[SVE] Add folds for truncation of vscale

[sdesmalen]

llvm/test/Transforms/InstCombine/vscale_trunc.ll

5

This test isn't really necessary, if you handle the case for i32 -> i8 for the two edge cases.

16

Three things:

1. Please inline the vscale attribute here, that makes the test more more readable.
2. Please test the edge cases:

vscale_range(0, 256) // can't fold.
vscale_range(0, 255) // can be folded.

3. This name seems wrong: vscale_SExt_i32toi8? (same for the use of SExt in the test below)

28

This test isn't really necessary, if you handle the case for i32 -> i8 for the two edge cases.

39

I'm missing a test for the case where there is no attribute specified.

[sdesmalen]

llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp

971

This also requires a check that MaxVScale > 0, because according to the langref (a value of 0 means unbounded).

[DylanFleming-arm]

Added check for MaxVScale > 0

Removed redundant test and added extra case for no attribute

[david-arm]

LGTM!

llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp

if (match(Src, m_OneUse(m_Intrinsic<Intrinsic::ctlz>(m_ZExt(m_Value(A)),
     unsigned AWidth = A->getType()->getScalarSizeInBits();
     if (AWidth == DestWidth && AWidth > Log2_32(SrcWidth)) {
       Value *WidthDiff = ConstantInt::get(A->getType(), SrcWidth - AWidth);
       Value *NarrowCtlz =
           Builder.CreateIntrinsic(Intrinsic::ctlz, {Trunc.getType()}, {A, B});
       return BinaryOperator::CreateAdd(NarrowCtlz, WidthDiff);
     }
   }
+
+  if (match(Src, m_VScale(DL))) {
+    if (Trunc.getFunction()->hasFnAttribute(Attribute::VScaleRange)) {
+      unsigned MaxVScale = Trunc.getFunction()
+                               ->getFnAttribute(Attribute::VScaleRange)
+                               .getVScaleRangeArgs()
+                               .second;
+      if (MaxVScale > 0 && Log2_32(MaxVScale) < DestWidth) {
+        Value *VScale = Builder.CreateVScale(ConstantInt::get(DestTy, 1));
+        return replaceInstUsesWith(Trunc, VScale);
+      }
+    }
+  }
+
   return nullptr;
 }
 
 /// Transform (zext icmp) to bitwise / integer operations in order to
 /// eliminate it. If DoTransform is false, just test whether the given
 /// (zext icmp) can be transformed.
 Instruction *InstCombinerImpl::transformZExtICmp(ICmpInst *Cmp, ZExtInst &Zext,
                                                  bool DoTransform) {

llvm/test/Transforms/InstCombine/vscale_trunc.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -instcombine -S | FileCheck %s
+
+define i8 @vscale_trunc_i32toi8() vscale_range(0, 255) {
+; CHECK-LABEL: @vscale_trunc_i32toi8(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = call i8 @llvm.vscale.i8()
+; CHECK-NEXT:    ret i8 [[TMP0]]
+entry:
+  %0 = call i32 @llvm.vscale.i32()
+  %1 = trunc i32 %0 to i8
+  ret i8 %1
+}
+
+
+define i8 @vscale_trunc_i32toi8_poison() vscale_range(0, 256) {
+; CHECK-LABEL: @vscale_trunc_i32toi8_poison(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = call i32 @llvm.vscale.i32()
+; CHECK-NEXT:    [[TMP1:%.*]] = trunc i32 [[TMP0]] to i8
+; CHECK-NEXT:    ret i8 [[TMP1]]
+  entry:
+  %0 = call i32 @llvm.vscale.i32()
+  %1 = trunc i32 %0 to i8
+  ret i8 %1
+}
+
+define i8 @vscale_trunc_i32toi8_noAttr() {
+; CHECK-LABEL: @vscale_trunc_i32toi8_noAttr(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = call i32 @llvm.vscale.i32()
+; CHECK-NEXT:    [[TMP1:%.*]] = trunc i32 [[TMP0]] to i8
+; CHECK-NEXT:    ret i8 [[TMP1]]
+  entry:
+  %0 = call i32 @llvm.vscale.i32()
+  %1 = trunc i32 %0 to i8
+  ret i8 %1
+}
+
+declare i32 @llvm.vscale.i32()