https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110237
Bug ID: 110237
Summary: gcc.dg/torture/pr58955-2.c is miscompiled by RTL
scheduling after reload
Product: gcc
Version: 14.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: rtl-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: rguenth at gcc dot gnu.org
Target Milestone: ---
When compiling the testcase with fully masked AVX512 vectorization,
-march=znver4 --param=vect-partial-vector-usage=2 -fdiagnostics-plain-output
-O3 -fomit-frame-pointer -funroll-loops -fpeel-loops -ftracer
-finline-functions
RTL sched2 is presented with
(insn 38 35 39 3 (set (mem:V16SI (plus:DI (reg:DI 40 r12 [orig:90 _22 ] [90])
(const:DI (plus:DI (symbol_ref:DI ("b") [flags 0x2] <var_decl
0x7fda7413dd80 b>)
(const_int -4 [0xfffffffffffffffc])))) [1 MEM
<vector(16) int> [(int *)vectp_b.12_28]+0 S64 A32])
(vec_merge:V16SI (reg:V16SI 21 xmm1 [118])
(mem:V16SI (plus:DI (reg:DI 40 r12 [orig:90 _22 ] [90])
(const:DI (plus:DI (symbol_ref:DI ("b") [flags 0x2]
<var_decl 0x7fda7413dd80 b>)
(const_int -4 [0xfffffffffffffffc])))) [1 MEM
<vector(16) int> [(int *)vectp_b.12_28]+0 S64 A32])
(reg:HI 69 k1 [116])))
"/space/rguenther/src/gcc11queue/gcc/testsuite/gcc.dg/torture/pr58955-2.c":12:12
1942 {avx512f_storev16si_mask}
(expr_list:REG_DEAD (reg:HI 69 k1 [116])
(expr_list:REG_DEAD (reg:DI 40 r12 [orig:90 _22 ] [90])
(expr_list:REG_DEAD (reg:V16SI 21 xmm1 [118])
(nil)))))
...
(insn 41 39 42 3 (set (reg:CCZ 17 flags)
(compare:CCZ (mem/c:SI (const:DI (plus:DI (symbol_ref:DI ("b") [flags
0x2] <var_decl 0x7fda7413dd80 b>)
(const_int 4 [0x4]))) [1 b[1]+0 S4 A32])
(const_int 1 [0x1])))
"/space/rguenther/src/gcc11queue/gcc/testsuite/gcc.dg/torture/pr58955-2.c":15:6
11 {*cmpsi_1}
(nil))
and it moves the masked store across the load of one of the destinations
elements:
- 32: xmm0:V16QI=vec_duplicate(bx:QI)
- REG_DEAD bx:QI
- 33: NOTE_INSN_DELETED
- 34: k1:HI=unspec[xmm0:V16QI,[`*.LC0'],0x6] 146
- REG_DEAD xmm0:V16QI
36: cx:SI=0x1
REG_EQUIV 0x1
+ 41: flags:CCZ=cmp([const(`b'+0x4)],0x1)
+ 32: xmm0:V16QI=vec_duplicate(bx:QI)
+ REG_DEAD bx:QI
35: xmm1:V16SI=vec_duplicate(cx:SI)
REG_DEAD cx:SI
REG_EQUIV const_vector
+ 34: k1:HI=unspec[xmm0:V16QI,[`*.LC0'],0x6] 146
+ REG_DEAD xmm0:V16QI
+ 39: [`a']=0x2
38:
[r12:DI+const(`b'-0x4)]=vec_merge(xmm1:V16SI,[r12:DI+const(`b'-0x4)],k1:HI)
REG_DEAD k1:HI
REG_DEAD r12:DI
REG_DEAD xmm1:V16SI
- 39: [`a']=0x2
- 41: flags:CCZ=cmp([const(`b'+0x4)],0x1)
the address of the masked store is computed oddly though:
14: r12:DI=dx:DI<<0x2+0x4
REG_DEAD dx:DI
and in the end the assembly contains
leaq 4(,%rdx,4), %r12
...
cmpl $1, b+4(%rip)
...
vmovdqu32 %zmm1, b-4(%r12){%k1}
(%rdx is zero)