llvm.noalias - InstCombiner::FoldGEPICmp should look through them

[hfinkel]

Summary

This is part of the series started by D9375, and teaches InstCombiner::FoldGEPICmp to look through calls to llvm.noalias.

[reames]

Resigning as a reviewer to get a very stale review off my list of blocking tasks in phabricator. Please readd when desired.

[hfinkel]

This change should not be needed if noalias intrinsics are stripped by default.

lib/Transforms/InstCombine/InstCombineCompares.cpp

Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
   // e.g. "&foo[0] <s &foo[1]" can't be folded to "true" because "foo" could be
   // the maximum signed value for the pointer type.
   if (ICmpInst::isSigned(Cond))
     return nullptr;
 
   // Look through bitcasts and addrspacecasts. We do not however want to remove
   // 0 GEPs.
   if (!isa<GetElementPtrInst>(RHS))
-    RHS = RHS->stripPointerCasts();
+    RHS = RHS->stripPointerCasts(/*LookThroughNoAlias*/true);
 
   Value *PtrBase = GEPLHS->getOperand(0);
   if (PtrBase == RHS && GEPLHS->isInBounds()) {
     // ((gep Ptr, OFFSET) cmp Ptr)   ---> (OFFSET cmp 0).
     // This transformation (ignoring the base and scales) is valid because we
     // know pointers can't overflow since the gep is inbounds.  See if we can
     // output an optimized form.
     Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, *this, DL);
if (PtrBase != GEPRHS->getOperand(0)) {
         return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0));
 
       // If we're comparing GEPs with two base pointers that only differ in type
       // and both GEPs have only constant indices or just one use, then fold
       // the compare with the adjusted indices.
       if (GEPLHS->isInBounds() && GEPRHS->isInBounds() &&
           (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) &&
           (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) &&
-          PtrBase->stripPointerCasts() ==
-              GEPRHS->getOperand(0)->stripPointerCasts()) {
+          PtrBase->stripPointerCasts(/*LookThroughNoAlias*/true) ==
+              GEPRHS->getOperand(0)->stripPointerCasts(true)) {
         Value *LOffset = EmitGEPOffset(GEPLHS);
         Value *ROffset = EmitGEPOffset(GEPRHS);
 
         // If we looked through an addrspacecast between different sized address
         // spaces, the LHS and RHS pointers are different sized
         // integers. Truncate to the smaller one.
         Type *LHSIndexTy = LOffset->getType();
         Type *RHSIndexTy = ROffset->getType();

test/Transforms/InstCombine/icmp.ll

define i1 @test60(i8* %foo, i64 %i, i64 %j) {
   %cmp = icmp ult i8* %cast1, %gep2
   ret i1 %cmp
 ; CHECK-LABEL: @test60(
 ; CHECK-NEXT: %gep1.idx = shl nuw i64 %i, 2
 ; CHECK-NEXT: icmp slt i64 %gep1.idx, %j
 ; CHECK-NEXT: ret i1
 }
 
+define i1 @test60a(i8* %foo, i64 %i, i64 %j) {
+  %bit = bitcast i8* %foo to i32*
+  %gep1 = getelementptr inbounds i32, i32* %bit, i64 %i
+  %fooa = call i8* @llvm.noalias.p0i8(i8* %foo, metadata !1)
+  %gep2 = getelementptr inbounds i8, i8* %fooa, i64 %j
+  %cast1 = bitcast i32* %gep1 to i8*
+  %cmp = icmp ult i8* %cast1, %gep2
+  ret i1 %cmp
+; CHECK-LABEL: @test60a(
+; CHECK-NEXT: %gep1.idx = shl nuw i64 %i, 2
+; CHECK-NEXT: icmp slt i64 %gep1.idx, %j
+; CHECK-NEXT: ret i1
+}
+
 define i1 @test60_as1(i8 addrspace(1)* %foo, i64 %i, i64 %j) {
   %bit = bitcast i8 addrspace(1)* %foo to i32 addrspace(1)*
   %gep1 = getelementptr inbounds i32, i32 addrspace(1)* %bit, i64 %i
   %gep2 = getelementptr inbounds i8, i8 addrspace(1)* %foo, i64 %j
   %cast1 = bitcast i32 addrspace(1)* %gep1 to i8 addrspace(1)*
   %cmp = icmp ult i8 addrspace(1)* %cast1, %gep2
   ret i1 %cmp
 ; CHECK: @test60_as1

 ; CHECK-NEXT: ret i32 %s
 define i32 @f7(i32 %a, i32 %b) {
   %sext = shl i32 %a, 23
   %sext6 = shl i32 %b, 23
   %cmp = icmp ne i32 %sext, %sext6
   %s = select i1 %cmp, i32 10000, i32 0
   ret i32 %s
 }
+
+declare i8* @llvm.noalias.p0i8(i8*, metadata) nounwind
+
+!0 = !{!0, !"some domain"}
+!1 = !{!1, !0, !"some scope"}
+