RFR: 8266609: AArch64: include FP/LR space in LIR_Assembler::initial_frame_size_in_bytes()

[Andrew Haley]

On Thu, 6 May 2021 10:03:14 GMT, Nick Gasson <ngasson at openjdk.org> wrote:

> LIR_Assembler::initial_frame_size_in_bytes() returns the frame size
> without the additional 2*wordSize needed to store FP/LR (i.e. just the
> space required for the C1 locals).  When we pass this value to
> MacroAssembler build_frame and remove_frame we need to remember to add
> back the 2*wordSize.  This patch changes initial_frame_size_in_bytes()
> to return the full frame size including the space for FP/LR.  The PPC
> and S390 ports already do this, and it also matches the behaviour of
> C->output()->frame_size_in_bytes() in C2.
> 
> This change may seem a bit trivial in mainline but the Valhalla lworld
> branch makes more calls to build_frame/remove_frame in C1 and keeping
> track of whether "framesize" is with or without FP/LR quickly becomes
> confusing.
> 
> Tested tier1 on AArch64.  The only use of this function is as input to
> C1_MacroAssembler build_frame() and remove_frame() so seems safe.

This calculation is the same as x86. Can you explain why it has to be changed on AArch64, but should not be changed on x86?

src/hotspot/cpu/aarch64/c1_LIRAssembler_aarch64.cpp line 377:

> 375:   // if rounding, must let FrameMap know!
> 376: 
> 377:   return in_bytes(frame_map()->framesize_in_bytes());

Umm, really? `framesize_in_bytes()` returns a value that has to be converted to bytes?

-------------

PR: https://git.openjdk.java.net/jdk/pull/3898