Home | History | Annotate | Download | only in InstCombine
      1 ; RUN: opt < %s -instcombine -S | FileCheck %s
      2 
      3 ; testing-case "float fold(float a) { return 1.2f * a * 2.3f; }"
      4 ; 1.2f and 2.3f is supposed to be fold.
      5 define float @fold(float %a) {
      6   %mul = fmul fast float %a, 0x3FF3333340000000
      7   %mul1 = fmul fast float %mul, 0x4002666660000000
      8   ret float %mul1
      9 ; CHECK-LABEL: @fold(
     10 ; CHECK: fmul fast float %a, 0x4006147AE0000000
     11 }
     12 
     13 ; Same testing-case as the one used in fold() except that the operators have
     14 ; fixed FP mode.
     15 define float @notfold(float %a) {
     16 ; CHECK-LABEL: @notfold(
     17 ; CHECK: %mul = fmul fast float %a, 0x3FF3333340000000
     18   %mul = fmul fast float %a, 0x3FF3333340000000
     19   %mul1 = fmul float %mul, 0x4002666660000000
     20   ret float %mul1
     21 }
     22 
     23 define float @fold2(float %a) {
     24 ; CHECK-LABEL: @fold2(
     25 ; CHECK: fmul fast float %a, 0x4006147AE0000000
     26   %mul = fmul float %a, 0x3FF3333340000000
     27   %mul1 = fmul fast float %mul, 0x4002666660000000
     28   ret float %mul1
     29 }
     30 
     31 ; C * f1 + f1 = (C+1) * f1
     32 define double @fold3(double %f1) {
     33   %t1 = fmul fast double 2.000000e+00, %f1
     34   %t2 = fadd fast double %f1, %t1
     35   ret double %t2
     36 ; CHECK-LABEL: @fold3(
     37 ; CHECK: fmul fast double %f1, 3.000000e+00
     38 }
     39 
     40 ; (C1 - X) + (C2 - Y) => (C1+C2) - (X + Y)
     41 define float @fold4(float %f1, float %f2) {
     42   %sub = fsub float 4.000000e+00, %f1
     43   %sub1 = fsub float 5.000000e+00, %f2
     44   %add = fadd fast float %sub, %sub1
     45   ret float %add
     46 ; CHECK-LABEL: @fold4(
     47 ; CHECK: %1 = fadd fast float %f1, %f2
     48 ; CHECK: fsub fast float 9.000000e+00, %1
     49 }
     50 
     51 ; (X + C1) + C2 => X + (C1 + C2)
     52 define float @fold5(float %f1, float %f2) {
     53   %add = fadd float %f1, 4.000000e+00
     54   %add1 = fadd fast float %add, 5.000000e+00
     55   ret float %add1
     56 ; CHECK-LABEL: @fold5(
     57 ; CHECK: fadd fast float %f1, 9.000000e+00
     58 }
     59 
     60 ; (X + X) + X => 3.0 * X
     61 define float @fold6(float %f1) {
     62   %t1 = fadd fast float %f1, %f1
     63   %t2 = fadd fast float %f1, %t1
     64   ret float %t2
     65 ; CHECK-LABEL: @fold6(
     66 ; CHECK: fmul fast float %f1, 3.000000e+00
     67 }
     68 
     69 ; C1 * X + (X + X) = (C1 + 2) * X
     70 define float @fold7(float %f1) {
     71   %t1 = fmul fast float %f1, 5.000000e+00
     72   %t2 = fadd fast float %f1, %f1
     73   %t3 = fadd fast float %t1, %t2
     74   ret float %t3
     75 ; CHECK-LABEL: @fold7(
     76 ; CHECK: fmul fast float %f1, 7.000000e+00
     77 }
     78 
     79 ; (X + X) + (X + X) => 4.0 * X
     80 define float @fold8(float %f1) {
     81   %t1 = fadd fast float %f1, %f1
     82   %t2 = fadd fast float %f1, %f1
     83   %t3 = fadd fast float %t1, %t2
     84   ret float %t3
     85 ; CHECK: fold8
     86 ; CHECK: fmul fast float %f1, 4.000000e+00
     87 }
     88 
     89 ; X - (X + Y) => 0 - Y
     90 define float @fold9(float %f1, float %f2) {
     91   %t1 = fadd float %f1, %f2
     92   %t3 = fsub fast float %f1, %t1
     93   ret float %t3
     94 
     95 ; CHECK-LABEL: @fold9(
     96 ; CHECK: fsub fast float 0.000000e+00, %f2
     97 }
     98 
     99 ; Let C3 = C1 + C2. (f1 + C1) + (f2 + C2) => (f1 + f2) + C3 instead of
    100 ; "(f1 + C3) + f2" or "(f2 + C3) + f1". Placing constant-addend at the
    101 ; top of resulting simplified expression tree may potentially reveal some
    102 ; optimization opportunities in the super-expression trees.
