clang/www/builtins.py

   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
   8 # doesn't implement the builtin. (Either because the vector operation
   9 # works without a builtin, or for other reasons.)
  10
  11 repl_map = {
  12     "__builtin_ia32_addps": "_mm_add_ps",
  13     "__builtin_ia32_addsd": "_mm_add_sd",
  14     "__builtin_ia32_addpd": "_mm_add_pd",
  15     "__builtin_ia32_addss": "_mm_add_ss",
  16     "__builtin_ia32_paddb128": "_mm_add_epi8",
  17     "__builtin_ia32_paddw128": "_mm_add_epi16",
  18     "__builtin_ia32_paddd128": "_mm_add_epi32",
  19     "__builtin_ia32_paddq128": "_mm_add_epi64",
  20     "__builtin_ia32_subps": "_mm_sub_ps",
  21     "__builtin_ia32_subsd": "_mm_sub_sd",
  22     "__builtin_ia32_subpd": "_mm_sub_pd",
  23     "__builtin_ia32_subss": "_mm_sub_ss",
  24     "__builtin_ia32_psubb128": "_mm_sub_epi8",
  25     "__builtin_ia32_psubw128": "_mm_sub_epi16",
  26     "__builtin_ia32_psubd128": "_mm_sub_epi32",
  27     "__builtin_ia32_psubq128": "_mm_sub_epi64",
  28     "__builtin_ia32_mulsd": "_mm_mul_sd",
  29     "__builtin_ia32_mulpd": "_mm_mul_pd",
  30     "__builtin_ia32_mulps": "_mm_mul_ps",
  31     "__builtin_ia32_mulss": "_mm_mul_ss",
  32     "__builtin_ia32_pmullw128": "_mm_mullo_epi16",
  33     "__builtin_ia32_divsd": "_mm_div_sd",
  34     "__builtin_ia32_divpd": "_mm_div_pd",
  35     "__builtin_ia32_divps": "_mm_div_ps",
  36     "__builtin_ia32_subss": "_mm_div_ss",
  37     "__builtin_ia32_andpd": "_mm_and_pd",
  38     "__builtin_ia32_andps": "_mm_and_ps",
  39     "__builtin_ia32_pand128": "_mm_and_si128",
  40     "__builtin_ia32_andnpd": "_mm_andnot_pd",
  41     "__builtin_ia32_andnps": "_mm_andnot_ps",
  42     "__builtin_ia32_pandn128": "_mm_andnot_si128",
  43     "__builtin_ia32_orpd": "_mm_or_pd",
  44     "__builtin_ia32_orps": "_mm_or_ps",
  45     "__builtin_ia32_por128": "_mm_or_si128",
  46     "__builtin_ia32_xorpd": "_mm_xor_pd",
  47     "__builtin_ia32_xorps": "_mm_xor_ps",
  48     "__builtin_ia32_pxor128": "_mm_xor_si128",
  49     "__builtin_ia32_cvtps2dq": "_mm_cvtps_epi32",
  50     "__builtin_ia32_cvtsd2ss": "_mm_cvtsd_ss",
  51     "__builtin_ia32_cvtsi2sd": "_mm_cvtsi32_sd",
  52     "__builtin_ia32_cvtss2sd": "_mm_cvtss_sd",
  53     "__builtin_ia32_cvttsd2si": "_mm_cvttsd_si32",
  54     "__builtin_ia32_vec_ext_v2df": "_mm_cvtsd_f64",
  55     "__builtin_ia32_loadhpd": "_mm_loadh_pd",
  56     "__builtin_ia32_loadlpd": "_mm_loadl_pd",
  57     "__builtin_ia32_loadlv4si": "_mm_loadl_epi64",
  58     "__builtin_ia32_cmpeqps": "_mm_cmpeq_ps",
