GenXDeadVectorRemoval predicate fix

Do not modify "all true" predicates even if a part of vector is dead.
This can result in creation of unnecessary predicated instructions

(cherry picked from commit 62df6fe)

Author: mshelego

IGC/VectorCompiler/lib/GenXCodeGen/GenXDeadVectorRemoval.cpp

@@ -151,6 +151,11 @@ Constant *GenXDeadVectorRemoval::trySimplify(Constant *C,
   if (!LiveElems.isAnyDead())
     return C;
 
+  // Do not modify "all true" predicates to avoid creation of redundant
+  // execution masks
+  if (C->getType()->getScalarType()->isIntegerTy(1) && C->isAllOnesValue())
+    return C;
+
   if (auto CA = dyn_cast<ConstantAggregate>(C))
     return trySimplify(CA, LiveElems);
 

IGC/VectorCompiler/test/DeadVectorRemoval/const_wrregion.ll

@@ -27,3 +27,13 @@ define <16 x i8> @test2(<16 x i8> %val) {
   %.sum = add <16 x i8> %.region1, %.region2
   ret <16 x i8> %.sum
 }
+
+declare <4 x i32> @llvm.genx.wrregioni.v4i32.v4i32.i16.v4i1(<4 x i32>, <4 x i32>, i32, i32, i32, i16, i32, <4 x i1>) readnone nounwind
+declare i32 @llvm.genx.rdregioni.i32.v4i32.i16(<4 x i32>, i32, i32, i32, i16, i32) readnone nounwind
+
+define i32 @test3(<4 x i32> %val) {
+; CHECK: %1 = call <4 x i32> @llvm.genx.wrregioni.v4i32.v4i32.i16.v4i1(<4 x i32> undef, <4 x i32> %val, i32 0, i32 4, i32 1, i16 0, i32 undef, <4 x i1> <i1 true, i1 true, i1 true, i1 true>)
+  %1 = call <4 x i32> @llvm.genx.wrregioni.v4i32.v4i32.i16.v4i1(<4 x i32> <i32 0, i32 1, i32 2, i32 3>, <4 x i32> %val, i32 0, i32 4, i32 1, i16 0, i32 undef, <4 x i1> <i1 true, i1 true, i1 true, i1 true>)
+  %.region = tail call i32 @llvm.genx.rdregioni.i32.v4i32.i16(<4 x i32> %1, i32 0, i32 1, i32 0, i16 4, i32 undef)
+  ret i32 %.region
+}