    103 ;
    104 define float @fold10(float %f1, float %f2) {
    105   %t1 = fadd fast float 2.000000e+00, %f1
    106   %t2 = fsub fast float %f2, 3.000000e+00
    107   %t3 = fadd fast float %t1, %t2
    108   ret float %t3
    109 ; CHECK-LABEL: @fold10(
    110 ; CHECK: %t3 = fadd fast float %t2, -1.000000e+00
    111 ; CHECK: ret float %t3
    112 }
    113 
    114 ; once cause Crash/miscompilation
    115 define float @fail1(float %f1, float %f2) {
    116   %conv3 = fadd fast float %f1, -1.000000e+00
    117   %add = fadd fast float %conv3, %conv3
    118   %add2 = fadd fast float %add, %conv3
    119   ret float %add2
    120 ; CHECK-LABEL: @fail1(
    121 ; CHECK: ret
    122 }
    123 
    124 define double @fail2(double %f1, double %f2) {
    125   %t1 = fsub fast double %f1, %f2
    126   %t2 = fadd fast double %f1, %f2
    127   %t3 = fsub fast double %t1, %t2
    128   ret double %t3
    129 ; CHECK-LABEL: @fail2(
    130 ; CHECK: ret
    131 }
    132 
    133 ; c1 * x - x => (c1 - 1.0) * x
    134 define float @fold13(float %x) {
    135   %mul = fmul fast float %x, 7.000000e+00
    136   %sub = fsub fast float %mul, %x
    137   ret float %sub
    138 ; CHECK: fold13
    139 ; CHECK: fmul fast float %x, 6.000000e+00
    140 ; CHECK: ret
    141 }
    142 
    143 ; -x + y => y - x
    144 define float @fold14(float %x, float %y) {
    145   %neg = fsub fast float -0.0, %x
    146   %add = fadd fast float %neg, %y
    147   ret float %add
    148 ; CHECK: fold14
    149 ; CHECK: fsub fast float %y, %x
    150 ; CHECK: ret
    151 }
    152 
    153 ; x + -y => x - y
    154 define float @fold15(float %x, float %y) {
    155   %neg = fsub fast float -0.0, %y
    156   %add = fadd fast float %x, %neg
    157   ret float %add
    158 ; CHECK: fold15
    159 ; CHECK: fsub fast float %x, %y
    160 ; CHECK: ret
    161 }
    162 
    163 ; (select X+Y, X-Y) => X + (select Y, -Y)
    164 define float @fold16(float %x, float %y) {
    165   %cmp = fcmp ogt float %x, %y
    166   %plus = fadd fast float %x, %y
    167   %minus = fsub fast float %x, %y
    168   %r = select i1 %cmp, float %plus, float %minus
    169   ret float %r
    170 ; CHECK: fold16
    171 ; CHECK: fsub fast float
    172 ; CHECK: select
    173 ; CHECK: fadd fast float
    174 ; CHECK: ret
    175 }
    176 
    177 
    178 
    179 ; =========================================================================
    180 ;
    181 ;   Testing-cases about fmul begin
    182 ;
    183 ; =========================================================================
    184 
    185 ; ((X*C1) + C2) * C3 => (X * (C1*C3)) + (C2*C3) (i.e. distribution)
    186 define float @fmul_distribute1(float %f1) {
    187   %t1 = fmul float %f1, 6.0e+3
    188   %t2 = fadd float %t1, 2.0e+3
    189   %t3 = fmul fast float %t2, 5.0e+3
    190   ret float %t3
    191 ; CHECK-LABEL: @fmul_distribute1(
    192 ; CHECK: %1 = fmul fast float %f1, 3.000000e+07
    193 ; CHECK: %t3 = fadd fast float %1, 1.000000e+07
    194 }
    195 
    196 ; (X/C1 + C2) * C3 => X/(C1/C3) + C2*C3
    197 define double @fmul_distribute2(double %f1, double %f2) {
    198   %t1 = fdiv double %f1, 3.0e+0
    199   %t2 = fadd double %t1, 5.0e+1
    200   ; 0x10000000000000 = DBL_MIN
