I experimented with having the compiler spill the two registers that
are clobbered by the write barrier fast path, but it slightly slows
down compilebench, which is a good write barrier benchmark:
name old time/op new time/op delta
Template 175ms ± 0% 176ms ± 1% ~ (p=0.393 n=10+10)
Unicode 83.6ms ± 1% 85.1ms ± 2% +1.79% (p=0.000 n=9+10)
GoTypes 585ms ± 0% 588ms ± 1% ~ (p=0.173 n=8+10)
Compiler 2.78s ± 1% 2.81s ± 2% +0.81% (p=0.023 n=10+10)
SSA 7.11s ± 1% 7.15s ± 1% +0.59% (p=0.029 n=10+10)
Flate 115ms ± 1% 116ms ± 2% ~ (p=0.853 n=10+10)
GoParser 144ms ± 2% 145ms ± 2% ~ (p=1.000 n=10+10)
Reflect 389ms ± 1% 390ms ± 1% ~ (p=0.481 n=10+10)
Tar 185ms ± 2% 185ms ± 2% ~ (p=0.529 n=10+10)
XML 205ms ± 0% 207ms ± 2% ~ (p=0.065 n=9+10)
Since this didn't pan out, remove the TODO.
Change-Id: I2186942c6d1ba10585a5da03cd7c1d26ce906273
Reviewed-on: https://go-review.googlesource.com/78034
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
Reviewed-by: Rick Hudson <rlh@golang.org>
// It clobbers FLAGS. It does not clobber any general-purpose registers,
// but may clobber others (e.g., SSE registers).
TEXT runtime·gcWriteBarrier(SB),NOSPLIT,$120
- // Save the registers clobbered by the fast path.
- //
- // TODO: Teach the register allocator that this clobbers some registers
- // so we don't always have to save them? Use regs it's least likely to
- // care about.
+ // Save the registers clobbered by the fast path. This is slightly
+ // faster than having the caller spill these.
MOVQ R14, 104(SP)
MOVQ R13, 112(SP)
// TODO: Consider passing g.m.p in as an argument so they can be shared