  59     "__builtin_ia32_cmpltps": "_mm_cmplt_ps",
  60     "__builtin_ia32_cmpleps": "_mm_cmple_ps",
  61     "__builtin_ia32_cmpgtps": "_mm_cmpgt_ps",
  62     "__builtin_ia32_cmpgeps": "_mm_cmpge_ps",
  63     "__builtin_ia32_cmpunordps": "_mm_cmpunord_ps",
  64     "__builtin_ia32_cmpneqps": "_mm_cmpneq_ps",
  65     "__builtin_ia32_cmpnltps": "_mm_cmpnlt_ps",
  66     "__builtin_ia32_cmpnleps": "_mm_cmpnle_ps",
  67     "__builtin_ia32_cmpngtps": "_mm_cmpngt_ps",
  68     "__builtin_ia32_cmpordps": "_mm_cmpord_ps",
  69     "__builtin_ia32_cmpeqss": "_mm_cmpeq_ss",
  70     "__builtin_ia32_cmpltss": "_mm_cmplt_ss",
  71     "__builtin_ia32_cmpless": "_mm_cmple_ss",
  72     "__builtin_ia32_cmpunordss": "_mm_cmpunord_ss",
  73     "__builtin_ia32_cmpneqss": "_mm_cmpneq_ss",
  74     "__builtin_ia32_cmpnltss": "_mm_cmpnlt_ss",
  75     "__builtin_ia32_cmpnless": "_mm_cmpnle_ss",
  76     "__builtin_ia32_cmpngtss": "_mm_cmpngt_ss",
  77     "__builtin_ia32_cmpngess": "_mm_cmpnge_ss",
  78     "__builtin_ia32_cmpordss": "_mm_cmpord_ss",
  79     "__builtin_ia32_movss": "_mm_move_ss",
  80     "__builtin_ia32_movsd": "_mm_move_sd",
  81     "__builtin_ia32_movhlps": "_mm_movehl_ps",
  82     "__builtin_ia32_movlhps": "_mm_movelh_ps",
  83     "__builtin_ia32_movqv4si": "_mm_move_epi64",
  84     "__builtin_ia32_unpckhps": "_mm_unpackhi_ps",
  85     "__builtin_ia32_unpckhpd": "_mm_unpackhi_pd",
  86     "__builtin_ia32_punpckhbw128": "_mm_unpackhi_epi8",
  87     "__builtin_ia32_punpckhwd128": "_mm_unpackhi_epi16",
  88     "__builtin_ia32_punpckhdq128": "_mm_unpackhi_epi32",
  89     "__builtin_ia32_punpckhqdq128": "_mm_unpackhi_epi64",
  90     "__builtin_ia32_unpcklps": "_mm_unpacklo_ps",
  91     "__builtin_ia32_unpcklpd": "_mm_unpacklo_pd",
  92     "__builtin_ia32_punpcklbw128": "_mm_unpacklo_epi8",
  93     "__builtin_ia32_punpcklwd128": "_mm_unpacklo_epi16",
  94     "__builtin_ia32_punpckldq128": "_mm_unpacklo_epi32",
  95     "__builtin_ia32_punpcklqdq128": "_mm_unpacklo_epi64",
  96     "__builtin_ia32_cmpeqpd": "_mm_cmpeq_pd",
  97     "__builtin_ia32_cmpltpd": "_mm_cmplt_pd",
  98     "__builtin_ia32_cmplepd": "_mm_cmple_pd",
  99     "__builtin_ia32_cmpgtpd": "_mm_cmpgt_pd",
 100     "__builtin_ia32_cmpgepd": "_mm_cmpge_pd",
 101     "__builtin_ia32_cmpunordpd": "_mm_cmpunord_pd",
 102     "__builtin_ia32_cmpneqpd": "_mm_cmpneq_pd",
 103     "__builtin_ia32_cmpnltpd": "_mm_cmpnlt_pd",
 104     "__builtin_ia32_cmpnlepd": "_mm_cmpnle_pd",
 105     "__builtin_ia32_cmpngtpd": "_mm_cmpngt_pd",
