On 11/14/19 7:34 AM, Kwok Cheung Yeung wrote:
Hello
Currently, when choosing a spill register, GCC just picks the first
available register in the register class returned by the
TAQRGET_SPILL_CLASS hook that doesn't conflict.
On AMD GCN this can cause problems as DImode values stored in SGPRs
must start on an even register number and TImode values on a multiple
of 4. This is enforced by defining TARGET_HARD_REGNO_MODE_OK to be
false when this condition is not satisfied. However,
assign_spill_hard_regs does not check TARGET_HARD_REGNO_MODE_OK, so it
can assign an unsuitable hard register for the mode. I have fixed this
by rejecting spill registers that do not satisfy
TARGET_HARD_REGNO_MODE_OK for the largest mode of the spilt register.
Built and tested for an AMD GCN target. This fixes failures in:
gcc.dg/vect/no-scevccp-outer-9.c
gcc.dg/vect/no-scevccp-outer-10.c
I have also ensured that the code bootstraps on x86_64, though it
currently does not use spill registers.
Okay for trunk?
Spilling into registers of another class was an experimental feature.
On some x86-64 micro-architectures, it resulted in a worse performance
code. But I guess in your case, it is profitable. I hope after your
patches it can be switched on for some x86-64 micro-architectures.
As for the patch, it is OK to commit. Thank you.
2019-11-14 Kwok Cheung Yeung <k...@codesourcery.com>
gcc/
* lra-spills.c (assign_spill_hard_regs): Check that the spill
register is suitable for the mode.
---
gcc/lra-spills.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/gcc/lra-spills.c b/gcc/lra-spills.c
index 54f76cc..8fbd3a8 100644
--- a/gcc/lra-spills.c
+++ b/gcc/lra-spills.c
@@ -283,7 +283,8 @@ assign_spill_hard_regs (int *pseudo_regnos, int n)
for (k = 0; k < spill_class_size; k++)
{
hard_regno = ira_class_hard_regs[spill_class][k];
- if (TEST_HARD_REG_BIT (eliminable_regset, hard_regno))
+ if (TEST_HARD_REG_BIT (eliminable_regset, hard_regno)
+ || !targetm.hard_regno_mode_ok (hard_regno, mode))
continue;
if (! overlaps_hard_reg_set_p (conflict_hard_regs, mode,
hard_regno))
break;
