diff options
author | Andrew Pinski <apinski@marvell.com> | 2022-08-19 17:46:40 +0000 |
---|---|---|
committer | Andrew Pinski <apinski@marvell.com> | 2022-08-24 11:30:44 -0700 |
commit | e5e6983c3da53729e58a32af1d531ea74b3dbf5d (patch) | |
tree | 1de40f51a76673e5c4953b1a48e4a9e6bda3bd00 /gcc/config | |
parent | cb2daf5acce003300ee948a89860c0d13ebcae79 (diff) | |
download | gcc-e5e6983c3da53729e58a32af1d531ea74b3dbf5d.zip gcc-e5e6983c3da53729e58a32af1d531ea74b3dbf5d.tar.gz gcc-e5e6983c3da53729e58a32af1d531ea74b3dbf5d.tar.bz2 |
Fix PR 106601: __builtin_bswap16 code gen could be improved with ZBB enabled
The default expansion for bswap16 is two extractions (shift/and)
followed by an insertation (ior) and then a zero extend. This can be improved
with ZBB enabled to just full byteswap followed by a (logical) shift right.
This patch adds a new pattern for this which does that.
OK? Built and tested on riscv32-linux-gnu and riscv64-linux-gnu.
gcc/ChangeLog:
PR target/106601
* config/riscv/bitmanip.md (bswaphi2): New pattern.
gcc/testsuite/ChangeLog:
PR target/106601
* gcc.target/riscv/zbb_32_bswap-2.c: New test.
* gcc.target/riscv/zbb_bswap-2.c: New test.
Diffstat (limited to 'gcc/config')
-rw-r--r-- | gcc/config/riscv/bitmanip.md | 24 |
1 files changed, 24 insertions, 0 deletions
diff --git a/gcc/config/riscv/bitmanip.md b/gcc/config/riscv/bitmanip.md index c7ba667..c438328 100644 --- a/gcc/config/riscv/bitmanip.md +++ b/gcc/config/riscv/bitmanip.md @@ -276,6 +276,30 @@ "rev8\t%0,%1" [(set_attr "type" "bitmanip")]) +;; HI bswap can be emulated using SI/DI bswap followed +;; by a logical shift right +;; SI bswap for TARGET_64BIT is already similarly in +;; the common code. +(define_expand "bswaphi2" + [(set (match_operand:HI 0 "register_operand" "=r") + (bswap:HI (match_operand:HI 1 "register_operand" "r")))] + "TARGET_ZBB" +{ + rtx tmp = gen_reg_rtx (word_mode); + rtx newop1 = gen_lowpart (word_mode, operands[1]); + if (TARGET_64BIT) + emit_insn (gen_bswapdi2 (tmp, newop1)); + else + emit_insn (gen_bswapsi2 (tmp, newop1)); + rtx tmp1 = gen_reg_rtx (word_mode); + if (TARGET_64BIT) + emit_insn (gen_lshrdi3 (tmp1, tmp, GEN_INT (64 - 16))); + else + emit_insn (gen_lshrsi3 (tmp1, tmp, GEN_INT (32 - 16))); + emit_move_insn (operands[0], gen_lowpart (HImode, tmp1)); + DONE; +}) + (define_insn "<bitmanip_optab><mode>3" [(set (match_operand:X 0 "register_operand" "=r") (bitmanip_minmax:X (match_operand:X 1 "register_operand" "r") |