 106     "__builtin_ia32_cmpngepd": "_mm_cmpnge_pd",
 107     "__builtin_ia32_cmpordpd": "_mm_cmpord_pd",
 108     "__builtin_ia32_cmpeqsd": "_mm_cmpeq_sd",
 109     "__builtin_ia32_cmpltsd": "_mm_cmplt_sd",
 110     "__builtin_ia32_cmplesd": "_mm_cmple_sd",
 111     "__builtin_ia32_cmpunordsd": "_mm_cmpunord_sd",
 112     "__builtin_ia32_cmpneqsd": "_mm_cmpneq_sd",
 113     "__builtin_ia32_cmpnltsd": "_mm_cmpnlt_sd",
 114     "__builtin_ia32_cmpnlesd": "_mm_cmpnle_sd",
 115     "__builtin_ia32_cmpordsd": "_mm_cmpord_sd",
 116     "__builtin_ia32_cvtsi642ss": "_mm_cvtsi64_ss",
 117     "__builtin_ia32_cvttss2si64": "_mm_cvtss_si64",
 118     "__builtin_ia32_shufps": "_mm_shuffle_ps",
 119     "__builtin_ia32_shufpd": "_mm_shuffle_pd",
 120     "__builtin_ia32_pshufhw": "_mm_shufflehi_epi16",
 121     "__builtin_ia32_pshuflw": "_mm_shufflelo_epi16",
 122     "__builtin_ia32_pshufd": "_mm_shuffle_epi32",
 123     "__builtin_ia32_movshdup": "_mm_movehdup_ps",
 124     "__builtin_ia32_movsldup": "_mm_moveldup_ps",
 125     "__builtin_ia32_maxps": "_mm_max_ps",
 126     "__builtin_ia32_pslldi128": "_mm_slli_epi32",
 127     "__builtin_ia32_vec_set_v16qi": "_mm_insert_epi8",
 128     "__builtin_ia32_vec_set_v8hi": "_mm_insert_epi16",
 129     "__builtin_ia32_vec_set_v4si": "_mm_insert_epi32",
 130     "__builtin_ia32_vec_set_v2di": "_mm_insert_epi64",
 131     "__builtin_ia32_vec_set_v4hi": "_mm_insert_pi16",
 132     "__builtin_ia32_vec_ext_v16qi": "_mm_extract_epi8",
 133     "__builtin_ia32_vec_ext_v8hi": "_mm_extract_epi16",
 134     "__builtin_ia32_vec_ext_v4si": "_mm_extract_epi32",
 135     "__builtin_ia32_vec_ext_v2di": "_mm_extract_epi64",
 136     "__builtin_ia32_vec_ext_v4hi": "_mm_extract_pi16",
 137     "__builtin_ia32_vec_ext_v4sf": "_mm_extract_ps",
 138     # Removed MMX builtins
 139     "__builtin_ia32_paddb": "_mm_add_pi8",
 140     "__builtin_ia32_paddw": "_mm_add_pi16",
 141     "__builtin_ia32_paddd": "_mm_add_pi32",
 142     "__builtin_ia32_paddsb": "_mm_adds_pi8",
 143     "__builtin_ia32_paddsw": "_mm_adds_pi16",
 144     "__builtin_ia32_paddusb": "_mm_adds_pu8",
 145     "__builtin_ia32_paddusw": "_mm_adds_pu16",
 146     "__builtin_ia32_psubb": "_mm_sub_pi8",
 147     "__builtin_ia32_psubw": "_mm_sub_pi16",
 148     "__builtin_ia32_psubd": "_mm_sub_pi32",
 149     "__builtin_ia32_psubsb": "_mm_subs_pi8",
 150     "__builtin_ia32_psubsw": "_mm_subs_pi16",
 151     "__builtin_ia32_psubusb": "_mm_subs_pu8",
 152     "__builtin_ia32_psubusw": "_mm_subs_pu16",
 153     "__builtin_ia32_pmulhw": "_mm_mulhi_pi16",