    201   %t3 = fmul fast double %t2, 0x10000000000000
    202   ret double %t3
    203 
    204 ; CHECK-LABEL: @fmul_distribute2(
    205 ; CHECK: %1 = fdiv fast double %f1, 0x7FE8000000000000
    206 ; CHECK: fadd fast double %1, 0x69000000000000
    207 }
    208 
    209 ; 5.0e-1 * DBL_MIN yields denormal, so "(f1*3.0 + 5.0e-1) * DBL_MIN" cannot
    210 ; be simplified into f1 * (3.0*DBL_MIN) + (5.0e-1*DBL_MIN)
    211 define double @fmul_distribute3(double %f1) {
    212   %t1 = fdiv double %f1, 3.0e+0
    213   %t2 = fadd double %t1, 5.0e-1
    214   %t3 = fmul fast double %t2, 0x10000000000000
    215   ret double %t3
    216 
    217 ; CHECK-LABEL: @fmul_distribute3(
    218 ; CHECK: fmul fast double %t2, 0x10000000000000
    219 }
    220 
    221 ; ((X*C1) + C2) * C3 => (X * (C1*C3)) + (C2*C3) (i.e. distribution)
    222 define float @fmul_distribute4(float %f1) {
    223   %t1 = fmul float %f1, 6.0e+3
    224   %t2 = fsub float 2.0e+3, %t1
    225   %t3 = fmul fast float %t2, 5.0e+3
    226   ret float %t3
    227 ; CHECK-LABEL: @fmul_distribute4(
    228 ; CHECK: %1 = fmul fast float %f1, 3.000000e+07
    229 ; CHECK: %t3 = fsub fast float 1.000000e+07, %1
    230 }
    231 
    232 ; C1/X * C2 => (C1*C2) / X
    233 define float @fmul2(float %f1) {
    234   %t1 = fdiv float 2.0e+3, %f1
    235   %t3 = fmul fast float %t1, 6.0e+3
    236   ret float %t3
    237 ; CHECK-LABEL: @fmul2(
    238 ; CHECK: fdiv fast float 1.200000e+07, %f1
    239 }
    240 
    241 ; X/C1 * C2 => X * (C2/C1) is disabled if X/C1 has multiple uses
    242 @fmul2_external = external global float
    243 define float @fmul2_disable(float %f1) {
    244   %div = fdiv fast float 1.000000e+00, %f1 
    245   store float %div, float* @fmul2_external
    246   %mul = fmul fast float %div, 2.000000e+00
    247   ret float %mul
    248 ; CHECK-LABEL: @fmul2_disable
    249 ; CHECK: store
    250 ; CHECK: fmul fast
    251 }
    252 
    253 ; X/C1 * C2 => X * (C2/C1) (if C2/C1 is normal Fp)
    254 define float @fmul3(float %f1, float %f2) {
    255   %t1 = fdiv float %f1, 2.0e+3
    256   %t3 = fmul fast float %t1, 6.0e+3
    257   ret float %t3
    258 ; CHECK-LABEL: @fmul3(
    259 ; CHECK: fmul fast float %f1, 3.000000e+00
    260 }
    261 
    262 define <4 x float> @fmul3_vec(<4 x float> %f1, <4 x float> %f2) {
    263   %t1 = fdiv <4 x float> %f1, <float 2.0e+3, float 3.0e+3, float 2.0e+3, float 1.0e+3>
    264   %t3 = fmul fast <4 x float> %t1, <float 6.0e+3, float 6.0e+3, float 2.0e+3, float 1.0e+3>
    265   ret <4 x float> %t3
    266 ; CHECK-LABEL: @fmul3_vec(
    267 ; CHECK: fmul fast <4 x float> %f1, <float 3.000000e+00, float 2.000000e+00, float 1.000000e+00, float 1.000000e+00>
    268 }
    269 
    270 ; Rule "X/C1 * C2 => X * (C2/C1) is not applicable if C2/C1 is either a special
    271 ; value of a denormal. The 0x3810000000000000 here take value FLT_MIN
    272 ;
    273 define float @fmul4(float %f1, float %f2) {
    274   %t1 = fdiv float %f1, 2.0e+3
    275   %t3 = fmul fast float %t1, 0x3810000000000000
    276   ret float %t3
    277 ; CHECK-LABEL: @fmul4(
    278 ; CHECK: fmul fast float %t1, 0x3810000000000000
    279 }
    280 
    281 ; X / C1 * C2 => X / (C2/C1) if  C1/C2 is either a special value of a denormal,
    282 ;  and C2/C1 is a normal value.
