[Bug target/118794] The attached c++ openmp offload code fails, because the c sqrt function makes nonlocal gotos..
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118794 Thomas Schwinge changed: What|Removed |Added Status|ASSIGNED|RESOLVED Target Milestone|--- |15.0 Resolution|--- |FIXED --- Comment #13 from Thomas Schwinge --- This code now works with (the upcoming) GCC 15, without requiring optimizations enabled, or host-side '-fno-exceptions'. Tested host, Nvidia GPU offloading, AMD GPU offloading.
[Bug target/118794] The attached c++ openmp offload code fails, because the c sqrt function makes nonlocal gotos..
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118794 --- Comment #12 from GCC Commits --- The trunk branch has been updated by Thomas Schwinge : https://gcc.gnu.org/g:fe283dba774be57b705a7a871b000d2894d2e553 commit r15-9470-gfe283dba774be57b705a7a871b000d2894d2e553 Author: Thomas Schwinge Date: Fri Mar 28 09:20:49 2025 +0100 GCN, nvptx: Support '-mfake-exceptions', and use it for offloading compilation [PR118794] With '-mfake-exceptions' enabled, the user-visible behavior in presence of exception handling constructs changes such that the compile-time 'sorry, unimplemented: exception handling not supported' is skipped, code generation proceeds, and instead, exception handling constructs 'abort' at run time. (..., or don't, if they're in dead code.) PR target/118794 gcc/ * config/gcn/gcn.opt (-mfake-exceptions): Support. * config/nvptx/nvptx.opt (-mfake-exceptions): Likewise. * config/gcn/gcn.md (define_expand "exception_receiver"): Use it. * config/nvptx/nvptx.md (define_expand "exception_receiver"): Likewise. * config/gcn/mkoffload.cc (main): Set it. * config/nvptx/mkoffload.cc (main): Likewise. * config/nvptx/nvptx.cc (nvptx_assemble_integer) : Special handling for 'SYMBOL_REF's. * except.cc (expand_dw2_landing_pad_for_region): Don't generate bogus code for (default) '#define EH_RETURN_DATA_REGNO(N) INVALID_REGNUM'. libgcc/ * config/gcn/unwind-gcn.c (_Unwind_Resume): New. * config/nvptx/unwind-nvptx.c (_Unwind_Resume): Likewise. gcc/testsuite/ * g++.target/gcn/exceptions-bad_cast-2.C: Set '-mno-fake-exceptions'. * g++.target/gcn/exceptions-pr118794-1.C: Likewise. * g++.target/gcn/exceptions-throw-2.C: Likewise. * g++.target/nvptx/exceptions-bad_cast-2.C: Likewise. * g++.target/nvptx/exceptions-pr118794-1.C: Likewise. * g++.target/nvptx/exceptions-throw-2.C: Likewise. * g++.target/gcn/exceptions-bad_cast-2_-mfake-exceptions.C: New. * g++.target/gcn/exceptions-pr118794-1_-mfake-exceptions.C: Likewise. * g++.target/gcn/exceptions-throw-2_-mfake-exceptions.C: Likewise. * g++.target/nvptx/exceptions-bad_cast-2_-mfake-exceptions.C: Likewise. * g++.target/nvptx/exceptions-pr118794-1_-mfake-exceptions.C: Likewise. * g++.target/nvptx/exceptions-throw-2_-mfake-exceptions.C: Likewise. libgomp/ * testsuite/libgomp.c++/target-exceptions-bad_cast-2-offload-sorry-GCN.C: Set '-foffload-options=-mno-fake-exceptions'. * testsuite/libgomp.c++/target-exceptions-bad_cast-2-offload-sorry-nvptx.C: Likewise. * testsuite/libgomp.c++/target-exceptions-pr118794-1-offload-sorry-GCN.C: Likewise. * testsuite/libgomp.c++/target-exceptions-pr118794-1-offload-sorry-nvptx.C: Likewise. * testsuite/libgomp.c++/target-exceptions-throw-2-offload-sorry-GCN.C: Likewise. * testsuite/libgomp.c++/target-exceptions-throw-2-offload-sorry-nvptx.C: Likewise. * testsuite/libgomp.oacc-c++/exceptions-bad_cast-2-offload-sorry-GCN.C: Likewise. * testsuite/libgomp.oacc-c++/exceptions-bad_cast-2-offload-sorry-nvptx.C: Likewise. * testsuite/libgomp.oacc-c++/exceptions-throw-2-offload-sorry-GCN.C: Likewise. * testsuite/libgomp.oacc-c++/exceptions-throw-2-offload-sorry-nvptx.C: Likewise. * testsuite/libgomp.c++/target-exceptions-bad_cast-2.C: Adjust. * testsuite/libgomp.c++/target-exceptions-pr118794-1.C: Likewise. * testsuite/libgomp.c++/target-exceptions-throw-2.C: Likewise. * testsuite/libgomp.oacc-c++/exceptions-bad_cast-2.C: Likewise. * testsuite/libgomp.oacc-c++/exceptions-throw-2.C: Likewise. * testsuite/libgomp.c++/target-exceptions-throw-2-O0.C: New.
