A fairly small repro is gfx_unittests:

$ out/release/gfx_unittests --gtest_filter=Matrix3fTest.Inverse
IMPORTANT DEBUGGING NOTE: batches of tests are run inside their
own process. For debugging a test inside a debugger, use the
--gtest_filter=<your_test_name> flag along with
--single-process-tests.
Using sharding settings from environment. This is shard 0/1
Using 1 parallel jobs.
Note: Google Test filter = Matrix3fTest.Inverse
[==========] Running 1 test from 1 test case.
[----------] Global test environment set-up.
[----------] 1 test from Matrix3fTest
[ RUN      ] Matrix3fTest.Inverse
../../ui/gfx/geometry/matrix3_unittest.cc:76: Failure
Value of: singular.Inverse()
  Actual: 36-byte object <69-33 EC-BD 00-00 00-80 1D-2C 21-3E D4-7F 90-3C 74-13 C8-3D D4-7F 10-BD 97-0E 96-3E 1D-2C A1-BD 81-4E DE-BC>
Expected: Matrix3F::Zeros()
Which is: 36-byte object <00-00 00-00 00-00 00-00 00-00 00-00 00-00 00-00 00-00 00-00 00-00 00-00 00-00 00-00 00-00 00-00 00-00 00-00>
../../ui/gfx/geometry/matrix3_unittest.cc:86: Failure
Value of: regular.IsNear(inv_regular, 0.00001f)
  Actual: false
Expected: true
[  FAILED  ] Matrix3fTest.Inverse (0 ms)
[----------] 1 test from Matrix3fTest (0 ms total)

[----------] Global test environment tear-down
[==========] 1 test from 1 test case ran. (0 ms total)
[  PASSED  ] 0 tests.
[  FAILED  ] 1 test, listed below:
[  FAILED  ] Matrix3fTest.Inverse

 1 FAILED TEST
[1/1] Matrix3fTest.Inverse (0 ms)
1 test failed:
    Matrix3fTest.Inverse (../../ui/gfx/geometry/matrix3_unittest.cc:66)
Tests took 0 seconds.


That test is mostly exercising ui/gfx/geometry/matrix3_f.cc

Attaching a preprocessed version of that; it can be built with:

clang++ -fno-strict-aliasing --param=ssp-buffer-size=4 -fstack-protector -funwind-tables -fPIC -pipe -m64 -march=x86-64 -pthread -O2 -fvisibility=hidden -fvisibility-inlines-hidden -std=gnu++11 -fno-rtti -fno-exceptions -w -c /tmp/matrix3_f.ii -o /tmp/matrix3_f.o


Diffing the asm before r296575 and after shows changes in _ZNK3gfx8Matrix3F7InverseEv (source here: https://cs.chromium.org/chromium/src/ui/gfx/geometry/matrix3_f.cc?dr&l=101)

