<html>
<head>
<base href="https://bugs.llvm.org/">
</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 - narrow casted FMA op?"
href="https://bugs.llvm.org/show_bug.cgi?id=43841">43841</a>
</td>
</tr>
<tr>
<th>Summary</th>
<td>narrow casted FMA op?
</td>
</tr>
<tr>
<th>Product</th>
<td>libraries
</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>enhancement
</td>
</tr>
<tr>
<th>Priority</th>
<td>P
</td>
</tr>
<tr>
<th>Component</th>
<td>Scalar Optimizations
</td>
</tr>
<tr>
<th>Assignee</th>
<td>unassignedbugs@nondot.org
</td>
</tr>
<tr>
<th>Reporter</th>
<td>spatel+llvm@rotateright.com
</td>
</tr>
<tr>
<th>CC</th>
<td>llvm-bugs@lists.llvm.org
</td>
</tr></table>
<p>
<div>
<pre>float fmaf(float x, float y, float z) {
return (((double)x * (double)y) + (double)z);
}
$ clang -O3 -march=haswell -ffp-contract=on
fmaf(float, float, float):
vcvtss2sd xmm0, xmm0, xmm0
vcvtss2sd xmm1, xmm1, xmm1
vcvtss2sd xmm2, xmm2, xmm2
vfmadd231sd xmm2, xmm0, xmm1 # xmm2 = (xmm0 * xmm1) + xmm2
vcvtsd2ss xmm0, xmm2, xmm2
ret
Or as IR:
define float @fmaf(float %0, float %1, float %2) {
%4 = fpext float %0 to double
%5 = fpext float %1 to double
%6 = fpext float %2 to double
%7 = tail call double @llvm.fmuladd.f64(double %4, double %5, double %6)
%8 = fptrunc double %7 to float
ret float %8
}
declare double @llvm.fmuladd.f64(double, double, double)
<a href="https://gcc.godbolt.org/z/FxOJlh">https://gcc.godbolt.org/z/FxOJlh</a>
-------------------------------------------------------------------------------
Can we remove all of the casts and reduce this to a float (f32 - "vfmadd231ss")
math instruction because the intermediate rounding steps can be proven to not
affect the result?
InstCombiner::visitFPTrunc() has a section dedicated to this kind of transform,
but it does not include FMA. It only includes the FP binops (fmul, fadd, fsub,
fdiv, frem).
The existing transforms reference:
Figueroa's 2000 PhD thesis, "A Rigorous Framework for Fully Supporting the IEEE
Standard ..."</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>