[PATCH] D159205: [SLP]Try to vectorize scalars, being vectorized already, but do not need to be scheduled.

[Bogdan Graur via Phabricator via llvm-commits]

bgraur added a comment.

We found another issue where the compiler enters an infinite loop at this revision.
The infinite loop reproduces up until revision 7db87a66b0206af027724fc1704e5e5ce7772a45 <https://reviews.llvm.org/rG7db87a66b0206af027724fc1704e5e5ce7772a45> where the compiler exits with error:

  Instruction does not dominate all uses!
    %7 = shufflevector <2 x double> %6, <2 x double> poison, <2 x i32> zeroinitializer
    %5 = fmul <2 x double> %7, <double 0x3FD4F2094F2094F2, double 0x3FE5555555555555>
  in function aa
  fatal error: error in backend: Broken function found, compilation aborted!

There's no change at the revision fixing the previous report (ebed4692f8160e65c5a04c3edfb61b82ca037574 <https://reviews.llvm.org/rGebed4692f8160e65c5a04c3edfb61b82ca037574>)

Reduced repro:

  int a;
  double b, c, d, e, f, g, h, i, j, k, l, m, n, o, p, q, r, s, t, u, v, w, x, y,
      z;
  void aa() {
    double ab, ac, ad = ab = c ? h + d : f * g;
    double un = 0.18e2 / 0.55e2 * ab - 0.1e2 / 0.33e2 * ab - 0.2e1 / 0.3e1 * ab;
    if (a)
      t = s;
    ac = v + 0.2e1 / 0.3e1 * ab * d - w - 0.2e1 / 0.3e1 * ab * 0 + x + i - y - j +
         e + k - e - l + e + m - e - n;
    ad = d + 0.18e2 / 0.55e2 * ab * o - e - 0.1e2 / 0.33e2 * ab * p + f + b - d -
         q + e + b - e - r + f + b - e - b;
    z = c ? u + ac + ad : 0;
  }

Compile command:

  $ clang -cc1 -triple x86_64-pc-linux-gnu -S \
     -fmath-errno -target-feature +sse4.2 \
     -O3 -vectorize-slp -x c repro.c \
     -o /tmp/out

@ABataev  can you please take a look? This second issue is much more spread and causes lots of breakages for us.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D159205/new/

https://reviews.llvm.org/D159205