Extended Description

I added a test case to the nightly test suite: SingleSource/Benchmarks/Misc/matmul_f64_4x4.c

A function multiplies two matrices and puts the result in a local buffer before copying it out:

static void mul4(double *Out, const double A[4][4], const double B[4][4]) {
double Res[16];
Res[0] = ...
...
Res[15] = ...
for (n = 0; n < 16; ++n)
Out[n] = Res[n];
}

SROA converts the double[16] buffer to an i1024:

%188 = fadd double %184, %187
%189 = bitcast double %188 to i64
%190 = zext i64 %189 to i1024
%191 = shl nuw nsw i1024 %190, 768
…
%ins = or i1024 %mask, %191
%222 = bitcast double* %c to i1024*
store i1024 %ins, i1024* %222, align 4
ret void

This completely confuses ARM codegen which is unable to recover the original stores. Instead it copies the doubles to the GPRs and stores i32s:

    vmov    r12, r9, d13
    str.w   r12, [r0, #​96]
    str.w   r9, [r0, #​100]

We wanted a simple f64 store:

    vstr   d13, [r0, #​96]

The problem seems to begin with LegalizeTypes. It breaks the i1024 store into i32 stores because i64 is not a legal ARM type.