[PATCH] support llvm.vectorization.vector_width metadata

Redmond, Paul paul.redmond at intel.com
Mon May 27 12:42:52 PDT 2013


Attached is an updated patch that:

- documents llvm.vectorizer.* metadata
- documents llvm.loop and updates llvm.mem.parallel_loop_access
- moves loop vectorizer metadata handling into utility class LoopVectorizerHints
- moves getting/setting of llvm.vectorizer.already_vectorized into LoopVectorizerHints
- adds support for llvm.vectorizer.width, llvm.vectorizer.unroll, and llvm.vectorizer.disable
  - llvm.vectorizer.force and llvm.vectorizer.max_iterations will be addressed later.
- updated tests

paul


On 2013-05-24, at 2:07 PM, Nadav Rotem wrote:

Hi Paul,

Thanks for working on this.  Can you also refactor the code that sets metadata on vectorized loops (so that they are not re-vectorized) ?


-bool Loop::isAnnotatedParallel() const {
+MDNode *Loop::getLoopID() const {
+  MDNode *LoopID = 0;

Please add braces below.

+  if (isLoopSimplifyForm())
+    LoopID = getLoopLatch()->getTerminator()->getMetadata("llvm.loop");
+  else {
+    // Go through each predecessor of the loop header and check the
+    // terminator for the metadata.
+    BasicBlock *H


Please add a string for the asserts below.

+void Loop::setLoopID(MDNode *LoopID) const {
+  assert(LoopID);
+  assert(LoopID->getNumOperands() > 0);
+  assert(LoopID->getOperand(0) == LoopID);


Can you outline the code below to a function and use it around the code that checks for the command line flags ?

+    unsigned UserVF = VectorizationFactor;
+
+    // Check the loop metadata for vectorization hints.
+    if (const MDNode *LoopID = L->getLoopID()) {
+      assert(LoopID->getNumOperands() > 0);
+      assert(LoopID->getOperand(0) == LoopID);
+      for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
+        const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
+        if (!MD)
+          continue;

Use early exits (continue) here:

+        if (MD->getNumOperands() == 2) {
+          if (const MDString *S = dyn_cast<MDString>(MD->getOperand(0))) {
+            if (S->getString() == "llvm.vectorization.vector_width") {
+              const ConstantInt *C = dyn_cast<ConstantInt>(MD->getOperand(1));
+              UserVF = C ? C->getZExtValue() : UserVF;


Because of this:

+            }
+          }
+        }
+      }
+    }
+


Thanks,
Nadav


On May 24, 2013, at 10:49 AM, "Redmond, Paul" <paul.redmond at intel.com<mailto:paul.redmond at intel.com>> wrote:

Hi,

The attached patch modifies LoopVectorizer to recognize llvm.vectorization.vector_width metadata attached to llvm.loop metadata.

The llvm.loop.parallel metadata has been renamed to llvm.loop to be more generic. Loop::isAnnotatedParallel now looks for llvm.loop and associated llvm.mem.parallel_loop_access. If we decide that we still need an explicit llvm.loop.parallel metadata then it can be added as a child of llvm.loop.

Loop::setLoopID was added for symmetry and it is used in a forthcoming patch which preserves llvm.loop metadata in loop passes.

paul

<Mail Attachment>_______________________________________________
llvm-commits mailing list
llvm-commits at cs.uiuc.edu<mailto:llvm-commits at cs.uiuc.edu>
http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits


-------------- next part --------------
A non-text attachment was scrubbed...
Name: vectorizer_metadata.diff
Type: application/octet-stream
Size: 30919 bytes
Desc: vectorizer_metadata.diff
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20130527/d924b808/attachment.obj>


More information about the llvm-commits mailing list