Even when a field is marked const, we may emit multiple consecutive in-literal stores to that field. That is, in 'JSNativeContextSpecialization::BuildPropertyStore', when the access mode is 'kStoreInLiteral' and we are accessing a const field, we may produce a StoreField node, even though another StoreField (that stores something other than 'Uninitialized') to the same const field dominates it. This appears to be sound, since earlier stores to literals cannot be observed anyways.
Unfortunately this behavior conflicts with the double const store invariant in load elimination: Roughly speaking, we assume that load elimination may never observe two consecutive const stores to the same field on the same object.
The apparent solution would be to treat 'kStoreInLiteral' accesses like regular 'kStore' accesses: For consecutive stores to const properties we don't emit StoreField, but instead emit code that checks whether the value about to be written is equivalent to the previously written one, and otherwise deopt ('DeoptimizeReason::kWrongValue'). Unfortunately this turns out impractical, since for 'kStoreInLiteral' accesses we can't easily decide whether we're dealing with the first such store or one of the consecutive ones. Also see this abandoned CL: https://chromium-review.googlesource.com/c/v8/v8/+/1762020.
This CL instead adds an exception to the invariant in load elimination. We track whether a store arose from a 'kStoreInLiteral' access, and use this information when visiting StoreField nodes in load elimination.
R=neis@chromium.org, tebbi@chromium.org
Bug: chromium:987205
Change-Id: I8829752aa0637e9599677d20aad2d706d40d7fe6
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/1763535
Reviewed-by: Tobias Tebbi <tebbi@chromium.org>
Reviewed-by: Georg Neis <neis@chromium.org>
Commit-Queue: Georg Schmid <gsps@google.com>
Cr-Commit-Position: refs/heads/master@{#63385}
7fd1922823