[LLVM][CodeGen] Lower ConstantInt vectors like shufflevector base splats. (#144395)

ConstantInt vectors utilise DAG.getConstant() when constructing the
initial DAG. This can have the effect of legalising the constant before
the DAG combiner is run, significant altering the generated code. To
mitigate this (hopefully as a temporary measure) we instead try to
construct the DAG in the same way as shufflevector based splats.

Author: paulwalker-arm

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

@@ -1791,8 +1791,26 @@ SDValue SelectionDAGBuilder::getValueImpl(const Value *V) {
   if (const Constant *C = dyn_cast<Constant>(V)) {
     EVT VT = TLI.getValueType(DAG.getDataLayout(), V->getType(), true);
 
-    if (const ConstantInt *CI = dyn_cast<ConstantInt>(C))
-      return DAG.getConstant(*CI, getCurSDLoc(), VT);
+    if (const ConstantInt *CI = dyn_cast<ConstantInt>(C)) {
+      SDLoc DL = getCurSDLoc();
+
+      // DAG.getConstant() may attempt to legalise the vector constant which can
+      // significantly change the combines applied to the DAG. To reduce the
+      // divergence when enabling ConstantInt based vectors we try to construct
+      // the DAG in the same way as shufflevector based splats. TODO: The
+      // divergence sometimes leads to better optimisations. Ideally we should
+      // prevent DAG.getConstant() from legalising too early but there are some
+      // degradations preventing this.
+      if (VT.isScalableVector())
+        return DAG.getNode(
+            ISD::SPLAT_VECTOR, DL, VT,
+            DAG.getConstant(CI->getValue(), DL, VT.getVectorElementType()));
+      if (VT.isFixedLengthVector())
+        return DAG.getSplatBuildVector(
+            VT, DL,
+            DAG.getConstant(CI->getValue(), DL, VT.getVectorElementType()));
+      return DAG.getConstant(*CI, DL, VT);
+    }
 
     if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
       return DAG.getGlobalAddress(GV, getCurSDLoc(), VT);

llvm/test/CodeGen/AArch64/sve-expand-div.ll

@@ -1,5 +1,6 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -mtriple=aarch64-linux-gnu < %s | FileCheck %s
+; RUN: llc -mtriple=aarch64-linux-gnu -use-constant-int-for-scalable-splat < %s | FileCheck %s
 
 ; Check that expensive divides are expanded into a more performant sequence
 

llvm/test/CodeGen/AArch64/sve-fixed-length-sdiv-pow2.ll

@@ -3,6 +3,7 @@
 ; RUN: llc -aarch64-sve-vector-bits-min=256  < %s | FileCheck %s -check-prefixes=CHECK,VBITS_GE_256
 ; RUN: llc -aarch64-sve-vector-bits-min=512  < %s | FileCheck %s -check-prefixes=CHECK,VBITS_GE_512
 ; RUN: llc -aarch64-sve-vector-bits-min=2048 < %s | FileCheck %s -check-prefixes=CHECK,VBITS_GE_512
+; RUN: llc -aarch64-sve-vector-bits-min=128 -use-constant-int-for-fixed-length-splat < %s | FileCheck %s -check-prefixes=CHECK,VBITS_GE_128
 
 target triple = "aarch64-unknown-linux-gnu"
 

llvm/test/CodeGen/AArch64/sve-sdiv-pow2.ll

@@ -1,5 +1,6 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s | FileCheck %s
+; RUN: llc -use-constant-int-for-scalable-splat < %s | FileCheck %s
 
 target triple = "aarch64-unknown-linux-gnu"