added shift

builtins/util-nvptx.m4

@@ -764,97 +764,6 @@ define <WIDTH x $1> @__broadcast_$1(<WIDTH x $1>, i32) nounwind readnone alwaysi
   ret <WIDTH x $1> %broadcast
 }
 
-define <WIDTH x $1> @__rotate_$1(<WIDTH x $1>, i32) nounwind readnone alwaysinline {
-  %isc = call i1 @__is_compile_time_constant_uniform_int32(i32 %1)
-  br i1 %isc, label %is_const, label %not_const
-
-is_const:
-  ; though verbose, this turms into tight code if %1 is a constant
-forloop(i, 0, eval(WIDTH-1), `  
-  %delta_`'i = add i32 %1, i
-  %delta_clamped_`'i = and i32 %delta_`'i, eval(WIDTH-1)
-  %v_`'i = extractelement <WIDTH x $1> %0, i32 %delta_clamped_`'i')
-
-  %ret_0 = insertelement <WIDTH x $1> undef, $1 %v_0, i32 0
-forloop(i, 1, eval(WIDTH-1), `  %ret_`'i = insertelement <WIDTH x $1> %ret_`'eval(i-1), $1 %v_`'i, i32 i
-')
-  ret <WIDTH x $1> %ret_`'eval(WIDTH-1)
-
-not_const:
-  ; store two instances of the vector into memory
-  %ptr = alloca <WIDTH x $1>, i32 2
-  %ptr0 = getelementptr <WIDTH x $1> * %ptr, i32 0
-  store <WIDTH x $1> %0, <WIDTH x $1> * %ptr0
-  %ptr1 = getelementptr <WIDTH x $1> * %ptr, i32 1
-  store <WIDTH x $1> %0, <WIDTH x $1> * %ptr1
-
-  ; compute offset in [0,vectorwidth-1], then index into the doubled-up vector
-  %offset = and i32 %1, eval(WIDTH-1)
-  %ptr_as_elt_array = bitcast <WIDTH x $1> * %ptr to [eval(2*WIDTH) x $1] *
-  %load_ptr = getelementptr [eval(2*WIDTH) x $1] * %ptr_as_elt_array, i32 0, i32 %offset
-  %load_ptr_vec = bitcast $1 * %load_ptr to <WIDTH x $1> *
-  %result = load <WIDTH x $1> * %load_ptr_vec, align $2
-  ret <WIDTH x $1> %result
-}
-
-define <WIDTH x $1> @__shift_$1(<WIDTH x $1>, i32) nounwind readnone alwaysinline {
-  %ptr = alloca <WIDTH x $1>, i32 3
-  %ptr0 = getelementptr <WIDTH x $1> * %ptr, i32 0
-  store <WIDTH x $1> zeroinitializer, <WIDTH x $1> * %ptr0
-  %ptr1 = getelementptr <WIDTH x $1> * %ptr, i32 1
-  store <WIDTH x $1> %0, <WIDTH x $1> * %ptr1
-  %ptr2 = getelementptr <WIDTH x $1> * %ptr, i32 2
-  store <WIDTH x $1> zeroinitializer, <WIDTH x $1> * %ptr2
-
-  %offset = add i32 %1, WIDTH
-  %ptr_as_elt_array = bitcast <WIDTH x $1> * %ptr to [eval(3*WIDTH) x $1] *
-  %load_ptr = getelementptr [eval(3*WIDTH) x $1] * %ptr_as_elt_array, i32 0, i32 %offset
-  %load_ptr_vec = bitcast $1 * %load_ptr to <WIDTH x $1> *
-  %result = load <WIDTH x $1> * %load_ptr_vec, align $2
-  ret <WIDTH x $1> %result
-}
-
-
-define <WIDTH x $1> @__shuffle2_$1(<WIDTH x $1>, <WIDTH x $1>, <WIDTH x i32>) nounwind readnone alwaysinline {
-  %v2 = shufflevector <WIDTH x $1> %0, <WIDTH x $1> %1, <eval(2*WIDTH) x i32> <
-      forloop(i, 0, eval(2*WIDTH-2), `i32 i, ') i32 eval(2*WIDTH-1)
-  >
-forloop(i, 0, eval(WIDTH-1), `  
-  %index_`'i = extractelement <WIDTH x i32> %2, i32 i')
-
-  %isc = call i1 @__is_compile_time_constant_varying_int32(<WIDTH x i32> %2)
-  br i1 %isc, label %is_const, label %not_const
-
-is_const:
-  ; extract from the requested lanes and insert into the result; LLVM turns
-  ; this into good code in the end
-forloop(i, 0, eval(WIDTH-1), `  
-  %v_`'i = extractelement <eval(2*WIDTH) x $1> %v2, i32 %index_`'i')
-
-  %ret_0 = insertelement <WIDTH x $1> undef, $1 %v_0, i32 0
-forloop(i, 1, eval(WIDTH-1), `  %ret_`'i = insertelement <WIDTH x $1> %ret_`'eval(i-1), $1 %v_`'i, i32 i
-')
-  ret <WIDTH x $1> %ret_`'eval(WIDTH-1)
-
-not_const:
-  ; otherwise store the two vectors onto the stack and then use the given
-  ; permutation vector to get indices into that array...
-  %ptr = alloca <eval(2*WIDTH) x $1>
-  store <eval(2*WIDTH) x $1> %v2, <eval(2*WIDTH) x $1> * %ptr
-  %baseptr = bitcast <eval(2*WIDTH) x $1> * %ptr to $1 *
-
-  %ptr_0 = getelementptr $1 * %baseptr, i32 %index_0
-  %val_0 = load $1 * %ptr_0
-  %result_0 = insertelement <WIDTH x $1> undef, $1 %val_0, i32 0
-
-forloop(i, 1, eval(WIDTH-1), `  
-  %ptr_`'i = getelementptr $1 * %baseptr, i32 %index_`'i
-  %val_`'i = load $1 * %ptr_`'i
-  %result_`'i = insertelement <WIDTH x $1> %result_`'eval(i-1), $1 %val_`'i, i32 i
-')
-
-  ret <WIDTH x $1> %result_`'eval(WIDTH-1)
-}
 ')
 
 define(`define_shuffles',`