[Bug target/118794] The attached c++ openmp offload code fails, because the c sqrt function makes nonlocal gotos..
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118794 --- Comment #11 from GCC Commits --- The trunk branch has been updated by Thomas Schwinge : https://gcc.gnu.org/g:aa3e72f943032e5f074b2bd2fd06d130dda8760b commit r15-9463-gaa3e72f943032e5f074b2bd2fd06d130dda8760b Author: Thomas Schwinge Date: Thu Mar 27 23:06:37 2025 +0100 Add test cases for exception handling constructs in dead code for GCN, nvptx target and OpenMP 'target' offloading [PR118794] PR target/118794 gcc/testsuite/ * g++.target/gcn/exceptions-pr118794-1.C: New. * g++.target/nvptx/exceptions-pr118794-1.C: Likewise. libgomp/ * testsuite/libgomp.c++/target-exceptions-pr118794-1.C: New. * testsuite/libgomp.c++/target-exceptions-pr118794-1-offload-sorry-GCN.C: Likewise. * testsuite/libgomp.c++/target-exceptions-pr118794-1-offload-sorry-nvptx.C: Likewise.
[Bug target/118794] The attached c++ openmp offload code fails, because the c sqrt function makes nonlocal gotos..
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118794 Thomas Schwinge changed: What|Removed |Added Last reconfirmed||2025-03-27 Keywords||openacc, openmp Ever confirmed|0 |1 Target|nvptx |GCN, nvptx Assignee|unassigned at gcc dot gnu.org |tschwinge at gcc dot gnu.org CC||tschwinge at gcc dot gnu.org Status|UNCONFIRMED |ASSIGNED --- Comment #10 from Thomas Schwinge --- Indeed for '-O0' compilation, you currently need host-side '-fno-exceptions' to prevent GCN as well as nvptx offloading compilation to fail with 'sorry, unimplemented: exception handling not supported', which I'm aware we should resolve properly. With optimizations enabled, or '-O0 -fno-exceptions', this code now works with (the upcoming) GCC 15: compiles, executes successfully (tested Nvidia GPU), produces output (that I've not verified in detail, but have 'diff'ed to host output: matches, apart from a few numeric differences for E-14 etc. numbers; probably rounding differences).
[Bug target/118794] The attached c++ openmp offload code fails, because the c sqrt function makes nonlocal gotos..
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118794 Richard Biener changed: What|Removed |Added Target||nvptx --- Comment #9 from Richard Biener --- Hmm, try -fno-exceptions? It might be because of SJLJ EH - what target are you offloading to? I suppose nvptx? Try declaring sqrt as noexcept (or nothrow, don't know C++ enough off-head).
[Bug target/118794] The attached c++ openmp offload code fails, because the c sqrt function makes nonlocal gotos..
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118794
--- Comment #8 from Benjamin Schulz ---
with this here it is satisfied:
// Normalize v
T norm=0;
// T norm=fabs(gpu_dot_product_w(v,v));
T normc= sqrt(norm);
// const T normc=norm;
#pragma omp parallel for
for (size_t i = 0; i < pext0; ++i)
{
v(i,pstrv0)= v(i,pstrv0)/normc;
}
However, gpu_dot_product_w is called before asT
dot_pr=gpu_dot_product_w(u,v);
which does not throw any nonlocal gotos. it is defined like this:
template
inline T gpu_dot_product_w( const datastruct& vec1, const datastruct
&vec2)
{
const size_t n=vec1.pextents[0];
const size_t strv1=vec1.pstrides[0];
const size_t strv2=vec2.pstrides[0];
T result=0;
#pragma omp parallel for reduction(+:result)
for (size_t i = 0; i < n; ++i)
{
result += vec1(i,strv1) * vec2(i,strv2);
}
return result;
}
and the operators are these:
#pragma omp begin declare target
template
inline T& datastruct::operator()(const size_t row, const size_t stride)
{
return pdata[row * stride];
}
#pragma omp end declare target
none of this has anything to do with gotos pstrides are pointers to non-stl
arrays
i do not know what is going on here...
