[SandboxVec] Add option -sbvec-allow-file for bisection debugging

[vporpo]

This new option lets you specify an allow-list of source files and disables vectorization if the IR is not in the list. This can be used for debugging miscompiles.

[llvmbot]

@llvm/pr-subscribers-vectorizers

@llvm/pr-subscribers-llvm-transforms

Author: vporpo (vporpo)

Changes

This new option lets you specify an allow-list of source files and disables vectorization if the IR is not in the list. This can be used for debugging miscompiles.

---

Full diff: https://github.com/llvm/llvm-project/pull/129127.diff

3 Files Affected:

• (modified) llvm/include/llvm/Transforms/Vectorize/SandboxVectorizer/SandboxVectorizer.h (+3)
• (modified) llvm/lib/Transforms/Vectorize/SandboxVectorizer/SandboxVectorizer.cpp (+42)
• (added) llvm/test/Transforms/SandboxVectorizer/allow_files.ll (+39)

diff --git a/llvm/include/llvm/Transforms/Vectorize/SandboxVectorizer/SandboxVectorizer.h b/llvm/include/llvm/Transforms/Vectorize/SandboxVectorizer/SandboxVectorizer.h
index 7ea9386f08bee..fea53329719b9 100644
--- a/llvm/include/llvm/Transforms/Vectorize/SandboxVectorizer/SandboxVectorizer.h
+++ b/llvm/include/llvm/Transforms/Vectorize/SandboxVectorizer/SandboxVectorizer.h
@@ -37,6 +37,9 @@ class SandboxVectorizerPass : public PassInfoMixin<SandboxVectorizerPass> {
   // within FPM may register/unregister callbacks, so they need access to
   // Context.
   sandboxir::FunctionPassManager FPM;
+  /// \Returns true if we should attempt to vectorize \p SrcFilePath based on
+  /// `AllowFiles` option.
+  bool allowFile(const std::string &SrcFilePath);
 
   bool runImpl(Function &F);
 
diff --git a/llvm/lib/Transforms/Vectorize/SandboxVectorizer/SandboxVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SandboxVectorizer/SandboxVectorizer.cpp
index 5837cc16fcbac..bffb9f187e882 100644
--- a/llvm/lib/Transforms/Vectorize/SandboxVectorizer/SandboxVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/SandboxVectorizer/SandboxVectorizer.cpp
@@ -8,9 +8,11 @@
 
 #include "llvm/Transforms/Vectorize/SandboxVectorizer/SandboxVectorizer.h"
 #include "llvm/Analysis/TargetTransformInfo.h"
+#include "llvm/IR/Module.h"
 #include "llvm/SandboxIR/Constant.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Transforms/Vectorize/SandboxVectorizer/SandboxVectorizerPassBuilder.h"
+#include <regex>
 
 using namespace llvm;
 
@@ -29,6 +31,22 @@ static cl::opt<std::string> UserDefinedPassPipeline(
     cl::desc("Comma-separated list of vectorizer passes. If not set "
              "we run the predefined pipeline."));
 
+// This option is useful for bisection debugging.
+// For example you may use it to figure out which filename is the one causing a
+// miscompile. You can specify a regex for the filename like: "/[a-m][^/]*"
+// which will enable any file name starting with 'a' to 'm' and disable the
+// rest. If the miscompile goes away, then we try "/[n-z][^/]*" for the other
+// half of the range, from 'n' to 'z'. If we can reproduce the miscompile then
+// we can keep looking in [n-r] and [s-z] and so on, in a binary-search fashion.
+//
+// Please note that we are using [^/]* and not .* to make sure that we are
+// matching the actual filename and not some other directory in the path.
+cl::opt<std::string> AllowFiles(
+    "sbvec-allow-files", cl::init(".*"), cl::Hidden,
+    cl::desc("Run the vectorizer only on file paths that match any in the "
+             "list of comma-separated regex's."));
+static constexpr const char AllowFilesDelim = ',';
+
 SandboxVectorizerPass::SandboxVectorizerPass() : FPM("fpm") {
   if (UserDefinedPassPipeline == DefaultPipelineMagicStr) {
     // TODO: Add passes to the default pipeline. It currently contains:
@@ -66,6 +84,23 @@ PreservedAnalyses SandboxVectorizerPass::run(Function &F,
   return PA;
 }
 
+bool SandboxVectorizerPass::allowFile(const std::string &SrcFilePath) {
+  // Iterate over all files in AllowFiles separated by `AllowFilesDelim`.
+  size_t DelimPos = 0;
+  do {
+    size_t LastPos = DelimPos != 0 ? DelimPos + 1 : DelimPos;
+    DelimPos = AllowFiles.find(AllowFilesDelim, LastPos);
+    auto FileNameToMatch = AllowFiles.substr(LastPos, DelimPos - LastPos);
+    if (FileNameToMatch.empty())
+      return false;
+    // Note: This only runs when debugging so its OK not to reuse the regex.
+    std::regex FileNameRegex(std::string(".*") + FileNameToMatch);
+    if (std::regex_match(SrcFilePath, FileNameRegex))
+      return true;
+  } while (DelimPos != std::string::npos);
+  return false;
+}
+
 bool SandboxVectorizerPass::runImpl(Function &LLVMF) {
   if (Ctx == nullptr)
     Ctx = std::make_unique<sandboxir::Context>(LLVMF.getContext());
@@ -75,6 +110,13 @@ bool SandboxVectorizerPass::runImpl(Function &LLVMF) {
     return false;
   }
 
