[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

2 files changed, 101 insertions, 33 deletions

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp b/llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
index c391034dc00..945664de686 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
@@ -167,7 +167,8 @@ Instruction *InstCombiner::scalarizePHI(ExtractElementInst &EI, PHINode *PN) {
 }
 
 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 @@ static Instruction *foldBitcastExtElt(ExtractElementInst &Ext,
     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 @@ Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
       }
     }
 
-    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
diff --git a/llvm/test/Transforms/InstCombine/extractelement.ll b/llvm/test/Transforms/InstCombine/extractelement.ll
index f29c020b2d2..65ef373dd48 100644
--- a/llvm/test/Transforms/InstCombine/extractelement.ll
+++ b/llvm/test/Transforms/InstCombine/extractelement.ll
@@ -42,11 +42,14 @@ define i64 @test2(i64 %in) {
 }
 
 define i32 @bitcasted_inselt_wide_source_zero_elt(i64 %x) {
-; ANY-LABEL: @bitcasted_inselt_wide_source_zero_elt(
-; ANY-NEXT:    [[I:%.*]] = insertelement <2 x i64> undef, i64 [[X:%.*]], i32 0
-; ANY-NEXT:    [[B:%.*]] = bitcast <2 x i64> [[I]] to <4 x i32>
-; ANY-NEXT:    [[R:%.*]] = extractelement <4 x i32> [[B]], i32 0
-; ANY-NEXT:    ret i32 [[R]]
+; LE-LABEL: @bitcasted_inselt_wide_source_zero_elt(
+; LE-NEXT:    [[R:%.*]] = trunc i64 [[X:%.*]] to i32
+; LE-NEXT:    ret i32 [[R]]
+;
+; BE-LABEL: @bitcasted_inselt_wide_source_zero_elt(
+; BE-NEXT:    [[TMP1:%.*]] = lshr i64 [[X:%.*]], 32
+; BE-NEXT:    [[R:%.*]] = trunc i64 [[TMP1]] to i32
+; BE-NEXT:    ret i32 [[R]]
 ;
   %i = insertelement <2 x i64> zeroinitializer, i64 %x, i32 0
   %b = bitcast <2 x i64> %i to <4 x i32>
@@ -55,11 +58,14 @@ define i32 @bitcasted_inselt_wide_source_zero_elt(i64 %x) {
 }
 
 define i16 @bitcasted_inselt_wide_source_modulo_elt(i64 %x) {
-; ANY-LABEL: @bitcasted_inselt_wide_source_modulo_elt(
-; ANY-NEXT:    [[I:%.*]] = insertelement <2 x i64> undef, i64 [[X:%.*]], i32 1
-; ANY-NEXT:    [[B:%.*]] = bitcast <2 x i64> [[I]] to <8 x i16>
-; ANY-NEXT:    [[R:%.*]] = extractelement <8 x i16> [[B]], i32 4
-; ANY-NEXT:    ret i16 [[R]]
+; LE-LABEL: @bitcasted_inselt_wide_source_modulo_elt(
+; LE-NEXT:    [[R:%.*]] = trunc i64 [[X:%.*]] to i16
+; LE-NEXT:    ret i16 [[R]]
+;
+; BE-LABEL: @bitcasted_inselt_wide_source_modulo_elt(
+; BE-NEXT:    [[TMP1:%.*]] = lshr i64 [[X:%.*]], 48
+; BE-NEXT:    [[R:%.*]] = trunc i64 [[TMP1]] to i16
+; BE-NEXT:    ret i16 [[R]]
 ;
   %i = insertelement <2 x i64> undef, i64 %x, i32 1
   %b = bitcast <2 x i64> %i to <8 x i16>
@@ -68,11 +74,14 @@ define i16 @bitcasted_inselt_wide_source_modulo_elt(i64 %x) {
 }
 
 define i32 @bitcasted_inselt_wide_source_not_modulo_elt(i64 %x) {
-; ANY-LABEL: @bitcasted_inselt_wide_source_not_modulo_elt(
-; ANY-NEXT:    [[I:%.*]] = insertelement <2 x i64> undef, i64 [[X:%.*]], i32 0
-; ANY-NEXT:    [[B:%.*]] = bitcast <2 x i64> [[I]] to <4 x i32>
-; ANY-NEXT:    [[R:%.*]] = extractelement <4 x i32> [[B]], i32 1
-; ANY-NEXT:    ret i32 [[R]]
+; LE-LABEL: @bitcasted_inselt_wide_source_not_modulo_elt(
+; LE-NEXT:    [[TMP1:%.*]] = lshr i64 [[X:%.*]], 32
+; LE-NEXT:    [[R:%.*]] = trunc i64 [[TMP1]] to i32
+; LE-NEXT:    ret i32 [[R]]
+;
+; BE-LABEL: @bitcasted_inselt_wide_source_not_modulo_elt(
+; BE-NEXT:    [[R:%.*]] = trunc i64 [[X:%.*]] to i32
+; BE-NEXT:    ret i32 [[R]]
 ;
   %i = insertelement <2 x i64> undef, i64 %x, i32 0
   %b = bitcast <2 x i64> %i to <4 x i32>
@@ -81,11 +90,15 @@ define i32 @bitcasted_inselt_wide_source_not_modulo_elt(i64 %x) {
 }
 
