diff mbox

[12/15,AArch64] Add vcvt(_high)?_f32_f16 intrinsics, with BE RTL fix

Message ID 1440499622-24989-1-git-send-email-alan.lawrence@arm.com
State New
Headers show

Commit Message

Alan Lawrence Aug. 25, 2015, 10:47 a.m. UTC 
James Greenhalgh wrote:
>>
>> -  VAR1 (UNOP, vec_unpacks_hi_, 10, v4sf)
>> +  VAR2 (UNOP, vec_unpacks_hi_, 10, v4sf, v8hf)
>
> Should this not use the appropriate "BUILTIN_..." iterator?

Indeed; BUILTIN_VQ_HSF it is.

>>    VAR1 (BINOP, float_truncate_hi_, 0, v4sf)
>>    VAR1 (BINOP, float_truncate_hi_, 0, v8hf)

I could also use BUILTIN_VQ_HSF here (these two were added in a previous patch, before the VQ_HSF iterator was introduced). However, that goes against the principle that we should use the same iterator as the pattern (the pattern uses the <Vdbl> attribute of the VDF iterator), so I'm not sure whether that would be preferable (i.e. as a separate patch)?

>> -  VAR1 (UNOP, float_extend_lo_, 0, v2df)
>> +  VAR2 (UNOP, float_extend_lo_, 0, v2df, v4sf)
>
> Likewise.

Similarly, the required iterator does not exist, as float_extend_lo_ is named
after the <Vwide> attribute of the VDF iterator. The nearest equivalents I can
see use two VAR1's rather than a VAR2, so I've updated the patch to do that too.

OK with those two changes? (patch attached and bootstrapped+check-gcc on aarch64-none-linux-gnu)

Thanks, Alan
---
 gcc/config/aarch64/aarch64-simd-builtins.def |  3 +-
 gcc/config/aarch64/aarch64-simd.md           | 63 ++++++++++++++++++----------
 gcc/config/aarch64/arm_neon.h                | 16 +++++--
 gcc/config/aarch64/iterators.md              | 18 +++++---
 4 files changed, 69 insertions(+), 31 deletions(-)

Comments

James Greenhalgh Sept. 4, 2015, 9:50 a.m. UTC | #1 
On Tue, Aug 25, 2015 at 11:47:02AM +0100, Alan Lawrence wrote:
> James Greenhalgh wrote:
> >>
> >> -  VAR1 (UNOP, vec_unpacks_hi_, 10, v4sf)
> >> +  VAR2 (UNOP, vec_unpacks_hi_, 10, v4sf, v8hf)
> >
> > Should this not use the appropriate "BUILTIN_..." iterator?
> 
> Indeed; BUILTIN_VQ_HSF it is.
> 
> >>    VAR1 (BINOP, float_truncate_hi_, 0, v4sf)
> >>    VAR1 (BINOP, float_truncate_hi_, 0, v8hf)
> 
> I could also use BUILTIN_VQ_HSF here (these two were added in a previous
> patch, before the VQ_HSF iterator was introduced). However, that goes against
> the principle that we should use the same iterator as the pattern (the
> pattern uses the <Vdbl> attribute of the VDF iterator), so I'm not sure
> whether that would be preferable (i.e. as a separate patch)?
> 
> >> -  VAR1 (UNOP, float_extend_lo_, 0, v2df)
> >> +  VAR2 (UNOP, float_extend_lo_, 0, v2df, v4sf)
> >
> > Likewise.
> 
> Similarly, the required iterator does not exist, as float_extend_lo_ is named
> after the <Vwide> attribute of the VDF iterator. The nearest equivalents I can
> see use two VAR1's rather than a VAR2, so I've updated the patch to do that too.
> 
> OK with those two changes? (patch attached and bootstrapped+check-gcc on
> aarch64-none-linux-gnu)

OK.

Thanks,
James
diff mbox

Patch

diff --git a/gcc/config/aarch64/aarch64-simd-builtins.def b/gcc/config/aarch64/aarch64-simd-builtins.def
index c5b46aa..2c13cfb 100644
--- a/gcc/config/aarch64/aarch64-simd-builtins.def
+++ b/gcc/config/aarch64/aarch64-simd-builtins.def
@@ -361,11 +361,12 @@ 
   BUILTIN_VSDQ_I_DI (UNOP, abs, 0)
   BUILTIN_VDQF (UNOP, abs, 2)
 