+  // This is used for debugging.
+  if (LLVM_UNLIKELY(AllowFiles != ".*")) {
+    const auto &SrcFilePath = LLVMF.getParent()->getSourceFileName();
+    if (!allowFile(SrcFilePath))
+      return false;
+  }
+
   // If the target claims to have no vector registers early return.
   if (!TTI->getNumberOfRegisters(TTI->getRegisterClassForType(true))) {
     LLVM_DEBUG(dbgs() << "SBVec: Target has no vector registers, return.\n");
diff --git a/llvm/test/Transforms/SandboxVectorizer/allow_files.ll b/llvm/test/Transforms/SandboxVectorizer/allow_files.ll
new file mode 100644
index 0000000000000..0929eca6a1047
--- /dev/null
+++ b/llvm/test/Transforms/SandboxVectorizer/allow_files.ll
@@ -0,0 +1,39 @@
+; RUN: opt -passes=sandbox-vectorizer -sbvec-vec-reg-bits=1024 -sbvec-allow-non-pow2 -sbvec-passes="seed-collection<tr-save,bottom-up-vec,tr-accept>" -sbvec-allow-files="some_other_file" %s -S | FileCheck %s --check-prefix=ALLOW_OTHER
+; RUN: opt -passes=sandbox-vectorizer -sbvec-vec-reg-bits=1024 -sbvec-allow-non-pow2 -sbvec-passes="seed-collection<tr-save,bottom-up-vec,tr-accept>" -sbvec-allow-files="allow_files.ll" %s -S | FileCheck %s --check-prefix=ALLOW_THIS
+; RUN: opt -passes=sandbox-vectorizer -sbvec-vec-reg-bits=1024 -sbvec-allow-non-pow2 -sbvec-passes="seed-collection<tr-save,bottom-up-vec,tr-accept>" -sbvec-allow-files="al.*_files.ll" %s -S | FileCheck %s --check-prefix=ALLOW_REGEX
+; RUN: opt -passes=sandbox-vectorizer -sbvec-vec-reg-bits=1024 -sbvec-allow-non-pow2 -sbvec-passes="seed-collection<tr-save,bottom-up-vec,tr-accept>" -sbvec-allow-files="some_file,.*_files.ll,some_other_file" %s -S | FileCheck %s --check-prefix=ALLOW_REGEX_CSV
+; RUN: opt -passes=sandbox-vectorizer -sbvec-vec-reg-bits=1024 -sbvec-allow-non-pow2 -sbvec-passes="seed-collection<tr-save,bottom-up-vec,tr-accept>" -sbvec-allow-files="allow" %s -S | FileCheck %s --check-prefix=ALLOW_BAD_REGEX
+; RUN: opt -passes=sandbox-vectorizer -sbvec-vec-reg-bits=1024 -sbvec-allow-non-pow2 -sbvec-passes="seed-collection<tr-save,bottom-up-vec,tr-accept>" -sbvec-allow-files="some_file,some_other_file1,some_other_file2" %s -S | FileCheck %s --check-prefix=ALLOW_OTHER_CSV
+; RUN: opt -passes=sandbox-vectorizer -sbvec-vec-reg-bits=1024 -sbvec-allow-non-pow2 -sbvec-passes="seed-collection<tr-save,bottom-up-vec,tr-accept>" -sbvec-allow-files="" %s -S | FileCheck %s --check-prefix=ALLOW_EMPTY
+; RUN: opt -passes=sandbox-vectorizer -sbvec-vec-reg-bits=1024 -sbvec-allow-non-pow2 -sbvec-passes="seed-collection<tr-save,bottom-up-vec,tr-accept>" %s -S | FileCheck %s --check-prefix=DEFAULT
+
+; Checks the command-line option `-sbvec-allow-files`.
+define void @widen(ptr %ptr) {
+; ALLOW_OTHER:     store float {{%.*}}, ptr {{%.*}}, align 4
+; ALLOW_OTHER:     store float {{%.*}}, ptr {{%.*}}, align 4
+;
+; ALLOW_THIS:      store <2 x float> {{%.*}}, ptr {{%.*}}, align 4
+;
+; ALLOW_REGEX:     store <2 x float> {{%.*}}, ptr {{%.*}}, align 4
+;
+; ALLOW_REGEX_CSV: store <2 x float> {{%.*}}, ptr {{%.*}}, align 4
+;
+; ALLOW_BAD_REGEX: store float {{%.*}}, ptr {{%.*}}, align 4
+; ALLOW_BAD_REGEX: store float {{%.*}}, ptr {{%.*}}, align 4
+;
+; ALLOW_OTHER_CSV: store float {{%.*}}, ptr {{%.*}}, align 4
+; ALLOW_OTHER_CSV: store float {{%.*}}, ptr {{%.*}}, align 4
+;
+; ALLOW_EMPTY:     store float {{%.*}}, ptr {{%.*}}, align 4
+; ALLOW_EMPTY:     store float {{%.*}}, ptr {{%.*}}, align 4
+;
+; DEFAULT:         store <2 x float> {{%.*}}, ptr {{%.*}}, align 4
+;
+  %ptr0 = getelementptr float, ptr %ptr, i32 0
+  %ptr1 = getelementptr float, ptr %ptr, i32 1
+  %ld0 = load float, ptr %ptr0
+  %ld1 = load float, ptr %ptr1
+  store float %ld0, ptr %ptr0
+  store float %ld1, ptr %ptr1
+  ret void
+}