[AVR] Optimize int16 shift operation for shift amount greater than 8

[benshi001]

Summary

Skip operation on the lower byte in int16 logical left shift when
shift amount is greater than 8.

Skip operation on the higher byte in int16 logical & arithmetic
right shift when shift amount is greater than 8.

[aykevl]

Is there a reason why you introduced 3 new pseudo instructions? It seems to me that the existing 16-bit shift instructions (LSLWN, LSRWN, ASRWN) should already be able to support this.

[aykevl]

Ok, I think I see why you did it the way you did. Which makes sense given the current architecture.

I suspect shift lowering might be simpler when it is all done in the "expand pseudo instructions" pass instead of splitting it early in ISelLowering. But most of the logic is already in ISelLowering so it makes sense to keep it there, at least for now.

llvm/lib/Transforms/Utils/RelLookupTableConverter.cpp

static void convertToRelLookupTable(GlobalVariable &LookupTable) {
 
   // Place new instruction sequence before GEP.
   Builder.SetInsertPoint(GEP);
   Value *Index = GEP->getOperand(2);
   IntegerType *IntTy = cast<IntegerType>(Index->getType());
   Value *Offset =
       Builder.CreateShl(Index, ConstantInt::get(IntTy, 2), "reltable.shift");
 
+  // Insert the call to load.relative instrinsic before LOAD.
+  // GEP might not be immediately followed by a LOAD, like it can be hoisted
+  // outside the loop or another instruction might be inserted them in between.
+  Builder.SetInsertPoint(Load);
   Function *LoadRelIntrinsic = llvm::Intrinsic::getDeclaration(
       &M, Intrinsic::load_relative, {Index->getType()});
   Value *Base = Builder.CreateBitCast(RelLookupTable, Builder.getInt8PtrTy());
 
   // Create a call to load.relative intrinsic that computes the target address
   // by adding base address (lookup table address) and relative offset.
   Value *Result = Builder.CreateCall(LoadRelIntrinsic, {Base, Offset},
                                      "reltable.intrinsic");

llvm/test/Transforms/RelLookupTableConverter/X86/relative_lookup_table.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; REQUIRES: x86-registered-target
 ; RUN: opt < %s -passes=rel-lookup-table-converter -relocation-model=pic -S | FileCheck %s
 target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
 target triple = "x86_64-unknown-linux-gnu"
 
 @.str = private unnamed_addr constant [5 x i8] c"zero\00", align 1
 @.str.1 = private unnamed_addr constant [4 x i8] c"one\00", align 1
 @.str.2 = private unnamed_addr constant [4 x i8] c"two\00", align 1
 @.str.3 = private unnamed_addr constant [8 x i8] c"default\00", align 1
 @.str.4 = private unnamed_addr constant [6 x i8] c"three\00", align 1
 @.str.5 = private unnamed_addr constant [5 x i8] c"str1\00", align 1
 @.str.6 = private unnamed_addr constant [5 x i8] c"str2\00", align 1
 @.str.7 = private unnamed_addr constant [12 x i8] c"singlevalue\00", align 1
+@.str.8 = private unnamed_addr constant [2 x i8] c"a\00", align 1
+@.str.9 = private unnamed_addr constant [2 x i8] c"b\00", align 1
+@.str.10 = private unnamed_addr constant [2 x i8] c"c\00", align 1
 
 @a1 = external global i32, align 4
 @b1 = external global i32, align 4
 @c1 = external global i32, align 4
 @d1 = external global i32, align 4
 
 @a2 = internal constant i32 0, align 4
 @b2 = internal constant i32 0, align 4

 
 @user_defined_lookup_table.table = internal unnamed_addr constant [3 x i8*]
                                    [
                                     i8* getelementptr inbounds ([5 x i8], [5 x i8]* @.str, i32 0, i32 0),
                                     i8* getelementptr inbounds ([4 x i8], [4 x i8]* @.str.1, i32 0, i32 0),
                                     i8* getelementptr inbounds ([4 x i8], [4 x i8]* @.str.2, i32 0, i32 0)
                                    ], align 16
 
