[llvm-commits] [PATCH] Mips backend. Emit $gp restore instructions.

[Bruno Cardoso Lopes]

On Tue, May 3, 2011 at 8:18 PM, Akira Hatanaka <ahatanak at gmail.com> wrote:
> This patch prevents instructions using $gp from being placed between a jalr
> and the instruction that restores the clobbered $gp (please see the
> following example).
>
> The current implementation creates load nodes that restore $gp when call
> instructions are lowered in MipsISelLowering::LowerCall. However, it does
> not add chains or glues that are needed to make sure load/store nodes that
> have $gp as an operand are not scheduled between JALR nodes and nodes that
> restore $gp after the call.
>
> This patch runs a pre-regalloc pass that emits $gp restore instructions
> instead.
>
> - bitcode
> @p = external global i32
> ...
> tail call void (...)* @f1() nounwind
> %tmp = load i32* @p, align 4, !tbaa !0
>
> - currently produces following code:
> ...
> lw  $25, %call16(f1)($gp)
> jalr  $25                        # calls f1. $gp is clobbered.
> nop
> lw  $3, 16($sp)              # load $gp from stack.
> lw  $2, %got(p)($gp)      # this instr uses a clobbered $gp.
> addu  $gp, $zero, $3     # $gp is restored here.
> ...
>
> - produces following code after patch is applied.
> ...
> lw  $25, %call16(f1)($gp)
> lw  $16, %got(p)($gp)  # $gp has correct value here.
> jalr  $25                     # calls f1. clobbers $gp.
> nop
> lw  $gp, 16($sp)          # $gp restored.
> ...

I've tried other approaches to solve this issue but they we're not
clean enough, I can't figure out now another way besides doing a
separate pass. If someone else knows a better a approach we can go for
it, but in the meantime go for it. Please commit.



-- 
Bruno Cardoso Lopes
http://www.brunocardoso.cc