As Jaiwei noted, many (all?) of the rvv/base/spill tests started failing
after the introduction of shrink-wrapping.
The core issue is we're expecting the frame to have a constant size, but
it doesn't. So when using the to_constant method we abort.
The safest thing to do is to set no shrink-wrapping components when the
frame size is not fixed. We might be able to do better later -- iff we
know the offset to the GPRs/FPRs is fixed and fits into the appropriate
number of bits.
Bootstrapped and regression tested (C-only) on riscv64-linux-gnu. As
expected, it fixes a bucketload of failures in rvv/base/spill-*.c.
Installed on the trunk,
Jeff
commit ca73d4c80ea06087d9dd22594e5670bb15e21066
Author: Jeff Law <jeffreya...@gmail.com>
Date: Tue Nov 22 18:12:45 2022 -0500
Fix recent rvv/base/spill testcase failures
he core issue is we're expecting the frame to have a constant size, but it
doesn't. So when using the to_constant method we abort.
The safest thing to do is to set no shrink-wrapping components when the
frame size is not fixed. We might be able to do better later -- iff we
know the offset to the GPRs/FPRs is fixed and fits into the appropriate
number of bits.
Bootstrapped and regression tested (C-only) on riscv64-linux-gnu. As
expected, it fixes a bucketload of failures in rvv/base/spill-*.c.
gcc/
* config/riscv/riscv.cc (riscv_get_separate_components): Do not
do shrink-wrapping for a frame with a variable size.
diff --git a/gcc/config/riscv/riscv.cc b/gcc/config/riscv/riscv.cc
index 7ec4ce97e6c..7bfc0e9f595 100644
--- a/gcc/config/riscv/riscv.cc
+++ b/gcc/config/riscv/riscv.cc
@@ -5340,7 +5340,8 @@ riscv_get_separate_components (void)
bitmap_clear (components);
if (riscv_use_save_libcall (&cfun->machine->frame)
- || cfun->machine->interrupt_handler_p)
+ || cfun->machine->interrupt_handler_p
+ || !cfun->machine->frame.gp_sp_offset.is_constant ())
return components;
offset = cfun->machine->frame.gp_sp_offset.to_constant ();