public inbox for gcc-cvs@sourceware.org help / color / mirror / Atom feed
From: Kyrylo Tkachov <ktkachov@gcc.gnu.org> To: gcc-cvs@gcc.gnu.org Subject: [gcc r14-62] aarch64: PR target/108840 Simplify register shift RTX costs and eliminate shift amount masking Date: Wed, 19 Apr 2023 08:35:49 +0000 (GMT) [thread overview] Message-ID: <20230419083549.4FD7F3858CDA@sourceware.org> (raw) https://gcc.gnu.org/g:136330bf637b50a4f10ace017a4316541386b9c0 commit r14-62-g136330bf637b50a4f10ace017a4316541386b9c0 Author: Kyrylo Tkachov <kyrylo.tkachov@arm.com> Date: Wed Apr 19 09:34:40 2023 +0100 aarch64: PR target/108840 Simplify register shift RTX costs and eliminate shift amount masking In this PR we fail to eliminate explicit &31 operations for variable shifts such as in: void bar (int x[3], int y) { x[0] <<= (y & 31); x[1] <<= (y & 31); x[2] <<= (y & 31); } This is rejected by RTX costs that end up giving too high a cost for: (set (reg:SI 96) (ashift:SI (reg:SI 98) (subreg:QI (and:SI (reg:SI 99) (const_int 31 [0x1f])) 0))) There is code to handle the AND-31 case in rtx costs, but it gets confused by the subreg. It's easy enough to fix by looking inside the subreg when costing the expression. While doing that I noticed that the ASHIFT case and the other shift-like cases are almost identical and we should just merge them. This code will only be used for valid insns anyway, so the code after this patch should do the Right Thing (TM) for all such shift cases. With this patch there are no more "and wn, wn, 31" instructions left in the testcase. Bootstrapped and tested on aarch64-none-linux-gnu. PR target/108840 gcc/ChangeLog: * config/aarch64/aarch64.cc (aarch64_rtx_costs): Merge ASHIFT and ROTATE, ROTATERT, LSHIFTRT, ASHIFTRT cases. Handle subregs in op1. gcc/testsuite/ChangeLog: * gcc.target/aarch64/pr108840.c: New test. Diff: --- gcc/config/aarch64/aarch64.cc | 63 +++++------------------------ gcc/testsuite/gcc.target/aarch64/pr108840.c | 38 +++++++++++++++++ 2 files changed, 49 insertions(+), 52 deletions(-) diff --git a/gcc/config/aarch64/aarch64.cc b/gcc/config/aarch64/aarch64.cc index adbdaaf8603..0d7470c05a1 100644 --- a/gcc/config/aarch64/aarch64.cc +++ b/gcc/config/aarch64/aarch64.cc @@ -14678,6 +14678,10 @@ cost_plus: } return false; + case ROTATE: + case ROTATERT: + case LSHIFTRT: + case ASHIFTRT: case ASHIFT: op0 = XEXP (x, 0); op1 = XEXP (x, 1); @@ -14693,8 +14697,8 @@ cost_plus: } else { - /* LSL (immediate), UBMF, UBFIZ and friends. These are all - aliases. */ + /* LSL (immediate), ASR (immediate), UBMF, UBFIZ and friends. + These are all aliases. */ *cost += extra_cost->alu.shift; } } @@ -14718,9 +14722,13 @@ cost_plus: else { if (speed) - /* LSLV. */ + /* LSLV, ASRV. */ *cost += extra_cost->alu.shift_reg; + /* The register shift amount may be in a shorter mode expressed + as a lowpart SUBREG. For costing purposes just look inside. */ + if (SUBREG_P (op1) && subreg_lowpart_p (op1)) + op1 = SUBREG_REG (op1); if (GET_CODE (op1) == AND && REG_P (XEXP (op1, 0)) && CONST_INT_P (XEXP (op1, 1)) && known_eq (INTVAL (XEXP (op1, 1)), @@ -14735,55 +14743,6 @@ cost_plus: return false; /* All arguments need to be in registers. */ } - case ROTATE: - case ROTATERT: - case LSHIFTRT: - case ASHIFTRT: - op0 = XEXP (x, 0); - op1 = XEXP (x, 1); - - if (CONST_INT_P (op1)) - { - /* ASR (immediate) and friends. */ - if (speed) - { - if (VECTOR_MODE_P (mode)) - *cost += extra_cost->vect.alu; - else - *cost += extra_cost->alu.shift; - } - - *cost += rtx_cost (op0, mode, (enum rtx_code) code, 0, speed); - return true; - } - else - { - if (VECTOR_MODE_P (mode)) - { - if (speed) - /* Vector shift (register). */ - *cost += extra_cost->vect.alu; - } - else - { - if (speed) - /* ASR (register) and friends. */ - *cost += extra_cost->alu.shift_reg; - - if (GET_CODE (op1) == AND && REG_P (XEXP (op1, 0)) - && CONST_INT_P (XEXP (op1, 1)) - && known_eq (INTVAL (XEXP (op1, 1)), - GET_MODE_BITSIZE (mode) - 1)) - { - *cost += rtx_cost (op0, mode, (rtx_code) code, 0, speed); - /* We already demanded XEXP (op1, 0) to be REG_P, so - don't recurse into it. */ - return true; - } - } - return false; /* All arguments need to be in registers. */ - } - case SYMBOL_REF: if (aarch64_cmodel == AARCH64_CMODEL_LARGE diff --git a/gcc/testsuite/gcc.target/aarch64/pr108840.c b/gcc/testsuite/gcc.target/aarch64/pr108840.c new file mode 100644 index 00000000000..804c1cd9156 --- /dev/null +++ b/gcc/testsuite/gcc.target/aarch64/pr108840.c @@ -0,0 +1,38 @@ +/* PR target/108840. Check that the explicit &31 is eliminated. */ +/* { dg-do compile } */ +/* { dg-options "-O" } */ + +int +foo (int x, int y) +{ + return x << (y & 31); +} + +void +bar (int x[3], int y) +{ + x[0] <<= (y & 31); + x[1] <<= (y & 31); + x[2] <<= (y & 31); +} + +void +baz (int x[3], int y) +{ + y &= 31; + x[0] <<= y; + x[1] <<= y; + x[2] <<= y; +} + +void corge (int, int, int); + +void +qux (int x, int y, int z, int n) +{ + n &= 31; + corge (x << n, y << n, z >> n); +} + +/* { dg-final { scan-assembler-not {and\tw[0-9]+, w[0-9]+, 31} } } */ +
reply other threads:[~2023-04-19 8:35 UTC|newest] Thread overview: [no followups] expand[flat|nested] mbox.gz Atom feed
Reply instructions: You may reply publicly to this message via plain-text email using any one of the following methods: * Save the following mbox file, import it into your mail client, and reply-to-all from there: mbox Avoid top-posting and favor interleaved quoting: https://en.wikipedia.org/wiki/Posting_style#Interleaved_style * Reply using the --to, --cc, and --in-reply-to switches of git-send-email(1): git send-email \ --in-reply-to=20230419083549.4FD7F3858CDA@sourceware.org \ --to=ktkachov@gcc.gnu.org \ --cc=gcc-cvs@gcc.gnu.org \ /path/to/YOUR_REPLY https://kernel.org/pub/software/scm/git/docs/git-send-email.html * If your mail client supports setting the In-Reply-To header via mailto: links, try the mailto: linkBe sure your reply has a Subject: header at the top and a blank line before the message body.
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).