Unable to build with -O0 -mllvm -enzyme-auto-sparsity=1 #2366
Description
I am unable to build with-O0 -mllvm -enzyme-auto-sparsity=1. Higher optimization levels work.
A simple reproducer is ringspring.cpp in the sparse integration tests.
The error message is:
/app/Enzyme/enzyme/Enzyme/FunctionUtils.cpp:8041: void replaceToDense(llvm::CallBase*, bool, llvm::Function*, const llvm::DataLayout&): Assertion `SI->getValueOperand() != val' failed.
I'm including a link to compiler explorer, the source code and a the detailed error message below.
Source code for reference
// Compilation flags
// -O0 -fno-vectorize -ffast-math -fno-unroll-loops -mllvm -enable-load-pre=0 -mllvm -enzyme-auto-sparsity=1
// Compilation fails with -O0. Works with higher optimization levels, e.g., -01.
#include <stdio.h>
#include <assert.h>
#include <vector>
#include <math.h>
#include <cmath>
#include <sys/time.h>
float tdiff(struct timeval *start, struct timeval *end) {
return (end->tv_sec-start->tv_sec) + 1e-6*(end->tv_usec-start->tv_usec);
}
template<typename T>
struct Triple {
size_t row;
size_t col;
T val;
Triple(Triple&&) = default;
Triple(size_t row, size_t col, T val) : row(row), col(col), val(val) {}
};
__attribute__((enzyme_sparse_accumulate))
static void inner_storeflt(int64_t row, int64_t col, float val, std::vector<Triple<float>> &triplets) {
#ifdef BENCHMARK
if (val == 0.0) return;
#else
#warning "Compiling for debug/verfication, performance may be slowed"
#endif
triplets.emplace_back(row, col, val);
}
__attribute__((enzyme_sparse_accumulate))
static void inner_storedbl(int64_t row, int64_t col, double val, std::vector<Triple<double>> &triplets) {
#ifdef BENCHMARK
if (val == 0.0) return;
#else
#warning "Compiling for debug/verfication, performance may be slowed"
#endif
triplets.emplace_back(row, col, val);
}
template<typename T>
__attribute__((always_inline))
static void sparse_store(T val, int64_t idx, size_t i, std::vector<Triple<T>> &triplets) {
if (val == 0.0) return;
idx /= sizeof(T);
if constexpr (sizeof(T) == 4)
inner_storeflt(i, idx, val, triplets);
else
inner_storedbl(i, idx, val, triplets);
}
template<typename T>
__attribute__((always_inline))
static T sparse_load(int64_t idx, size_t i, std::vector<Triple<T>> &triplets) {
return 0.0;
}
template<typename T>
__attribute__((always_inline))
static void ident_store(T, int64_t idx, size_t i) {
assert(0 && "should never load");
}
template<typename T>
__attribute__((always_inline))
static T ident_load(int64_t idx, size_t i) {
idx /= sizeof(T);
return (T)(idx == i);// ? 1.0 : 0.0;
}
extern int enzyme_width;
extern int enzyme_dup;
extern int enzyme_dupv;
extern int enzyme_const;
extern int enzyme_dupnoneed;
template <typename T, typename... Tys>
extern T __enzyme_autodiff(void*, Tys...) noexcept;
template <typename T, typename... Tys>
extern T __enzyme_fwddiff(void *, Tys...) noexcept;
template <typename T, typename... Tys>
extern T __enzyme_todense(Tys...) noexcept;
template <typename T, typename... Tys>
extern T __enzyme_post_sparse_todense(Tys...) noexcept;
template<typename T, size_t n>
__attribute__((always_inline))
static void elementwise_difference(T (&out)[n], const T x[n], const T y[n]) {
#pragma clang loop unroll(full)
for (int i=0; i<n; i++)
out[i] = x[i] - y[i];
}
template<typename T, size_t n>
__attribute__((always_inline))
static void elementwise_sum(T (&out)[n], const T x[n], const T y[n]) {
#pragma clang loop unroll(full)
