Convert an APInt to int64_t properly in TTI::getGEPCost().

[jlebar]

Summary

If the pointer width is 32 bits and the calculated GEP offset is
negative, we call APInt::getLimitedValue(), which does a
*zero*-extension of the offset. That's wrong -- we should do an sext.

Fixes a bug introduced in rL314362 and found by Evgeny Astigeevich.

[efriedma]

llvm/include/llvm/Analysis/TargetTransformInfoImpl.h

726

This doesn't actually clamp the way you want it to; suppose BaseOffset is less than int64_t::min.

I'd suggest just BaseOffset.sextOrTrunc(64).getSExtValue(); not precisely the same thing, but good enough until someone comes along with an architecture which actually needs 128-bit pointer offsets.

[jlebar]

Address efriedma's comment.

llvm/include/llvm/Analysis/TargetTransformInfoImpl.h

726

Indeed, that's better, thank you.

[efriedma]

LGTM

[jlebar]

Thank you for the review.

llvm/include/llvm/Analysis/TargetTransformInfoImpl.h

for (auto I = Operands.begin(); I != Operands.end(); ++I, ++GTI) {
           Scale = ElementSize;
         }
       }
     }
 
     // Assumes the address space is 0 when Ptr is nullptr.
     unsigned AS =
         (Ptr == nullptr ? 0 : Ptr->getType()->getPointerAddressSpace());
+
+    // Convert BaseOffset to an int64_t, clamping to int64_t::max if necessary.
+    int64_t BaseOffsetInt64 =
+        BaseOffset.sgt(std::numeric_limits<int64_t>::max())
+            ? std::numeric_limits<int64_t>::max()
+            : BaseOffset.getSExtValue();
     if (static_cast<T *>(this)->isLegalAddressingMode(
-            TargetType, const_cast<GlobalValue *>(BaseGV),
-            static_cast<int64_t>(BaseOffset.getLimitedValue()), HasBaseReg,
-            Scale, AS))
+            TargetType, const_cast<GlobalValue *>(BaseGV), BaseOffsetInt64,
+            HasBaseReg, Scale, AS))
       return TTI::TCC_Free;
     return TTI::TCC_Basic;
   }
 
   int getGEPCost(const GEPOperator *GEP, ArrayRef<const Value *> Operands) {
     if (!isa<Instruction>(GEP))
       return TTI::TCC_Basic;
 

llvm/test/Analysis/CostModel/ARM/gep.ll

;CHECK: cost of 1 for instruction: {{.*}} getelementptr inbounds <4 x i32>, <4 x i32>*
   %c9 = getelementptr inbounds <4 x i32>, <4 x i32>* undef, i32 %i
 ;CHECK: cost of 1 for instruction: {{.*}} getelementptr inbounds <4 x i64>, <4 x i64>*
   %c10 = getelementptr inbounds <4 x i64>, <4 x i64>* undef, i32 %i
 ;CHECK: cost of 1 for instruction: {{.*}} getelementptr inbounds <4 x float>, <4 x float>*
   %c11 = getelementptr inbounds <4 x float>, <4 x float>* undef, i32 %i
 ;CHECK: cost of 1 for instruction: {{.*}} getelementptr inbounds <4 x double>, <4 x double>*
   %c12 = getelementptr inbounds <4 x double>, <4 x double>* undef, i32 %i
 
+;CHECK: cost of 0 for instruction: {{.*}} getelementptr inbounds i8, i8*
+  %d0 = getelementptr inbounds i8, i8* undef, i32 -1
+
   ret void
 }