;;
;; target-neon-common.ll
;;
;;  Copyright(c) 2013-2015 Google, Inc.
;;
;;  All rights reserved.
;;
;;  Redistribution and use in source and binary forms, with or without
;;  modification, are permitted provided that the following conditions are
;;  met:
;;
;;    * Redistributions of source code must retain the above copyright
;;      notice, this list of conditions and the following disclaimer.
;;
;;    * Redistributions in binary form must reproduce the above copyright
;;      notice, this list of conditions and the following disclaimer in the
;;      documentation and/or other materials provided with the distribution.
;;
;;    * Neither the name of Matt Pharr nor the names of its
;;      contributors may be used to endorse or promote products derived from
;;      this software without specific prior written permission.
;;
;;
;;   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
;;   IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
;;   TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
;;   PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
;;   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
;;   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
;;   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
;;   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
;;   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
;;   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
;;   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.  

target datalayout = "e-p:32:32:32-S32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f16:16:16-f32:32:32-f64:32:64-f128:128:128-v64:32:64-v128:32:128-a0:0:64-n32"

stdlib_core()
scans()
reduce_equal(WIDTH)
rdrand_decls()
define_shuffles()
aossoa()
ctlztz()

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; half conversion routines

declare <4 x i16> @llvm.arm.neon.vcvtfp2hf(<4 x float>) nounwind readnone
declare <4 x float> @llvm.arm.neon.vcvthf2fp(<4 x i16>) nounwind readnone

define float @__half_to_float_uniform(i16 %v) nounwind readnone alwaysinline {
  %v1 = bitcast i16 %v to <1 x i16>
  %vec = shufflevector <1 x i16> %v1, <1 x i16> undef, 
           <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %h = call <4 x float> @llvm.arm.neon.vcvthf2fp(<4 x i16> %vec)
  %r = extractelement <4 x float> %h, i32 0
  ret float %r
}

define i16 @__float_to_half_uniform(float %v) nounwind readnone alwaysinline {
  %v1 = bitcast float %v to <1 x float>
  %vec = shufflevector <1 x float> %v1, <1 x float> undef, 
           <4 x i32> <i32 0, i32 0, i32 0, i32 0>
  %h = call <4 x i16> @llvm.arm.neon.vcvtfp2hf(<4 x float> %vec)
  %r = extractelement <4 x i16> %h, i32 0
  ret i16 %r
}

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; math

declare i32 @llvm.arm.get.fpscr() nounwind
declare void @llvm.arm.set.fpscr(i32) nounwind

define void @__fastmath() nounwind alwaysinline {
  %x = call i32 @llvm.arm.get.fpscr()
  ; Turn on FTZ (bit 24) and default NaN (bit 25)
  %y = or i32 %x, 50331648
  call void @llvm.arm.set.fpscr(i32 %y)
  ret void
}

;; round/floor/ceil

;; FIXME: grabbed these from the sse2 target, which does not have native
;; instructions for these.  Is there a better approach for NEON?

define float @__round_uniform_float(float) nounwind readonly alwaysinline {
  %float_to_int_bitcast.i.i.i.i = bitcast float %0 to i32
  %bitop.i.i = and i32 %float_to_int_bitcast.i.i.i.i, -2147483648
  %bitop.i = xor i32 %bitop.i.i, %float_to_int_bitcast.i.i.i.i
  %int_to_float_bitcast.i.i40.i = bitcast i32 %bitop.i to float
  %binop.i = fadd float %int_to_float_bitcast.i.i40.i, 8.388608e+06
  %binop21.i = fadd float %binop.i, -8.388608e+06
  %float_to_int_bitcast.i.i.i = bitcast float %binop21.i to i32
  %bitop31.i = xor i32 %float_to_int_bitcast.i.i.i, %bitop.i.i
  %int_to_float_bitcast.i.i.i = bitcast i32 %bitop31.i to float
  ret float %int_to_float_bitcast.i.i.i
}

define float @__floor_uniform_float(float) nounwind readonly alwaysinline {
  %calltmp.i = tail call float @__round_uniform_float(float %0) nounwind
  %bincmp.i = fcmp ogt float %calltmp.i, %0
  %selectexpr.i = sext i1 %bincmp.i to i32
  %bitop.i = and i32 %selectexpr.i, -1082130432
  %int_to_float_bitcast.i.i.i = bitcast i32 %bitop.i to float
  %binop.i = fadd float %calltmp.i, %int_to_float_bitcast.i.i.i
  ret float %binop.i
}

define float @__ceil_uniform_float(float) nounwind readonly alwaysinline {
  %calltmp.i = tail call float @__round_uniform_float(float %0) nounwind
  %bincmp.i = fcmp olt float %calltmp.i, %0
  %selectexpr.i = sext i1 %bincmp.i to i32
  %bitop.i = and i32 %selectexpr.i, 1065353216
  %int_to_float_bitcast.i.i.i = bitcast i32 %bitop.i to float
  %binop.i = fadd float %calltmp.i, %int_to_float_bitcast.i.i.i
  ret float %binop.i
}

