[arm64] Fix backwards branch ranges.
The `Instruction::IsValidImmPCOffset()` method was taking an `offset` argument in numbers of *instructions* while we were passing it numbers of *bytes*. See `Instruction::IsTargetInImmPCOffsetRange()` and `MacroAssembler::NeedExtraInstructionsOrRegisterBranch()`. As a result, we were 4 times too conservative when computing branch ranges going backwards, forcing us to generate the following sequence for TBZ more often than needed: ``` TBNZ <skip> B <target> skip: ``` This happened rarely for loops, but a lot when doing an early return from out-of-line calls to write barriers. Since out-of-line code is easily out of range of 8K, although the real range of TBZ is 32K. This fixes it by changing this method to take a byte offset instead of instructions, as this is more intuitive and in line with similar methods. For instance, `Instruction::ImmPcOffset()` returns an offset in bytes. The tests are adapted so that they would have caught such a bug: * TEST(far_branch_backward): This test used to only check the code worked if the branch was very far away, but it didn't test the range was correct. So this test was changed to check each branch type separately, and test in-range and out-of-range cases separately too. * TEST(far_branch_veneer_broken_link_chain): Because of the backwards range bug, this test wasn't actually testing what it should. The idea of the test is to make sure the MacroAssembler can still cope when the chain of links is broken after a veneer was emitted. But no veneers were ever emitted. Change-Id: Iddb5c683a71147455175f38fa7ae57da0a3e7337 Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/1781058Reviewed-by: Ulan Degenbaev <ulan@chromium.org> Reviewed-by: Sigurd Schneider <sigurds@chromium.org> Commit-Queue: Sigurd Schneider <sigurds@chromium.org> Cr-Commit-Position: refs/heads/master@{#63518}
Showing
This diff is collapsed.
Please
register
or
sign in
to comment