--- /tmp/matrix3_f.good.s       2017-03-01 10:42:02.875397777 -0800
+++ /tmp/matrix3_f.bad.s        2017-03-01 10:42:11.071318686 -0800
@@ -239,14 +239,14 @@
        movups  16(%rsi), %xmm12
        movss   32(%rsi), %xmm3         # xmm3 = mem[0],zero,zero,zero
        movaps  %xmm12, %xmm7
-       shufps  $229, %xmm7, %xmm7      # xmm7 = xmm7[1,1,2,3]
+       shufps  $231, %xmm7, %xmm7      # xmm7 = xmm7[3,1,2,3]
        cvtss2sd        %xmm7, %xmm5
        movaps  %xmm12, %xmm4
-       shufps  $227, %xmm4, %xmm4      # xmm4 = xmm4[3,0,2,3]
+       shufps  $226, %xmm4, %xmm4      # xmm4 = xmm4[2,0,2,3]
        cvtps2pd        %xmm4, %xmm1
        cvtss2sd        %xmm10, %xmm9
        movaps  %xmm12, %xmm4
-       movhlps %xmm4, %xmm4            # xmm4 = xmm4[1,1]
+       shufps  $229, %xmm4, %xmm4      # xmm4 = xmm4[1,1,2,3]
        xorps   %xmm0, %xmm0
        cvtss2sd        %xmm4, %xmm0
        movapd  %xmm5, %xmm4
@@ -281,7 +281,7 @@
        ja      .LBB8_2
 # BB#1:                                 # %if.end
        movaps  %xmm12, %xmm4
-       shufps  $231, %xmm4, %xmm4      # xmm4 = xmm4[3,1,2,3]
+       movhlps %xmm4, %xmm4            # xmm4 = xmm4[1,1]
        movaps  %xmm12, %xmm0
        shufps  $156, %xmm0, %xmm0      # xmm0 = xmm0[0,3,1,2]
        movaps  %xmm3, %xmm1
@@ -290,7 +290,7 @@
        shufps  $48, %xmm1, %xmm2       # xmm2 = xmm2[0,0],xmm1[3,0]
        shufps  $132, %xmm2, %xmm1      # xmm1 = xmm1[0,1],xmm2[0,2]
        movaps  %xmm12, %xmm2
-       shufps  $33, %xmm1, %xmm2       # xmm2 = xmm2[1,0],xmm1[2,0]
+       shufps  $35, %xmm1, %xmm2       # xmm2 = xmm2[3,0],xmm1[2,0]
        shufps  $36, %xmm2, %xmm1       # xmm1 = xmm1[0,1],xmm2[2,0]
        mulps   %xmm0, %xmm1
        movaps  %xmm10, %xmm0
@@ -318,7 +318,7 @@
        unpcklps        %xmm8, %xmm0    # xmm0 = xmm0[0],xmm8[0],xmm0[1],xmm8[1]
        unpcklps        %xmm13, %xmm0   # xmm0 = xmm0[0],xmm13[0],xmm0[1],xmm13[1]
        movaps  %xmm12, %xmm1
-       shufps  $34, %xmm0, %xmm1       # xmm1 = xmm1[2,0],xmm0[2,0]
+       shufps  $33, %xmm0, %xmm1       # xmm1 = xmm1[1,0],xmm0[2,0]
        shufps  $36, %xmm1, %xmm0       # xmm0 = xmm0[0,1],xmm1[2,0]
        unpcklps        %xmm8, %xmm3    # xmm3 = xmm3[0],xmm8[0],xmm3[1],xmm8[1]
        shufps  $48, %xmm3, %xmm4       # xmm4 = xmm4[0,0],xmm3[3,0]
@@ -332,7 +332,7 @@
        unpcklps        %xmm0, %xmm0    # xmm0 = xmm0[0,0,1,1]
        unpcklps        %xmm0, %xmm13   # xmm13 = xmm13[0],xmm0[0],xmm13[1],xmm0[1]
        movaps  %xmm12, %xmm0
-       shufps  $230, %xmm0, %xmm0      # xmm0 = xmm0[2,1,2,3]
+       shufps  $157, %xmm0, %xmm0      # xmm0 = xmm0[1,3,1,2]
        mulps   %xmm13, %xmm0
        subps   %xmm0, %xmm3
        cvtps2pd        %xmm3, %xmm0
@@ -923,5 +923,5 @@
        .globl  _ZN3gfx8Matrix3FD1Ev
        .type   _ZN3gfx8Matrix3FD1Ev,@function
 _ZN3gfx8Matrix3FD1Ev = _ZN3gfx8Matrix3FD2Ev
-       .ident  "clang version 5.0.0 (trunk 296632) (llvm/trunk 296642)"
+       .ident  "clang version 5.0.0 (trunk 296632) (llvm/trunk 296634)"
        .section        ".note.GNU-stack","",@progbits

Reverted in LLVM r296654

Oops, actually attaching the preprocessed source this time.

Deleted: matrix3_f.ii 1.2 MB

matrix3_f.ii 1.2 MB Download

Reduced test case:

extern "C" int memcmp(__const void *, __const void *, long);
class Matrix3F {
public:
  ~Matrix3F();
  bool m_fn1(const Matrix3F &) const;
  void m_fn2(float p1, float p2, float p3, float p4, float p5, float p6,
             float p7, float p8) {
    data_[0] = p1;
    data_[1] = p2;
    data_[2] = p3;
    data_[4] = p4;
    data_[5] = p5;
    data_[6] = p6;
    data_[7] = p7;
    data_[8] = p8;
  }
  Matrix3F m_fn3() const;
  Matrix3F();
  float data_[9];
};

enum { M00, M01, M02, M10, M11, M12, M20, M21, M22 };
Matrix3F g;
Matrix3F::Matrix3F() {}

Matrix3F::~Matrix3F() {}

bool Matrix3F::m_fn1(const Matrix3F &p1) const {
  return 0 == memcmp(data_, p1.data_, sizeof data_);
}

Matrix3F Matrix3F::m_fn3() const {
  double a, b;
  Matrix3F c, d;
  d.m_fn2(0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f);
  c = d;
  Matrix3F e = c;
  const float *f = data_;
  b = f[M00] * f[M11] * f[M22] - f[M12] * f[M21] + f[M01] * f[M12] * f[M20] +
      f[M02] * (f[M10] - static_cast<double>(f[M11]) * f[M20]);
  a = __builtin_fabs(b);
  if (1.19209290e-7F > a)
    return e;
  e.m_fn2(b, b, data_[M01] / b, (data_[M00] * data_[M22] - data_[M20]) / b,
          (data_[M02] * data_[M10] - data_[M12]) / b,
          (data_[M10] * data_[M21] - data_[M20]) / b,
          (data_[M01] * data_[M20] - data_[M21]) / b, b);
  return e;
}

int main() {
  Matrix3F h, i, j;
  g.m_fn2(1.0f, 3.0f, 4.0f, 11.0f, 5.0f, 0.5f, 1.5f, 2.0f);
  i = g.m_fn3();
  j.m_fn2(0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f);
  h = j;
  if (h.m_fn1(i))
    return 0;
  return 1;
}

$ bin/clang.good -O2 a.ii && ./a.out && echo ok
ok
$ bin/clang.bad -O2 a.ii && ./a.out && echo ok

This was fixed a few weeks ago.

It was filed upstream in llvm.org/PR32109 and fixed by a revert in r296654.