+@table = internal constant [2 x i8*] [
+  i8* getelementptr inbounds ([2 x i8], [2 x i8]* @.str.8, i32 0, i32 0),
+  i8* getelementptr inbounds ([2 x i8], [2 x i8]* @.str.9, i32 0, i32 0)
+], align 16
+
+@table2 = internal constant [2 x i8*] [
+  i8* getelementptr inbounds ([2 x i8], [2 x i8]* @.str.8, i32 0, i32 0),
+  i8* getelementptr inbounds ([2 x i8], [2 x i8]* @.str.9, i32 0, i32 0)
+], align 16
+
 ; Lookup table check for integer pointers that have external linkage
 ; CHECK: @switch.table.external_linkage = private unnamed_addr constant [3 x i32*] [i32* @a1, i32* @b1, i32* @c1], align
 
 ; Lookup table check for integer pointers that have internal linkage
 ; CHECK: @reltable.internal_linkage = private unnamed_addr constant [3 x i32]
 ; CHECK-SAME: [
 ; CHECK-SAME: i32 trunc (i64 sub (i64 ptrtoint (i32* @a2 to i64), i64 ptrtoint ([3 x i32]* @reltable.internal_linkage to i64)) to i32),
 ; CHECK-SAME: i32 trunc (i64 sub (i64 ptrtoint (i32* @b2 to i64), i64 ptrtoint ([3 x i32]* @reltable.internal_linkage to i64)) to i32),

 ; CHECK: @reltable.single_value = private unnamed_addr constant [3 x i32]
 ; CHECK-SAME: [
 ; CHECK-SAME: i32 trunc (i64 sub (i64 ptrtoint ([5 x i8]* @.str to i64), i64 ptrtoint ([3 x i32]* @reltable.single_value to i64)) to i32),
 ; CHECK-SAME: i32 trunc (i64 sub (i64 ptrtoint ([4 x i8]* @.str.1 to i64), i64 ptrtoint ([3 x i32]* @reltable.single_value to i64)) to i32),
 ; CHECK-SAME: i32 trunc (i64 sub (i64 ptrtoint ([4 x i8]* @.str.2 to i64), i64 ptrtoint ([3 x i32]* @reltable.single_value to i64)) to i32)
 ; CHECK-SAME: ], align 4
 ;
 
+; Relative lookup table for the loop hoist check test
+; CHECK: @reltable.loop_hoist = internal unnamed_addr constant [2 x i32]
+; CHECK-SAME: [
+; CHECK-SAME: i32 trunc (i64 sub (i64 ptrtoint ([2 x i8]* @.str.8 to i64), i64 ptrtoint ([2 x i32]* @reltable.loop_hoist to i64)) to i32),
+; CHECK-SAME: i32 trunc (i64 sub (i64 ptrtoint ([2 x i8]* @.str.9 to i64), i64 ptrtoint ([2 x i32]* @reltable.loop_hoist to i64)) to i32)
+; CHECK-SAME: ], align 4
+
+; Relative look up table for the test where gep is not immediately followed by a load check
+; CHECK: @reltable.gep_is_not_imm_followed_by_load = internal unnamed_addr constant [2 x i32]
+; CHECK-SAME: [
+; CHECK-SAME: i32 trunc (i64 sub (i64 ptrtoint ([2 x i8]* @.str.8 to i64), i64 ptrtoint ([2 x i32]* @reltable.gep_is_not_imm_followed_by_load to i64)) to i32),
+; CHECK-SAME: i32 trunc (i64 sub (i64 ptrtoint ([2 x i8]* @.str.9 to i64), i64 ptrtoint ([2 x i32]* @reltable.gep_is_not_imm_followed_by_load to i64)) to i32)
+; CHECK-SAME: ], align 4
+
 ; Lookup table check for integer pointers that have external linkage
 define i32* @external_linkage(i32 %cond) {
 ; CHECK-LABEL: @external_linkage(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[TMP0:%.*]] = icmp ult i32 [[COND:%.*]], 3
 ; CHECK-NEXT:    br i1 [[TMP0]], label [[SWITCH_LOOKUP:%.*]], label [[RETURN:%.*]]
 ; CHECK:       switch.lookup:
 ; CHECK-NEXT:    [[SWITCH_GEP:%.*]] = getelementptr inbounds [3 x i32*], [3 x i32*]* @switch.table.external_linkage, i32 0, i32 [[COND:%.*]]
cond.false:                                       ; preds = %entry
   %0 = load i8*, i8** %arrayidx, align 8, !tbaa !4
   br label %cond.end
 
 cond.end:                                         ; preds = %entry, %cond.false
   %cond1 = phi i8* [ %0, %cond.false ], [ getelementptr inbounds ([8 x i8], [8 x i8]* @.str.3, i64 0, i64 0), %entry ]
   ret i8* %cond1
 }
 
