[InstCombine] Rewrite multi-use GEPs when simplifying comparison (llvm#146100)

We already do this when both sides are a GEP, but not if only one is.
This ensures that the offset arithmetic is not duplicated.

Author: nikic

llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

@@ -711,7 +711,7 @@ Instruction *InstCombinerImpl::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
   Value *PtrBase = GEPLHS->getOperand(0);
   if (PtrBase == RHS && CanFold(GEPLHS->getNoWrapFlags())) {
     // ((gep Ptr, OFFSET) cmp Ptr)   ---> (OFFSET cmp 0).
-    Value *Offset = EmitGEPOffset(GEPLHS);
+    Value *Offset = EmitGEPOffset(GEPLHS, /*RewriteGEP=*/true);
     return NewICmp(GEPLHS->getNoWrapFlags(), Offset,
                    Constant::getNullValue(Offset->getType()));
   }

llvm/test/Transforms/InstCombine/icmp-gep.ll

@@ -221,6 +221,34 @@ define i1 @eq_base_inbounds_commute_use(i64 %y) {
   ret i1 %r
 }
 
+define i1 @ne_base_inbounds_use_scaled(ptr %x, i64 %y) {
+; CHECK-LABEL: @ne_base_inbounds_use_scaled(
+; CHECK-NEXT:    [[G_IDX:%.*]] = shl nsw i64 [[Y:%.*]], 3
+; CHECK-NEXT:    [[G:%.*]] = getelementptr inbounds i8, ptr [[X:%.*]], i64 [[G_IDX]]
+; CHECK-NEXT:    call void @use(ptr [[G]])
+; CHECK-NEXT:    [[R:%.*]] = icmp ne i64 [[Y]], 0
+; CHECK-NEXT:    ret i1 [[R]]
+;
+  %g = getelementptr inbounds i64, ptr %x, i64 %y
+  call void @use(ptr %g)
+  %r = icmp ne ptr %g, %x
+  ret i1 %r
+}
+
+define i1 @ne_base_use_scaled(ptr %x, i64 %y) {
+; CHECK-LABEL: @ne_base_use_scaled(
+; CHECK-NEXT:    [[G_IDX_MASK:%.*]] = shl i64 [[Y:%.*]], 3
+; CHECK-NEXT:    [[G:%.*]] = getelementptr i8, ptr [[X:%.*]], i64 [[G_IDX_MASK]]
+; CHECK-NEXT:    call void @use(ptr [[G]])
+; CHECK-NEXT:    [[R:%.*]] = icmp ne i64 [[G_IDX_MASK]], 0
+; CHECK-NEXT:    ret i1 [[R]]
+;
+  %g = getelementptr i64, ptr %x, i64 %y
+  call void @use(ptr %g)
+  %r = icmp ne ptr %g, %x
+  ret i1 %r
+}
+
 define i1 @eq_bitcast_base(ptr %p, i64 %x) {
 ; CHECK-LABEL: @eq_bitcast_base(
 ; CHECK-NEXT:    [[GEP_IDX_MASK:%.*]] = and i64 [[X:%.*]], 9223372036854775807

llvm/test/Transforms/PhaseOrdering/loop-access-checks.ll

@@ -24,7 +24,8 @@ define void @test_fill_with_foreach([2 x i64] %elems.coerce) {
 ; CHECK-NEXT:    [[ELEMS_COERCE_FCA_0_EXTRACT:%.*]] = extractvalue [2 x i64] [[ELEMS_COERCE]], 0
 ; CHECK-NEXT:    [[TMP0:%.*]] = inttoptr i64 [[ELEMS_COERCE_FCA_0_EXTRACT]] to ptr
 ; CHECK-NEXT:    [[ELEMS_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [2 x i64] [[ELEMS_COERCE]], 1
-; CHECK-NEXT:    [[ADD_PTR_I:%.*]] = getelementptr inbounds i32, ptr [[TMP0]], i64 [[ELEMS_COERCE_FCA_1_EXTRACT]]
+; CHECK-NEXT:    [[ADD_PTR_I_IDX:%.*]] = shl nsw i64 [[ELEMS_COERCE_FCA_1_EXTRACT]], 2
+; CHECK-NEXT:    [[ADD_PTR_I:%.*]] = getelementptr inbounds i8, ptr [[TMP0]], i64 [[ADD_PTR_I_IDX]]
 ; CHECK-NEXT:    [[CMP_NOT_I_I_I_I:%.*]] = icmp slt i64 [[ELEMS_COERCE_FCA_1_EXTRACT]], 0
 ; CHECK-NEXT:    br i1 [[CMP_NOT_I_I_I_I]], label [[ERROR:%.*]], label [[FOR_COND_PREHEADER_SPLIT:%.*]]
 ; CHECK:       for.cond.preheader.split: