From patchwork Tue Jan 31 14:29:20 2012 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Andrew Stubbs X-Patchwork-Id: 6469 Return-Path: X-Original-To: patchwork@peony.canonical.com Delivered-To: patchwork@peony.canonical.com Received: from fiordland.canonical.com (fiordland.canonical.com [91.189.94.145]) by peony.canonical.com (Postfix) with ESMTP id DB94923EAE for ; Tue, 31 Jan 2012 14:29:45 +0000 (UTC) Received: from mail-bk0-f52.google.com (mail-bk0-f52.google.com [209.85.214.52]) by fiordland.canonical.com (Postfix) with ESMTP id 8C938A18396 for ; Tue, 31 Jan 2012 14:29:45 +0000 (UTC) Received: by bkar19 with SMTP id r19so10223bka.11 for ; Tue, 31 Jan 2012 06:29:45 -0800 (PST) Received: by 10.205.139.12 with SMTP id iu12mr10614555bkc.2.1328020170147; Tue, 31 Jan 2012 06:29:30 -0800 (PST) X-Forwarded-To: linaro-patchwork@canonical.com X-Forwarded-For: patch@linaro.org linaro-patchwork@canonical.com Delivered-To: patches@linaro.org Received: by 10.204.130.220 with SMTP id u28cs182897bks; Tue, 31 Jan 2012 06:29:29 -0800 (PST) Received: by 10.224.188.209 with SMTP id db17mr28008895qab.60.1328020168265; Tue, 31 Jan 2012 06:29:28 -0800 (PST) Received: from relay1.mentorg.com (relay1.mentorg.com. [192.94.38.131]) by mx.google.com with ESMTPS id l5si16700304qap.25.2012.01.31.06.29.27 (version=TLSv1/SSLv3 cipher=OTHER); Tue, 31 Jan 2012 06:29:28 -0800 (PST) Received-SPF: neutral (google.com: 192.94.38.131 is neither permitted nor denied by best guess record for domain of Andrew_Stubbs@mentor.com) client-ip=192.94.38.131; Authentication-Results: mx.google.com; spf=neutral (google.com: 192.94.38.131 is neither permitted nor denied by best guess record for domain of Andrew_Stubbs@mentor.com) smtp.mail=Andrew_Stubbs@mentor.com Received: from svr-orw-exc-10.mgc.mentorg.com ([147.34.98.58]) by relay1.mentorg.com with esmtp id 1RsEi5-0003TY-RU from Andrew_Stubbs@mentor.com ; Tue, 31 Jan 2012 06:29:25 -0800 Received: from SVR-IES-FEM-01.mgc.mentorg.com ([137.202.0.104]) by SVR-ORW-EXC-10.mgc.mentorg.com with Microsoft SMTPSVC(6.0.3790.4675); Tue, 31 Jan 2012 06:29:03 -0800 Received: from [172.30.10.94] (137.202.0.76) by SVR-IES-FEM-01.mgc.mentorg.com (137.202.0.104) with Microsoft SMTP Server id 14.1.289.1; Tue, 31 Jan 2012 14:29:23 +0000 Message-ID: <4F27FAC0.1010107@codesourcery.com> Date: Tue, 31 Jan 2012 14:29:20 +0000 From: Andrew Stubbs User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:9.0) Gecko/20111229 Thunderbird/9.0 MIME-Version: 1.0 To: Richard Earnshaw CC: "gcc-patches@gcc.gnu.org" , "patches@linaro.org" Subject: Re: [PATCH][ARM] Improve 64-bit shifts (non-NEON) References: <4F22CBB2.7010101@codesourcery.com> <4F26B687.8020506@arm.com> In-Reply-To: <4F26B687.8020506@arm.com> X-OriginalArrivalTime: 31 Jan 2012 14:29:03.0968 (UTC) FILETIME=[AE9A1600:01CCE024] On 30/01/12 15:25, Richard Earnshaw wrote: > What's the impact of this on -Os? At present we fall back to the > libcalls, but I can't immediately see how the new code would do