    283 ;
    284 define float @fmul5(float %f1, float %f2) {
    285   %t1 = fdiv float %f1, 3.0e+0
    286   %t3 = fmul fast float %t1, 0x3810000000000000
    287   ret float %t3
    288 ; CHECK-LABEL: @fmul5(
    289 ; CHECK: fdiv fast float %f1, 0x47E8000000000000
    290 }
    291 
    292 ; (X*Y) * X => (X*X) * Y
    293 define float @fmul6(float %f1, float %f2) {
    294   %mul = fmul float %f1, %f2
    295   %mul1 = fmul fast float %mul, %f1
    296   ret float %mul1
    297 ; CHECK-LABEL: @fmul6(
    298 ; CHECK: fmul fast float %f1, %f1
    299 }
    300 
    301 ; "(X*Y) * X => (X*X) * Y" is disabled if "X*Y" has multiple uses
    302 define float @fmul7(float %f1, float %f2) {
    303   %mul = fmul float %f1, %f2
    304   %mul1 = fmul fast float %mul, %f1
    305   %add = fadd float %mul1, %mul
    306   ret float %add
    307 ; CHECK-LABEL: @fmul7(
    308 ; CHECK: fmul fast float %mul, %f1
    309 }
    310 
    311 ; =========================================================================
    312 ;
    313 ;   Testing-cases about negation
    314 ;
    315 ; =========================================================================
    316 define float @fneg1(float %f1, float %f2) {
    317   %sub = fsub float -0.000000e+00, %f1
    318   %sub1 = fsub nsz float 0.000000e+00, %f2
    319   %mul = fmul float %sub, %sub1
    320   ret float %mul
    321 ; CHECK-LABEL: @fneg1(
    322 ; CHECK: fmul float %f1, %f2
    323 }
    324 
    325 ; =========================================================================
    326 ;
    327 ;   Testing-cases about div
    328 ;
    329 ; =========================================================================
    330 
    331 ; X/C1 / C2 => X * (1/(C2*C1))
    332 define float @fdiv1(float %x) {
    333   %div = fdiv float %x, 0x3FF3333340000000
    334   %div1 = fdiv fast float %div, 0x4002666660000000
    335   ret float %div1
    336 ; 0x3FF3333340000000 = 1.2f
    337 ; 0x4002666660000000 = 2.3f
    338 ; 0x3FD7303B60000000 = 0.36231884057971014492
    339 ; CHECK-LABEL: @fdiv1(
    340 ; CHECK: fmul fast float %x, 0x3FD7303B60000000
    341 }
    342 
    343 ; X*C1 / C2 => X * (C1/C2)
    344 define float @fdiv2(float %x) {
    345   %mul = fmul float %x, 0x3FF3333340000000
    346   %div1 = fdiv fast float %mul, 0x4002666660000000
    347   ret float %div1
    348 
    349 ; 0x3FF3333340000000 = 1.2f
    350 ; 0x4002666660000000 = 2.3f
    351 ; 0x3FE0B21660000000 = 0.52173918485641479492
    352 ; CHECK-LABEL: @fdiv2(
    353 ; CHECK: fmul fast float %x, 0x3FE0B21660000000
    354 }
    355 
    356 define <2 x float> @fdiv2_vec(<2 x float> %x) {
    357   %mul = fmul <2 x float> %x, <float 6.0, float 9.0>
    358   %div1 = fdiv fast <2 x float> %mul, <float 2.0, float 3.0>
    359   ret <2 x float> %div1
    360 
    361 ; CHECK-LABEL: @fdiv2_vec(
