qemu-ppc
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [PATCH v2 2/5] target/ppc/power8-pmu-insn-cnt: introduce inc_spr_if_


From: Richard Henderson
Subject: Re: [PATCH v2 2/5] target/ppc/power8-pmu-insn-cnt: introduce inc_spr_if_cond()
Date: Thu, 23 Dec 2021 13:14:02 -0800
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:78.0) Gecko/20100101 Thunderbird/78.14.0

On 12/23/21 12:18 PM, Daniel Henrique Barboza wrote:
The code that increments a PMC is repetitive: check if a given register
has a bit/mask set or cleared and increment the counter.

inc_spr_if_cond() will help deal with this repetition. This patch also
gives a sample of how the function works by incrementing PMC5, which is
supposed to be incremented only if MMCR0_FC56 is not set.

We've also removing the call from the helper since that would cause
PMC5 to be counted twice.

Signed-off-by: Daniel Henrique Barboza <danielhb413@gmail.com>
---
  target/ppc/power8-pmu-insn-cnt.c.inc | 43 ++++++++++++++++++++++------
  1 file changed, 34 insertions(+), 9 deletions(-)

diff --git a/target/ppc/power8-pmu-insn-cnt.c.inc 
b/target/ppc/power8-pmu-insn-cnt.c.inc
index 6cdf2d2d88..3cfb801c69 100644
--- a/target/ppc/power8-pmu-insn-cnt.c.inc
+++ b/target/ppc/power8-pmu-insn-cnt.c.inc
@@ -10,6 +10,38 @@
   * See the COPYING file in the top-level directory.
   */
+#if defined(TARGET_PPC64) && !defined(CONFIG_USER_ONLY)
+/*
+ * This function increments a SPR 'spr' by 'inc_val' if a given
+ * register 'reg' has a bitmask 'mask' set (cond = TCG_COND_EQ) or
+ * not set (TCG_COND_NE).
+ */
+static void inc_spr_if_cond(int reg, uint64_t mask, TCGCond cond,
+                            int spr, int inc_val)
+{
+    TCGCond exit_cond = tcg_invert_cond(cond);
+    TCGLabel *l_exit;
+    TCGv t0, t1;
+
+    l_exit = gen_new_label();
+
+    t0 = tcg_temp_new();
+    gen_load_spr(t0, reg);
+    tcg_gen_andi_tl(t0, t0, mask);
+    tcg_gen_brcondi_tl(exit_cond, t0, mask, l_exit);

When testing a single bit, compare against 0, not the bit.

+    t1 = tcg_temp_new();
+    gen_load_spr(t1, spr);
+    tcg_gen_addi_tl(t1, t1, inc_val);
+    gen_store_spr(spr, t1);

It will probably perform better to make this a true conditional add.
I.e.

    gen_load_spr(t0, spr);
    tcg_gen_addi_tl(t1, t0, inc);
    tcg_gen_movcond_tl(cond, t0, reg, zero, t1, t0);
    gen_store_spr(spr, t0);

-    /*
-     * The PMU insns_inc() helper stops the internal PMU timer if a
-     * counter overflows happens. In that case, if the guest is
-     * running with icount and we do not handle it beforehand,
-     * the helper can trigger a 'bad icount read'.
-     */
-    gen_icount_io_start(ctx);

Removing this is incorrect.

-    gen_helper_insns_inc(cpu_env, tcg_constant_i32(ctx->base.num_insns));
+    inc_spr_if_cond(SPR_POWER_MMCR0, MMCR0_FC56, TCG_COND_NE,
+                    SPR_POWER_PMC5, ctx->base.num_insns);

This is non-bisectable. You're removing support for all registers and only adding back PMC5. You add them all back before the end of the series, but the middle patches do not behave correctly.


r~



reply via email to

[Prev in Thread] Current Thread [Next in Thread]