-  VAR1 (UNOP, vec_unpacks_hi_, 10, v4sf)
+  BUILTIN_VQ_HSF (UNOP, vec_unpacks_hi_, 10)
   VAR1 (BINOP, float_truncate_hi_, 0, v4sf)
   VAR1 (BINOP, float_truncate_hi_, 0, v8hf)
 
   VAR1 (UNOP, float_extend_lo_, 0, v2df)
+  VAR1 (UNOP, float_extend_lo_,  0, v4sf)
   BUILTIN_VDF (UNOP, float_truncate_lo_, 0)
 
   /* Implemented by aarch64_ld1<VALL_F16:mode>.  */
diff --git a/gcc/config/aarch64/aarch64-simd.md b/gcc/config/aarch64/aarch64-simd.md
index f8754cd..160acf9 100644
--- a/gcc/config/aarch64/aarch64-simd.md
+++ b/gcc/config/aarch64/aarch64-simd.md
@@ -1691,36 +1691,57 @@ 
 
 ;; Float widening operations.
 
-(define_insn "vec_unpacks_lo_v4sf"
-  [(set (match_operand:V2DF 0 "register_operand" "=w")
-	(float_extend:V2DF
-	  (vec_select:V2SF
-	    (match_operand:V4SF 1 "register_operand" "w")
-	    (parallel [(const_int 0) (const_int 1)])
-	  )))]
+(define_insn "aarch64_simd_vec_unpacks_lo_<mode>"
+  [(set (match_operand:<VWIDE> 0 "register_operand" "=w")
+        (float_extend:<VWIDE> (vec_select:<VHALF>
+			       (match_operand:VQ_HSF 1 "register_operand" "w")
+			       (match_operand:VQ_HSF 2 "vect_par_cnst_lo_half" "")
+			    )))]
   "TARGET_SIMD"
-  "fcvtl\\t%0.2d, %1.2s"
+  "fcvtl\\t%0.<Vwtype>, %1.<Vhalftype>"
   [(set_attr "type" "neon_fp_cvt_widen_s")]
 )
 
-(define_insn "aarch64_float_extend_lo_v2df"
-  [(set (match_operand:V2DF 0 "register_operand" "=w")
-	(float_extend:V2DF
-	  (match_operand:V2SF 1 "register_operand" "w")))]
+(define_expand "vec_unpacks_lo_<mode>"
+  [(match_operand:<VWIDE> 0 "register_operand" "")
+   (match_operand:VQ_HSF 1 "register_operand" "")]
   "TARGET_SIMD"
-  "fcvtl\\t%0.2d, %1.2s"
+  {
+    rtx p = aarch64_simd_vect_par_cnst_half (<MODE>mode, false);
+    emit_insn (gen_aarch64_simd_vec_unpacks_lo_<mode> (operands[0],
+						       operands[1], p));
+    DONE;
+  }
+)
+
+(define_insn "aarch64_simd_vec_unpacks_hi_<mode>"
+  [(set (match_operand:<VWIDE> 0 "register_operand" "=w")
+        (float_extend:<VWIDE> (vec_select:<VHALF>
+			       (match_operand:VQ_HSF 1 "register_operand" "w")
+			       (match_operand:VQ_HSF 2 "vect_par_cnst_hi_half" "")
+			    )))]
+  "TARGET_SIMD"
+  "fcvtl2\\t%0.<Vwtype>, %1.<Vtype>"
   [(set_attr "type" "neon_fp_cvt_widen_s")]
 )
 
-(define_insn "vec_unpacks_hi_v4sf"
-  [(set (match_operand:V2DF 0 "register_operand" "=w")
-	(float_extend:V2DF
-	  (vec_select:V2SF
-	    (match_operand:V4SF 1 "register_operand" "w")
-	    (parallel [(const_int 2) (const_int 3)])
-	  )))]
+(define_expand "vec_unpacks_hi_<mode>"
+  [(match_operand:<VWIDE> 0 "register_operand" "")
+   (match_operand:VQ_HSF 1 "register_operand" "")]
+  "TARGET_SIMD"
+  {
+    rtx p = aarch64_simd_vect_par_cnst_half (<MODE>mode, true);
+    emit_insn (gen_aarch64_simd_vec_unpacks_lo_<mode> (operands[0],
+						       operands[1], p));
+    DONE;
+  }
+)
+(define_insn "aarch64_float_extend_lo_<Vwide>"
+  [(set (match_operand:<VWIDE> 0 "register_operand" "=w")
+	(float_extend:<VWIDE>
+	  (match_operand:VDF 1 "register_operand" "w")))]
   "TARGET_SIMD"
-  "fcvtl2\\t%0.2d, %1.4s"
+  "fcvtl\\t%0<Vmwtype>, %1<Vmtype>"
   [(set_attr "type" "neon_fp_cvt_widen_s")]
 )
 
diff --git a/gcc/config/aarch64/arm_neon.h b/gcc/config/aarch64/arm_neon.h
index c7bfadf..91ada61 100644
--- a/gcc/config/aarch64/arm_neon.h
+++ b/gcc/config/aarch64/arm_neon.h
@@ -6025,10 +6025,6 @@  vaddlvq_u32 (uint32x4_t a)
        result;                                                          \
      })
 