for (int i=0; i<n; i++)
out[i] = x[i] + y[i];
}
template<typename T, size_t n>
__attribute__((always_inline))
static T dot_product(const T a[n], const T b[n]) {
T result = 0.0;
#pragma clang loop unroll(full)
for (size_t i = 0; i < n; ++i) {
result += a[i] * b[i];
}
return result;
}
template<typename T, size_t n>
__attribute__((always_inline))
static T norm(const T v[n]) {
T sum_squares = 0.0;
#pragma clang loop unroll(full)
for (size_t i=0; i<n; i++) {
T val = v[i];
sum_squares += val * val;
}
return std::sqrt(sum_squares);
}
template<typename T, size_t n, size_t m>
__attribute__((always_inline))
static void transpose(T (&out)[n][m], const T in[m][n]) {
#pragma clang loop unroll(full)
for (int i=0; i<n; i++)
#pragma clang loop unroll(full)
for (int j=0; j<m; j++)
out[i][j] = in[j][i];
}
template<typename T, size_t m, size_t n, size_t k>
__attribute__((always_inline))
static void matrix_multiply(T (&result)[m][k], const T matrix1[m][n], const T matrix2[n][k]) {
#pragma clang loop unroll(full)
for (int i = 0; i < m; ++i) {
#pragma clang loop unroll(full)
for (int j = 0; j < k; ++j) {
result[i][j] = 0.0;
#pragma clang loop unroll(full)
for (int z = 0; z < n; ++z) {
result[i][j] += matrix1[i][z] * matrix2[z][j];
}
}
}
}
template<typename T>
__attribute__((always_inline))
static void inv(T (&out)[3][3], const T (&F)[3][3]) {
T det = F[0][0] * (F[1][1] * F[2][2] - F[1][2] * F[2][1])
- F[0][1] * (F[1][0] * F[2][2] - F[1][2] * F[2][0])
+ F[0][2] * (F[1][0] * F[2][1] - F[1][1] * F[2][0]);
T inv_det = 1 / det;
out[0][0] = (F[1][1] * F[2][2] - F[1][2] * F[2][1]) * inv_det;
out[0][1] = (F[0][2] * F[2][1] - F[0][1] * F[2][2]) * inv_det;
out[0][2] = (F[0][1] * F[1][2] - F[0][2] * F[1][1]) * inv_det;
out[1][0] = (F[1][2] * F[2][0] - F[1][0] * F[2][2]) * inv_det;
out[1][1] = (F[0][0] * F[2][2] - F[0][2] * F[2][0]) * inv_det;
out[1][2] = (F[0][2] * F[1][0] - F[0][0] * F[1][2]) * inv_det;
out[2][0] = (F[1][0] * F[2][1] - F[1][1] * F[2][0]) * inv_det;
out[2][1] = (F[0][1] * F[2][0] - F[0][0] * F[2][1]) * inv_det;
out[2][2] = (F[0][0] * F[1][1] - F[0][1] * F[1][0]) * inv_det;
}
template<typename T>
__attribute__((always_inline))
static void inv(T (&out)[2][2], const T (&F)[2][2]) {
T det = F[0][0] * F[1][1] - F[0][1] * F[1][0];
T inv_det = 1 / det;
out[0][0] = F[1][1] * inv_det;
out[0][1] = -F[0][1] * inv_det;
out[1][0] = -F[1][0] * inv_det;
out[1][1] = F[0][0] * inv_det;
}
template<typename T, size_t m, size_t n>
__attribute__((always_inline))
static void pseudo_inverse(T (&matTsqrinv)[n][m], const T mat[m][n]) {
T matT[n][m];
transpose(matT, mat);
T matmatT[m][m];
matrix_multiply(matmatT, mat, matT);
T sqrinv[m][m];
inv(sqrinv, matmatT);
matrix_multiply(matTsqrinv, matT, sqrinv);
}
// m is 2 n is 3
template<typename T, int n, int m>
__attribute__((always_inline))
static void get_pos(
T (&__restrict__ out)[n][m],
const float *__restrict__ pos,
const int idx[n]) {
static_assert(m == 3, "Only Vector3 is supported");
// extract the 3d points at idx[0], idx[1], idx[2], idx[3]
#pragma clang loop unroll(full)
for (int i = 0; i < n; ++i) {
out[i][0] = pos[m * idx[i]];
out[i][1] = pos[m * idx[i] + 1];
out[i][2] = pos[m * idx[i] + 2];
}
}
// m is 2 n is 3
template<typename T, int n, int m>
__attribute__((always_inline))
static void get_pos_affine(
T (&__restrict__ out)[n][m],
const float *__restrict__ pos) {
static_assert(m == 3, "Only Vector3 is supported");
