[Bug target/81471] internal compiler error: in curr_insn_transform, at lra-constraints.c:3495

[gcp at sjeng dot org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81471

--- Comment #4 from Gian-Carlo Pascutto  ---
Further reduced testcase:

#include 

uint64_t f(uint64_t x) {
return ((uint32_t)x << 55) | ((uint32_t)x >> -23);
}

This makes it more clear the code is UB, but AFAIK a compiler ICE doesn't fall
under allowable UB :-)

[Bug target/81471] internal compiler error: in curr_insn_transform, at lra-constraints.c:3495

[gcp at sjeng dot org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81471

--- Comment #3 from Gian-Carlo Pascutto  ---
Note the flags, -march=native in this case was Intel Haswell.

-O3 -march=haswell is required to trigger this.

[Bug target/81471] internal compiler error: in curr_insn_transform, at lra-constraints.c:3495

[gcp at sjeng dot org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81471

--- Comment #2 from Gian-Carlo Pascutto  ---
#include 

inline uint32_t rotl(const uint32_t x, const int k) {
return (x << k) | (x >> (32 - k));
}

uint64_t s[2];

uint64_t random(void) {
const uint64_t s0 = s[0];
uint64_t s1 = s[1];
const uint64_t result = s0 + s1;

s1 ^= s0;
s[0] = rotl(s0, 55) ^ s1 ^ (s1 << 14); 
s[1] = rotl(s1, 36);

return result;
}

int main(void) {
  int x = random();
  return 0;
}

According to https://godbolt.org/g/8CVJ9a this is broken in all current gcc
versions, not just 5.4.0.

[Bug c/81471] New: internal compiler error: in curr_insn_transform, at lra-constraints.c:3495

[gcp at sjeng dot org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81471

Bug ID: 81471
   Summary: internal compiler error: in curr_insn_transform, at
lra-constraints.c:3495
   Product: gcc
   Version: unknown
Status: UNCONFIRMED
  Severity: normal
  Priority: P3
 Component: c
  Assignee: unassigned at gcc dot gnu.org
  Reporter: gcp at sjeng dot org
  Target Milestone: ---

gcc version 5.4.0 20160609 (Ubuntu 5.4.0-6ubuntu1~16.04.4) 

-Wall -Wextra -pipe -O3 -g -ffast-math -march=native -flto -fopenmp -std=c++11
-DNDEBUG -D_CONSOLE

UCTSearch.cpp:130:1: error: unable to generate reloads for:
 }
 ^
(insn 1085 1084 1086 88 (set (reg:DI 538 [ D.12393 ])
(zero_extend:DI (rotatert:SI (subreg:SI (reg/v:DI 321 [ s0 ]) 0)
(const_int -23 [0xffe9] Random.cpp:30 590
{*bmi2_rorxsi3_1_zext}
 (nil))
UCTSearch.cpp:130:1: internal compiler error: in curr_insn_transform, at
lra-constraints.c:3495
Please submit a full bug report,
with preprocessed source if appropriate.
See  for instructions.
lto-wrapper: fatal error: /usr/bin/g++-5.real returned 1 exit status
compilation terminated.
/usr/bin/ld.bfd.real: error: lto-wrapper failed

The offending function is:

uint64 Random::random(void) {
const uint64 s0 = s[0];
uint64 s1 = s[1];
const uint64 result = s0 + s1;

s1 ^= s0;
s[0] = Utils::rotl(s0, 55) ^ s1 ^ (s1 << 14);<<<<<<<<<<<<<<<
s[1] = Utils::rotl(s1, 36);

return result;
}

In another module there is:

inline uint32 rotl(const uint32 x, const int k) {
return (x << k) | (x >> (32 - k));
}

Note that the code is buggy, the rotl should be using 64-bit integers. But gcc
shouldn't ICE.

[Bug lto/65950] Loop is not vectorized with lto.

[gcp at sjeng dot org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65950

Gian-Carlo Pascutto  changed:

   What|Removed |Added

 CC||gcp at sjeng dot org

--- Comment #6 from Gian-Carlo Pascutto  ---
I'm seeing similar behavior in gcc-4.9.2 (debian stable) and gcc-5.3.1 (Ubuntu
16.04 prelease).

Adding -flto causes certain loops not to get vectorized, adding -fno-lto to
affected files fixes it.

Unfortunately producing a reduced testcase is not so easy, but program source
is available to gcc devs on request.