-/* vcvt_f32_f16 not supported */
-
-/* vcvt_high_f32_f16 not supported */
-
 #define vcvt_n_f32_s32(a, b)                                            \
   __extension__                                                         \
     ({                                                                  \
@@ -13436,6 +13432,12 @@  vcvt_high_f32_f64 (float32x2_t __a, float64x2_t __b)
 
 /* vcvt (float -> double).  */
 
+__extension__ static __inline float32x4_t __attribute__ ((__always_inline__))
+vcvt_f32_f16 (float16x4_t __a)
+{
+  return __builtin_aarch64_float_extend_lo_v4sf (__a);
+}
+
 __extension__ static __inline float64x2_t __attribute__ ((__always_inline__))
 vcvt_f64_f32 (float32x2_t __a)
 {
@@ -13443,6 +13445,12 @@  vcvt_f64_f32 (float32x2_t __a)
   return __builtin_aarch64_float_extend_lo_v2df (__a);
 }
 
+__extension__ static __inline float32x4_t __attribute__ ((__always_inline__))
+vcvt_high_f32_f16 (float16x8_t __a)
+{
+  return __builtin_aarch64_vec_unpacks_hi_v8hf (__a);
+}
+
 __extension__ static __inline float64x2_t __attribute__ ((__always_inline__))
 vcvt_high_f64_f32 (float32x4_t __a)
 {
diff --git a/gcc/config/aarch64/iterators.md b/gcc/config/aarch64/iterators.md
index 3c1c30f..c2af1de 100644
--- a/gcc/config/aarch64/iterators.md
+++ b/gcc/config/aarch64/iterators.md
@@ -91,6 +91,9 @@ 
 ;; Vector single Float modes.
 (define_mode_iterator VDQSF [V2SF V4SF])
 
+;; Quad vector Float modes with half/single elements.
+(define_mode_iterator VQ_HSF [V8HF V4SF])
+
 ;; Modes suitable to use as the return type of a vcond expression.
 (define_mode_iterator VDQF_COND [V2SF V2SI V4SF V4SI V2DF V2DI])
 
@@ -492,14 +495,18 @@ 
 			 (V2SI "V2DI") (V16QI "V8HI") 
 			 (V8HI "V4SI") (V4SI "V2DI")
 			 (HI "SI")     (SI "DI")
+			 (V8HF "V4SF") (V4SF "V2DF")
 			 (V4HF "V4SF") (V2SF "V2DF")]
-
 )
 
-;; Widened mode register suffixes for VD_BHSI/VQW.
+;; Widened modes of vector modes, lowercase
+(define_mode_attr Vwide [(V2SF "v2df") (V4HF "v4sf")])
+
+;; Widened mode register suffixes for VD_BHSI/VQW/VQ_HSF.
 (define_mode_attr Vwtype [(V8QI "8h") (V4HI "4s")
 			  (V2SI "2d") (V16QI "8h") 
-			  (V8HI "4s") (V4SI "2d")])
+			  (V8HI "4s") (V4SI "2d")
+			  (V8HF "4s") (V4SF "2d")])
 
 ;; Widened mode register suffixes for VDW/VQW.
 (define_mode_attr Vmwtype [(V8QI ".8h") (V4HI ".4s")
@@ -508,9 +515,10 @@ 
 			   (V4HF ".4s") (V2SF ".2d")
 			   (SI   "")    (HI   "")])
 
-;; Lower part register suffixes for VQW.
+;; Lower part register suffixes for VQW/VQ_HSF.
 (define_mode_attr Vhalftype [(V16QI "8b") (V8HI "4h")
-			     (V4SI "2s")])
+			     (V4SI "2s") (V8HF "4h")
+			     (V4SF "2s")])
 
 ;; Define corresponding core/FP element mode for each vector mode.
 (define_mode_attr vw   [(V8QI "w") (V16QI "w")