x86_64: Fix up -fpic -mcmodel=large -fno-plt [PR98063]
authorJakub Jelinek <jakub@redhat.com>
Tue, 1 Dec 2020 09:44:40 +0000 (10:44 +0100)
committerJakub Jelinek <jakub@redhat.com>
Tue, 1 Dec 2020 09:44:40 +0000 (10:44 +0100)
commit69157fe75823fc34f1e3265345f2d2b99cd8d380
tree5847d5a087f989264c63d24f34215e294ff8787c
parent57fcbe579c2f2b0919486b658a5bec8f4e1ef42b
x86_64: Fix up -fpic -mcmodel=large -fno-plt [PR98063]

On the following testcase with -fpic -mcmodel=large -fno-plt we emit
call puts@GOTPCREL(%rip)
but that is not really appropriate for CM_LARGE_PIC, the .text can be larger
than 2GB in that case and the .got slot further away from %rip than what can
fit into the signed 32-bit immediate.

The following patch computes the address of the .got slot the way it is
computed for that model for function pointer loads, and calls that.

2020-12-01  Jakub Jelinek  <jakub@redhat.com>

PR target/98063
* config/i386/i386-expand.c (ix86_expand_call): Handle non-plt
CM_LARGE_PIC calls.

* gcc.target/i386/pr98063.c: New test.
gcc/config/i386/i386-expand.c
gcc/testsuite/gcc.target/i386/pr98063.c [new file with mode: 0644]