Table 2-4 of the 21264/21364 Compiler Writer's Guide gives itof a
latency of 4, but ev6.md has used 3.
Measured on an EV68AL, itof and ftoi can only be chained through one
another, so I separated them by inserting a known number of FP adds. A
chain of itof, fadd, ftoi costs 13.105 cycles and itof, fadd, fadd, ftoi
costs 17.128, against 4 + 6 + 3 = 13 and 4 + 4 + 6 + 3 = 17 predicted by
the guide with an itof latency of 4.
An itof result consumed directly by an ftoi costs more than the two
latencies imply. A chain alternating the two takes 11.083 cycles per
round trip, and ftoi is 3, so the itof to ftoi edge is 8 rather than 4.
Routing the value through any FP operate removes the penalty, which is
why the chains above land on the latencies the guide predicts. Add a
bypass for the direct case.
Section 3.7.1 of the guide recommends ftoi and itof over bouncing a
value through memory, to avoid a store-load replay trap. With this
penalty the register path is only 1.20x faster than the memory path
(11.083 against 13.335 cycles).
gcc/ChangeLog:
* config/alpha/ev6.md (ev6_itof): Correct the latency to 4.
Add a bypass from ev6_itof to ev6_ftoi.
---
gcc/config/alpha/ev6.md | 9 +++++++--
1 file changed, 7 insertions(+), 2 deletions(-)
diff --git a/gcc/config/alpha/ev6.md b/gcc/config/alpha/ev6.md
index 6e9317230c7..3c1e6cafb38 100644
--- a/gcc/config/alpha/ev6.md
+++ b/gcc/config/alpha/ev6.md
@@ -126,8 +126,9 @@
(eq_attr "type" "jsr"))
"ev6_l0")
-; Ftoi/itof only issue to lower pipes.
-(define_insn_reservation "ev6_itof" 3
+; Ftoi/itof only issue to lower pipes. Table 2-4 of the Compiler Writer's
+; Guide gives itof a latency of 4.
+(define_insn_reservation "ev6_itof" 4
(and (eq_attr "tune" "ev6")
(eq_attr "type" "itof"))
"ev6_l")
@@ -137,6 +138,10 @@
(eq_attr "type" "ftoi"))
"ev6_fst,nothing,ev6_l")
+; An itof feeding an ftoi directly costs 8 cycles rather than 4. Routing the
+; value through a floating-point operate avoids the penalty.
+(define_bypass 8 "ev6_itof" "ev6_ftoi")
+
(define_insn_reservation "ev6_fmul" 4
(and (eq_attr "tune" "ev6")
(eq_attr "type" "fmul"))
--
2.56.1
