Extended Description

--
$ cat memcmp.ll
target datalayout = "e-i8:8:8-i16:16:16"
target triple = "x86_64-unknown-unknown"
declare i32 @​memcmp(i8* nocapture, i8* nocapture, i64)

define i32 @​cmp2(i8* nocapture readonly %x, i8* nocapture readonly %y) {
%call = tail call i32 @​memcmp(i8* %x, i8* %y, i64 2)
ret i32 %call
}

$ ./llvm/bin/opt -expandmemcmp -S -o - memcmp.ll
; ModuleID = 'memcmp.ll'
source_filename = "memcmp.ll"
target datalayout = "e-i8:8:8-i16:16:16"
target triple = "x86_64-unknown-unknown"

declare i32 @​memcmp(i8* nocapture, i8* nocapture, i64)

define i32 @​cmp2(i8* nocapture readonly %x, i8* nocapture readonly %y) {
%1 = bitcast i8* %x to i16*
%2 = bitcast i8* %y to i16*
%3 = load i16, i16* %1
%4 = load i16, i16* %2
%5 = call i16 @​llvm.bswap.i16(i16 %3)
%6 = call i16 @​llvm.bswap.i16(i16 %4)
%7 = zext i16 %5 to i32
%8 = zext i16 %6 to i32
%9 = sub i32 %7, %8
ret i32 %9
}

; Function Attrs: nounwind readnone speculatable willreturn
declare i16 @​llvm.bswap.i16(i16) #​0

attributes #​0 = { nounwind readnone speculatable willreturn }

This is incorrect because the loads in the output has align 2 (omitted alignment in loads mean they have the ABI alignment). If %x or %y is not 2-byte aligned, the optimized code raises undefined behavior, where as the source wouldn't.