https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63986
Bug ID: 63986
Summary: [5 Regression][SH] gcc.target/sh/pr51244-15.c failures
Product: gcc
Version: 5.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: olegendo at gcc dot gnu.org
Target: sh
Starting with r217496 the SH specific test cases in gcc.target/sh/pr51244-15.c
are failing.
The reason for this is that, e.g. a function such as
typedef char bool;
int
test_0 (int a, int b, int c, int* d)
{
bool x = a == 0;
d[2] = !x;
return x ? b : c;
}
with -O2 -m4 compiles to:
tst r4,r4 <<< test against zero
mov #-1,r1
negc r1,r1
tst r4,r4 <<< redundant test against zero
bt/s .L5
mov.l r1,@(8,r7)
mov r6,r5
.L5:
rts
mov r5,r0
and the expected result is:
tst r4,r4 <<< only one test against zero
movt r0
xor #1,r0
bt/s .L4
mov.l r0,@(8,r7)
mov r6,r5
.L4:
rts
mov r5,r0
Before r217496 there are the following insns before the combine pass that are
used to do the test against zero and store the inverted and non-inverted test
result in a reg:
(insn 11 6 14 2 (set (reg:SI 147 t)
(eq:SI (reg/v:SI 167 [ a ])
(const_int 0 [0]))) sh_tmp.cpp:55 17 {cmpeqsi_t}
(expr_list:REG_DEAD (reg/v:SI 167 [ a ])
(nil)))
(insn 14 11 15 2 (set (reg:SI 163 [ D.1480 ])
(zero_extend:SI (subreg:QI (reg:SI 147 t) 0))) sh_tmp.cpp:55 398 {movt}
(expr_list:REG_DEAD (reg:SI 147 t)
(nil)))
(insn 15 14 16 2 (set (reg:SI 173)
(xor:SI (reg:SI 163 [ D.1480 ])
(const_int 1 [0x1]))) sh_tmp.cpp:56 138 {*xorsi3_compact}
(nil))
(insn 16 15 17 2 (set (reg:SI 174)
(zero_extend:SI (subreg:QI (reg:SI 173) 0))) sh_tmp.cpp:56 224
{*zero_extendqisi2_compact}
(expr_list:REG_DEAD (reg:SI 173)
(nil)))
combine then uses the following pattern from sh.md:
Trying 14 -> 15:
Successfully matched this instruction:
(parallel [
(set (reg:SI 173)
(xor:SI (reg:SI 147 t)
(const_int 1 [0x1])))
(set (reg:SI 163 [ D.1480 ])
(zero_extend:SI (subreg:QI (reg:SI 147 t) 0)))
])
which is then split into the appropriate/expected insn sequence.
After r217496 the insns before combine look like:
(insn 11 6 12 2 (set (reg:SI 147 t)
(eq:SI (reg/v:SI 166 [ a ])
(const_int 0 [0]))) sh_tmp.cpp:56 17 {cmpeqsi_t}
(nil))
(insn 13 12 14 2 (parallel [
(set (reg:SI 171 [ D.1480 ])
(xor:SI (reg:SI 147 t)
(const_int 1 [0x1])))
(set (reg:SI 147 t)
(const_int 1 [0x1]))
(use (reg:SI 172))
]) sh_tmp.cpp:56 400 {movrt_negc}
(expr_list:REG_DEAD (reg:SI 172)
(expr_list:REG_UNUSED (reg:SI 147 t)
(nil))))
(insn 15 14 7 2 (set (reg:SI 147 t)
(eq:SI (reg/v:SI 166 [ a ])
(const_int 0 [0]))) sh_tmp.cpp:57 17 {cmpeqsi_t}
(expr_list:REG_DEAD (reg/v:SI 166 [ a ])
(nil)))
and combine does not try the parallel pattern above.
I've briefly tried working around the issue with some combine bridge patterns,
but without any success.
A possible solution could be adding some extra code to the movrt_negc splitter
which looks at the surrounding insn and splits into a movt-xor sequence if
required. However, I'd like to wait a bit until other pending combine patches
have been installed.
