public inbox for gcc-cvs@sourceware.org help / color / mirror / Atom feed
From: Hongyu Wang <hongyuw@gcc.gnu.org> To: gcc-cvs@gcc.gnu.org Subject: [gcc r14-6255] [APX NDD] Support TImode shift for NDD Date: Thu, 7 Dec 2023 01:40:30 +0000 (GMT) [thread overview] Message-ID: <20231207014030.E011A3857B8B@sourceware.org> (raw) https://gcc.gnu.org/g:3ba505c7b17a208b1c546b7a974a02e8003b60ef commit r14-6255-g3ba505c7b17a208b1c546b7a974a02e8003b60ef Author: Hongyu Wang <hongyu.wang@intel.com> Date: Sat Dec 2 12:55:59 2023 +0800 [APX NDD] Support TImode shift for NDD For TImode shifts, they are splitted by splitter functions, which assume operands[0] and operands[1] to be the same. For the NDD alternative the assumption may not be true so add split functions for NDD to emit the NDD form instructions, and omit the handling of !64bit target split. Although the NDD form allows memory src, for post-reload splitter there are no extra register to accept NDD form shift, especially shld/shrd. So only accept register alternative for shift src under NDD. gcc/ChangeLog: * config/i386/i386-expand.cc (ix86_split_ashl_ndd): New function to split NDD form lshift. (ix86_split_rshift_ndd): Likewise for l/ashiftrt. * config/i386/i386-protos.h (ix86_split_ashl_ndd): New prototype. (ix86_split_rshift_ndd): Likewise. * config/i386/i386.md (ashl<mode>3_doubleword): Add NDD alternative, call ndd split function when operands[0] not equal to operands[1]. (define_split for doubleword lshift): Likewise. (define_peephole for doubleword lshift): Likewise. (<insn><mode>3_doubleword): Likewise for l/ashiftrt. (define_split for doubleword l/ashiftrt): Likewise. (define_peephole for doubleword l/ashiftrt): Likewise. gcc/testsuite/ChangeLog: * gcc.target/i386/apx-ndd-ti-shift.c: New test. Diff: --- gcc/config/i386/i386-expand.cc | 136 +++++++++++++++++++++++ gcc/config/i386/i386-protos.h | 2 + gcc/config/i386/i386.md | 56 ++++++++-- gcc/testsuite/gcc.target/i386/apx-ndd-ti-shift.c | 91 +++++++++++++++ 4 files changed, 273 insertions(+), 12 deletions(-) diff --git a/gcc/config/i386/i386-expand.cc b/gcc/config/i386/i386-expand.cc index d4bbd33ce07..a53d69d5400 100644 --- a/gcc/config/i386/i386-expand.cc +++ b/gcc/config/i386/i386-expand.cc @@ -6678,6 +6678,142 @@ ix86_split_lshr (rtx *operands, rtx scratch, machine_mode mode) } } +/* Helper function to split TImode ashl under NDD. */ +void +ix86_split_ashl_ndd (rtx *operands, rtx scratch) +{ + gcc_assert (TARGET_APX_NDD); + int half_width = GET_MODE_BITSIZE (TImode) >> 1; + + rtx low[2], high[2]; + int count; + + split_double_mode (TImode, operands, 2, low, high); + if (CONST_INT_P (operands[2])) + { + count = INTVAL (operands[2]) & (GET_MODE_BITSIZE (TImode) - 1); + + if (count >= half_width) + { + count = count - half_width; + if (count == 0) + { + if (!rtx_equal_p (high[0], low[1])) + emit_move_insn (high[0], low[1]); + } + else if (count == 1) + emit_insn (gen_adddi3 (high[0], low[1], low[1])); + else + emit_insn (gen_ashldi3 (high[0], low[1], GEN_INT (count))); + + ix86_expand_clear (low[0]); + } + else if (count == 1) + { + rtx x3 = gen_rtx_REG (CCCmode, FLAGS_REG); + rtx