https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105504
Bug ID: 105504
Summary: Fails to break dependency for vcvtss2sd xmm, xmm, mem
Product: gcc
Version: unknown
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: amonakov at gcc dot gnu.org
Target Milestone: ---
Created attachment 52933
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=52933&action=edit
testcase
Hit by core-math team at
https://gcc.gnu.org/pipermail/gcc-help/2022-May/141480.html
Compile the attached testcase with -O2 -march=haswell (other AVX-capable Intel
families except Alderlake are affected too) and observe that the big basic
block begins with
.L6:
vcvtss2sd xmm1, xmm1, DWORD PTR [rsp-4]
This creates a false dependency on the previous assignment into xmm1, resulting
in wildly varying (and suboptimal) throughput figures depending on how long the
CPU stalls waiting for the previous assignment to complete.
GCC has code to emit such instructions in a manner that avoids false
dependencies (see e.g. PR89071), but here it doesn't seem to work.
Also there's a potentially related issue that GCC copies the initial xmm0 value
to eax via stack in the beginning of the function:
cr_exp10f:
vmovss DWORD PTR [rsp-4], xmm0
mov eax, DWORD PTR [rsp-4]
This seems wrong since xmm-reg moves on Haswell are 1 cycle afaict.