 154     "__builtin_ia32_pmullw": "_mm_mullo_pi16",
 155     "__builtin_ia32_pmaddwd": "_mm_madd_pi16",
 156     "__builtin_ia32_pand": "_mm_and_si64",
 157     "__builtin_ia32_pandn": "_mm_andnot_si64",
 158     "__builtin_ia32_por": "_mm_or_si64",
 159     "__builtin_ia32_pxor": "_mm_xor_si64",
 160     "__builtin_ia32_psllw": "_mm_sll_pi16",
 161     "__builtin_ia32_pslld": "_mm_sll_pi32",
 162     "__builtin_ia32_psllq": "_mm_sll_si64",
 163     "__builtin_ia32_psrlw": "_mm_srl_pi16",
 164     "__builtin_ia32_psrld": "_mm_srl_pi32",
 165     "__builtin_ia32_psrlq": "_mm_srl_si64",
 166     "__builtin_ia32_psraw": "_mm_sra_pi16",
 167     "__builtin_ia32_psrad": "_mm_sra_pi32",
 168     "__builtin_ia32_psllwi": "_mm_slli_pi16",
 169     "__builtin_ia32_pslldi": "_mm_slli_pi32",
 170     "__builtin_ia32_psllqi": "_mm_slli_si64",
 171     "__builtin_ia32_psrlwi": "_mm_srli_pi16",
 172     "__builtin_ia32_psrldi": "_mm_srli_pi32",
 173     "__builtin_ia32_psrlqi": "_mm_srli_si64",
 174     "__builtin_ia32_psrawi": "_mm_srai_pi16",
 175     "__builtin_ia32_psradi": "_mm_srai_pi32",
 176     "__builtin_ia32_packsswb": "_mm_packs_pi16",
 177     "__builtin_ia32_packssdw": "_mm_packs_pi32",
 178     "__builtin_ia32_packuswb": "_mm_packs_pu16",
 179     "__builtin_ia32_punpckhbw": "_mm_unpackhi_pi8",
 180     "__builtin_ia32_punpckhwd": "_mm_unpackhi_pi16",
 181     "__builtin_ia32_punpckhdq": "_mm_unpackhi_pi32",
 182     "__builtin_ia32_punpcklbw": "_mm_unpacklo_pi8",
 183     "__builtin_ia32_punpcklwd": "_mm_unpacklo_pi16",
 184     "__builtin_ia32_punpckldq": "_mm_unpacklo_pi32",
 185     "__builtin_ia32_pcmpeqb": "_mm_cmpeq_pi8",
 186     "__builtin_ia32_pcmpeqw": "_mm_cmpeq_pi16",
 187     "__builtin_ia32_pcmpeqd": "_mm_cmpeq_pi32",
 188     "__builtin_ia32_pcmpgtb": "_mm_cmpgt_pi8",
 189     "__builtin_ia32_pcmpgtw": "_mm_cmpgt_pi16",
 190     "__builtin_ia32_pcmpgtd": "_mm_cmpgt_pi32",
 191     "__builtin_ia32_maskmovq": "_mm_maskmove_si64",
 192     "__builtin_ia32_movntq": "_mm_stream_pi",
 193     "__builtin_ia32_vec_init_v2si": "_mm_setr_pi32",
 194     "__builtin_ia32_vec_init_v4hi": "_mm_setr_pi16",
 195     "__builtin_ia32_vec_init_v8qi": "_mm_setr_pi8",
 196     "__builtin_ia32_cvtpi2ps": "_mm_cvtpi32_ps",
 197     "__builtin_ia32_cvtps2pi": "_mm_cvtps_pi32",
 198     "__builtin_ia32_cvttps2pi": "_mm_cvttps_pi32",
 199     "__builtin_ia32_pavgb": "_mm_avg_pu8",
 200     "__builtin_ia32_pavgw": "_mm_avg_pu16",
 201     "__builtin_ia32_pmaxsw": "_mm_max_pi16",
 202     "__builtin_ia32_pmaxub": "_mm_max_pu8",
