we expect. */
+
+float64x1_t
+test_vmaxnm_f64 (float64x1_t a, float64x1_t b)
+{
+ /* { dg-final { scan-assembler-times "fmaxnm\td0, d0, d1" 1 } } */
+ return vmaxnm_f64 (a, b);
+}
+
+float64x1_t
+test_vminnm_f64 (float64x1_t a, float64x1_t b)
+{
+ /* { dg-final { scan-assembler-times "fminnm\td0, d0, d1" 1 } } */
+ return vminnm_f64 (a, b);
+}
+
+float64x1_t
+test_vmax_f64 (float64x1_t a, float64x1_t b)
+{
+ /* { dg-final { scan-assembler-times "fmax\td0, d0, d1" 1 } } */
+ return vmax_f64 (a, b);
+}
+
+float64x1_t
+test_vmin_f64 (float64x1_t a, float64x1_t b)
+{
+ /* { dg-final { scan-assembler-times "fmin\td0, d0, d1" 1 } } */
+ return vmin_f64 (a, b);
+}
\ No newline at end of file