https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128169
Bug ID: 128169
Summary: [17 Regression] x86: vec_extract from V32BF now goes
through V2BF/V4BF subreg, emits vpsrldq $0
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: lin1.hu at intel dot com
Target Milestone: ---
Since r17-5483-gd5ac76d8da8 ("expand: Improve tuple-extraction handling"),
gcc.target/i386/vect-bfloat16-2b.c fails with -m32:
FAIL: gcc.target/i386/vect-bfloat16-2b.c scan-assembler-times vpinsrw 68
(reported in
https://gcc.gnu.org/pipermail/gcc-patches/2026-October/735820.html)
The count drops 68 -> 63 (-m32) and 36 -> 32 (-m64, not scanned). Only
vec_extract_v32bf_{5,8,14,16,24,28} change. Testcase:
typedef __bf16 v32bf __attribute__ ((__vector_size__ (64)));
__attribute__((noipa)) __bf16 f (v32bf v) { return v[8]; }
-O2 -mavx512f -mno-avx512bw
-m64, before:
vmovdqa64 %zmm0, -64(%rsp)
vpinsrw $0, -48(%rsp), %xmm0, %xmm0
-m64, after:
vmovdqa64 %zmm0, -64(%rsp)
vmovlps -48(%rsp), %xmm1, %xmm1 # *movv4bf_internal/13
vpsrldq $0, %xmm1, %xmm0 # *mmx_pextrwv4bf/5
-m32, before:
vmovdqa64 %zmm0, (%esp)
vpinsrw $0, 16(%esp), %xmm0, %xmm0
-m32, after:
vmovdqa64 %zmm0, (%esp)
vmovss 16(%esp), %xmm1 # *movv2bf_internal/4
vpsrldq $0, %xmm1, %xmm0 # *pextrwv2bf/4
