public inbox for gcc-cvs@sourceware.org
help / color / mirror / Atom feed
* [gcc(refs/vendors/riscv/heads/gcc-13-with-riscv-opts)] RISC-V: elide unnecessary sign extend when expanding cmp_and_jump
@ 2023-11-21 4:09 Jeff Law
0 siblings, 0 replies; only message in thread
From: Jeff Law @ 2023-11-21 4:09 UTC (permalink / raw)
To: gcc-cvs
https://gcc.gnu.org/g:26d3ce5d39af033b91b3be3a06038846017abac1
commit 26d3ce5d39af033b91b3be3a06038846017abac1
Author: Vineet Gupta <vineetg@rivosinc.com>
Date: Tue Oct 24 20:38:49 2023 -0700
RISC-V: elide unnecessary sign extend when expanding cmp_and_jump
RV64 compare and branch instructions only support 64-bit operands.
At Expand time, the backend conservatively zero/sign extends
its operands even if not needed, such as incoming function args
which ABI/ISA guarantee to be sign-extended already (this is true for
SI, HI, QI operands)
And subsequently REE fails to eliminate them as
"missing defintion(s)" or "multiple definition(s)
since function args don't have explicit definition.
So during expand riscv_extend_comparands (), if an operand is a
subreg-promoted SI with inner DI, which is representative of a function
arg, just peel away the subreg to expose the DI, eliding the sign
extension. As Jeff noted this routine is also used in if-conversion so
potentially can also help there.
Note there's currently patches floating around to improve REE and also a
new pass to eliminate unneccesary extensions, but it is still beneficial
to not generate those extra extensions in first place. It is obviously
less work for post-reload passes such as REE, but even for earlier
passes, such as combine, having to deal with one less thing and ensuing
fewer combinations is a win too.
Way too many existing tests used to observe this issue.
e.g. gcc.c-torture/compile/20190827-1.c -O2 -march=rv64gc
It elimiates the SEXT.W
gcc/ChangeLog:
* config/riscv/riscv.cc (riscv_sign_extend_if_not_subreg_prom): New.
* (riscv_extend_comparands): Call New function on operands.
Tested-by: Patrick O'Neill <patrick@rivosinc.com> # pre-commit-CI #676
Signed-off-by: Vineet Gupta <vineetg@rivosinc.com>
(cherry picked from commit fb4e2c1648ea4cf1b1d34205a1da519702328d92)
Diff:
---
gcc/config/riscv/riscv.cc | 23 +++++++++++++++++++++--
1 file changed, 21 insertions(+), 2 deletions(-)
diff --git a/gcc/config/riscv/riscv.cc b/gcc/config/riscv/riscv.cc
index c9035d7c424..6a13c4b68b5 100644
--- a/gcc/config/riscv/riscv.cc
+++ b/gcc/config/riscv/riscv.cc
@@ -3695,6 +3695,24 @@ riscv_zero_if_equal (rtx cmp0, rtx cmp1)
cmp0, cmp1, 0, 0, OPTAB_DIRECT);
}
+/* Helper function for riscv_extend_comparands to Sign-extend the OP.
+ However if the OP is SI subreg promoted with an inner DI, such as
+ (subreg/s/v:SI (reg/v:DI) 0)
+ just peel off the SUBREG to get DI, avoiding extraneous extension. */
+
+static void
+riscv_sign_extend_if_not_subreg_prom (rtx *op)
+{
+ if (GET_CODE (*op) == SUBREG
+ && SUBREG_PROMOTED_VAR_P (*op)
+ && SUBREG_PROMOTED_SIGNED_P (*op)
+ && (GET_MODE_SIZE (GET_MODE (XEXP (*op, 0))).to_constant ()
+ == GET_MODE_SIZE (word_mode)))
+ *op = XEXP (*op, 0);
+ else
+ *op = gen_rtx_SIGN_EXTEND (word_mode, *op);
+}
+
/* Sign- or zero-extend OP0 and OP1 for integer comparisons. */
static void
@@ -3724,9 +3742,10 @@ riscv_extend_comparands (rtx_code code, rtx *op0, rtx *op1)
}
else
{
- *op0 = gen_rtx_SIGN_EXTEND (word_mode, *op0);
+ riscv_sign_extend_if_not_subreg_prom (op0);
+
if (*op1 != const0_rtx)
- *op1 = gen_rtx_SIGN_EXTEND (word_mode, *op1);
+ riscv_sign_extend_if_not_subreg_prom (op1);
}
}
}
^ permalink raw reply [flat|nested] only message in thread
only message in thread, other threads:[~2023-11-21 4:09 UTC | newest]
Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2023-11-21 4:09 [gcc(refs/vendors/riscv/heads/gcc-13-with-riscv-opts)] RISC-V: elide unnecessary sign extend when expanding cmp_and_jump Jeff Law
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).