[Bug target/118794] The attached c++ openmp offload code fails, because the c sqrt function makes nonlocal gotos..
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118794
--- Comment #7 from Benjamin Schulz ---
with this here it is satisfied:
// Normalize v
T norm=0;
// T norm=fabs(gpu_dot_product_w(v,v));
T normc= sqrt(norm);
// const T normc=norm;
#pragma omp parallel for
for (size_t i = 0; i < pext0; ++i)
{
v(i,pstrv0)= v(i,pstrv0)/normc;
}
However, gpu_dot_product_w is called before asT
dot_pr=gpu_dot_product_w(u,v);
which does not throw any nonlocal gotos. it is defined like this:
template
inline T gpu_dot_product_w( const datastruct& vec1, const datastruct
&vec2)
{
const size_t n=vec1.pextents[0];
const size_t strv1=vec1.pstrides[0];
const size_t strv2=vec2.pstrides[0];
T result=0;
#pragma omp parallel for reduction(+:result)
for (size_t i = 0; i < n; ++i)
{
result += vec1(i,strv1) * vec2(i,strv2);
}
return result;
}
and the operators are these:
#pragma omp begin declare target
template
inline T& datastruct::operator()(const size_t row, const size_t stride)
{
return pdata[row * stride];
}
#pragma omp end declare target
none of this has anything to do with gotos pstrides are pointers to non-stl
arrays
i do not know what is going on here...
[Bug target/118794] The attached c++ openmp offload code fails, because the c sqrt function makes nonlocal gotos..
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118794
--- Comment #6 from Benjamin Schulz ---
Hi thanks for the fast reply. Unfortunately none of these works...
(yes, putting in the -fno-math-errno option also raises this error, even if i
put it into -offload...
even if i try -foffload= -fno-math-errno the assert also does not work. and the
builtin unreachable option does also not work.
One problem is that there is no unsigned double in c++...
Strangely, not even this here compiles:
T norm=fabs(gpu_dot_product_w(v,v));
T normc= sqrt(norm);
// const T normc=norm;
#pragma omp parallel for
for (size_t i = 0; i < pext0; ++i)
{
v(i,pstrv0)= v(i,pstrv0)/normc;
}
Is there another division by zero problem?
I do not really know what is going on here..
[Bug target/118794] The attached c++ openmp offload code fails, because the c sqrt function makes nonlocal gotos..
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118794 --- Comment #4 from Andrew Pinski --- Try -fno-math-errno ? Or add: [[assert(norm>=0)]]; or: if (norm>=0) __builtin_unreachable();
[Bug target/118794] The attached c++ openmp offload code fails, because the c sqrt function makes nonlocal gotos..
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118794 --- Comment #5 from Andrew Pinski --- (In reply to Andrew Pinski from comment #4) > Try -fno-math-errno ? Or add: > [[assert(norm>=0)]]; > > or: > if (norm>=0) > __builtin_unreachable(); Sorry the if statement should have been: if (norm<0) __builtin_unreachable(); But you should get the idea.
[Bug target/118794] The attached c++ openmp offload code fails, because the c sqrt function makes nonlocal gotos..
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118794 --- Comment #3 from Benjamin Schulz --- gcc --version gcc (Gentoo 14.2.1_p20241221 p7) 14.2.1 20241221
[Bug target/118794] The attached c++ openmp offload code fails, because the c sqrt function makes nonlocal gotos..
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118794 --- Comment #2 from Benjamin Schulz --- Created attachment 60424 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=60424&action=edit cmakelists.txt
[Bug target/118794] The attached c++ openmp offload code fails, because the c sqrt function makes nonlocal gotos..
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118794 --- Comment #1 from Benjamin Schulz --- Created attachment 60423 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=60423&action=edit main.cpp
