diff options
author | Juzhe-Zhong <juzhe.zhong@rivai.ai> | 2023-09-19 10:59:22 +0800 |
---|---|---|
committer | Lehua Ding <lehua.ding@rivai.ai> | 2023-09-19 20:06:23 +0800 |
commit | e1817426e85f2181a97ac4f71a87baaa9645b1b5 (patch) | |
tree | 658dd172b79d208de57932041a96ea9ff9998ab1 | |
parent | 005888670a1a0e26d818449a0fbbc5eb3541e303 (diff) | |
download | gcc-e1817426e85f2181a97ac4f71a87baaa9645b1b5.zip gcc-e1817426e85f2181a97ac4f71a87baaa9645b1b5.tar.gz gcc-e1817426e85f2181a97ac4f71a87baaa9645b1b5.tar.bz2 |
RISC-V: Fix RVV can change mode class bug
After support the VLS mode conversion, current case triggers a latent bug that we are
lucky we didn't encounter.
This is a real bug in 'cprop_hardreg':
orig:RVVMF8BI,16,16
new:V32BI,32,0
during RTL pass: cprop_hardreg
auto.c: In function 'main':
auto.c:79:1: internal compiler error: in partial_subreg_p, at rtl.h:3186
79 | }
| ^
0x10979a7 partial_subreg_p(machine_mode, machine_mode)
../../../../gcc/gcc/rtl.h:3186
0x1723eda mode_change_ok
../../../../gcc/gcc/regcprop.cc:402
0x1724007 maybe_mode_change
../../../../gcc/gcc/regcprop.cc:436
0x172445d find_oldest_value_reg
../../../../gcc/gcc/regcprop.cc:489
0x172534d copyprop_hardreg_forward_1
../../../../gcc/gcc/regcprop.cc:808
0x1727017 cprop_hardreg_bb
../../../../gcc/gcc/regcprop.cc:1358
0x17272f7 execute
../../../../gcc/gcc/regcprop.cc:1425
When trying to do reg copy propagation between RVVMF8BI (precision = 16,16)
and V32BI (precision = 32,0).
The assertion failed in partial_subreg_p:
gcc_checking_assert (ordered_p (outer_prec, inner_prec));
In regcprop.cc:
if (partial_subreg_p (orig_mode, new_mode))
return false;
If orig_mode (RVVMF8BI) smaller than new_mode (V32BI), we don't do the hard reg propogation.
However, the 'partial_subreg_p' cause ICE since gcc_checking_assert (ordered_p (outer_prec, inner_prec)).
After analysis in aarch64.cc, they do careful block in 'TARGET_CAN_CHANGE_MODE_CLASS'.
So it's reasonable block regcprop when old mode size maybe_lt than new mode size since we won't do the
copy propgation.
gcc/ChangeLog:
* config/riscv/riscv.cc (riscv_can_change_mode_class): Block unordered VLA and VLS modes.
-rw-r--r-- | gcc/config/riscv/riscv.cc | 20 |
1 files changed, 17 insertions, 3 deletions
diff --git a/gcc/config/riscv/riscv.cc b/gcc/config/riscv/riscv.cc index 8c766e2..f1b721d 100644 --- a/gcc/config/riscv/riscv.cc +++ b/gcc/config/riscv/riscv.cc @@ -8536,12 +8536,26 @@ riscv_slow_unaligned_access (machine_mode, unsigned int) /* Implement TARGET_CAN_CHANGE_MODE_CLASS. */ static bool -riscv_can_change_mode_class (machine_mode, machine_mode, reg_class_t rclass) -{ +riscv_can_change_mode_class (machine_mode from, machine_mode to, + reg_class_t rclass) +{ + /* We have RVV VLS modes and VLA modes sharing same REG_CLASS. + In 'cprop_hardreg' stage, we will try to do hard reg copy propagation + between wider mode (FROM) and narrow mode (TO). + + E.g. We should not allow copy propagation + - RVVMF8BI (precision = [16, 16]) -> V32BI (precision = [32, 0]) + since we can't order their size which will cause ICE in regcprop. + + TODO: Even though they are have different size, they always change + the whole register. We may enhance such case in regcprop to optimize + it in the future. */ + if (reg_classes_intersect_p (V_REGS, rclass) + && !ordered_p (GET_MODE_PRECISION (from), GET_MODE_PRECISION (to))) + return false; return !reg_classes_intersect_p (FP_REGS, rclass); } - /* Implement TARGET_CONSTANT_ALIGNMENT. */ static HOST_WIDE_INT |