that. > > Gut feeling is that shift by a constant is always worth inlining at -Os, > but shift by a register isn't. Ah, I hadn't considered that. Good point! This updated patch causes it to fall back to the old behaviour in optimize_size mode. This should do what you want. OK? Andrew 2012-01-31 Andrew Stubbs * config/arm/arm-protos.h (arm_emit_coreregs_64bit_shift): New prototype. * config/arm/arm.c (arm_emit_coreregs_64bit_shift): New function. * config/arm/arm.md (ashldi3): Use arm_emit_coreregs_64bit_shift. (ashrdi3,lshrdi3): Likewise. --- gcc/config/arm/arm-protos.h | 3 + gcc/config/arm/arm.c | 198 +++++++++++++++++++++++++++++++++++++++++++ gcc/config/arm/arm.md | 102 ++++++++++++++++------ 3 files changed, 276 insertions(+), 27 deletions(-) diff --git a/gcc/config/arm/arm-protos.h b/gcc/config/arm/arm-protos.h index 296550a..df8d7a9 100644 --- a/gcc/config/arm/arm-protos.h +++ b/gcc/config/arm/arm-protos.h @@ -242,6 +242,9 @@ struct tune_params extern const struct tune_params *current_tune; extern int vfp3_const_double_for_fract_bits (rtx); + +extern void arm_emit_coreregs_64bit_shift (enum rtx_code, rtx, rtx, rtx, rtx, + rtx); #endif /* RTX_CODE */ #endif /* ! GCC_ARM_PROTOS_H */ diff --git a/gcc/config/arm/arm.c b/gcc/config/arm/arm.c index 0bded8d..eefc45c 100644 --- a/gcc/config/arm/arm.c +++ b/gcc/config/arm/arm.c @@ -25139,5 +25139,203 @@ vfp3_const_double_for_fract_bits (rtx operand) return 0; } +/* The default expansion of general 64-bit shifts in core-regs is suboptimal + on ARM, since we know that shifts by negative amounts are no-ops. + + It's safe for the input and output to be the same register, but + early-clobber rules apply for the shift amount and scratch registers. + + Shift by register requires both scratch registers. Shift by a constant + less than 32 in Thumb2 mode requires SCRATCH1 only. In all other cases + the scratch registers may be NULL. + + Additionally, ashiftrt by a register also clobbers the CC register. */ +void +arm_emit_coreregs_64bit_shift (enum rtx_code code, rtx out, rtx in, + rtx amount, rtx scratch1, rtx scratch2) +{ + rtx out_high = gen_highpart (SImode, out); + rtx out_low = gen_lowpart (SImode, out); + rtx in_high = gen_highpart (SImode, in); + rtx in_low = gen_lowpart (SImode, in); + + /* Bits flow from up-stream to down-stream. */ + rtx out_up = code == ASHIFT ? out_low : out_high; + rtx out_down = code == ASHIFT ? out_high : out_low; + rtx in_up = code == ASHIFT ? in_low : in_high; + rtx in_down = code == ASHIFT ? in_high : in_low; + + gcc_assert (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT); + gcc_assert (out + && (REG_P (out) || GET_CODE (out) == SUBREG) + && GET_MODE (out) == DImode); + gcc_assert (in + && (REG_P (in) || GET_CODE (in) == SUBREG) + && GET_MODE (in) == DImode); + gcc_assert (amount + && (((REG_P (amount) || GET_CODE (amount) == SUBREG) + && GET_MODE (amount) == SImode) + || CONST_INT_P (amount))); + gcc_assert (scratch1 == NULL + || (GET_CODE (scratch1) == SCRATCH) + || (GET_MODE (scratch1) == SImode + && REG_P (scratch1))); + gcc_assert (scratch2 == NULL + || (GET_CODE (scratch2) == SCRATCH) + || (GET_MODE (scratch2) == SImode + && REG_P (scratch2))); + gcc_assert (!REG_P (out) || !REG_P (amount) + || !HARD_REGISTER_P (out) + || (REGNO (out) != REGNO (amount) + && REGNO (out) + 1 != REGNO (amount))); + + /* Macros to make following code more readable. */ + #define SUB_32(DEST,SRC) \ + gen_addsi3 ((DEST), (SRC), gen_rtx_CONST_INT (VOIDmode, -32)) + #define RSB_32(DEST,SRC) \ + gen_subsi3 ((DEST), gen_rtx_CONST_INT (VOIDmode, 32), (SRC)) + #define SUB_S_32(DEST,SRC) \ + gen_addsi3_compare0 ((DEST), (SRC), \ + gen_rtx_CONST_INT (VOIDmode, -32)) + #define SET(DEST,SRC) \ + gen_rtx_SET (SImode, (DEST), (SRC)) + #define SHIFT(CODE,SRC,AMOUNT) \ + gen_rtx_fmt_ee ((CODE), SImode, (SRC), (AMOUNT)) + #define LSHIFT(CODE,SRC,AMOUNT) \ + gen_rtx_fmt_ee ((CODE) == ASHIFT ? ASHIFT : LSHIFTRT, \ + SImode, (SRC), (AMOUNT)) + #define REV_LSHIFT(CODE,SRC,AMOUNT) \ + gen_rtx_fmt_ee ((CODE) == ASHIFT ? LSHIFTRT : ASHIFT, \ + SImode, (SRC), (AMOUNT)) + #define ORR(A,B) \ + gen_rtx_IOR (SImode, (A), (B)) + #define IF(COND,RTX) \ + gen_rtx_COND_EXEC (VOIDmode, \ + gen_rtx_ ## COND (CCmode, cc_reg, \ + const0_rtx), \ + (RTX)) + + if (CONST_INT_P (amount)) + { + /* Shifts by a constant amount. */ + if (INTVAL (amount) <= 0) + /* Match what shift-by-register would do. */ + emit_insn (gen_movdi (out, in)); + else if (INTVAL (amount) >= 64) + { + /* Match what shift-by-register would do. */ + if (code == ASHIFTRT) + { + rtx const31_rtx = gen_rtx_CONST_INT (VOIDmode, 31); + emit_insn (SET (out_down, SHIFT (code, in_up, const31_rtx))); + emit_insn (SET (out_up, SHIFT (code, in_up, const31_rtx))); + } + else + emit_insn (gen_movdi (out, const0_rtx)); + } + else if (INTVAL (amount) < 32) + { + /* Shifts by a constant less than 32. */ + rtx reverse_amount = gen_rtx_CONST_INT (VOIDmode, + 32 - INTVAL (amount)); + + emit_insn (SET (out_down, LSHIFT (code, in_down, amount))); + emit_insn (SET (out_down, + ORR (REV_LSHIFT (code, in_up, reverse_amount), + out_down))); + emit_insn (SET (out_up, SHIFT (code, in_up, amount))); + } + else + { + /* Shifts by a constant greater than 31. */ + rtx adj_amount = gen_rtx_CONST_INT (VOIDmode, INTVAL (amount) - 32); + + emit_insn (SET (out_down, SHIFT (code, in_up, adj_amount))); + if (code == ASHIFTRT) + emit_insn (gen_ashrsi3 (out_up, in_up, + gen_rtx_CONST_INT (VOIDmode, 31))); + else + emit_insn (SET (out_up, const0_rtx)); + } + } + else + { + /* Shifts by a variable amount. */ + rtx cc_reg = gen_rtx_REG (CC_NCVmode, CC_REGNUM); + + gcc_assert (scratch1 && REG_P (scratch1)); + gcc_assert (scratch2 && REG_P (scratch2)); + + switch (code) + { + case ASHIFT: + emit_insn (SUB_32 (scratch1, amount)); + emit_insn (RSB_32 (scratch2, amount)); + break; + case ASHIFTRT: + emit_insn (RSB_32 (scratch1, amount)); + emit_insn (SUB_S_32 (scratch2, amount)); + break; + case LSHIFTRT: + emit_insn (RSB_32 (scratch1, amount)); + emit_insn (SUB_32 (scratch2, amount)); + break; + default: + gcc_unreachable (); + } + + emit_insn (SET (out_down, LSHIFT (code, in_down, amount))); + + if (!TARGET_THUMB2) + { + /* If this were only called during expand we could just use the else + case and let combine deal with it, but this can also be called + from post-reload splitters. */ + emit_insn (SET (out_down, + ORR (SHIFT (ASHIFT, in_up, scratch1), out_down))); + if (code == ASHIFTRT) + { + emit_insn (IF (GE, + SET (out_down, + ORR (SHIFT (ASHIFTRT, in_up, scratch2), + out_down)))); + } + else + emit_insn (SET (out_down, ORR (SHIFT (LSHIFTRT, in_up, scratch2), + out_down))); + } + else + { + /* Thumb2 can't do shift and or in one insn. */ + emit_insn (SET (scratch1, SHIFT (ASHIFT, in_up, scratch1))); + emit_insn (gen_iorsi3 (out_down, out_down, scratch1)); + + if (code == ASHIFTRT) + { + emit_insn (IF (GE, SET (scratch2, + SHIFT (ASHIFTRT, in_up, scratch2)))); + emit_insn (IF (GE, SET (out_down, ORR (out_down, scratch2)))); + } + else + { + emit_insn (SET (scratch2, SHIFT (LSHIFTRT, in_up, scratch2))); + emit_insn (gen_iorsi3 (out_down, out_down, scratch2)); + } + } + + emit_insn (SET (out_up, SHIFT (code, in_up, amount))); + } + + #undef SUB_32 + #undef RSB_32 + #undef SUB_S_32 + #undef SET + #undef SHIFT + #undef LSHIFT + #undef REV_LSHIFT + #undef ORR + #undef IF +} + #include "gt-arm.h" diff --git a/gcc/config/arm/arm.md b/gcc/config/arm/arm.md index 751997f..e814dc4 100644 --- a/gcc/config/arm/arm.md +++ b/gcc/config/arm/arm.md @@ -3466,21 +3466,37 @@ (match_operand:SI 2 "reg_or_int_operand" "")))] "TARGET_32BIT" " - if (GET_CODE (operands[2]) == CONST_INT) + if (!CONST_INT_P (operands[2]) + && (TARGET_REALLY_IWMMXT || (TARGET_HARD_FLOAT && TARGET_MAVERICK))) + ; /* No special preparation statements; expand pattern as above. */ + else { - if ((HOST_WIDE_INT) INTVAL (operands[2]) == 1) + rtx scratch1, scratch2; + + if (GET_CODE (operands[2]) == CONST_INT + && (HOST_WIDE_INT) INTVAL (operands[2]) == 1) { emit_insn (gen_arm_ashldi3_1bit (operands[0], operands[1])); DONE; } - /* Ideally we shouldn't fail here if we could know that operands[1] - ends up already living in an iwmmxt register. Otherwise it's - cheaper to have the alternate code being generated than moving - values to iwmmxt regs and back. */ - FAIL; + + /* Ideally we should use iwmmxt here if we could know that operands[1] + ends up already living in an iwmmxt register. Otherwise it's + cheaper to have the alternate code being generated than moving + values to iwmmxt regs and back. */ + + /* If we're optimizing for size, we prefer the libgcc calls. */ + if (optimize_size) + FAIL; + + /* Expand operation using core-registers. + 'FAIL' would achieve the same thing, but this is a bit smarter. */ + scratch1 = gen_reg_rtx (SImode); + scratch2 = gen_reg_rtx (SImode); + arm_emit_coreregs_64bit_shift (ASHIFT, operands[0], operands[1], + operands[2], scratch1, scratch2); + DONE; } - else if (!TARGET_REALLY_IWMMXT && !