 203     "__builtin_ia32_pminsw": "_mm_min_pi16",
 204     "__builtin_ia32_pminub": "_mm_min_pu8",
 205     "__builtin_ia32_pmovmskb": "_mm_movemask_pi8",
 206     "__builtin_ia32_pmulhuw": "_mm_mulhi_pu16",
 207     "__builtin_ia32_psadbw": "_mm_sad_pu8",
 208     "__builtin_ia32_pshufw": "_mm_shuffle_pi16",
 209     "__builtin_ia32_cvtpd2pi": "_mm_cvtpd_pi32",
 210     "__builtin_ia32_cvtpi2pd": "_mm_cvtpi32_pd",
 211     "__builtin_ia32_cvttpd2pi": "_mm_cvttpd_pi32",
 212     "__builtin_ia32_paddq": "_mm_add_si64",
 213     "__builtin_ia32_pmuludq": "_mm_mul_su32",
 214     "__builtin_ia32_psubq": "_mm_sub_si64",
 215     "__builtin_ia32_pabsb": "_mm_abs_pi8",
 216     "__builtin_ia32_pabsd": "_mm_abs_pi32",
 217     "__builtin_ia32_pabsw": "_mm_abs_pi16",
 218     "__builtin_ia32_palignr": "_mm_alignr_pi8",
 219     "__builtin_ia32_phaddd": "_mm_hadd_pi32",
 220     "__builtin_ia32_phaddsw": "_mm_hadds_pi16",
 221     "__builtin_ia32_phaddw": "_mm_hadd_pi16",
 222     "__builtin_ia32_phsubd": "_mm_hsub_pi32",
 223     "__builtin_ia32_phsubsw": "_mm_hsubs_pi16",
 224     "__builtin_ia32_phsubw": "_mm_hsub_pi16",
 225     "__builtin_ia32_pmaddubsw": "_mm_maddubs_pi16",
 226     "__builtin_ia32_pmulhrsw": "_mm_mulhrs_pi16",
 227     "__builtin_ia32_pshufb": "_mm_shuffle_pi8",
 228     "__builtin_ia32_psignw": "_mm_sign_pi16",
 229     "__builtin_ia32_psignb": "_mm_sign_pi8",
 230     "__builtin_ia32_psignd": "_mm_sign_pi32",
 231 }
 232
 233 # Special unhandled cases:
 234 #   __builtin_ia32_vec_ext_*(__P, idx) -> _mm_store_sd/_mm_storeh_pd
 235 #     depending on index. No abstract insert/extract for these oddly.
 236 unhandled = [
 237     "__builtin_ia32_vec_ext_v2df",
 238     "__builtin_ia32_vec_ext_v2si",
 239 ]
 240
 241
 242 def report_repl(builtin, repl):
 243     sys.stderr.write(
 244         "%s:%d: x86 builtin %s used, replaced with %s\n"
 245         % (fileinput.filename(), fileinput.filelineno(), builtin, repl)
 246     )
 247
 248
 249 def report_cant(builtin):
 250     sys.stderr.write(
 251         "%s:%d: x86 builtin %s used, too many replacements\n"
 252         % (fileinput.filename(), fileinput.filelineno(), builtin)
 253     )
 254
 255
 256 for line in fileinput.input(inplace=1):
 257     for builtin, repl in repl_map.items():
 258         if builtin in line:
 259             line = line.replace(builtin, repl)
 260             report_repl(builtin, repl)
 261     for unh in unhandled:
 262         if unh in line:
 263             report_cant(unh)
 264     sys.stdout.write(line)
 265
 266 sys.exit(err)