;; FIXME: rounding doubles and double vectors needs to be implemented
declare double @__round_uniform_double(double) nounwind readnone 
declare double @__floor_uniform_double(double) nounwind readnone 
declare double @__ceil_uniform_double(double) nounwind readnone 

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; min/max

define float @__max_uniform_float(float, float) nounwind readnone alwaysinline {
  %cmp = fcmp ugt float %0, %1
  %r = select i1 %cmp, float %0, float %1
  ret float %r
}

define float @__min_uniform_float(float, float) nounwind readnone alwaysinline {
  %cmp = fcmp ult float %0, %1
  %r = select i1 %cmp, float %0, float %1
  ret float %r
}

define i32 @__min_uniform_int32(i32, i32) nounwind readnone alwaysinline {
  %cmp = icmp slt i32 %0, %1
  %r = select i1 %cmp, i32 %0, i32 %1
  ret i32 %r
}

define i32 @__max_uniform_int32(i32, i32) nounwind readnone alwaysinline {
  %cmp = icmp sgt i32 %0, %1
  %r = select i1 %cmp, i32 %0, i32 %1
  ret i32 %r
}

define i32 @__min_uniform_uint32(i32, i32) nounwind readnone alwaysinline {
  %cmp = icmp ult i32 %0, %1
  %r = select i1 %cmp, i32 %0, i32 %1
  ret i32 %r
}

define i32 @__max_uniform_uint32(i32, i32) nounwind readnone alwaysinline {
  %cmp = icmp ugt i32 %0, %1
  %r = select i1 %cmp, i32 %0, i32 %1
  ret i32 %r
}

define i64 @__min_uniform_int64(i64, i64) nounwind readnone alwaysinline {
  %cmp = icmp slt i64 %0, %1
  %r = select i1 %cmp, i64 %0, i64 %1
  ret i64 %r
}

define i64 @__max_uniform_int64(i64, i64) nounwind readnone alwaysinline {
  %cmp = icmp sgt i64 %0, %1
  %r = select i1 %cmp, i64 %0, i64 %1
  ret i64 %r
}

define i64 @__min_uniform_uint64(i64, i64) nounwind readnone alwaysinline {
  %cmp = icmp ult i64 %0, %1
  %r = select i1 %cmp, i64 %0, i64 %1
  ret i64 %r
}

define i64 @__max_uniform_uint64(i64, i64) nounwind readnone alwaysinline {
  %cmp = icmp ugt i64 %0, %1
  %r = select i1 %cmp, i64 %0, i64 %1
  ret i64 %r
}

define double @__min_uniform_double(double, double) nounwind readnone alwaysinline {
  %cmp = fcmp olt double %0, %1
  %r = select i1 %cmp, double %0, double %1
  ret double %r
}

define double @__max_uniform_double(double, double) nounwind readnone alwaysinline {
  %cmp = fcmp ogt double %0, %1
  %r = select i1 %cmp, double %0, double %1
  ret double %r
}

define <WIDTH x i64> @__min_varying_int64(<WIDTH x i64>, <WIDTH x i64>) nounwind readnone alwaysinline {
  %m = icmp slt <WIDTH x i64> %0, %1
  %r = select <WIDTH x i1> %m, <WIDTH x i64> %0, <WIDTH x i64> %1
  ret <WIDTH x i64> %r
}

define <WIDTH x i64> @__max_varying_int64(<WIDTH x i64>, <WIDTH x i64>) nounwind readnone alwaysinline {
  %m = icmp sgt <WIDTH x i64> %0, %1
  %r = select <WIDTH x i1> %m, <WIDTH x i64> %0, <WIDTH x i64> %1
  ret <WIDTH x i64> %r
}

define <WIDTH x i64> @__min_varying_uint64(<WIDTH x i64>, <WIDTH x i64>) nounwind readnone alwaysinline {
  %m = icmp ult <WIDTH x i64> %0, %1
  %r = select <WIDTH x i1> %m, <WIDTH x i64> %0, <WIDTH x i64> %1
  ret <WIDTH x i64> %r
}

define <WIDTH x i64> @__max_varying_uint64(<WIDTH x i64>, <WIDTH x i64>) nounwind readnone alwaysinline {
  %m = icmp ugt <WIDTH x i64> %0, %1
  %r = select <WIDTH x i1> %m, <WIDTH x i64> %0, <WIDTH x i64> %1
  ret <WIDTH x i64> %r
}

define <WIDTH x double> @__min_varying_double(<WIDTH x double>,
                                              <WIDTH x double>) nounwind readnone alwaysinline {
  %m = fcmp olt <WIDTH x double> %0, %1
  %r = select <WIDTH x i1> %m, <WIDTH x double> %0, <WIDTH x double> %1
  ret <WIDTH x double> %r
}

define <WIDTH x double> @__max_varying_double(<WIDTH x double>,
                                              <WIDTH x double>) nounwind readnone alwaysinline {
  %m = fcmp ogt <WIDTH x double> %0, %1
  %r = select <WIDTH x i1> %m, <WIDTH x double> %0, <WIDTH x double> %1
  ret <WIDTH x double> %r
}

