https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123735
Bug ID: 123735
Summary: x86: switch + tailcall generates unnecessary jump
Product: gcc
Version: 15.2.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: hpa at zytor dot com
Target Milestone: ---
Created attachment 63432
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=63432&action=edit
Test case with output from gcc 15.2.1 and clang 21.1.8
The Linux kernel are these days generally compiled with -fno-jump-tables due to
indirect branch prediction concerns. This means that there are some large and
very performance-sensitive switch trees in Linux; the most critical one being
the system call dispatch. In that tree, the switch contains about 500 entries,
each of which tailcalls another function with a uniform prototype.
Unfortunately, gcc 15.2.1 generates unnecessary jcc-to-jmp for each case, which
at the very best means additional icache pressure. In contrast, clang 21.1.8
does not.
I have attached a simple test case as well as the .i, .s and .o output from gcc
and clang.
