Compiler in infinite loop after 6fd229a

[dyung]

One of our internal tests started to time out and when I looked into it, the test actually appears to be hitting an infinite loop in the compiler. I was able to bisect the cause back to commit 6fd229a. Note that in #109215 it mentions ec78f0d as a possible fix, but I tried that and the compiler still seems to be stuck in an infinite loop.

Consider the following code:

long b;
struct {
  unsigned long a : 43;
} c;
void d() {
  long e = b & b - 1, f = c.a = e;
  c.a = 0;
  volatile long g(~f & g);
}

If compiled with optimizations and the BMI attributes with a compiler that does not include 6fd229a, the compilation succeeds quickly:

$ time ~/src/upstream/b9cae45b63bc9c44521cc28b4a381afec6181f54-linux/bin/clang -c -O2 -mbmi repro.cpp -o repro.o

real    0m0.046s
user    0m0.012s
sys     0m0.035s

However, the same compilation with a compiler built from ec78f0d still seems to be stuck in an infinite loop somewhere:

$ time /home/dyung/src/upstream/ec78f0da0e9b1b8e2b2323e434ea742e272dd913-linux/bin/clang -c -O2 -mbmi repro.cpp -o repro.o
^C

real    5m6.955s
user    5m6.594s
sys     0m0.337s

Godbolt link: https://godbolt.org/z/7deanqrnn

[Saldivarcher]

Here is reduced llvm-ir:

$ cat reduced.ll 
define void @_Z1dv(i64 %0) {
entry:
  %and = and i64 %0, 8796093022206
  %bf.value = and i64 8796093022206, 0
  %not = xor i64 %and, -1
  %and4 = and i64 %0, %not
  store i64 %and4, ptr null, align 8
  ret void
}

$ llc reduced.ll -mtriple=x86_64-- -mattr=+bmi
^C

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: None (dyung)

One of our internal tests started to time out and when I looked into it, the test actually appears to be hitting an infinite loop in the compiler. I was able to bisect the cause back to commit 6fd229a. Note that in #109215 it mentions ec78f0d as a possible fix, but I tried that and the compiler still seems to be stuck in an infinite loop.

Consider the following code:

long b;
struct {
  unsigned long a : 43;
} c;
void d() {
  long e = b & b - 1, f = c.a = e;
  c.a = 0;
  volatile long g(~f & g);
}

If compiled with optimizations and the BMI attributes with a compiler that does not include 6fd229a, the compilation succeeds quickly:

$ time ~/src/upstream/b9cae45b63bc9c44521cc28b4a381afec6181f54-linux/bin/clang -c -O2 -mbmi repro.cpp -o repro.o

real    0m0.046s
user    0m0.012s
sys     0m0.035s

However, the same compilation with a compiler built from ec78f0d still seems to be stuck in an infinite loop somewhere:

$ time /home/dyung/src/upstream/ec78f0da0e9b1b8e2b2323e434ea742e272dd913-linux/bin/clang -c -O2 -mbmi repro.cpp -o repro.o
^C

real    5m6.955s
user    5m6.594s
sys     0m0.337s

Godbolt link: https://godbolt.org/z/7deanqrnn