    362 ; CHECK: fmul fast <2 x float> %x, <float 3.000000e+00, float 3.000000e+00>
    363 }
    364 
    365 ; "X/C1 / C2 => X * (1/(C2*C1))" is disabled (for now) is C2/C1 is a denormal
    366 ;
    367 define float @fdiv3(float %x) {
    368   %div = fdiv float %x, 0x47EFFFFFE0000000
    369   %div1 = fdiv fast float %div, 0x4002666660000000
    370   ret float %div1
    371 ; CHECK-LABEL: @fdiv3(
    372 ; CHECK: fdiv float %x, 0x47EFFFFFE0000000
    373 }
    374 
    375 ; "X*C1 / C2 => X * (C1/C2)" is disabled if C1/C2 is a denormal
    376 define float @fdiv4(float %x) {
    377   %mul = fmul float %x, 0x47EFFFFFE0000000
    378   %div = fdiv float %mul, 0x3FC99999A0000000
    379   ret float %div
    380 ; CHECK-LABEL: @fdiv4(
    381 ; CHECK: fmul float %x, 0x47EFFFFFE0000000
    382 }
    383 
    384 ; (X/Y)/Z = > X/(Y*Z)
    385 define float @fdiv5(float %f1, float %f2, float %f3) {
    386   %t1 = fdiv float %f1, %f2
    387   %t2 = fdiv fast float %t1, %f3
    388   ret float %t2
    389 ; CHECK-LABEL: @fdiv5(
    390 ; CHECK: fmul float %f2, %f3
    391 }
    392 
    393 ; Z/(X/Y) = > (Z*Y)/X
    394 define float @fdiv6(float %f1, float %f2, float %f3) {
    395   %t1 = fdiv float %f1, %f2
    396   %t2 = fdiv fast float %f3, %t1
    397   ret float %t2
    398 ; CHECK-LABEL: @fdiv6(
    399 ; CHECK: fmul float %f3, %f2
    400 }
    401 
    402 ; C1/(X*C2) => (C1/C2) / X
    403 define float @fdiv7(float %x) {
    404   %t1 = fmul float %x, 3.0e0
    405   %t2 = fdiv fast float 15.0e0, %t1
    406   ret float %t2
    407 ; CHECK-LABEL: @fdiv7(
    408 ; CHECK: fdiv fast float 5.000000e+00, %x
    409 }
    410 
    411 ; C1/(X/C2) => (C1*C2) / X
    412 define float @fdiv8(float %x) {
    413   %t1 = fdiv float %x, 3.0e0
    414   %t2 = fdiv fast float 15.0e0, %t1
    415   ret float %t2
    416 ; CHECK-LABEL: @fdiv8(
    417 ; CHECK: fdiv fast float 4.500000e+01, %x
    418 }
    419 
    420 ; C1/(C2/X) => (C1/C2) * X
    421 define float @fdiv9(float %x) {
    422   %t1 = fdiv float 3.0e0, %x
    423   %t2 = fdiv fast float 15.0e0, %t1
    424   ret float %t2
    425 ; CHECK-LABEL: @fdiv9(
    426 ; CHECK: fmul fast float %x, 5.000000e+00
    427 }
    428 
    429 ; =========================================================================
    430 ;
    431 ;   Testing-cases about factorization
    432 ;
    433 ; =========================================================================
    434 ; x*z + y*z => (x+y) * z
    435 define float @fact_mul1(float %x, float %y, float %z) {
    436   %t1 = fmul fast float %x, %z
    437   %t2 = fmul fast float %y, %z
    438   %t3 = fadd fast float %t1, %t2
    439   ret float %t3
    440 ; CHECK-LABEL: @fact_mul1(
    441 ; CHECK: fmul fast float %1, %z
    442 }
    443 
    444 ; z*x + y*z => (x+y) * z
    445 define float @fact_mul2(float %x, float %y, float %z) {
    446   %t1 = fmul fast float %z, %x
