On 11/18/23 22:42, Maciej W. Rozycki wrote:
We have floating-point coditional-set machine instructions for a subset
of FP comparisons, so avoid going through a comparison against constant
zero in `riscv_expand_float_scc' where not necessary, preventing an
extraneous RTL instruction from being produced that counts against the
cost of the replacement branchless code sequence in if-conversion, e.g.:
(insn 29 6 30 2 (set (reg:DI 142)
(ge:DI (reg/v:DF 135 [ w ])
(reg/v:DF 136 [ x ]))) 297 {*cstoredfdi4}
(nil))
(insn 30 29 31 2 (set (reg:DI 143)
(ne:DI (reg:DI 142)
(const_int 0 [0]))) 319 {*sne_zero_didi}
(nil))
(insn 31 30 32 2 (set (reg:DI 141)
(reg:DI 143)) 206 {*movdi_64bit}
(nil))
(insn 32 31 33 2 (set (reg:DI 144)
(neg:DI (reg:DI 141))) 15 {negdi2}
(nil))
(insn 33 32 34 2 (set (reg:DI 145)
(and:DI (reg:DI 144)
(reg/v:DI 137 [ y ]))) 102 {*anddi3}
(nil))
(insn 34 33 35 2 (set (reg:DI 146)
(not:DI (reg:DI 144))) 111 {one_cmpldi2}
(nil))
(insn 35 34 36 2 (set (reg:DI 147)
(and:DI (reg:DI 146)
(reg/v:DI 138 [ z ]))) 102 {*anddi3}
(nil))
(insn 36 35 21 2 (set (reg/v:DI 138 [ z ])
(ior:DI (reg:DI 145)
(reg:DI 147))) 105 {iordi3}
(nil))
where the second insn effectively just copies its input. This now gets
simplified to:
(insn 29 6 30 2 (set (reg:DI 141)
(ge:DI (reg/v:DF 135 [ w ])
(reg/v:DF 136 [ x ]))) 297 {*cstoredfdi4}
(nil))
(insn 30 29 31 2 (set (reg:DI 142)
(neg:DI (reg:DI 141))) 15 {negdi2}
(nil))
(insn 31 30 32 2 (set (reg:DI 143)
(and:DI (reg:DI 142)
(reg/v:DI 137 [ y ]))) 102 {*anddi3}
(nil))
(insn 32 31 33 2 (set (reg:DI 144)
(not:DI (reg:DI 142))) 111 {one_cmpldi2}
(nil))
(insn 33 32 34 2 (set (reg:DI 145)
(and:DI (reg:DI 144)
(reg/v:DI 138 [ z ]))) 102 {*anddi3}
(nil))
(insn 34 33 21 2 (set (reg/v:DI 138 [ z ])
(ior:DI (reg:DI 143)
(reg:DI 145))) 105 {iordi3}
(nil))
lowering the cost of the code sequence produced (even though combine
would swallow the second insn anyway).
We still need to produce a comparison against constant zero where the
instruction following a floating-point coditional-set operation is a
branch, so add canonicalization to `riscv_expand_conditional_branch'
instead.
gcc/
* config/riscv/riscv.cc (riscv_emit_float_compare) <NE>: Handle
separately.
<EQ, LE, LT, GE, GT>: Return operands supplied as is.
(riscv_emit_binary): Call `riscv_emit_binary' directly rather
than going through a temporary register for word-mode targets.
(riscv_expand_conditional_branch): Canonicalize the comparison
if not against constant zero.
OK
jeff
