[PATCH] D114734: [InstCombine] Allow fake vector insert folding to bit-logic only if the insert element is integer type

[Srividya Karumuri via Phabricator via llvm-commits]

Srividya-Karumuri created this revision.
Srividya-Karumuri added reviewers: spatel, qcolombet, dsanders.
Herald added a subscriber: hiraditya.
Srividya-Karumuri requested review of this revision.
Herald added a project: LLVM.
Herald added a subscriber: llvm-commits.

The below commit is causing assertion when insert element type is not integer
 type such as half. This is because the transformation is creating zext before
 doing bitwise OR, and the zext is supported only for integer types
https://github.com/llvm/llvm-project/commit/80ab06c599a0f5a90951c36a57b2a9b492b19d61


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D114734

Files:
  llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
  llvm/test/Transforms/InstCombine/bitcast-inselt-bitcast.ll


Index: llvm/test/Transforms/InstCombine/bitcast-inselt-bitcast.ll
===================================================================
--- llvm/test/Transforms/InstCombine/bitcast-inselt-bitcast.ll
+++ llvm/test/Transforms/InstCombine/bitcast-inselt-bitcast.ll
@@ -70,6 +70,29 @@
   ret i32 %r
 }
 
+; i32 is a common type, so we can convert independently of the data layout.
+; Endian determines if a shift is needed (and so the transform is avoided).
+; half type can not be used in zext instruction (and so the transform is avoided).
+
+define i32 @insert0_v2half(i32 %x, half %y) {
+; BE-LABEL: @insert0_v2half(
+; BE-NEXT:    [[V:%.*]] = bitcast i32 [[X:%.*]] to <2 x half>
+; BE-NEXT:    [[I:%.*]] = insertelement <2 x half> [[V]], half [[Y:%.*]], i8 0
+; BE-NEXT:    [[R:%.*]] = bitcast <2 x half> [[I]] to i32
+; BE-NEXT:    ret i32 [[R]]
+;
+; LE-LABEL: @insert0_v2half(
+; LE-NEXT:    [[V:%.*]] = bitcast i32 [[X:%.*]] to <2 x half>
+; LE-NEXT:    [[I:%.*]] = insertelement <2 x half> [[V]], half [[Y:%.*]], i8 0
+; LE-NEXT:    [[R:%.*]] = bitcast <2 x half> [[I]] to i32
+; LE-NEXT:    ret i32 [[R]]
+;
+  %v = bitcast i32 %x to <2 x half>
+  %i = insertelement <2 x half> %v, half %y, i8 0
+  %r = bitcast <2 x half> %i to i32
+  ret i32 %r
+}
+
 ; i64 is a legal type, so we can convert based on the data layout.
 ; Endian determines if a shift is needed (and so the transform is avoided).
 
Index: llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
===================================================================
--- llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
+++ llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
@@ -2771,7 +2771,7 @@
     if (match(Src, m_OneUse(m_InsertElt(m_OneUse(m_BitCast(m_Value(X))),
                                         m_Value(Y), m_ConstantInt(IndexC)))) &&
         DestTy->isIntegerTy() && X->getType() == DestTy &&
-        isDesirableIntType(BitWidth)) {
+        Y->getType()->isIntegerTy() && isDesirableIntType(BitWidth)) {
       // Adjust for big endian - the LSBs are at the high index.
       if (DL.isBigEndian())
         IndexC = SrcVTy->getNumElements() - 1 - IndexC;


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D114734.390417.patch
Type: text/x-patch
Size: 2148 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20211129/31d53ddc/attachment.bin>