    447   %t2 = fmul fast float %y, %z
    448   %t3 = fsub fast float %t1, %t2
    449   ret float %t3
    450 ; CHECK-LABEL: @fact_mul2(
    451 ; CHECK: fmul fast float %1, %z
    452 }
    453 
    454 ; z*x - z*y => (x-y) * z
    455 define float @fact_mul3(float %x, float %y, float %z) {
    456   %t2 = fmul fast float %z, %y
    457   %t1 = fmul fast float %z, %x
    458   %t3 = fsub fast float %t1, %t2
    459   ret float %t3
    460 ; CHECK-LABEL: @fact_mul3(
    461 ; CHECK: fmul fast float %1, %z
    462 }
    463 
    464 ; x*z - z*y => (x-y) * z
    465 define float @fact_mul4(float %x, float %y, float %z) {
    466   %t1 = fmul fast float %x, %z
    467   %t2 = fmul fast float %z, %y
    468   %t3 = fsub fast float %t1, %t2
    469   ret float %t3
    470 ; CHECK-LABEL: @fact_mul4(
    471 ; CHECK: fmul fast float %1, %z
    472 }
    473 
    474 ; x/y + x/z, no xform
    475 define float @fact_div1(float %x, float %y, float %z) {
    476   %t1 = fdiv fast float %x, %y
    477   %t2 = fdiv fast float %x, %z
    478   %t3 = fadd fast float %t1, %t2
    479   ret float %t3
    480 ; CHECK: fact_div1
    481 ; CHECK: fadd fast float %t1, %t2
    482 }
    483 
    484 ; x/y + z/x; no xform
    485 define float @fact_div2(float %x, float %y, float %z) {
    486   %t1 = fdiv fast float %x, %y
    487   %t2 = fdiv fast float %z, %x
    488   %t3 = fadd fast float %t1, %t2
    489   ret float %t3
    490 ; CHECK: fact_div2
    491 ; CHECK: fadd fast float %t1, %t2
    492 }
    493 
    494 ; y/x + z/x => (y+z)/x
    495 define float @fact_div3(float %x, float %y, float %z) {
    496   %t1 = fdiv fast float %y, %x
    497   %t2 = fdiv fast float %z, %x
    498   %t3 = fadd fast float %t1, %t2
    499   ret float %t3
    500 ; CHECK: fact_div3
    501 ; CHECK: fdiv fast float %1, %x
    502 }
    503 
    504 ; y/x - z/x => (y-z)/x
    505 define float @fact_div4(float %x, float %y, float %z) {
    506   %t1 = fdiv fast float %y, %x
    507   %t2 = fdiv fast float %z, %x
    508   %t3 = fsub fast float %t1, %t2
    509   ret float %t3
    510 ; CHECK: fact_div4
    511 ; CHECK: fdiv fast float %1, %x
    512 }
    513 
    514 ; y/x - z/x => (y-z)/x is disabled if y-z is denormal.
    515 define float @fact_div5(float %x) {
    516   %t1 = fdiv fast float 0x3810000000000000, %x
    517   %t2 = fdiv fast float 0x3800000000000000, %x
    518   %t3 = fadd fast float %t1, %t2
    519   ret float %t3
    520 ; CHECK: fact_div5
    521 ; CHECK: fdiv fast float 0x3818000000000000, %x
    522 }
    523 
    524 ; y/x - z/x => (y-z)/x is disabled if y-z is denormal.
    525 define float @fact_div6(float %x) {
    526   %t1 = fdiv fast float 0x3810000000000000, %x
    527   %t2 = fdiv fast float 0x3800000000000000, %x
    528   %t3 = fsub fast float %t1, %t2
    529   ret float %t3
    530 ; CHECK: fact_div6
    531 ; CHECK: %t3 = fsub fast float %t1, %t2
    532 }
    533