<html>
<head>
<base href="http://llvm.org/bugs/" />
</head>
<body><table border="1" cellspacing="0" cellpadding="8">
<tr>
<th>Bug ID</th>
<td><a class="bz_bug_link
bz_status_NEW "
title="NEW --- - Failure to simplify SIMD vector conversion."
href="http://llvm.org/bugs/show_bug.cgi?id=16739">16739</a>
</td>
</tr>
<tr>
<th>Summary</th>
<td>Failure to simplify SIMD vector conversion.
</td>
</tr>
<tr>
<th>Product</th>
<td>new-bugs
</td>
</tr>
<tr>
<th>Version</th>
<td>trunk
</td>
</tr>
<tr>
<th>Hardware</th>
<td>PC
</td>
</tr>
<tr>
<th>OS</th>
<td>All
</td>
</tr>
<tr>
<th>Status</th>
<td>NEW
</td>
</tr>
<tr>
<th>Severity</th>
<td>normal
</td>
</tr>
<tr>
<th>Priority</th>
<td>P
</td>
</tr>
<tr>
<th>Component</th>
<td>new bugs
</td>
</tr>
<tr>
<th>Assignee</th>
<td>unassignedbugs@nondot.org
</td>
</tr>
<tr>
<th>Reporter</th>
<td>silvas@purdue.edu
</td>
</tr>
<tr>
<th>CC</th>
<td>llvmbugs@cs.uiuc.edu
</td>
</tr>
<tr>
<th>Classification</th>
<td>Unclassified
</td>
</tr></table>
<p>
<div>
<pre>This code was reduced from a function that converted between SIMD vector
classes used by two different libraries; the source and destination vectors
have a <4 x float> underlying storage, but notionally hold only {x, y, z} (and
the destination duplicates z into the last lane; the source leaves it undefined
I think).
typedef float __m128 __attribute__((__vector_size__(16)));
union ElementWiseAccess {
ElementWiseAccess(__m128 v) : ReprM128(v) {}
__m128 ReprM128;
float ReprFloatArray[4];
float getAt(int i) const { return ReprFloatArray[i]; }
};
// Making this return `const ElementWiseAccess` instead of `const
ElementWiseAccess &`
// still results in a failure to optimize, but in a different way.
static const ElementWiseAccess &castToElementWiseAccess(const __m128 &t) {
return reinterpret_cast<const ElementWiseAccess &>(t);
}
__m128 ConvertVectors(const __m128 &V) {
// Replacing `castToElementWiseAccess` with directly calling
// `ElementWiseAccess` makes the issue go away.
return (__m128) { castToElementWiseAccess(V).getAt(0), //
castToElementWiseAccess(V).getAt(1), //
castToElementWiseAccess(V).getAt(2), //
castToElementWiseAccess(V).getAt(2) };
}
clang -O3 produces:
define <4 x float> @_Z14ConvertVectorsRKDv4_f(<4 x float>* nocapture readonly
%V) #0 {
%1 = bitcast <4 x float>* %V to [4 x float]*
%2 = getelementptr inbounds <4 x float>* %V, i64 0, i64 0
%3 = load float* %2, align 4, !tbaa !0
%4 = insertelement <4 x float> undef, float %3, i32 0
%5 = getelementptr inbounds [4 x float]* %1, i64 0, i64 1
%6 = load float* %5, align 4, !tbaa !0
%7 = insertelement <4 x float> %4, float %6, i32 1
%8 = getelementptr inbounds [4 x float]* %1, i64 0, i64 2
%9 = load float* %8, align 4, !tbaa !0
%10 = insertelement <4 x float> %7, float %9, i32 2
%11 = insertelement <4 x float> %10, float %9, i32 3
ret <4 x float> %11
}
It appears that something is interfering with folding the load/insertelement
sequence into a vector load + shufflevector.
Making the modification indicated in the comments of having
`castToElementWiseAccess` return by value instead of by reference results in:
define <4 x float> @_Z14ConvertVectorsRKDv4_f(<4 x float>* nocapture readonly
%V) #0 {
%1 = bitcast <4 x float>* %V to i8*
%2 = bitcast <4 x float>* %V to double*
%3 = load double* %2, align 16
%4 = getelementptr inbounds i8* %1, i64 8
%5 = bitcast i8* %4 to double*
%6 = bitcast double %3 to i64
%trunc = trunc i64 %6 to i32
%bitcast = bitcast i32 %trunc to float
%7 = insertelement <4 x float> undef, float %bitcast, i32 0
%8 = lshr i64 %6, 32
%9 = trunc i64 %8 to i32
%10 = bitcast i32 %9 to float
%11 = insertelement <4 x float> %7, float %10, i32 1
%12 = load double* %5, align 8
%13 = bitcast double %12 to i64
%trunc6 = trunc i64 %13 to i32
%bitcast7 = bitcast i32 %trunc6 to float
%14 = insertelement <4 x float> %11, float %bitcast7, i32 2
%15 = insertelement <4 x float> %14, float %bitcast7, i32 3
ret <4 x float> %15
}
The issue in this case seems to be that clang lowers `castToElementWiseAccess`
as returning `{double, double}`, which then prevents a <4 x float> load being
generated.
Making the modification of replacing the call to `castToElementWiseAcess` with
directly invoking the constructor (e.g. `ElementWiseAccess(V).getAt(<<<n>>>)`)
results in the following code, which is the desired codegen for the initial
test case:
define <4 x float> @_Z14ConvertVectorsRKDv4_f(<4 x float>* nocapture readonly
%V) #0 {
%1 = load <4 x float>* %V, align 16, !tbaa !0
%2 = shufflevector <4 x float> %1, <4 x float> undef, <4 x i32> <i32 0, i32
1, i32 2, i32 2>
ret <4 x float> %2
}</pre>
</div>
</p>
<hr>
<span>You are receiving this mail because:</span>
<ul>
<li>You are on the CC list for the bug.</li>
</ul>
</body>
</html>