[SimplifyLibCalls] Retain attributes added by Builder.CreateMem*

[arichardson]

Summary

This currently does not make much of a difference (only one tests is
affected), but it is helpful e.g. for the out-of-tree CHERI target where
Builder.CreateMemCpy() can add attributes other than parameter alignment.

[nikic]

LGTM

This currently does not make much of a difference (only one tests is
affected), but it is helpful e.g. for the out-of-tree CHERI target where
Builder.CreateMemCpy() can add attributes other than parameter alignment.

Out of interest, what attributes does it add?

[arichardson]

In D135075#3830939, @nikic wrote:

LGTM

This currently does not make much of a difference (only one tests is
affected), but it is helpful e.g. for the out-of-tree CHERI target where
Builder.CreateMemCpy() can add attributes other than parameter alignment.

Out of interest, what attributes does it add?

CHERI has strict requirements on load/store alignment to preserve the hidden validity bits (you have to use aligned 16 byte loads/stores on CHERI-extended 64-bit architectures). This means we can't emit memcpy/memmove inline if the known alignment is only 8 bytes since 8 byte loads&stores will invalidate all validity bits. This means something like a 16 byte memcpy with 8 byte alignment has to be lowered to a memcpy() call so that it preserves validity bits if it is aligned to 16 bytes at runtime.

I added a new attribute "no_preserve_cheri_tags" that the frontend (as well as a few passes such as SLC) can use to indicate that the memory being copied will never have validity tags sets, which means we can inline the calls using smaller loads&stores.

[jrtc27]

(Well, you can emit it inline so long as you're happy to branch, at which point it's not really much of a gain over making a call to memcpy and has the issue of code bloat, but maybe for particularly hot places you'd want to still inline a partially-specialised version, and you have to for __builtin_memcpy_inline)

llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp

static Value *copyFlags(const CallInst &Old, Value *New) {
   assert(!Old.isMustTailCall() && "do not copy musttail call flags");
   assert(!Old.isNoTailCall() && "do not copy notail call flags");
   if (auto *NewCI = dyn_cast_or_null<CallInst>(New))
     NewCI->setTailCallKind(Old.getTailCallKind());
   return New;
 }
 
 static Value *mergeAttributesAndFlags(CallInst *NewCI, const CallInst &Old) {
-  NewCI->setAttributes(Old.getAttributes());
+  NewCI->setAttributes(AttributeList::get(
+      NewCI->getContext(), {NewCI->getAttributes(), Old.getAttributes()}));
   NewCI->removeRetAttrs(AttributeFuncs::typeIncompatible(NewCI->getType()));
   return copyFlags(Old, NewCI);
 }
 
 // Helper to avoid truncating the length if size_t is 32-bits.
 static StringRef substr(StringRef Str, uint64_t Len) {
   return Len >= Str.size() ? Str : Str.substr(0, Len);
 }

llvm/test/Transforms/CodeGenPrepare/X86/memset_chk-simplify-nobuiltin.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt -S -mtriple=x86_64 -disable-simplify-libcalls -codegenprepare < %s | FileCheck %s
 target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
 
 ; This is a workaround for PR23093: when building with -mkernel/-fno-builtin,
 ; we still generate fortified library calls.
 
 ; Check that we ignore two things:
 ; - attribute nobuiltin
 ; - TLI::has (always returns false thanks to -disable-simplify-libcalls)
 
 define void @test_nobuiltin(i8* %dst, i64 %len) {
 ; CHECK-LABEL: @test_nobuiltin(
-; CHECK-NEXT:    call void @llvm.memset.p0i8.i64(i8* [[DST:%.*]], i8 0, i64 [[LEN:%.*]], i1 false) #1
+; CHECK-NEXT:    call void @llvm.memset.p0i8.i64(i8* align 1 [[DST:%.*]], i8 0, i64 [[LEN:%.*]], i1 false)
 ; CHECK-NEXT:    ret void
 ;
   call i8* @__memset_chk(i8* %dst, i32 0, i64 %len, i64 -1) nobuiltin
   ret void
 }
 
 declare i8* @__memset_chk(i8*, i32, i64, i64)