SVE vectorization

To continue my vectorization spree, here is the version with ARM SVE instructions. Tested in QEMU 5.2. Compiler support is still a bit wonky -- GCC 10 generates very bloated assembly and produces incorrect code for non-native vector sizes, while Clang 11 misses some obvious optimizations.

pystencils/backends/arm_instruction_sets.py

@@ -65,12 +65,14 @@ def get_vector_instruction_set_arm(data_type='double', instruction_set='neon'):
     result = dict()
     result['bytes'] = bitwidth // 8
 
+    predicate = f'{prefix}whilelt_b{bits[data_type]}(0, {width})'
+    int_predicate = f'{prefix}whilelt_b{bits["int"]}(0, {intwidth})'
+
     for intrinsic_id, function_shortcut in base_names.items():
         function_shortcut = function_shortcut.strip()
         name = function_shortcut[:function_shortcut.index('[')]
 
-        arg_string = get_argument_string(function_shortcut, first=f'{prefix}ptrue_b{bits[data_type]}()'
-                                         if prefix == 'sv' else '')
+        arg_string = get_argument_string(function_shortcut, first=predicate if prefix == 'sv' else '')
         if prefix == 'sv' and not name.startswith('ld') and not name.startswith('st') and not name.startswith(cmp):
             undef = '_x'
         else:
@@ -86,20 +88,19 @@ def get_vector_instruction_set_arm(data_type='double', instruction_set='neon'):
         result['makeVecConstInt'] = f'svdup_s{bits["int"]}' + '({0})'
         result['makeVecIndex'] = f'svindex_s{bits["int"]}' + '({0}, {1})'
 
-        result['+int'] = f"svadd_s{bits['int']}_x(svptrue_b{bits['int']}(), " + "{0}, {1})"
+        result['+int'] = f"svadd_s{bits['int']}_x({int_predicate}, " + "{0}, {1})"
 
-        attr = f' __attribute__((arm_sve_vector_bits({bitwidth})))'
-        result[data_type] = f'svfloat{bits[data_type]}_t{attr}'
-        result['int'] = f'svint{bits["int"]}_t{attr}'
-        result['bool'] = f'svbool_t{attr}'
+        result[data_type] = f'svfloat{bits[data_type]}_st'
+        result['int'] = f'svint{bits["int"]}_st'
+        result['bool'] = f'svbool_st'
 
         result['headers'] = ['<arm_sve.h>', '"arm_neon_helpers.h"']
 
-        result['&'] = f'svand_b_z(svptrue_b{bits[data_type]}(),' + ' {0}, {1})'
-        result['|'] = f'svorr_b_z(svptrue_b{bits[data_type]}(),' + ' {0}, {1})'
+        result['&'] = f'svand_b_z({predicate},' + ' {0}, {1})'
+        result['|'] = f'svorr_b_z({predicate},' + ' {0}, {1})'
         result['blendv'] = f'svsel_f{bits[data_type]}' + '({2}, {1}, {0})'
-        result['any'] = f'svptest_any(svptrue_b{bits[data_type]}(), {{0}}) > 0'
-        result['all'] = f'svcntp_b{bits[data_type]}(svptrue_b{bits[data_type]}(), {{0}}) == {width}'
+        result['any'] = f'svptest_any({predicate}, {{0}})'
+        result['all'] = f'svcntp_b{bits[data_type]}({predicate}, {{0}}) == {width}'
 
         result['compile_flags'] = [f'-msve-vector-bits={bitwidth}']
     else: