Eli Friedman

This isn't a recent regression, no. Was reported to me against an old version, and didn't get around to reporting it here for a while. Compare against clang-8. I don't...

Maybe if instcombine lowers an llvm.memset.inline, it should mark the function with an attribute suppressing the automatic formation of llvm.memset?

I think "no-builtin-memset" technically works, but not sure it's the cleanest way to express the semantics we want.

Consider the following: ``` define i32 @f(i1 signext %0) { %r = sext i1 %0 to i32 ret i32 %r } define i32 @g() { %r = call i32 @f(i1...

It sounds like some specific pass is getting more expensive due to runtime checks, or something like that. You're going to have to figure out what that pass is. Maybe...

I saw what I think is the same issue on the reverse-iteration buildbot. Didn't realize it was impacting release testing. /cherry-pick 2c29268bfcc84c3b94bcb0aa34b7ef9c9bd9af01

/cherry-pick 2c29268bfcc84c3b94bcb0aa34b7ef9c9bd9af01

Let's keep this open until the fix is merged to 15.0.2

"String table overflow" has been reported before; see #57718/#51143 . You should be able to work around it by building with clang instead. For memory usage, I've noticed that building...

Even ignoring the poison vs undef thing, the transform isn't legal: the function `@bar` could check if it's passed a pointer to `@g`.