On 1/15/24 06:34, Richard Biener wrote:
When the x86 backend generates code for cpymem with the rep_8byte
strathegy for the 8 byte aligned main rep movq it needs to compute
an adjusted pointer to the source after doing a prologue aligning
the destination. It computes that via
src_ptr + (dest_ptr - orig_dest_ptr)
which is perfectly fine. On RTL this is then
8: r134:DI=const(`g'+0x44)
9: {r133:DI=frame:DI-0x4c;clobber flags:CC;}
REG_UNUSED flags:CC
56: r129:DI=const(`g'+0x4c)
57: {r129:DI=r129:DI&0xfffffffffffffff8;clobber flags:CC;}
REG_UNUSED flags:CC
REG_EQUAL const(`g'+0x4c)&0xfffffffffffffff8
58: {r118:DI=r134:DI-r129:DI;clobber flags:CC;}
REG_DEAD r134:DI
REG_UNUSED flags:CC
REG_EQUAL const(`g'+0x44)-r129:DI
59: {r119:DI=r133:DI-r118:DI;clobber flags:CC;}
REG_DEAD r133:DI
REG_UNUSED flags:CC
but as written find_base_term happily picks the first candidate
it finds for the MINUS which means it picks const(`g') rather
than the correct frame:DI. This way find_base_term (but also
the unfixed find_base_value used by init_alias_analysis to
initialize REG_BASE_VALUE) performs pointer analysis isn't
sound. The following restricts the handling of multi-operand
operations to the case we know only one can be a pointer.
This for example causes gcc.dg/tree-ssa/pr94969.c to miss some
RTL PRE (I've opened PR113395 for this). A more drastic patch,
removing base_alias_check results in only gcc.dg/guality/pr41447-1.c
regressing (so testsuite coverage is bad). I've looked at
gcc.dg/tree-ssa tests and mostly scheduling changes are present,
the cc1plus .text size is only 230 bytes worse. With the this
less drastic patch below most scheduling changes are gone.
x86_64 might not the very best target to test for impact, but
test coverage on other targets is unlikely to be very much better.
Bootstrapped and tested on x86_64-unknown-linux-gnu (together
with 2/2). Jeff, can you maybe throw this on your tester?
Jakub, you did the PR64025 fix which was for a similar issue.
No issues across the cross compilers with those two patches.
Jeff
