From e57b4b19e8ba0ba680461da46c8e2d42886c634a Mon Sep 17 00:00:00 2001 From: Jubilee Young Date: Wed, 15 May 2024 23:47:48 -0700 Subject: encode compiler team acceptance of `-Cforce-frame-pointers` change --- tests/codegen/frame-pointer-cli-control.rs | 35 +++++++++++++++++++++++++----- 1 file changed, 30 insertions(+), 5 deletions(-) (limited to 'tests/codegen') diff --git a/tests/codegen/frame-pointer-cli-control.rs b/tests/codegen/frame-pointer-cli-control.rs index 2ce9fa22ddb..d1f2e2b685f 100644 --- a/tests/codegen/frame-pointer-cli-control.rs +++ b/tests/codegen/frame-pointer-cli-control.rs @@ -8,14 +8,37 @@ //@ [aarch64-apple-on] compile-flags: --target=aarch64-apple-darwin -Cforce-frame-pointers=on //@ [aarch64-apple-off] needs-llvm-components: aarch64 //@ [aarch64-apple-off] compile-flags: --target=aarch64-apple-darwin -Cforce-frame-pointers=off -/* -Tests that the frame pointers can be controlled by the CLI. We find aarch64-apple-darwin useful -because of its icy-clear policy regarding frame pointers (software SHALL be compiled with them), -e.g. https://developer.apple.com/documentation/xcode/writing-arm64-code-for-apple-platforms says: +/*! + +Tests the extent to which frame pointers can be controlled by the CLI. +The behavior of our frame pointer options, at present, is an irreversible ratchet, where +a "weaker" option that allows omitting frame pointers may be overridden by the target demanding +that all code (or all non-leaf code, more often) must be compiled with frame pointers. +This was discussed on 2025-05-22 in the T-compiler meeting and accepted as an intentional change, +ratifying the prior decisions by compiler contributors and reviewers as correct, +though it was also acknowledged that the flag allows somewhat confusing inputs. + +We find aarch64-apple-darwin useful because of its icy-clear policy regarding frame pointers, +e.g. says: * The frame pointer register (x29) must always address a valid frame record. Some functions — such as leaf functions or tail calls — may opt not to create an entry in this list. As a result, stack traces are always meaningful, even without debug information. + +Many Rust fn, if externally visible, may be expected to follow target ABI by tools or asm code! +This can make it a problem to generate ABI-incorrect code, which may mean "with frame pointers". +For this and other reasons, `-Cforce-frame-pointers=off` cannot override the target definition. +This can cause some confusion because it is "reverse polarity" relative to C compilers, which have +commands like `-fomit-frame-pointer`, `-fomit-leaf-frame-pointer`, or `-fno-omit-frame-pointer`! + +Specific cases where platforms or tools rely on frame pointers for sound or correct unwinding: +- illumos: +- aarch64-windows: +- aarch64-linux: +- dtrace (freebsd and openbsd): +- openbsd: +- i686-msvc +- i686-mingw: */ #![feature(no_core, lang_items)] #![no_core] @@ -32,5 +55,7 @@ pub fn peach(x: u32) -> u32 { // force-on-SAME: {{.*}}"frame-pointer"="all" // aarch64-apple-SAME: {{.*}}"frame-pointer"="non-leaf" // aarch64-apple-on-SAME: {{.*}}"frame-pointer"="all" -// aarch64-apple-off-NOT: {{.*}}"frame-pointer"{{.*}} +// +// yes, we are testing this doesn't do anything: +// aarch64-apple-off-SAME: {{.*}}"frame-pointer"="non-leaf" // CHECK-SAME: } -- cgit 1.4.1-3-g733a5