[PATCH] D79921: [OpenMP] Fix omp and clang pragmas

[ISHIGURO, Hiroshi via Phabricator via cfe-commits]

hishiguro created this revision.
hishiguro added a reviewer: ABataev.
hishiguro added a project: clang.
Herald added subscribers: cfe-commits, guansong, yaxunl.
Herald added a reviewer: jdoerfert.

Fixes PR45753

When a program that contains a loop to which both "omp parallel for" pragma and "clang loop" pragma are associated is compiled with the -fopenmp option, "clang loop" pragma will not take effect.
The example below should not be vectorized by the "clang loop` pragma but is it actually vectorized.
The cause is that "llvm.loop.vectorize.width" is not output to the IR when -fopenmp is specified.
The fix attaches attributes if they exist in the loop.

[example.c]

  int a[100], b[100];
  int foo() {
  #pragma omp parallel for
  #pragma clang loop vectorize(disable)
  for (int i=0; i<100; i++)
    a[i]+=b[i]*i;
  }

[compile]
clang -O2 -fopenmp a.c -c -Rpass=vect
a.c:3:1: remark: vectorized loop (vectorization width: 4, interleaved count: 2) [-Rpass=loop-vectorize] #pragma omp parallel for ^

[IR]

- -fopenmp

$ clang -O2 a.c -S -emit-llvm -mllvm -disable-llvm-optzns -o - -fopenmp |grep "vectorize\.width"
$

- -fno-openmp

$ clang -O2 a.c -S -emit-llvm -mllvm -disable-llvm-optzns -o - -fno-openmp |grep "vectorize\.width"
!7 = !{!"llvm.loop.vectorize.width", i32 1}


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D79921

Files:
  clang/lib/CodeGen/CGStmtOpenMP.cpp


Index: clang/lib/CodeGen/CGStmtOpenMP.cpp
===================================================================
--- clang/lib/CodeGen/CGStmtOpenMP.cpp
+++ clang/lib/CodeGen/CGStmtOpenMP.cpp
@@ -1730,9 +1730,22 @@
   auto CondBlock = createBasicBlock("omp.inner.for.cond");
   EmitBlock(CondBlock);
   const SourceRange R = S.getSourceRange();
-  LoopStack.push(CondBlock, SourceLocToDebugLoc(R.getBegin()),
-                 SourceLocToDebugLoc(R.getEnd()));
 
+  // If attributes are attached, push to the basic block with them.
+  const AttributedStmt *AS = nullptr;
+  if (auto *OMPD = dyn_cast<OMPParallelForDirective>(&S)) {
+    const CapturedStmt *CS = OMPD->getCapturedStmt(OMPD_parallel);
+    const Stmt *SS = CS->getCapturedStmt();
+    AS = dyn_cast_or_null<AttributedStmt>(SS);
+  }
+  if (AS)
+    LoopStack.push(CondBlock, CGM.getContext(), CGM.getCodeGenOpts(),
+                   AS->getAttrs(), SourceLocToDebugLoc(R.getBegin()),
+                   SourceLocToDebugLoc(R.getEnd()));
+  else
+    LoopStack.push(CondBlock, SourceLocToDebugLoc(R.getBegin()),
+                   SourceLocToDebugLoc(R.getEnd()));
+  
   // If there are any cleanups between here and the loop-exit scope,
   // create a block to stage a loop exit along.
   llvm::BasicBlock *ExitBlock = LoopExit.getBlock();


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D79921.263918.patch
Type: text/x-patch
Size: 1306 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20200514/841b15e3/attachment.bin>