https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98140
Bug ID: 98140
Summary: Reused register by xsmincdp leads to wrong NaN
propagation on Power9
Product: gcc
Version: 8.3.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: alexander.gr...@tu-dresden.de
Target Milestone: ---
Created attachment 49679
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=49679&action=edit
(preprocessed) source code to reproduce issue
Summary: xsmincdp instructions are generated in a form like `xsmincdp b,a,b`
for code that looks like `(a > b) ? b : a`
I was debugging an issue in PyTorch
(https://github.com/pytorch/pytorch/issues/48591) where I encountered the
following problem:
A clamp function is used which looks like this:
c[i] = a[i] < min_vec[i] ? min_vec[i] : (a[i] > max_vec[i] ? max_vec[i] :
a[i]);
This is used in very complex code using multiple levels of C++ templates,
lambdas and such and uses a combination of manually unrolled loops and
unroll-friendly loops (i.e. the above is called in a loop with a fixed trip
count of 8)
The generated ASM code has this (using objdump):
c[i] = a[i] < min_vec[i] ? min_vec[i] : (a[i] > max_vec[i] ? max_vec[i] :
a[i]);
8e970: 20 00 fe cb lfd f31,32(r30)
8e974: 00 f8 9c ff fcmpu cr7,f28,f31
8e978: 0c 00 9c 41 blt cr7,8e984
8e97c: 40 00 fe cb lfd f31,64(r30)
8e980: 40 fc fc f3 xsmincdp vs31,vs28,vs31
8e984: 28 00 9e cb lfd f28,40(r30)
So I assume f28/vs28 contains a[i] and vs31 contains max_vec[i], so the
instruction generated looks like `xsmincdp max_vec,a,max_vec` which on NaN will
return max_vec. However in the source code a should be returned due to the
condition evaluating to false when a NaN is involved.
Reproducing this is tricky, as it depends on many conditions. From my
observations I assume some register pressure is required and even some other
function also calling that code, so maybe some side effects from there. Using
GCC 10.2.0 I wasn't able to reproduce this as the codegen is slightly
different: Seemingly it notices that max_vec contains the same value for all i
and reuses a single register:
324: 00 70 1f fc fcmpu cr0,f31,f14
328: 90 f8 a0 fe fmr f21,f31
32c: 08 00 81 41 bgt 334
330: 40 74 be f2 xsmincdp vs21,vs30,vs14
334: 00 78 1f fc fcmpu cr0,f31,f15
338: 90 f8 c0 fe fmr f22,f31
33c: 08 00 81 41 bgt 344
340: 40 7c de f2 xsmincdp vs22,vs30,vs15
I'm attaching some source code which can be compiled using PyTorch 1.7.0 and 2
examples of preprocessed code which yield the above when compiled using `g++
-mcpu=power9 -g -std=gnu++14 -O3`
