[PATCH] D60283: [clang-cl] Don't emit checksums when compiling a preprocessed CPP

[Alexandre Ganea via Phabricator via cfe-commits]

aganea created this revision.
aganea added reviewers: rnk, scott.linder, uabelho, aprantl.
Herald added a project: clang.
Herald added a subscriber: cfe-commits.

When compiling from an already preprocessed CPP, the checksums generated in the debug info would be those of the (input) preprocessed CPP, not those of the actual #includes.
Note how all the files have the same checksum below:

  ** Module: "E:\RD\tests\clang_hash\main.obj"
  
       0 E:\RD\tests\clang_hash\lib.h (MD5: 136293700AE501A1FB76EBD273C8D288)
       1 C:\Program Files (x86)\Windows Kits\10\Include\10.0.17763.0\ucrt\stdio.h (MD5: 136293700AE501A1FB76EBD273C8D288)
       2 E:\RD\tests\clang_hash\main.cpp (MD5: 136293700AE501A1FB76EBD273C8D288)
       3 C:\Program Files (x86)\Windows Kits\10\Include\10.0.17763.0\ucrt\corecrt_stdio_config.h (MD5: 136293700AE501A1FB76EBD273C8D288)

When debugging in Visual Studio an EXE linked with the OBJ above, Visual Studio complains about the source files being different from when they were compiled, since the sources are compared by hash.

This patch simply clears the checksums to match MSVC behavior. Visual Studio will simply bypass the hash checking in that case, and hapily open the source file.

  ** Module: "E:\RD\tests\clang_hash\main.obj"
  
       0 c:\program files (x86)\windows kits\10\include\10.0.17763.0\ucrt\stdio.h (None)
       1 c:\program files (x86)\windows kits\10\include\10.0.17763.0\ucrt\corecrt_wstdio.h (None)
       2 c:\program files (x86)\windows kits\10\include\10.0.17763.0\ucrt\corecrt_stdio_config.h (None)
       3 c:\program files (x86)\microsoft visual studio\2017\professional\vc\tools\msvc\14.16.27023\include\vcruntime_new.h (None)
       4 e:\rd\tests\clang_hash\main.cpp (None)
       5 e:\rd\tests\clang_hash\lib.h (None)

We could write more complicated code to open the files from the corresponding #line directives, but we have no guarantee the preprocessed CPP is compiled in the same conditions as when it was pre-processed. Actually, when using Fastbuild, the files are preprocessed locally on the user's PC; then the preprocessed content is sent and compiled remotely on another PC that does not have the source code.

Fixes PR41215


Repository:
  rC Clang

https://reviews.llvm.org/D60283

Files:
  lib/CodeGen/CGDebugInfo.cpp
  test/Driver/cl-preprocess-md5.cpp


Index: test/Driver/cl-preprocess-md5.cpp
===================================================================
--- test/Driver/cl-preprocess-md5.cpp
+++ test/Driver/cl-preprocess-md5.cpp
@@ -0,0 +1,15 @@
+// RUN: %clang_cl /c %s /Z7 /Fo%t.obj
+// RUN: llvm-pdbutil dump -l %t.obj | grep -E " \(MD5: ([0-9A-F]+)\)" | sed -n -e 's/^.*MD5: //p' | sort | uniq | wc -l | FileCheck %s --check-prefix=GOOD-SIZE
+// GOOD-SIZE-NOT: 1
+
+// RUN: %clang_cl /c %s /Z7 /E >%t.cpp
+// RUN: %clang_cl /c %t.cpp /Z7 /Fo%t.obj
+// RUN: llvm-pdbutil dump -l %t.obj | not grep -E " \(MD5: ([0-9A-F]+)\)"
+// RUN: llvm-pdbutil dump -l %t.obj | grep " (no checksum)"
+// 
+
+#include <string>
+int main() {
+  std::string A{"a"};
+  return 0;
+}
Index: lib/CodeGen/CGDebugInfo.cpp
===================================================================
--- lib/CodeGen/CGDebugInfo.cpp
+++ lib/CodeGen/CGDebugInfo.cpp
@@ -421,13 +421,18 @@
       return cast<llvm::DIFile>(V);
   }
 
+  const FileID foundIdFromLoc = SM.getFileID(Loc);
   SmallString<32> Checksum;
-  Optional<llvm::DIFile::ChecksumKind> CSKind =
-      computeChecksum(SM.getFileID(Loc), Checksum);
+  Optional<llvm::DIFile::ChecksumKind> CSKind;
   Optional<llvm::DIFile::ChecksumInfo<StringRef>> CSInfo;
-  if (CSKind)
-    CSInfo.emplace(*CSKind, Checksum);
-  return createFile(FileName, CSInfo, getSource(SM, SM.getFileID(Loc)));
+  const FileEntry *fileEntry = SM.getFileEntryForID(foundIdFromLoc);
+  if (!fileEntry || fileEntry->getName().empty() ||
+      fileEntry->getName().equals(FileName)) {
+    CSKind = computeChecksum(foundIdFromLoc, Checksum);
+    if (CSKind)
+      CSInfo.emplace(*CSKind, Checksum);
+  }
+  return createFile(FileName, CSInfo, getSource(SM, foundIdFromLoc));
 }
 
 llvm::DIFile *


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D60283.193764.patch
Type: text/x-patch
Size: 1759 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20190404/dc3c641b/attachment.bin>