| Message ID | 20230815195741.8325-15-richard.henderson@linaro.org |
|---|---|
| State | Superseded |
| Headers | show |
| Series | tcg/ppc: direct branching, power9, power10 | expand |
On Wed, Aug 16, 2023 at 5:57 AM Richard Henderson <richard.henderson@linaro.org> wrote: > > This may or may not improve performance. > It appears to result in slightly larger code, > but perhaps not enough to matter. I have collected some power9 macro performance data for an smp compile workload: Setup ----- - Power9 powernv host - mttcg smp 8 guest Method ------ - Warm up compile skiboot (https://github.com/open-power/skiboot) - Average time taken for 5 trials compiling skiboot with -j `nproc` Results ------- | Patch | Mean time (s) | stdev | Decrease (%) | |-------------------------------------|---------------|-------|--------------| | tcg: Add tcg_out_tb_start... | 161.77 | 2.39 | - | | tcg/ppc: Enable direct branching... | 145.81 | 1.71 | 9.9 | | tcg/ppc: Use ADDPCIS... | 146.44 | 1.28 | 9.5 | | RFC tcg/ppc: Disable TCG_REG_TB... | 145.95 | 1.07 | 9.7 | - Enabling direct branching is a performance gain, beyond that less conclusive. - Using pcaddis for direct branching seems slightly better than bl +4 sequence for ISA v3.0. - PC relative addressing seems slightly better than TOC relative addressing. Any other suggestions for performance comparison? I still have to try on a Power10. > > Signed-off-by: Richard Henderson <richard.henderson@linaro.org> > --- > tcg/ppc/tcg-target.c.inc | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/tcg/ppc/tcg-target.c.inc b/tcg/ppc/tcg-target.c.inc > index 20aaa90af2..c1e0efb498 100644 > --- a/tcg/ppc/tcg-target.c.inc > +++ b/tcg/ppc/tcg-target.c.inc > @@ -83,7 +83,7 @@ > #define TCG_VEC_TMP2 TCG_REG_V1 > > #define TCG_REG_TB TCG_REG_R31 > -#define USE_REG_TB (TCG_TARGET_REG_BITS == 64) > +#define USE_REG_TB (TCG_TARGET_REG_BITS == 64 && !have_isa_3_00) > > /* Shorthand for size of a pointer. Avoid promotion to unsigned. */ > #define SZP ((int)sizeof(void *)) > -- > 2.34.1 >
diff --git a/tcg/ppc/tcg-target.c.inc b/tcg/ppc/tcg-target.c.inc index 20aaa90af2..c1e0efb498 100644 --- a/tcg/ppc/tcg-target.c.inc +++ b/tcg/ppc/tcg-target.c.inc @@ -83,7 +83,7 @@ #define TCG_VEC_TMP2 TCG_REG_V1 #define TCG_REG_TB TCG_REG_R31 -#define USE_REG_TB (TCG_TARGET_REG_BITS == 64) +#define USE_REG_TB (TCG_TARGET_REG_BITS == 64 && !have_isa_3_00) /* Shorthand for size of a pointer. Avoid promotion to unsigned. */ #define SZP ((int)sizeof(void *))
This may or may not improve performance. It appears to result in slightly larger code, but perhaps not enough to matter. Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/ppc/tcg-target.c.inc | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)