[llvm] SimplifyLibCalls: Use the correct address space when computing integer widths. (PR #118586)

[Owen Anderson via llvm-commits]

================
@@ -590,26 +588,35 @@ Value *LibCallSimplifier::optimizeStrCmp(CallInst *CI, IRBuilderBase &B) {
   if (Len1 && Len2) {
     return copyFlags(
         *CI, emitMemCmp(Str1P, Str2P,
-                        ConstantInt::get(DL.getIntPtrType(CI->getContext()),
-                                         std::min(Len1, Len2)),
+                        ConstantInt::get(
+                            DL.getIntPtrType(
+                                CI->getContext(),
+                                Str1P->getType()->getPointerAddressSpace()),
+                            std::min(Len1, Len2)),
----------------
resistor wrote:

There's a few alternatives in here, none of them great:

* We can use the current state of this patch, which will have every target always use i64. As efriedma points out, this will probably work out because of backend legalization.
* We can thread TLI into `IRBuilder::CreateMemCpy` so that it can pick the right size. This requires updating a number of other passes throughout the tree, many of which don't depends on TargetLibraryInfo today.
* Same as above, but we add it in parallel to the existing overloads so that we can migrate users gradually.
* Build the types in SimplifyLibCalls as in the original version of this diff. Do some local cleanup to make it more readable per arsenm's feedback.

Thoughts?

https://github.com/llvm/llvm-project/pull/118586