// extract the 3d points at idx[0], idx[1], idx[2], idx[3]
#pragma clang loop unroll(full)
for (int i = 0; i < n; ++i) {
out[i][0] = pos[m * i];
out[i][1] = pos[m * i + 1];
out[i][2] = pos[m * i + 2];
}
}
template<typename T>
__attribute__((always_inline))
static void cross(T (&out)[3], const T v1[3], const T v2[3]) {
out[0] = v1[1] * v2[2] - v1[2] * v2[1];
out[1] = v1[2] * v2[0] - v1[0] * v2[2];
out[2] = v1[0] * v2[1] - v1[1] * v2[0];
}
template<typename T>
__attribute__((always_inline))
static T area(const T *__restrict__ u, const T *__restrict__ v, const T *__restrict__ w) {
T diff1[3];
elementwise_difference(diff1, v, u);
T diff2[3];
elementwise_difference(diff2, w, u);
T cross_product[3];
cross(cross_product, diff1, diff2);
return 0.5 * norm<T, 3>(cross_product);
}
template<typename T>
__attribute__((always_inline))
static T f(size_t N, T* input) {
double out = 0;
// __builtin_assume(!((N-1) == 0));
for (size_t i=0; i<N; i++) {
//double sub = input[i] - input[i+1];
// out += sub * sub;
T sub = (input[i+1] - input[i]) * (input[i+1] - input[i]);
out += (sqrt(sub) - 1)*(sqrt(sub) - 1);
}
return out;
}
template<typename T>
__attribute__((always_inline))
static void grad_f(size_t N, T* input, T* dinput) {
__enzyme_autodiff<void>((void*)f<T>, enzyme_const, N, enzyme_dup, input, dinput);
}
template<typename T>
__attribute__((always_inline))
double ringident_load(int64_t idx, size_t i, size_t N) {
idx /= sizeof(double);
// return (double)( ( (idx == N) ? 0 : idx) == i);
return (double)((idx != N && idx == i) || (idx == N && 0 == i));
// return (double)( idx % N == i);
}
template<typename T>
__attribute__((always_inline))
void never_store(T val, int64_t idx, T* input, size_t N) {
assert(0 && "this is a read only input, why are you storing here...");
}
template<typename T>
__attribute__((always_inline))
double mod_load(int64_t idx, double* input, size_t N) {
idx /= sizeof(double);
return input[idx % N];
}
template<typename T>
__attribute__((noinline))
std::vector<Triple<T>> hess_f(size_t N, T* input) {
std::vector<Triple<T>> triplets;
input = __enzyme_todense<T*>((void*)mod_load<T>, (void*)never_store<T>, input, N);
__builtin_assume(N > 0);
__builtin_assume(N != 1);
for (size_t i=0; i<N; i++) {
__builtin_assume(i < 100000000);
T* d_input = __enzyme_todense<T*>((void*)ringident_load<T>, (void*)never_store<T>, i, N);
T* d_dinput = __enzyme_todense<T*>((void*)sparse_load<T>, (void*)sparse_store<T>, i, &triplets);
__enzyme_fwddiff<void>((void*)grad_f<T>,
enzyme_const, N,
enzyme_dup, input, d_input,
enzyme_dupnoneed, (T*)0x1, d_dinput);
}
return triplets;
}
int main(int argc, char** argv) {
size_t N = 8;
if (argc >= 2) {
N = atoi(argv[1]);
}
double *x = (double*)malloc(sizeof(double) * N);
for (int i=0; i<N; i++) x[i] = (i + 1) * (i + 1);
struct timeval start, end;
gettimeofday(&start, NULL);
auto res = hess_f(N, x);
gettimeofday(&end, NULL);
printf("Number of elements %ld\n", res.size());
printf("Runtime %0.6f\n", tdiff(&start, &end));
if (N <= 30) {
for (auto & tup : res)
printf("%ld, %ld = %f\n", tup.row, tup.col, tup.val);
}
return 0;
}Error message details
clang++: /app/Enzyme/enzyme/Enzyme/FunctionUtils.cpp:8041: void replaceToDense(llvm::CallBase*, bool, llvm::Function*, const llvm::DataLayout&): Assertion `SI->getValueOperand() != val' failed.
PLEASE submit a bug report to https://github.com/llvm/llvm-project/issues/ and include the crash backtrace, preprocessed source, and associated run script.
Stack dump:
0. Program arguments: /opt/compiler-explorer/clang-16.0.0/bin/clang++ -gdwarf-4 -g -o /tmp/compiler-explorer-compiler2025528-98-1f78sgi.0wxqk/output.s -fno-verbose-asm -c -fpass-plugin=/opt/compiler-explorer/main/ClangEnzyme-16.so -Xclang -load -Xclang /opt/compiler-explorer/main/ClangEnzyme-16.so -fcolor-diagnostics -fno-crash-diagnostics -O0 -fno-vectorize -ffast-math -fno-unroll-loops -mllvm -enable-load-pre=0 -mllvm -enzyme-auto-sparsity=1
- parser at end of file
- Optimizer
#0 0x0000555b5e20a8df llvm::sys::PrintStackTrace(llvm::raw_ostream&, int) (/opt/compiler-explorer/clang-16.0.0/bin/clang+++0x3ba18df)
Build a Julia compatible plugin library #1 0x0000555b5e20890c llvm::sys::CleanupOnSignal(unsigned long) (/opt/compiler-explorer/clang-16.0.0/bin/clang+++0x3b9f90c)
Fix CMake to allow relative path and add version # to library #2 0x0000555b5e15b518 CrashRecoverySignalHandler(int) CrashRecoveryContext.cpp:0:0
Better CI for LLVM and Julia #3 0x00007f2513021520 (/lib/x86_64-linux-gnu/libc.so.6+0x42520)
Fix the incredibly stupid plug AssertingVH bug #4 0x00007f25130759fc pthread_kill (/lib/x86_64-linux-gnu/libc.so.6+0x969fc)
Tfk maxnum #5 0x00007f2513021476 gsignal (/lib/x86_64-linux-gnu/libc.so.6+0x42476)
Incorrect derivative on -O0 #6 0x00007f25130077f3 abort (/lib/x86_64-linux-gnu/libc.so.6+0x287f3)
Fix active variable bug #7 0x00007f251300771b (/lib/x86_64-linux-gnu/libc.so.6+0x2871b)
Compiler crash with cannot deal with ptr that isn't arg #8 0x00007f2513018e96 (/lib/x86_64-linux-gnu/libc.so.6+0x39e96)
Handle select of pointers #9 0x00007f2512a3265b replaceToDense(llvm::CallBase*, bool, llvm::Function*, llvm::DataLayout const&) /app/Enzyme/enzyme/Enzyme/FunctionUtils.cpp:8042:62
Disable preprocessing optimizations for most tests #10 0x00007f2512a3330b LowerSparsification(llvm::Function*, bool) /app/Enzyme/enzyme/Enzyme/FunctionUtils.cpp:8128:3
Add memcpy for pointers and floats #11 0x00007f25128e65e8 (anonymous namespace)::EnzymeBase::lowerEnzymeCalls(llvm::Function&, std::set<llvm::Function*, std::lessllvm::Function*, std::allocatorllvm::Function*>&) /app/Enzyme/enzyme/Enzyme/Enzyme.cpp:2926:47
More general for loop structure #12 0x00007f25128e73b5 (anonymous namespace)::EnzymeBase::run(llvm::Module&) /app/Enzyme/enzyme/Enzyme/Enzyme.cpp:3057:15
Cleanup files and split for organization #13 0x00007f25128efd92 EnzymeNewPM::run(llvm::Module&, llvm::AnalysisManagerllvm::Module&) /app/Enzyme/enzyme/Enzyme/Enzyme.cpp:3318:56
More loop stability (including multiple exit loops) #14 0x00007f2512932d4f llvm::detail::PassModel<llvm::Module, EnzymeNewPM, llvm::PreservedAnalyses, llvm::AnalysisManagerllvm::Module>::run(llvm::Module&, llvm::AnalysisManagerllvm::Module&) /opt/compiler-explorer/clang-16.0.0/include/llvm/IR/PassManagerInternal.h:90:3
add functional c tests via check-enzyme-c command #15 0x0000555b5dabbfb9 llvm::PassManager<llvm::Module, llvm::AnalysisManagerllvm::Module>::run(llvm::Module&, llvm::AnalysisManagerllvm::Module&) (/opt/compiler-explorer/clang-16.0.0/bin/clang+++0x3452fb9)
Fix calling convention and other small bugs #16 0x0000555b5e5edb44 (anonymous namespace)::EmitAssemblyHelper::RunOptimizationPipeline(clang::BackendAction, std::unique_ptr<llvm::raw_pwrite_stream, std::default_deletellvm::raw_pwrite_stream>&, std::unique_ptr<llvm::ToolOutputFile, std::default_deletellvm::ToolOutputFile>&) BackendUtil.cpp:0:0
Ensure increment of canonical IV dominates uses in loops #17 0x0000555b5e5f0bd6 clang::EmitBackendOutput(clang::DiagnosticsEngine&, clang::HeaderSearchOptions const&, clang::CodeGenOptions const&, clang::TargetOptions const&, clang::LangOptions const&, llvm::StringRef, llvm::Module*, clang::BackendAction, std::unique_ptr<llvm::raw_pwrite_stream, std::default_deletellvm::raw_pwrite_stream>) (/opt/compiler-explorer/clang-16.0.0/bin/clang+++0x3f87bd6)