+; Check to ensure that call @llvm.load.relative is inserted before load, not before gep.
+; When a lookup table is accessed inside a loop, and a gep is hosted outside the loop via licm,
+; make sure that call @llvm.load.relative is inserted before load.
+define i8* @loop_hoist(i32 %x) {
+; CHECK-LABEL: @loop_hoist(i32 %x)
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = icmp sgt i32 [[X:%.*]], 1
+; CHECK-NEXT:    [[TMP1:%.*]] = getelementptr inbounds [2 x i8], [2 x i8]* @.str.10, i32 0, i32 0
+; CHECK-NEXT:    [[RELTABLE_SHIFT:%.*]] = shl i32 [[X:%.*]], 2
+; CHECK-NEXT:    br i1 [[TMP0]], label %if.done, label %if.false
+; CHECK:       if.false:
+; CHECK-NEXT:    [[RELTABLE_INTRINSIC:%.*]] = call i8* @llvm.load.relative.i32(i8* bitcast ([2 x i32]* @reltable.loop_hoist to i8*), i32 [[RELTABLE_SHIFT]])
+; CHECK-NEXT:    br label %if.done
+; CHECK:       if.done:
+; CHECK-NEXT:    [[TMP2:%.*]] = phi i8* [ [[TMP1]], %entry ], [ [[RELTABLE_INTRINSIC]], %if.false ]
+; CHECK-NEXT:    ret i8* [[TMP2]]
+;
+entry:
+  %0 = icmp sgt i32 %x, 1
+  %1 = getelementptr inbounds [2 x i8], [2 x i8]* @.str.10, i32 0, i32 0
+  %2 = getelementptr [2 x i8*], [2 x i8*]* @table, i32 0, i32 %x
+  br i1 %0, label %if.done, label %if.false
+
+if.false:
+  %3 = load i8*, i8** %2
+  br label %if.done
+
+if.done:
+  %4 = phi i8* [ %1, %entry ], [ %3, %if.false ]
+  ret i8* %4
+}
+
+; Another check to ensure that call @llvm.load.relative is inserted before load but not before gep.
+; When a lookup table is accessed, and gep is not immediately followed by a load (like if there is a function call
+; or an exception in between), make sure that call @llvm.load.relative is inserted before load.
+; CHECK-LABEL: @may_not_return()
+declare void @may_not_return()
+
+define i8* @gep_is_not_imm_followed_by_load(i32 %x) {
+; CHECK-LABEL: @gep_is_not_imm_followed_by_load(i32 %x)
+; CHECK:       entry:
+; CHECK-NEXT:    [[RELTABLE_SHIFT:%.*]] = shl i32 [[X:%.*]], 2
+; CHECK-NEXT:    call void @may_not_return()
+; CHECK-NEXT:    [[RELTABLE_INTRINSIC:%.*]] = call i8* @llvm.load.relative.i32(i8* bitcast ([2 x i32]* @reltable.gep_is_not_imm_followed_by_load to i8*), i32 [[RELTABLE_SHIFT]])
+; CHECK-NEXT:    ret i8* [[RELTABLE_INTRINSIC]]
+;
+entry:
+  %0 = getelementptr [2 x i8*], [2 x i8*]* @table2, i32 0, i32 %x
+  call void @may_not_return()
+  %1 = load i8*, i8** %0
+  ret i8* %1
+}
+
 !llvm.module.flags = !{!0, !1}
 !0 = !{i32 7, !"PIC Level", i32 2}
 !1 = !{i32 1, !"Code Model", i32 1}
 !4 = !{!"any pointer", !5, i64 0}
 !5 = !{!"omnipotent char", !6, i64 0}
 !6 = !{!"Simple C/C++ TBAA"}