<html>
<head>
<base href="https://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 --- - Additional xmm register used for nothing in one of two test cases"
href="https://llvm.org/bugs/show_bug.cgi?id=27596">27596</a>
</td>
</tr>
<tr>
<th>Summary</th>
<td>Additional xmm register used for nothing in one of two test cases
</td>
</tr>
<tr>
<th>Product</th>
<td>libraries
</td>
</tr>
<tr>
<th>Version</th>
<td>3.5
</td>
</tr>
<tr>
<th>Hardware</th>
<td>PC
</td>
</tr>
<tr>
<th>OS</th>
<td>Linux
</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>Loop Optimizer
</td>
</tr>
<tr>
<th>Assignee</th>
<td>unassignedbugs@nondot.org
</td>
</tr>
<tr>
<th>Reporter</th>
<td>g1mat000@gmail.com
</td>
</tr>
<tr>
<th>CC</th>
<td>llvm-bugs@lists.llvm.org
</td>
</tr>
<tr>
<th>Classification</th>
<td>Unclassified
</td>
</tr></table>
<p>
<div>
<pre>I think I found a bug where a register is used "for nothing"
Here's the output of the assembly with the bug:
movaps %xmm0, %xmm3
mulss %xmm3, %xmm3
xorps %xmm4, %xmm4 <- here
addss %xmm4, %xmm3
movaps %xmm2, %xmm5
mulss %xmm5, %xmm5
addss %xmm3, %xmm5
movaps %xmm1, %xmm3
mulss %xmm3, %xmm3
addss %xmm5, %xmm3
sqrtss %xmm3, %xmm3
mulss .LCPI1_0(%rip), %xmm3
movss .LCPI1_1(%rip), %xmm5
maxss %xmm3, %xmm5
mulss .LCPI1_2(%rip), %xmm0
addss %xmm4, %xmm0
addss %xmm2, %xmm2
addss %xmm0, %xmm2
mulss .LCPI1_3(%rip), %xmm1
addss %xmm2, %xmm1
divss %xmm5, %xmm1
movss .LCPI1_4(%rip), %xmm0
ucomiss %xmm1, %xmm0
ja .LBB1_2
Unless i'm wrong, xmm4 is never used for the computation.
And with another test case (same computation however) we have no additional
register:
movaps %xmm0, %xmm3
mulss %xmm3, %xmm3
movaps %xmm2, %xmm4
mulss %xmm4, %xmm4
addss %xmm3, %xmm4
movaps %xmm1, %xmm3
mulss %xmm3, %xmm3
addss %xmm4, %xmm3
sqrtss %xmm3, %xmm3
mulss .LCPI2_0(%rip), %xmm3
movss .LCPI2_1(%rip), %xmm4
maxss %xmm3, %xmm4
mulss .LCPI2_2(%rip), %xmm0
addss %xmm2, %xmm2
addss %xmm0, %xmm2
mulss .LCPI2_3(%rip), %xmm1
addss %xmm2, %xmm1
divss %xmm4, %xmm1
movss .LCPI2_4(%rip), %xmm0
ucomiss %xmm1, %xmm0
ja .LBB2_2
The first function in CPP looks like this:
float Do(int ac)
{
return Vec3(5, 2, -7).angleBetween(Vec3(ac, ac * 2, ac * -3)).count();
// count() <==> static_cast<float>
}
template<typename Quantity, size_t Size>
class Vector
{
explicit constexpr Vector(Arg&&... args) :
m_values{Quantity(args)...}
{
}
...
inline Angle angleBetween(const Vector& vec)
{
SquaredQuantity lenProduct = length() * vec.length();
if (lenProduct < SquaredQuantity(1e-6f))
lenProduct = SquaredQuantity(1e-6f);
Constant rotation = Clamp(dotProduct(vec) / lenProduct, -1_c, 1_c);
return Acos(rotation);
}
length(), dotProduct(), Clamp and Acos are also inline functions
using fixed-size loops like:
constexpr SquaredQuantity dotProduct(const Vector& vec) const
{
SquaredQuantity product(0.0);
for (size_t i = 0; i < Size; ++i)
product += m_values[i] * vec.at(i);
return product;
}
...
private:
Quantity m_values[Size];
}
Quantity is a class template wrapping a float, only used for strict typing.
