[InstCombine] try to convert vector insert+extract to trunc; 2nd try
This was originally committed at rL343407, but reverted at
rL343458 because it crashed trying to handle a case where
the destination type is FP. This version of the patch adds
a check for that possibility. Tests added at rL343480.
Original commit message:
This transform is requested for the backend in:
https://bugs.llvm.org/show_bug.cgi?id=39016
...but I figured it was worth doing in IR too, and it's probably
easier to implement here, so that's this patch.
In the simplest case, we are just truncating a scalar value. If the
extract index doesn't correspond to the LSBs of the scalar, then we
have to shift-right before the truncate. Endian-ness makes this tricky,
but hopefully the ASCII-art helps visualize the transform.
Differential Revision: https://reviews.llvm.org/D52439
llvm-svn: 343482
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp b/llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
index c391034..945664d 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
@@ -167,7 +167,8 @@
}
static Instruction *foldBitcastExtElt(ExtractElementInst &Ext,
- InstCombiner::BuilderTy &Builder) {
+ InstCombiner::BuilderTy &Builder,
+ bool IsBigEndian) {
Value *X;
uint64_t ExtIndexC;
if (!match(Ext.getVectorOperand(), m_BitCast(m_Value(X))) ||
@@ -186,6 +187,49 @@
if (Value *Elt = findScalarElement(X, ExtIndexC))
return new BitCastInst(Elt, DestTy);
+ // If the source elements are wider than the destination, try to shift and
+ // truncate a subset of scalar bits of an insert op.
+ // TODO: This is limited to integer types, but we could bitcast to/from FP.
+ if (NumSrcElts < NumElts && SrcTy->getScalarType()->isIntegerTy() &&
+ DestTy->getScalarType()->isIntegerTy()) {
+ Value *Scalar;
+ uint64_t InsIndexC;
+ if (!match(X, m_InsertElement(m_Value(), m_Value(Scalar),
+ m_ConstantInt(InsIndexC))))
+ return nullptr;
+
+ // The extract must be from the subset of vector elements that we inserted
+ // into. Example: if we inserted element 1 of a <2 x i64> and we are
+ // extracting an i16 (narrowing ratio = 4), then this extract must be from 1
+ // of elements 4-7 of the bitcasted vector.
+ unsigned NarrowingRatio = NumElts / NumSrcElts;
+ if (ExtIndexC / NarrowingRatio != InsIndexC)
+ return nullptr;
+
+ // We are extracting part of the original scalar. How that scalar is
+ // inserted into the vector depends on the endian-ness. Example:
+ // Vector Byte Elt Index: 0 1 2 3 4 5 6 7
+ // +--+--+--+--+--+--+--+--+
+ // inselt <2 x i32> V, <i32> S, 1: |V0|V1|V2|V3|S0|S1|S2|S3|
+ // extelt <4 x i16> V', 3: | |S2|S3|
+ // +--+--+--+--+--+--+--+--+
+ // If this is little-endian, S2|S3 are the MSB of the 32-bit 'S' value.
+ // If this is big-endian, S2|S3 are the LSB of the 32-bit 'S' value.
+ // In this example, we must right-shift little-endian. Big-endian is just a
+ // truncate.
+ unsigned Chunk = ExtIndexC % NarrowingRatio;
+ if (IsBigEndian)
+ Chunk = NarrowingRatio - 1 - Chunk;
+ unsigned ShAmt = Chunk * DestTy->getPrimitiveSizeInBits();
+ if (ShAmt) {
+ // Bail out if we could end with more instructions than we started with.
+ if (!Ext.getVectorOperand()->hasOneUse())
+ return nullptr;
+ Scalar = Builder.CreateLShr(Scalar, ShAmt);
+ }
+ return new TruncInst(Scalar, DestTy);
+ }
+
return nullptr;
}
@@ -224,7 +268,7 @@
}
}
- if (Instruction *I = foldBitcastExtElt(EI, Builder))
+ if (Instruction *I = foldBitcastExtElt(EI, Builder, DL.isBigEndian()))
return I;
// If there's a vector PHI feeding a scalar use through this extractelement