[llvm] [InstCombine] Optimize GEP comparisons with constant offsets (PR #208547)

[Ömer Sinan Ağacan via llvm-commits]

================
@@ -815,14 +815,35 @@ Instruction *InstCombinerImpl::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
       }
     }
 
-    if (Base.Ptr && CanFold(Base.LHSNW & Base.RHSNW) && !Base.isExpensive()) {
+    if (Base.Ptr && !Base.isExpensive()) {
       // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2)  --->  (OFFSET1 cmp OFFSET2)
-      Type *IdxTy = DL.getIndexType(GEPLHS->getType());
-      Value *L =
-          EmitGEPOffsets(Base.LHSGEPs, Base.LHSNW, IdxTy, /*RewriteGEP=*/true);
-      Value *R =
-          EmitGEPOffsets(Base.RHSGEPs, Base.RHSNW, IdxTy, /*RewriteGEP=*/true);
-      return NewICmp(Base.LHSNW & Base.RHSNW, L, R);
+      bool DoFold = CanFold(Base.LHSNW & Base.RHSNW);
+
+      if (!DoFold && Base.Ptr->getType()->isPointerTy()) {
+        // Without the flags, we can still fold if the offsets are constant and
+        // they cross the base's alignment boundary the same number of times, so
+        // either both arguments will wrap, or none of them will.
+        unsigned BW = DL.getIndexTypeSizeInBits(GEPLHS->getType());
+        APInt Alignment = APInt(BW, Base.Ptr->getPointerAlignment(DL).value());
+        APInt LOff(BW, 0);
+        APInt ROff(BW, 0);
+        if (GEPLHS->stripAndAccumulateConstantOffsets(
+                DL, LOff, /*AllowNonInbounds=*/true) == Base.Ptr &&
+            RHS->stripAndAccumulateConstantOffsets(
+                DL, ROff, /*AllowNonInbounds=*/true) == Base.Ptr)
+          DoFold =
+              APIntOps::RoundingSDiv(LOff, Alignment, APInt::Rounding::DOWN) ==
+              APIntOps::RoundingSDiv(ROff, Alignment, APInt::Rounding::DOWN);
+      }
+
+      if (DoFold) {
+        Type *IdxTy = DL.getIndexType(GEPLHS->getType());
+        Value *L = EmitGEPOffsets(Base.LHSGEPs, Base.LHSNW, IdxTy,
+                                  /*RewriteGEP=*/true);
+        Value *R = EmitGEPOffsets(Base.RHSGEPs, Base.RHSNW, IdxTy,
+                                  /*RewriteGEP=*/true);
+        return NewICmp(Base.LHSNW & Base.RHSNW, L, R);
----------------
osa1 wrote:

This block is the same as before, just indented.

(so transformation is the same, we just do it in more cases now)

https://github.com/llvm/llvm-project/pull/208547