 define i8 @bitcasted_inselt_wide_source_not_modulo_elt_not_half(i32 %x) {
-; ANY-LABEL: @bitcasted_inselt_wide_source_not_modulo_elt_not_half(
-; ANY-NEXT:    [[I:%.*]] = insertelement <2 x i32> undef, i32 [[X:%.*]], i32 0
-; ANY-NEXT:    [[B:%.*]] = bitcast <2 x i32> [[I]] to <8 x i8>
-; ANY-NEXT:    [[R:%.*]] = extractelement <8 x i8> [[B]], i32 2
-; ANY-NEXT:    ret i8 [[R]]
+; LE-LABEL: @bitcasted_inselt_wide_source_not_modulo_elt_not_half(
+; LE-NEXT:    [[TMP1:%.*]] = lshr i32 [[X:%.*]], 16
+; LE-NEXT:    [[R:%.*]] = trunc i32 [[TMP1]] to i8
+; LE-NEXT:    ret i8 [[R]]
+;
+; BE-LABEL: @bitcasted_inselt_wide_source_not_modulo_elt_not_half(
+; BE-NEXT:    [[TMP1:%.*]] = lshr i32 [[X:%.*]], 8
+; BE-NEXT:    [[R:%.*]] = trunc i32 [[TMP1]] to i8
+; BE-NEXT:    ret i8 [[R]]
 ;
   %i = insertelement <2 x i32> undef, i32 %x, i32 0
   %b = bitcast <2 x i32> %i to <8 x i8>
@@ -94,11 +107,15 @@ define i8 @bitcasted_inselt_wide_source_not_modulo_elt_not_half(i32 %x) {
 }
 
 define i3 @bitcasted_inselt_wide_source_not_modulo_elt_not_half_weird_types(i15 %x) {
-; ANY-LABEL: @bitcasted_inselt_wide_source_not_modulo_elt_not_half_weird_types(
-; ANY-NEXT:    [[I:%.*]] = insertelement <3 x i15> undef, i15 [[X:%.*]], i32 0
-; ANY-NEXT:    [[B:%.*]] = bitcast <3 x i15> [[I]] to <15 x i3>
-; ANY-NEXT:    [[R:%.*]] = extractelement <15 x i3> [[B]], i32 1
-; ANY-NEXT:    ret i3 [[R]]
+; LE-LABEL: @bitcasted_inselt_wide_source_not_modulo_elt_not_half_weird_types(
+; LE-NEXT:    [[TMP1:%.*]] = lshr i15 [[X:%.*]], 3
+; LE-NEXT:    [[R:%.*]] = trunc i15 [[TMP1]] to i3
+; LE-NEXT:    ret i3 [[R]]
+;
+; BE-LABEL: @bitcasted_inselt_wide_source_not_modulo_elt_not_half_weird_types(
+; BE-NEXT:    [[TMP1:%.*]] = lshr i15 [[X:%.*]], 9
+; BE-NEXT:    [[R:%.*]] = trunc i15 [[TMP1]] to i3
+; BE-NEXT:    ret i3 [[R]]
 ;
   %i = insertelement <3 x i15> undef, i15 %x, i32 0
   %b = bitcast <3 x i15> %i to <15 x i3>
@@ -125,12 +142,19 @@ define i8 @bitcasted_inselt_wide_source_wrong_insert(<2 x i32> %v, i32 %x) {
 declare void @use(<8 x i8>)
 
 define i8 @bitcasted_inselt_wide_source_uses(i32 %x) {
-; ANY-LABEL: @bitcasted_inselt_wide_source_uses(
-; ANY-NEXT:    [[I:%.*]] = insertelement <2 x i32> undef, i32 [[X:%.*]], i32 0
-; ANY-NEXT:    [[B:%.*]] = bitcast <2 x i32> [[I]] to <8 x i8>
-; ANY-NEXT:    call void @use(<8 x i8> [[B]])
-; ANY-NEXT:    [[R:%.*]] = extractelement <8 x i8> [[B]], i32 3
-; ANY-NEXT:    ret i8 [[R]]
+; LE-LABEL: @bitcasted_inselt_wide_source_uses(
+; LE-NEXT:    [[I:%.*]] = insertelement <2 x i32> undef, i32 [[X:%.*]], i32 0
+; LE-NEXT:    [[B:%.*]] = bitcast <2 x i32> [[I]] to <8 x i8>
+; LE-NEXT:    call void @use(<8 x i8> [[B]])
+; LE-NEXT:    [[R:%.*]] = extractelement <8 x i8> [[B]], i32 3
+; LE-NEXT:    ret i8 [[R]]
+;
+; BE-LABEL: @bitcasted_inselt_wide_source_uses(
+; BE-NEXT:    [[I:%.*]] = insertelement <2 x i32> undef, i32 [[X:%.*]], i32 0
+; BE-NEXT:    [[B:%.*]] = bitcast <2 x i32> [[I]] to <8 x i8>
+; BE-NEXT:    call void @use(<8 x i8> [[B]])
+; BE-NEXT:    [[R:%.*]] = trunc i32 [[X]] to i8
+; BE-NEXT:    ret i8 [[R]]
 ;
   %i = insertelement <2 x i32> undef, i32 %x, i32 0
   %b = bitcast <2 x i32> %i to <8 x i8>