The other which is bug-free is the same computation the C way:
struct V3
{
float tab[3];
};
float Do2(int ac)
{
V3 vec{{5.0f, 2.0f, -7.0f}};
V3 otherVec{{float(ac), float(ac * 2), float(ac * -3)}};
float lenProduct = std::sqrt(vec.tab[0] * vec.tab[0] + vec.tab[1] *
vec.tab[1] + vec.tab[2] * vec.tab[2]) *
std::sqrt(otherVec.tab[0] * otherVec.tab[0] + otherVec.tab[1] *
otherVec.tab[1] + otherVec.tab[2] * otherVec.tab[2]);
if (lenProduct < 1e-6f)
lenProduct = 1e-6f;
float product = vec.tab[0] * otherVec.tab[0] + vec.tab[1] * otherVec.tab[1]
+ vec.tab[2] * otherVec.tab[2];
float rotation = Clamp(product / lenProduct, -1.0f, 1.0f);
return std::acos(rotation);
}
Here's the LLVM code with the bug:
; Function Attrs: nounwind readnone uwtable
define float @_Z2Doi(i32 %ac) #4 {
%1 = shl nsw i32 %ac, 1
%2 = mul nsw i32 %ac, -3
%3 = sitofp i32 %ac to float
%4 = sitofp i32 %1 to float
%5 = sitofp i32 %2 to float
%6 = fmul float %3, %3
%7 = fadd float %6, 0.000000e+00 <- maybe here?
%8 = fmul float %4, %4
%9 = fadd float %7, %8
%10 = fmul float %5, %5
%11 = fadd float %10, %9
%12 = tail call float @sqrtf(float %11) #8
%13 = fmul float %12, 0x4021A9DC80000000
%14 = fcmp olt float %13, 0x3EB0C6F7A0000000
%15 = select i1 %14, float 0x3EB0C6F7A0000000, float %13
%16 = fmul float %3, 5.000000e+00
%17 = fadd float %16, 0.000000e+00
%18 = fmul float %4, 2.000000e+00
%19 = fadd float %17, %18
%20 = fmul float %5, -7.000000e+00
%21 = fadd float %20, %19
%22 = fdiv float %21, %15
%23 = fcmp olt float %22, -1.000000e+00
br i1 %23, label
%_ZN6studio6VectorINS_8QuantityILi0ELi1ELi0EEELm3EE12angleBetweenERKS3_.exit,
label %24
; <label>:24 ; preds = %0
%25 = fcmp ogt float %22, 1.000000e+00
%.pre.i = select i1 %25, float 1.000000e+00, float %22
br label
%_ZN6studio6VectorINS_8QuantityILi0ELi1ELi0EEELm3EE12angleBetweenERKS3_.exit
_ZN6studio6VectorINS_8QuantityILi0ELi1ELi0EEELm3EE12angleBetweenERKS3_.exit: ;
preds = %0, %24
%26 = phi float [ -1.000000e+00, %0 ], [ %.pre.i, %24 ]
%27 = tail call float @acosf(float %26) #8
ret float %27
}
And the LLVM code without:
; Function Attrs: nounwind readnone uwtable
define float @_Z3Do2i(i32 %ac) #4 {
%1 = sitofp i32 %ac to float
%2 = shl nsw i32 %ac, 1
%3 = sitofp i32 %2 to float
%4 = mul nsw i32 %ac, -3
%5 = sitofp i32 %4 to float
%6 = fmul float %1, %1
%7 = fmul float %3, %3
%8 = fadd float %6, %7
%9 = fmul float %5, %5
%10 = fadd float %9, %8
%11 = tail call float @sqrtf(float %10) #8
%12 = fmul float %11, 0x4021A9DC80000000
%13 = fcmp olt float %12, 0x3EB0C6F7A0000000
%lenProduct.0 = select i1 %13, float 0x3EB0C6F7A0000000, float %12
%14 = fmul float %1, 5.000000e+00
%15 = fmul float %3, 2.000000e+00
%16 = fadd float %14, %15
%17 = fmul float %5, -7.000000e+00
%18 = fadd float %17, %16
%19 = fdiv float %18, %lenProduct.0
%20 = fcmp olt float %19, -1.000000e+00
br i1 %20, label %_ZN6studio5ClampIfEERKT_S3_S3_S3_.exit, label %21
; <label>:21 ; preds = %0
%22 = fcmp ogt float %19, 1.000000e+00
%23 = select i1 %22, float 1.000000e+00, float %19
br label %_ZN6studio5ClampIfEERKT_S3_S3_S3_.exit
_ZN6studio5ClampIfEERKT_S3_S3_S3_.exit: ; preds = %0, %21
%24 = phi float [ -1.000000e+00, %0 ], [ %23, %21 ]
%25 = tail call float @acosf(float %24) #8
ret float %25
}
I'm not sure at all wether it comes from my code, the cpp templates or the
optimizations with inlined functions, or anything else.
But I'm guessing it's a bug since the register is never used, or it's just some
very minor issue due to unoptimized cpp code?</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>