https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100267
--- Comment #2 from Hongtao.liu <crazylht at gmail dot com> ---
(In reply to Konstantin Ananyev from comment #0)
> The code snippet below compiles ok with '-O2' for gcc-9.
> But with gcc-10 (and gcc-11) it generates -Wuninitialized warnings.
> Another thing (which is probably worse) 'gcc-10 -O2' generates code with
> unnecessary loads for ymm registers from the initiliazed portion of the
> stack.
> As I understand, thats where from these -Wuninitialized warnings come from:
> by some reason gcc-10 wants to put local '__m256i pdatap[2]' variables
> on the stack.
> Note that only '-O2' affected, '-O3' looks good for all versions I tried
> (gcc-9, gcc-10, gcc-11)..
>
> =====================
> $ cat tavx512u5.c
>
> #include <stddef.h>
> #include <stdint.h>
> #include <x86intrin.h>
>
>
> struct flow_avx512 {
> uint32_t num_packets;
> uint32_t total_packets;
> const uint8_t **idata;
> };
>
> static inline void
> start_flow_avx512x8(const struct flow_avx512 *flow, uint32_t num,
> uint32_t msk, __m256i pdata[2])
> {
> uint32_t n, m[2], nm[2];
> __m256i nd[2];
>
> m[0] = msk & 0xF;
> m[1] = msk >> 4;
>
> n = __builtin_popcount(m[0]);
> nm[0] = (1 << n) - 1;
> nm[1] = (1 << (num - n)) - 1;
>
> nd[0] = _mm256_maskz_loadu_epi64(nm[0],
> flow->idata + flow->num_packets);
> nd[1] = _mm256_maskz_loadu_epi64(nm[1],
> flow->idata + flow->num_packets + n);
>
> pdata[0] = _mm256_mask_expand_epi64(pdata[0], m[0], nd[0]);
> pdata[1] = _mm256_mask_expand_epi64(pdata[1], m[1], nd[1]);
> }
>
> __m256i
> dummyf1_avx512x8(const struct flow_avx512 *flow)
> {
> __m256i pdata[2];
>
> start_flow_avx512x8(flow, 8, 0xFF, pdata);
> return _mm256_add_epi64(pdata[0], pdata[1]);
> }
>
> ====================
> Good version (gcc-9) first:
> gcc-9 -m64 -mavx512f -mavx512vl -mavx512cd -mavx512bw -Wall -O2 -o
> tavx512u5.gcc9-O2.o -c tavx512u5.c
>
> $ objdump -d tavx512u5.gcc9-O2.o
>
> tavx512u5.gcc9-O2.o: file format elf64-x86-64
>
> Disassembly of section .text:
>
> 0000000000000000 <dummyf1_avx512x8>:
> 0: f3 0f 1e fa endbr64
> 4: 8b 17 mov (%rdi),%edx
> 6: 48 8b 47 08 mov 0x8(%rdi),%rax
> a: b9 0f 00 00 00 mov $0xf,%ecx
> f: c5 f8 92 c9 kmovw %ecx,%k1
> 13: 62 f2 fd a9 89 0c d0 vpexpandq (%rax,%rdx,8),%ymm1{%k1}{z}
> 1a: 62 f2 fd a9 89 44 d0 vpexpandq 0x20(%rax,%rdx,8),%ymm0{%k1}{z}
> 21: 04
> 22: c5 f5 d4 c0 vpaddq %ymm0,%ymm1,%ymm0
> 26: c3 retq
>
k1 is 0xf, so pdata is not used in _mm256_mask_expand_epi64, but gcc failed to
simplify vpexpandq (%rax,%rdx,8),%ymm1{%k1}{z} to vpexpandq (%rax,%rdx,8),
%ymm1 since we didn't support vpexpandq w/o mask? clang's codegen seems to be
optimal https://godbolt.org/z/d79v11Gz3
cut from sse.md, it seems we only support vpexpandq w/ mask since all
corresponding intrinsics are w/ mask.
----
(define_expand "<avx512>_expand<mode>_maskz"
[(set (match_operand:VI48F 0 "register_operand")
(unspec:VI48F
[(match_operand:VI48F 1 "nonimmediate_operand")
(match_operand:VI48F 2 "nonimm_or_0_operand")
(match_operand:<avx512fmaskmode> 3 "register_operand")]
UNSPEC_EXPAND))]
"TARGET_AVX512F"
"operands[2] = CONST0_RTX (<MODE>mode);")
(define_insn "<avx512>_expand<mode>_mask"
[(set (match_operand:VI48F 0 "register_operand" "=v,v")
(unspec:VI48F
[(match_operand:VI48F 1 "nonimmediate_operand" "v,m")
(match_operand:VI48F 2 "nonimm_or_0_operand" "0C,0C")
(match_operand:<avx512fmaskmode> 3 "register_operand" "Yk,Yk")]
UNSPEC_EXPAND))]
"TARGET_AVX512F"
"v<sseintprefix>expand<ssemodesuffix>\t{%1, %0%{%3%}%N2|%0%{%3%}%N2, %1}"
[(set_attr "type" "ssemov")
(set_attr "prefix" "evex")
(set_attr "memory" "none,load")
(set_attr "mode" "<sseinsnmode>")])
(define_insn "expand<mode>_mask"
[(set (match_operand:VI12_AVX512VLBW 0 "register_operand" "=v,v")
(unspec:VI12_AVX512VLBW
[(match_operand:VI12_AVX512VLBW 1 "nonimmediate_operand" "v,m")
(match_operand:VI12_AVX512VLBW 2 "nonimm_or_0_operand" "0C,0C")
(match_operand:<avx512fmaskmode> 3 "register_operand" "Yk,Yk")]
UNSPEC_EXPAND))]
"TARGET_AVX512VBMI2"
"v<sseintprefix>expand<ssemodesuffix>\t{%1, %0%{%3%}%N2|%0%{%3%}%N2, %1}"
[(set_attr "type" "ssemov")
(set_attr "prefix" "evex")
(set_attr "memory" "none,load")
(set_attr "mode" "<sseinsnmode>")])
----
