1 #!/usr/bin/env python 2 3 import sys, fileinput 4 5 err=0 6 7 # Giant associative set of builtin->intrinsic mappings where clang doesn't 8 # implement the builtin since the vector operation works by default. 9 10 repl_map = { 11 '__builtin_ia32_addps': '_mm_add_ps', 12 '__builtin_ia32_addsd': '_mm_add_sd', 13 '__builtin_ia32_addpd': '_mm_add_pd', 14 '__builtin_ia32_addss': '_mm_add_ss', 15 '__builtin_ia32_paddb128': '_mm_add_epi8', 16 '__builtin_ia32_paddw128': '_mm_add_epi16', 17 '__builtin_ia32_paddd128': '_mm_add_epi32', 18 '__builtin_ia32_paddq128': '_mm_add_epi64', 19 '__builtin_ia32_subps': '_mm_sub_ps', 20 '__builtin_ia32_subsd': '_mm_sub_sd', 21 '__builtin_ia32_subpd': '_mm_sub_pd', 22 '__builtin_ia32_subss': '_mm_sub_ss', 23 '__builtin_ia32_psubb128': '_mm_sub_epi8', 24 '__builtin_ia32_psubw128': '_mm_sub_epi16', 25 '__builtin_ia32_psubd128': '_mm_sub_epi32', 26 '__builtin_ia32_psubq128': '_mm_sub_epi64', 27 '__builtin_ia32_mulsd': '_mm_mul_sd', 28 '__builtin_ia32_mulpd': '_mm_mul_pd', 29 '__builtin_ia32_mulps': '_mm_mul_ps', 30 '__builtin_ia32_mulss': '_mm_mul_ss', 31 '__builtin_ia32_pmullw128': '_mm_mullo_epi16', 32 '__builtin_ia32_divsd': '_mm_div_sd', 33 '__builtin_ia32_divpd': '_mm_div_pd', 34 '__builtin_ia32_divps': '_mm_div_ps', 35 '__builtin_ia32_subss': '_mm_div_ss', 36 '__builtin_ia32_andpd': '_mm_and_pd', 37 '__builtin_ia32_andps': '_mm_and_ps', 38 '__builtin_ia32_pand128': '_mm_and_si128', 39 '__builtin_ia32_andnpd': '_mm_andnot_pd', 40 '__builtin_ia32_andnps': '_mm_andnot_ps', 41 '__builtin_ia32_pandn128': '_mm_andnot_si128', 42 '__builtin_ia32_orpd': '_mm_or_pd', 43 '__builtin_ia32_orps': '_mm_or_ps', 44 '__builtin_ia32_por128': '_mm_or_si128', 45 '__builtin_ia32_xorpd': '_mm_xor_pd', 46 '__builtin_ia32_xorps': '_mm_xor_ps', 47 '__builtin_ia32_pxor128': '_mm_xor_si128', 48 '__builtin_ia32_cvtps2dq': '_mm_cvtps_epi32', 49 '__builtin_ia32_cvtsd2ss': '_mm_cvtsd_ss', 50 '__builtin_ia32_cvtsi2sd': '_mm_cvtsi32_sd', 51 '__builtin_ia32_cvtss2sd': '_mm_cvtss_sd', 52 '__builtin_ia32_cvttsd2si': '_mm_cvttsd_si32', 53 '__builtin_ia32_vec_ext_v2df': '_mm_cvtsd_f64', 54 '__builtin_ia32_loadhpd': '_mm_loadh_pd', 55 '__builtin_ia32_loadlpd': '_mm_loadl_pd', 56 '__builtin_ia32_loadlv4si': '_mm_loadl_epi64', 57 '__builtin_ia32_cmpeqps': '_mm_cmpeq_ps', 58 '__builtin_ia32_cmpltps': '_mm_cmplt_ps', 59 '__builtin_ia32_cmpleps': '_mm_cmple_ps', 60 '__builtin_ia32_cmpgtps': '_mm_cmpgt_ps', 61 '__builtin_ia32_cmpgeps': '_mm_cmpge_ps', 62 '__builtin_ia32_cmpunordps': '_mm_cmpunord_ps', 63 '__builtin_ia32_cmpneqps': '_mm_cmpneq_ps', 64 '__builtin_ia32_cmpnltps': '_mm_cmpnlt_ps', 65 '__builtin_ia32_cmpnleps': '_mm_cmpnle_ps', 66 '__builtin_ia32_cmpngtps': '_mm_cmpngt_ps', 67 '__builtin_ia32_cmpordps': '_mm_cmpord_ps', 68 '__builtin_ia32_cmpeqss': '_mm_cmpeq_ss', 69 '__builtin_ia32_cmpltss': '_mm_cmplt_ss', 70 '__builtin_ia32_cmpless': '_mm_cmple_ss', 71 '__builtin_ia32_cmpunordss': '_mm_cmpunord_ss', 72 '__builtin_ia32_cmpneqss': '_mm_cmpneq_ss', 73 '__builtin_ia32_cmpnltss': '_mm_cmpnlt_ss', 74 '__builtin_ia32_cmpnless': '_mm_cmpnle_ss', 75 '__builtin_ia32_cmpngtss': '_mm_cmpngt_ss', 76 '__builtin_ia32_cmpngess': '_mm_cmpnge_ss', 77 '__builtin_ia32_cmpordss': '_mm_cmpord_ss', 78 '__builtin_ia32_movss': '_mm_move_ss', 79 '__builtin_ia32_movsd': '_mm_move_sd', 80 '__builtin_ia32_movhlps': '_mm_movehl_ps', 81 '__builtin_ia32_movlhps': '_mm_movelh_ps', 82 '__builtin_ia32_movqv4si': '_mm_move_epi64', 83 '__builtin_ia32_unpckhps': '_mm_unpackhi_ps', 84 '__builtin_ia32_unpckhpd': '_mm_unpackhi_pd', 85 '__builtin_ia32_punpckhbw128': '_mm_unpackhi_epi8', 86 '__builtin_ia32_punpckhwd128': '_mm_unpackhi_epi16', 87 '__builtin_ia32_punpckhdq128': '_mm_unpackhi_epi32', 88 '__builtin_ia32_punpckhqdq128': '_mm_unpackhi_epi64', 89 '__builtin_ia32_unpcklps': '_mm_unpacklo_ps', 90 '__builtin_ia32_unpcklpd': '_mm_unpacklo_pd', 91 '__builtin_ia32_punpcklbw128': '_mm_unpacklo_epi8', 92 '__builtin_ia32_punpcklwd128': '_mm_unpacklo_epi16', 93 '__builtin_ia32_punpckldq128': '_mm_unpacklo_epi32', 94 '__builtin_ia32_punpcklqdq128': '_mm_unpacklo_epi64', 95 '__builtin_ia32_cmpeqpd': '_mm_cmpeq_pd', 96 '__builtin_ia32_cmpltpd': '_mm_cmplt_pd', 97 '__builtin_ia32_cmplepd': '_mm_cmple_pd', 98 '__builtin_ia32_cmpgtpd': '_mm_cmpgt_pd', 99 '__builtin_ia32_cmpgepd': '_mm_cmpge_pd', 100 '__builtin_ia32_cmpunordpd': '_mm_cmpunord_pd', 101 '__builtin_ia32_cmpneqpd': '_mm_cmpneq_pd', 102 '__builtin_ia32_cmpnltpd': '_mm_cmpnlt_pd', 103 '__builtin_ia32_cmpnlepd': '_mm_cmpnle_pd', 104 '__builtin_ia32_cmpngtpd': '_mm_cmpngt_pd', 105 '__builtin_ia32_cmpngepd': '_mm_cmpnge_pd', 106 '__builtin_ia32_cmpordpd': '_mm_cmpord_pd', 107 '__builtin_ia32_cmpeqsd': '_mm_cmpeq_sd', 108 '__builtin_ia32_cmpltsd': '_mm_cmplt_sd', 109 '__builtin_ia32_cmplesd': '_mm_cmple_sd', 110 '__builtin_ia32_cmpunordsd': '_mm_cmpunord_sd', 111 '__builtin_ia32_cmpneqsd': '_mm_cmpneq_sd', 112 '__builtin_ia32_cmpnltsd': '_mm_cmpnlt_sd', 113 '__builtin_ia32_cmpnlesd': '_mm_cmpnle_sd', 114 '__builtin_ia32_cmpordsd': '_mm_cmpord_sd', 115 '__builtin_ia32_cvtsi642ss': '_mm_cvtsi64_ss', 116 '__builtin_ia32_cvttss2si64': '_mm_cvtss_si64', 117 '__builtin_ia32_shufps': '_mm_shuffle_ps', 118 '__builtin_ia32_shufpd': '_mm_shuffle_pd', 119 '__builtin_ia32_pshufhw': '_mm_shufflehi_epi16', 120 '__builtin_ia32_pshuflw': '_mm_shufflelo_epi16', 121 '__builtin_ia32_pshufd': '_mm_shuffle_epi32', 122 '__builtin_ia32_movshdup': '_mm_movehdup_ps', 123 '__builtin_ia32_movsldup': '_mm_moveldup_ps', 124 '__builtin_ia32_maxps': '_mm_max_ps', 125 '__builtin_ia32_pslldi128': '_mm_slli_epi32', 126 '__builtin_ia32_vec_set_v16qi': '_mm_insert_epi8', 127 '__builtin_ia32_vec_set_v8hi': '_mm_insert_epi16', 128 '__builtin_ia32_vec_set_v4si': '_mm_insert_epi32', 129 '__builtin_ia32_vec_set_v2di': '_mm_insert_epi64', 130 '__builtin_ia32_vec_set_v4hi': '_mm_insert_pi16', 131 '__builtin_ia32_vec_ext_v16qi': '_mm_extract_epi8', 132 '__builtin_ia32_vec_ext_v8hi': '_mm_extract_epi16', 133 '__builtin_ia32_vec_ext_v4si': '_mm_extract_epi32', 134 '__builtin_ia32_vec_ext_v2di': '_mm_extract_epi64', 135 '__builtin_ia32_vec_ext_v4hi': '_mm_extract_pi16', 136 '__builtin_ia32_vec_ext_v4sf': '_mm_extract_ps' 137 } 138 139 # Special unhandled cases: 140 # __builtin_ia32_vec_ext_*(__P, idx) -> _mm_store_sd/_mm_storeh_pd 141 # depending on index. No abstract insert/extract for these oddly. 142 unhandled = [ 143 '__builtin_ia32_vec_ext_v2df', 144 '__builtin_ia32_vec_ext_v2si', 145 ] 146 147 def report_repl(builtin, repl): 148 sys.stderr.write("%s:%d: x86 builtin %s used, replaced with %s\n" % (fileinput.filename(), fileinput.filelineno(), builtin, repl)) 149 150 def report_cant(builtin): 151 sys.stderr.write("%s:%d: x86 builtin %s used, too many replacements\n" % (fileinput.filename(), fileinput.filelineno(), builtin)) 152 153 for line in fileinput.input(inplace=1): 154 for builtin, repl in repl_map.iteritems(): 155 if builtin in line: 156 line = line.replace(builtin, repl) 157 report_repl(builtin, repl) 158 for unh in unhandled: 159 if unh in line: 160 report_cant(unh) 161 sys.stdout.write(line) 162 163 sys.exit(err) 164