x4 = gen_rtx_LTU (TImode, x3, const0_rtx); + emit_insn (gen_add3_cc_overflow_1 (DImode, low[0], + low[1], low[1])); + emit_insn (gen_add3_carry (DImode, high[0], high[1], high[1], + x3, x4)); + } + else + { + emit_insn (gen_x86_64_shld_ndd (high[0], high[1], low[1], + GEN_INT (count))); + emit_insn (gen_ashldi3 (low[0], low[1], GEN_INT (count))); + } + } + else + { + emit_insn (gen_x86_64_shld_ndd (high[0], high[1], low[1], + operands[2])); + emit_insn (gen_ashldi3 (low[0], low[1], operands[2])); + if (TARGET_CMOVE && scratch) + { + ix86_expand_clear (scratch); + emit_insn (gen_x86_shift_adj_1 + (DImode, high[0], low[0], operands[2], scratch)); + } + else + emit_insn (gen_x86_shift_adj_2 (DImode, high[0], low[0], operands[2])); + } +} + +/* Helper function to split TImode l/ashr under NDD. */ +void +ix86_split_rshift_ndd (enum rtx_code code, rtx *operands, rtx scratch) +{ + gcc_assert (TARGET_APX_NDD); + int half_width = GET_MODE_BITSIZE (TImode) >> 1; + bool ashr_p = code == ASHIFTRT; + rtx (*gen_shr)(rtx, rtx, rtx) = ashr_p ? gen_ashrdi3 + : gen_lshrdi3; + + rtx low[2], high[2]; + int count; + + split_double_mode (TImode, operands, 2, low, high); + if (CONST_INT_P (operands[2])) + { + count = INTVAL (operands[2]) & (GET_MODE_BITSIZE (TImode) - 1); + + if (ashr_p && (count == GET_MODE_BITSIZE (TImode) - 1)) + { + emit_insn (gen_shr (high[0], high[1], + GEN_INT (half_width - 1))); + emit_move_insn (low[0], high[0]); + } + else if (count >= half_width) + { + if (ashr_p) + emit_insn (gen_shr (high[0], high[1], + GEN_INT (half_width - 1))); + else + ix86_expand_clear (high[0]); + + if (count > half_width) + emit_insn (gen_shr (low[0], high[1], + GEN_INT (count - half_width))); + else + emit_move_insn (low[0], high[1]); + } + else + { + emit_insn (gen_x86_64_shrd_ndd (low[0], low[1], high[1], + GEN_INT (count))); + emit_insn (gen_shr (high[0], high[1], GEN_INT (count))); + } + } + else + { + emit_insn (gen_x86_64_shrd_ndd (low[0], low[1], high[1], + operands[2])); + emit_insn (gen_shr (high[0], high[1], operands[2])); + + if (TARGET_CMOVE && scratch) + { + if (ashr_p) + { + emit_move_insn (scratch, high[0]); + emit_insn (gen_shr (scratch, scratch, + GEN_INT (half_width - 1))); + } + else + ix86_expand_clear (scratch); + + emit_insn (gen_x86_shift_adj_1 + (DImode, low[0], high[0], operands[2], scratch)); + } + else if (ashr_p) + emit_insn (gen_x86_shift_adj_3 + (DImode, low[0], high[0], operands[2])); + else + emit_insn (gen_x86_shift_adj_2 + (DImode, low[0], high[0], operands[2])); + } +} + /* Expand move of V1TI mode register X to a new TI mode register. */ static rtx ix86_expand_v1ti_to_ti (rtx x) diff --git a/gcc/config/i386/i386-protos.h b/gcc/config/i386/i386-protos.h index fa952409729..56349064a6c 100644 --- a/gcc/config/i386/i386-protos.h +++ b/gcc/config/i386/i386-protos.h @@ -174,8 +174,10 @@ extern void x86_initialize_trampoline (rtx, rtx, rtx); extern rtx ix86_zero_extend_to_Pmode (rtx); extern void ix86_split_long_move (rtx[]); extern void ix86_split_ashl (rtx *, rtx, machine_mode); +extern void ix86_split_ashl_ndd (rtx *, rtx); extern void ix86_split_ashr (rtx *, rtx, machine_mode); extern void ix86_split_lshr (rtx *, rtx, machine_mode); +extern void ix86_split_rshift_ndd (enum rtx_code, rtx *, rtx); extern void ix86_expand_v1ti_shift (enum rtx_code, rtx[]); extern void ix86_expand_v1ti_rotate (enum rtx_code, rtx[]); extern void ix86_expand_v1ti_ashiftrt (rtx[]); diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md index 017ab720293..b4db50f61cd 100644 --- a/gcc/config/i386/i386.md +++ b/gcc/config/i386/i386.md @@ -14425,13 +14425,14 @@ }) (define_insn "ashl<mode>3_doubleword" - [(set (match_operand:DWI 0 "register_operand" "=&r") - (ashift:DWI (match_operand:DWI 1 "reg_or_pm1_operand" "0n") - (match_operand:QI 2 "nonmemory_operand" "<S>c"))) + [(set (match_operand:DWI 0 "register_operand" "=&r,&r") + (ashift:DWI (match_operand:DWI 1 "reg_or_pm1_operand" "0n,r") + (match_operand:QI 2 "nonmemory_operand" "<S>c,<S>c"))) (clobber (reg:CC FLAGS_REG))] "" "#" - [(set_attr "type" "multi")]) + [(set_attr "type" "multi") + (set_attr "isa" "*,apx_ndd")]) (define_split [(set (match_operand:DWI 0 "register_operand") @@ -14440,7 +14441,15 @@ (clobber (reg:CC FLAGS_REG))] "epilogue_completed" [(const_int 0)] - "ix86_split_ashl (operands, NULL_RTX, <MODE>mode); DONE;") +{ + if (TARGET_APX_NDD + && !rtx_equal_p (operands[0], operands[1]) + && REG_P (operands[1])) + ix86_split_ashl_ndd (operands, NULL_RTX); + else + ix86_split_ashl (operands, NULL_RTX, <MODE>mode); + DONE; +}) ;; By default we don't ask for a scratch register, because when DWImode ;; values are manipulated, registers are already at a premium. But if @@ -14456,7 +14465,15 @@ (match_dup 3)] "TARGET_CMOVE" [(const_int 0)] - "ix86_split_ashl (operands, operands[3], <DWI>mode); DONE;") +{ + if (TARGET_APX_NDD + && !rtx_equal_p (operands[0], operands[1]) + && (REG_P (operands[1]))) + ix86_split_ashl_ndd (operands, operands[3]); + else + ix86_split_ashl (operands, operands[3], <DWI>mode); + DONE; +}) (define_insn_and_split "*ashl<dwi>3_doubleword_highpart" [(set (match_operand:<DWI> 0 "register_operand" "=r") @@ -15713,16 +15730,24 @@ }) (define_insn_and_split "<insn><mode>3_doubleword" - [(set (match_operand:DWI 0 "register_operand" "=&r") - (any_shiftrt:DWI (match_operand:DWI 1 "register_operand" "0") - (match_operand:QI 2 "nonmemory_operand" "<S>c"))) + [(set (match_operand:DWI 0 "register_operand" "=&r,&r") + (any_shiftrt:DWI (match_operand:DWI 1 "register_operand" "0,r") + (match_operand:QI 2 "nonmemory_operand" "<S>c,<S>c"))) (clobber (reg:CC FLAGS_REG))] "" "#" "epilogue_completed" [(const_int 0)] - "ix86_split_<insn> (operands, NULL_RTX, <MODE>mode); DONE;" - [(set_attr "type" "multi")]) +{ + if (TARGET_APX_NDD + && !rtx_equal_p (operands[0], operands[1])) + ix86_split_rshift_ndd (<CODE>, operands, NULL_RTX); + else + ix86_split_<insn> (operands, NULL_RTX, <MODE>mode); + DONE; +} + [(set_attr "type" "multi") + (set_attr "isa" "*,apx_ndd")]) ;; By default we don't ask for a scratch register, because when DWImode ;; values are manipulated, registers are already at a premium. But if @@ -15738,7 +15763,14 @@ (match_dup 3)] "TARGET_CMOVE" [(const_int 0)] - "ix86_split_<insn> (operands, operands[3], <DWI>mode); DONE;") +{ + if (TARGET_APX_NDD + && !rtx_equal_p (operands[0], operands[1])) + ix86_split_rshift_ndd (<CODE>, operands, operands[3]); + else + ix86_split_<insn> (operands, operands[3], <DWI>mode); + DONE; +}) ;; Split truncations of double word right shifts into x86_shrd_1. (define_insn_and_split "<insn><dwi>3_doubleword_lowpart" diff --git a/gcc/testsuite/gcc.target/i386/apx-ndd-ti-shift.c b/gcc/testsuite/gcc.target/i386/apx-ndd-ti-shift.c new file mode 100644 index 00000000000..0489712b7f6 --- /dev/null +++ b/gcc/testsuite/gcc.target/i386/apx-ndd-ti-shift.c @@ -0,0 +1,91 @@ +/* { dg-do run { target { int128 && { ! ia32 } } } } */ +/* { dg-require-effective-target apxf } */ +/* { dg-options "-O2" } */ + +#include <stdlib.h> + +#define APX_TARGET __attribute__((noinline, target("apxf"))) +#define NO_APX __attribute__((noinline, target("no-apxf"))) +typedef __uint128_t u128; +typedef __int128 i128; + +#define TI_SHIFT_FUNC(TYPE, op, name) \ +APX_TARGET \ +TYPE apx_##name##TYPE (TYPE a, char b) \ +{ \ + return a op b; \ +} \ +TYPE noapx_##name##TYPE (TYPE a, char b) \ +{ \ + return a op b; \ +} \ + +#define TI_SHIFT_FUNC_CONST(TYPE, i, op, name) \ +APX_TARGET \ +TYPE apx_##name##TYPE##_const (TYPE a) \ +{ \ + return a op i; \ +} \ +NO_APX \ +TYPE noapx_##name##TYPE##_const (TYPE a) \ +{ \ + return a op i; \ +} + +#define TI_SHIFT_TEST(TYPE, name, val) \ +{\ + if (apx_##name##TYPE (val, b) != noapx_##name##TYPE (val, b)) \ + abort (); \ +} + +#define TI_SHIFT_CONST_TEST(TYPE, name, val) \ +{\ + if (apx_##name##1##TYPE##_const (val) \ + != noapx_##name##1##TYPE##_const (val)) \ + abort (); \ + if (apx_##name##2##TYPE##_const (val) \ + != noapx_##name##2##TYPE##_const (val)) \ + abort (); \ + if (apx_##name##3##TYPE##_const (val) \ + != noapx_##name##3##TYPE##_const (val)) \ + abort (); \ + if (apx_##name##4##TYPE##_const (val) \ + != noapx_##name##4##TYPE##_const (val)) \ + abort (); \ +} + +TI_SHIFT_FUNC(i128, <<, ashl) +TI_SHIFT_FUNC(i128, >>, ashr) +TI_SHIFT_FUNC(u128, >>, lshr) + +TI_SHIFT_FUNC_CONST(i128, 1, <<, ashl1) +TI_SHIFT_FUNC_CONST(i128, 65, <<, ashl2) +TI_SHIFT_FUNC_CONST(i128, 64, <<, ashl3) +TI_SHIFT_FUNC_CONST(i128, 87, <<, ashl4) +TI_SHIFT_FUNC_CONST(i128, 127, >>, ashr1) +TI_SHIFT_FUNC_CONST(i128, 87, >>, ashr2) +TI_SHIFT_FUNC_CONST(i128, 27, >>, ashr3) +TI_SHIFT_FUNC_CONST(i128, 64, >>, ashr4) +TI_SHIFT_FUNC_CONST(u128, 127, >>, lshr1) +TI_SHIFT_FUNC_CONST(u128, 87, >>, lshr2) +TI_SHIFT_FUNC_CONST(u128, 27, >>, lshr3) +TI_SHIFT_FUNC_CONST(u128, 64, >>, lshr4) + +int main (void) +{ + if (!__builtin_cpu_supports ("apxf")) + return 0; + + u128 ival = 0x123456788765432FLL; + u128 uval = 0xF234567887654321ULL; + char b = 28; + + TI_SHIFT_TEST(i128, ashl, ival) + TI_SHIFT_TEST(i128, ashr, ival) + TI_SHIFT_TEST(u128, lshr, uval) + TI_SHIFT_CONST_TEST(i128, ashl, ival) + TI_SHIFT_CONST_TEST(i128, ashr, ival) + TI_SHIFT_CONST_TEST(u128, lshr, uval) + + return 0; +}
reply other threads:[~2023-12-07 1:40 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=20231207014030.E011A3857B8B@sourceware.org \ --to=hongyuw@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).