cacheloads flag related fixes so insertsort_alt works with cacheloads=true #18 0x0000555b5f4fb238 clang::BackendConsumer::HandleTranslationUnit(clang::ASTContext&) (/opt/compiler-explorer/clang-16.0.0/bin/clang+++0x4e92238)
Fixtfkbranch #19 0x0000555b5ee10b50 clang::MultiplexConsumer::HandleTranslationUnit(clang::ASTContext&) (/opt/compiler-explorer/clang-16.0.0/bin/clang+++0x47a7b50)
[WIP] Handle Non-outermost Dynamic Loops #20 0x0000555b6071fd5d clang::ParseAST(clang::Sema&, bool, bool) (/opt/compiler-explorer/clang-16.0.0/bin/clang+++0x60b6d5d)
Selective cachereads #21 0x0000555b5f4fa865 clang::CodeGenAction::ExecuteAction() (/opt/compiler-explorer/clang-16.0.0/bin/clang+++0x4e91865)
Handle Triangular Loops #22 0x0000555b5edd6671 clang::FrontendAction::Execute() (/opt/compiler-explorer/clang-16.0.0/bin/clang+++0x476d671)
add readwriteread ll test #23 0x0000555b5ed5a993 clang::CompilerInstance::ExecuteAction(clang::FrontendAction&) (/opt/compiler-explorer/clang-16.0.0/bin/clang+++0x46f1993)
Global variables (no testcase yet) #24 0x0000555b5eebab3b clang::ExecuteCompilerInvocation(clang::CompilerInstance*) (/opt/compiler-explorer/clang-16.0.0/bin/clang+++0x4851b3b)
Differential returns that are pointer types #25 0x0000555b5b9e341c cc1_main(llvm::ArrayRef<char const*>, char const*, void*) (/opt/compiler-explorer/clang-16.0.0/bin/clang+++0x137a41c)
Active/Inactive determination (Tim, with input from Billy) #26 0x0000555b5b9debdc ExecuteCC1Tool(llvm::SmallVectorImpl<char const*>&) driver.cpp:0:0
Do not do replaceFunction on a pointer return type #27 0x0000555b5ebbbce9 void llvm::function_ref<void ()>::callback_fn<clang::driver::CC1Command::Execute(llvm::ArrayRef<std::optionalllvm::StringRef>, std::__cxx11::basic_string<char, std::char_traits, std::allocator>, bool) const::'lambda'()>(long) Job.cpp:0:0
Handle use of a better calling convention for combined forward/reverse #28 0x0000555b5e15b977 llvm::CrashRecoveryContext::RunSafely(llvm::function_ref<void ()>) (/opt/compiler-explorer/clang-16.0.0/bin/clang+++0x3af2977)
add differential pointer return c test #29 0x0000555b5ebbbf1c clang::driver::CC1Command::Execute(llvm::ArrayRef<std::optionalllvm::StringRef>, std::__cxx11::basic_string<char, std::char_traits, std::allocator>, bool) const (.part.0) Job.cpp:0:0
Better global handling and fix caching bug #30 0x0000555b5eb843fc clang::driver::Compilation::ExecuteCommand(clang::driver::Command const&, clang::driver::Command const*&, bool) const (/opt/compiler-explorer/clang-16.0.0/bin/clang+++0x451b3fc)
Intsecretpointerpartialfix #31 0x0000555b5eb84ded clang::driver::Compilation::ExecuteJobs(clang::driver::JobList const&, llvm::SmallVectorImpl<std::pair<int, clang::driver::Command const*>>&, bool) const (/opt/compiler-explorer/clang-16.0.0/bin/clang+++0x451bded)
Tx2 #32 0x0000555b5eb8f23c clang::driver::Driver::ExecuteCompilation(clang::driver::Compilation&, llvm::SmallVectorImpl<std::pair<int, clang::driver::Command const*>>&) (/opt/compiler-explorer/clang-16.0.0/bin/clang+++0x452623c)
More advanced activity analysis and many other bugfixes #33 0x0000555b5b9e1523 clang_main(int, char**) (/opt/compiler-explorer/clang-16.0.0/bin/clang+++0x1378523)
Large working changes #34 0x00007f2513008d90 (/lib/x86_64-linux-gnu/libc.so.6+0x29d90)
PHINode (and generally return value), doesn't presently assert failure if no arguments are active if phi is active #35 0x00007f2513008e40 __libc_start_main (/lib/x86_64-linux-gnu/libc.so.6+0x29e40)
Import Julia source and redo CI for 1.3 #36 0x0000555b5b9dad1e _start (/opt/compiler-explorer/clang-16.0.0/bin/clang+++0x1371d1e)
clang-16: error: clang frontend command failed with exit code 134 (use -v to see invocation)
Compiler returned: 134