diff options
author | Roger Sayle <roger@nextmovesoftware.com> | 2022-02-04 04:13:53 +0100 |
---|---|---|
committer | Tom de Vries <tdevries@suse.de> | 2022-02-10 09:01:54 +0100 |
commit | 9bacd7af2e3bba9ddad17e7de4e2d299419d819d (patch) | |
tree | dcaa5a85b4ee43e818edd22d0b05953252b4935f | |
parent | f68c3de7fc9065d8c9ac75b3736ea27abffdce45 (diff) | |
download | gcc-9bacd7af2e3bba9ddad17e7de4e2d299419d819d.zip gcc-9bacd7af2e3bba9ddad17e7de4e2d299419d819d.tar.gz gcc-9bacd7af2e3bba9ddad17e7de4e2d299419d819d.tar.bz2 |
PR target/104345: Use nvptx "set" instruction for cond ? -1 : 0
This patch addresses the "increased register pressure" regression on
nvptx-none caused by my change to transition the backend to a
STORE_FLAG_VALUE = 1 target. This improved code generation for the
more common case of producing 0/1 Boolean values, but unfortunately
made things marginally worse when a 0/-1 mask value is desired.
Unfortunately, nvptx kernels are extremely sensitive to changes in
register usage, which was observable in the reported PR.
This patch provides optimizations for -(cond ? 1 : 0), effectively
simplify this into cond ? -1 : 0, where these ternary operators are
provided by nvptx's selp instruction, and for the specific case of
SImode, using (restoring) nvptx's "set" instruction (which avoids
the need for a predicate register).
This patch has been tested on nvptx-none hosted on x86_64-pc-linux-gnu
with a "make" and "make -k check" with no new failures. Unfortunately,
the exact register usage of a nvptx kernel depends upon the version of
the Cuda drivers being used (and the hardware), but I believe this
change should resolve the PR (for Thomas) by improving code generation
for the cases that regressed.
gcc/ChangeLog:
PR target/104345
* config/nvptx/nvptx.md (sel_true<mode>): Fix indentation.
(sel_false<mode>): Likewise.
(define_code_iterator eqne): New code iterator for EQ and NE.
(*selp<mode>_neg_<code>): New define_insn_and_split to optimize
the negation of a selp instruction.
(*selp<mode>_not_<code>): New define_insn_and_split to optimize
the bitwise not of a selp instruction.
(*setcc_int<mode>): Use set instruction for neg:SI of a selp.
gcc/testsuite/ChangeLog:
PR target/104345
* gcc.target/nvptx/neg-selp.c: New test case.
-rw-r--r-- | gcc/config/nvptx/nvptx.md | 58 | ||||
-rw-r--r-- | gcc/testsuite/gcc.target/nvptx/neg-selp.c | 17 |
2 files changed, 71 insertions, 4 deletions
diff --git a/gcc/config/nvptx/nvptx.md b/gcc/config/nvptx/nvptx.md index 107df74..ad642e7 100644 --- a/gcc/config/nvptx/nvptx.md +++ b/gcc/config/nvptx/nvptx.md @@ -977,7 +977,7 @@ (define_insn "sel_true<mode>" [(set (match_operand:HSDIM 0 "nvptx_register_operand" "=R") - (if_then_else:HSDIM + (if_then_else:HSDIM (ne (match_operand:BI 1 "nvptx_register_operand" "R") (const_int 0)) (match_operand:HSDIM 2 "nvptx_nonmemory_operand" "Ri") (match_operand:HSDIM 3 "nvptx_nonmemory_operand" "Ri")))] @@ -986,7 +986,7 @@ (define_insn "sel_true<mode>" [(set (match_operand:SDFM 0 "nvptx_register_operand" "=R") - (if_then_else:SDFM + (if_then_else:SDFM (ne (match_operand:BI 1 "nvptx_register_operand" "R") (const_int 0)) (match_operand:SDFM 2 "nvptx_nonmemory_operand" "RF") (match_operand:SDFM 3 "nvptx_nonmemory_operand" "RF")))] @@ -995,7 +995,7 @@ (define_insn "sel_false<mode>" [(set (match_operand:HSDIM 0 "nvptx_register_operand" "=R") - (if_then_else:HSDIM + (if_then_else:HSDIM (eq (match_operand:BI 1 "nvptx_register_operand" "R") (const_int 0)) (match_operand:HSDIM 2 "nvptx_nonmemory_operand" "Ri") (match_operand:HSDIM 3 "nvptx_nonmemory_operand" "Ri")))] @@ -1004,13 +1004,63 @@ (define_insn "sel_false<mode>" [(set (match_operand:SDFM 0 "nvptx_register_operand" "=R") - (if_then_else:SDFM + (if_then_else:SDFM (eq (match_operand:BI 1 "nvptx_register_operand" "R") (const_int 0)) (match_operand:SDFM 2 "nvptx_nonmemory_operand" "RF") (match_operand:SDFM 3 "nvptx_nonmemory_operand" "RF")))] "" "%.\\tselp%t0\\t%0, %3, %2, %1;") +(define_code_iterator eqne [eq ne]) + +;; Split negation of a predicate into a conditional move. +(define_insn_and_split "*selp<mode>_neg_<code>" + [(set (match_operand:HSDIM 0 "nvptx_register_operand" "=R") + (neg:HSDIM (eqne:HSDIM + (match_operand:BI 1 "nvptx_register_operand" "R") + (const_int 0))))] + "" + "#" + "&& 1" + [(set (match_dup 0) + (if_then_else:HSDIM + (eqne (match_dup 1) (const_int 0)) + (const_int -1) + (const_int 0)))]) + +;; Split bitwise not of a predicate into a conditional move. +(define_insn_and_split "*selp<mode>_not_<code>" + [(set (match_operand:HSDIM 0 "nvptx_register_operand" "=R") + (not:HSDIM (eqne:HSDIM + (match_operand:BI 1 "nvptx_register_operand" "R") + (const_int 0))))] + "" + "#" + "&& 1" + [(set (match_dup 0) + (if_then_else:HSDIM + (eqne (match_dup 1) (const_int 0)) + (const_int -2) + (const_int -1)))]) + +(define_insn "*setcc_int<mode>" + [(set (match_operand:SI 0 "nvptx_register_operand" "=R") + (neg:SI + (match_operator:SI 1 "nvptx_comparison_operator" + [(match_operand:HSDIM 2 "nvptx_register_operand" "R") + (match_operand:HSDIM 3 "nvptx_nonmemory_operand" "Ri")])))] + "" + "%.\\tset%t0%c1\\t%0, %2, %3;") + +(define_insn "*setcc_int<mode>" + [(set (match_operand:SI 0 "nvptx_register_operand" "=R") + (neg:SI + (match_operator:SI 1 "nvptx_float_comparison_operator" + [(match_operand:SDFM 2 "nvptx_register_operand" "R") + (match_operand:SDFM 3 "nvptx_nonmemory_operand" "RF")])))] + "" + "%.\\tset%t0%c1\\t%0, %2, %3;") + (define_insn "setcc_float<mode>" [(set (match_operand:SF 0 "nvptx_register_operand" "=R") (match_operator:SF 1 "nvptx_comparison_operator" diff --git a/gcc/testsuite/gcc.target/nvptx/neg-selp.c b/gcc/testsuite/gcc.target/nvptx/neg-selp.c new file mode 100644 index 0000000..a8f0118 --- /dev/null +++ b/gcc/testsuite/gcc.target/nvptx/neg-selp.c @@ -0,0 +1,17 @@ +/* { dg-do compile } */ +/* { dg-options "-O2" } */ + +int neg(int x, int y) +{ + int t = (x == y) ? 1 : 0; + return -t; +} + +int not(int x, int y) +{ + int t = (x == y) ? 1 : 0; + return ~t; +} + +/* { dg-final { scan-assembler-not "neg.s32" } } */ +/* { dg-final { scan-assembler-not "not.b32" } } */ |