;; sqrt/rsqrt/rcp

declare float @llvm.sqrt.f32(float)

define float @__sqrt_uniform_float(float) nounwind readnone alwaysinline {
  %r = call float @llvm.sqrt.f32(float %0)
  ret float %r
}

declare double @llvm.sqrt.f64(double)

define double @__sqrt_uniform_double(double) nounwind readnone alwaysinline {
  %r = call double @llvm.sqrt.f64(double %0)
  ret double %r
}

;; bit ops

declare i32 @llvm.ctpop.i32(i32) nounwind readnone
declare i64 @llvm.ctpop.i64(i64) nounwind readnone

define i32 @__popcnt_int32(i32) nounwind readnone alwaysinline {
  %v = call i32 @llvm.ctpop.i32(i32 %0)
  ret i32 %v
}

define i64 @__popcnt_int64(i64) nounwind readnone alwaysinline {
  %v = call i64 @llvm.ctpop.i64(i64 %0)
  ret i64 %v
}

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; unaligned loads/loads+broadcasts

masked_load(i8,  1)
masked_load(i16, 2)
masked_load(i32, 4)
masked_load(float, 4)
masked_load(i64, 8)
masked_load(double, 8)

gen_masked_store(i8)
gen_masked_store(i16)
gen_masked_store(i32)
gen_masked_store(i64)
masked_store_float_double()

define void @__masked_store_blend_i8(<WIDTH x i8>* nocapture %ptr, <WIDTH x i8> %new,
                                     <WIDTH x MASK> %mask) nounwind alwaysinline {
  %old = load PTR_OP_ARGS(`<WIDTH x i8> ')  %ptr
  %mask1 = trunc <WIDTH x MASK> %mask to <WIDTH x i1>
  %result = select <WIDTH x i1> %mask1, <WIDTH x i8> %new, <WIDTH x i8> %old
  store <WIDTH x i8> %result, <WIDTH x i8> * %ptr
  ret void
}

define void @__masked_store_blend_i16(<WIDTH x i16>* nocapture %ptr, <WIDTH x i16> %new, 
                                      <WIDTH x MASK> %mask) nounwind alwaysinline {
  %old = load PTR_OP_ARGS(`<WIDTH x i16> ')  %ptr
  %mask1 = trunc <WIDTH x MASK> %mask to <WIDTH x i1>
  %result = select <WIDTH x i1> %mask1, <WIDTH x i16> %new, <WIDTH x i16> %old
  store <WIDTH x i16> %result, <WIDTH x i16> * %ptr
  ret void
}

define void @__masked_store_blend_i32(<WIDTH x i32>* nocapture %ptr, <WIDTH x i32> %new, 
                                      <WIDTH x MASK> %mask) nounwind alwaysinline {
  %old = load PTR_OP_ARGS(`<WIDTH x i32> ')  %ptr
  %mask1 = trunc <WIDTH x MASK> %mask to <WIDTH x i1>
  %result = select <WIDTH x i1> %mask1, <WIDTH x i32> %new, <WIDTH x i32> %old
  store <WIDTH x i32> %result, <WIDTH x i32> * %ptr
  ret void
}

define void @__masked_store_blend_i64(<WIDTH x i64>* nocapture %ptr,
                            <WIDTH x i64> %new, <WIDTH x MASK> %mask) nounwind alwaysinline {
  %old = load PTR_OP_ARGS(`<WIDTH x i64> ')  %ptr
  %mask1 = trunc <WIDTH x MASK> %mask to <WIDTH x i1>
  %result = select <WIDTH x i1> %mask1, <WIDTH x i64> %new, <WIDTH x i64> %old
  store <WIDTH x i64> %result, <WIDTH x i64> * %ptr
  ret void
}

;; yuck.  We need declarations of these, even though we shouldnt ever
;; actually generate calls to them for the NEON target...


include(`svml.m4')
svml_stubs(float,f,WIDTH)
svml_stubs(double,d,WIDTH)

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; gather

gen_gather_factored(i8)
gen_gather_factored(i16)
gen_gather_factored(i32)
gen_gather_factored(float)
gen_gather_factored(i64)
gen_gather_factored(double)

gen_scatter(i8)
gen_scatter(i16)
gen_scatter(i32)
gen_scatter(float)
gen_scatter(i64)
gen_scatter(double)

packed_load_and_store(4)

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; prefetch

define_prefetches()
declare_nvptx()