(TARGET_HARD_FLOAT && TARGET_MAVERICK)) - FAIL; " ) @@ -3525,21 +3541,37 @@ (match_operand:SI 2 "reg_or_int_operand" "")))] "TARGET_32BIT" " - if (GET_CODE (operands[2]) == CONST_INT) + if (!CONST_INT_P (operands[2]) + && (TARGET_REALLY_IWMMXT || (TARGET_HARD_FLOAT && TARGET_MAVERICK))) + ; /* No special preparation statements; expand pattern as above. */ + else { - if ((HOST_WIDE_INT) INTVAL (operands[2]) == 1) + rtx scratch1, scratch2; + + if (GET_CODE (operands[2]) == CONST_INT + && (HOST_WIDE_INT) INTVAL (operands[2]) == 1) { emit_insn (gen_arm_ashrdi3_1bit (operands[0], operands[1])); DONE; } - /* Ideally we shouldn't fail here if we could know that operands[1] - ends up already living in an iwmmxt register. Otherwise it's - cheaper to have the alternate code being generated than moving - values to iwmmxt regs and back. */ - FAIL; + + /* Ideally we should use iwmmxt here if we could know that operands[1] + ends up already living in an iwmmxt register. Otherwise it's + cheaper to have the alternate code being generated than moving + values to iwmmxt regs and back. */ + + /* If we're optimizing for size, we prefer the libgcc calls. */ + if (optimize_size) + FAIL; + + /* Expand operation using core-registers. + 'FAIL' would achieve the same thing, but this is a bit smarter. */ + scratch1 = gen_reg_rtx (SImode); + scratch2 = gen_reg_rtx (SImode); + arm_emit_coreregs_64bit_shift (ASHIFTRT, operands[0], operands[1], + operands[2], scratch1, scratch2); + DONE; } - else if (!TARGET_REALLY_IWMMXT) - FAIL; " ) @@ -3582,21 +3614,37 @@ (match_operand:SI 2 "reg_or_int_operand" "")))] "TARGET_32BIT" " - if (GET_CODE (operands[2]) == CONST_INT) + if (!CONST_INT_P (operands[2]) + && (TARGET_REALLY_IWMMXT || (TARGET_HARD_FLOAT && TARGET_MAVERICK))) + ; /* No special preparation statements; expand pattern as above. */ + else { - if ((HOST_WIDE_INT) INTVAL (operands[2]) == 1) + rtx scratch1, scratch2; + + if (GET_CODE (operands[2]) == CONST_INT + && (HOST_WIDE_INT) INTVAL (operands[2]) == 1) { emit_insn (gen_arm_lshrdi3_1bit (operands[0], operands[1])); DONE; } - /* Ideally we shouldn't fail here if we could know that operands[1] - ends up already living in an iwmmxt register. Otherwise it's - cheaper to have the alternate code being generated than moving - values to iwmmxt regs and back. */ - FAIL; + + /* Ideally we should use iwmmxt here if we could know that operands[1] + ends up already living in an iwmmxt register. Otherwise it's + cheaper to have the alternate code being generated than moving + values to iwmmxt regs and back. */ + + /* If we're optimizing for size, we prefer the libgcc calls. */ + if (optimize_size) + FAIL; + + /* Expand operation using core-registers. + 'FAIL' would achieve the same thing, but this is a bit smarter. */ + scratch1 = gen_reg_rtx (SImode); + scratch2 = gen_reg_rtx (SImode); + arm_emit_coreregs_64bit_shift (LSHIFTRT, operands[0], operands[1], + operands[2], scratch1, scratch2); + DONE; } - else if (!TARGET_REALLY_IWMMXT) - FAIL; " )