[InstCombine] Simplify llvm.vscale when vscale_range attribute exists

[junparser]

Summary

Reduce llvm.vscale to constant based on vscale_range attribute.

[david-arm]

Hi @junparser, thanks for looking into this performance improvement! However, I personally think it makes more sense to just always replace a call to vscale directly with a constant if you know the value of vscale at runtime - possibly done in InstSimplify or something like that? That way you don't need to add lots of different instcombine changes to support all the different folds that can happen. Constant folds would just kick in automatically.

[junparser]

Address comment. Move change to InstSimplify Pass

[david-arm]

LGTM! Thanks a lot for making the changes. :)

[frasercrmck]

llvm/lib/Analysis/InstructionSimplify.cpp

5787

I'd say that SVE isn't a good variable name here since RISC-V will also use this attribute at some point.

[david-arm]

llvm/lib/Analysis/InstructionSimplify.cpp

5787

Ah, good spot @frasercrmck ! Sorry I missed that. :(

[spatel]

llvm/test/Transforms/InstSimplify/fold-vscale.ll

10

The shifts in these tests are not necessary to show the transform, so I'd take them out.

[junparser]

Address comments.

[david-arm]

llvm/lib/Analysis/InstructionSimplify.cpp

5787

nit: I think 'Scalar' here is a bit confusing as that's kind of the opposite of 'Vector'. Perhaps 'Scalable' is better?

[frasercrmck]

llvm/lib/Analysis/InstructionSimplify.cpp

5787

Yeah. Or VScaleMin and VScaleMax?

[sdesmalen]

llvm/lib/Analysis/InstructionSimplify.cpp

5787

Or just Min and Max?

llvm/test/Transforms/InstSimplify/fold-vscale.ll

4

nit: it would be useful to add _range_<min>_<max> to the name as well, since the attributes are at the bottom, making it slightly easier to read, e.g.@ vscale_i64_range_1_1().

[junparser]

OK， I'll fix these with an NFC patch. Thanks!

[junparser]

Address comments in https://reviews.llvm.org/rGe2fe26e77b02633207bdd36a14e2dcf210cc2fe7

llvm/lib/Analysis/InstructionSimplify.cpp

default:
     break;
   }
 
   return nullptr;
 }
 
 static Value *simplifyIntrinsic(CallBase *Call, const SimplifyQuery &Q) {
 
-  // Intrinsics with no operands have some kind of side effect. Don't simplify.
   unsigned NumOperands = Call->getNumArgOperands();
-  if (!NumOperands)
-    return nullptr;
-
   Function *F = cast<Function>(Call->getCalledFunction());
   Intrinsic::ID IID = F->getIntrinsicID();
+
+  // Most of the intrinsics with no operands have some kind of side effect.
+  // Don't simplify.
+  if (!NumOperands) {
+    switch (IID) {
+    case Intrinsic::vscale: {
+      auto Attr = Call->getFunction()->getFnAttribute(Attribute::VScaleRange);
+      if (!Attr.isValid())
+        return nullptr;
+      unsigned MinScalarVectorSize, MaxScalarVectorSize;
+      std::tie(MinScalarVectorSize, MaxScalarVectorSize) =
+          Attr.getVScaleRangeArgs();
+      if (MinScalarVectorSize == MaxScalarVectorSize &&
+          MaxScalarVectorSize != 0)
+        return ConstantInt::get(F->getReturnType(), MinScalarVectorSize);
+      return nullptr;
+    }
+    default:
+      return nullptr;
+    }
+  }
+
   if (NumOperands == 1)
     return simplifyUnaryIntrinsic(F, Call->getArgOperand(0), Q);
 
   if (NumOperands == 2)
     return simplifyBinaryIntrinsic(F, Call->getArgOperand(0),
                                    Call->getArgOperand(1), Q);
 
   // Handle intrinsics with 3 or more arguments.

llvm/test/Transforms/InstSimplify/fold-vscale.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -instsimplify -S | FileCheck %s
+
+define i64 @vscale_i64() #1 {
+; CHECK-LABEL: @vscale_i64(
+; CHECK-NEXT:    ret i64 1
+;
+  %out = call i64 @llvm.vscale.i64()
+  ret i64 %out
+}
+
+define i32 @vscale_i32() #2 {
+; CHECK-LABEL: @vscale_i32(
+; CHECK-NEXT:    ret i32 2
+;
+  %out = call i32 @llvm.vscale.i32()
+  ret i32 %out
+}
+
+define i64 @vscale_i64_diff() #3 {
+; CHECK-LABEL: @vscale_i64_diff(
+; CHECK-NEXT:    [[OUT:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    ret i64 [[OUT]]
+;
+  %out = call i64 @llvm.vscale.i64()
+  ret i64 %out
+}
+
+; Function Attrs: nofree nosync nounwind readnone willreturn
+declare i64 @llvm.vscale.i64() #0
+
+; Function Attrs: nofree nosync nounwind readnone willreturn
+declare i32 @llvm.vscale.i32() #0
+
+attributes #0 = { nofree nosync nounwind readnone willreturn }
+attributes #1 = { mustprogress nofree nosync nounwind uwtable vscale_range(1,1) }
+attributes #2 = { mustprogress nofree nosync nounwind uwtable vscale_range(2,2) }
+attributes #3 = { mustprogress nofree nosync nounwind uwtable vscale_range(2,4) }
+
+