From patchwork Tue Nov 13 13:50:12 2012 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit Subject: [ARM] Improved core -> NEON extend Date: Tue, 13 Nov 2012 03:50:12 -0000 From: Ulrich Weigand X-Patchwork-Id: 198681 Message-Id: <201211131350.qADDoCIc011668@d06av02.portsmouth.uk.ibm.com> To: gcc-patches@gcc.gnu.org Cc: ramrad01@arm.com Hello, here's another of Andrew's patches to improve NEON usage. This one was originally posted here: http://gcc.gnu.org/ml/gcc-patches/2012-02/msg01213.html The idea to improve SImode to DImode extends that also move from core registers to NEON registers. In this situation, the compiler used to perform the extension in core registers first and then moves to NEON, wasting a core register in the process. The patch changes this to move to NEON first and extend there. [ This patch requires both the NEON shift and the lower-subreg patches, both of which are now in mainline, so this patch is ready to merge as well at this point. ] Tested on arm-linux-gnueabi. OK for mainline? Bye, Ulrich 2012-11-13 Andrew Stubbs Ulrich Weigand gcc/ * config/arm/arm.md (zero_extenddi2): Add extra alternatives for NEON registers. Add alternative for one-instruction extend-in-place. (extenddi2): Likewise. Add constraints for Thumb-mode memory loads. Prevent extend splitters doing NEON alternatives. * config/arm/iterators.md (qhs_extenddi_cstr, qhs_zextenddi_cstr): Adjust constraints to add new alternatives. * config/arm/neon.md: Add splitters for zero- and sign-extend. gcc/testsuite/ * gcc.target/arm/neon-extend-1.c: New file. * gcc.target/arm/neon-extend-2.c: New file. === modified file 'gcc/config/arm/arm.md' --- gcc/config/arm/arm.md 2012-09-19 12:57:52 +0000 +++ gcc/config/arm/arm.md 2012-09-19 13:19:31 +0000 @@ -4567,33 +4567,36 @@ ;; Zero and sign extension instructions. (define_insn "zero_extenddi2" - [(set (match_operand:DI 0 "s_register_operand" "=r") + [(set (match_operand:DI 0 "s_register_operand" "=w,r,?r") (zero_extend:DI (match_operand:QHSI 1 "" "")))] "TARGET_32BIT " "#" - [(set_attr "length" "8") + [(set_attr "length" "8,4,8") (set_attr "ce_count" "2") (set_attr "predicable" "yes")] ) (define_insn "extenddi2" - [(set (match_operand:DI 0 "s_register_operand" "=r") + [(set (match_operand:DI 0 "s_register_operand" "=w,r,?r,?r") (sign_extend:DI (match_operand:QHSI 1 "" "")))] "TARGET_32BIT " "#" - [(set_attr "length" "8") + [(set_attr "length" "8,4,8,8") (set_attr "ce_count" "2") (set_attr "shift" "1") - (set_attr "predicable" "yes")] + (set_attr "predicable" "yes") + (set_attr "arch" "*,*,a,t")] ) ;; Splits for all extensions to DImode (define_split [(set (match_operand:DI 0 "s_register_operand" "") (zero_extend:DI (match_operand 1 "nonimmediate_operand" "")))] - "TARGET_32BIT" + "TARGET_32BIT && (!TARGET_NEON + || (reload_completed + && !(IS_VFP_REGNUM (REGNO (operands[0])))))" [(set (match_dup 0) (match_dup 1))] { rtx lo_part = gen_lowpart (SImode, operands[0]); @@ -4619,7 +4622,9 @@ (define_split [(set (match_operand:DI 0 "s_register_operand" "") (sign_extend:DI (match_operand 1 "nonimmediate_operand" "")))] - "TARGET_32BIT" + "TARGET_32BIT && (!TARGET_NEON + || (reload_completed + && !(IS_VFP_REGNUM (REGNO (operands[0])))))" [(set (match_dup 0) (ashiftrt:SI (match_dup 1) (const_int 31)))] { rtx lo_part = gen_lowpart (SImode, operands[0]); === modified file 'gcc/config/arm/iterators.md' --- gcc/config/arm/iterators.md 2012-09-19 12:57:52 +0000 +++ gcc/config/arm/iterators.md 2012-09-19 13:19:31 +0000 @@ -412,8 +412,8 @@ (define_mode_attr qhs_extenddi_op [(SI "s_register_operand") (HI "nonimmediate_operand") (QI "arm_reg_or_extendqisi_mem_op")]) -(define_mode_attr qhs_extenddi_cstr [(SI "r") (HI "rm") (QI "rUq")]) -(define_mode_attr qhs_zextenddi_cstr [(SI "r") (HI "rm") (QI "rm")]) +(define_mode_attr qhs_extenddi_cstr [(SI "r,0,r,r") (HI "r,0,rm,rm") (QI "r,0,rUq,rm")]) +(define_mode_attr qhs_zextenddi_cstr [(SI "r,0,r") (HI "r,0,rm") (QI "r,0,rm")]) ;; Mode attributes used for fixed-point support. (define_mode_attr qaddsub_suf [(V4UQQ "8") (V2UHQ "16") (UQQ "8") (UHQ "16") === modified file 'gcc/config/arm/neon.md' --- gcc/config/arm/neon.md 2012-09-19 12:57:52 +0000 +++ gcc/config/arm/neon.md 2012-09-19 13:19:31 +0000 @@ -5878,3 +5878,65 @@ (const_string "neon_fp_vadd_qqq_vabs_qq")) (const_string "neon_int_5")))] ) + +;; Copy from core-to-neon regs, then extend, not vice-versa + +(define_split + [(set (match_operand:DI 0 "s_register_operand" "") + (sign_extend:DI (match_operand:SI 1 "s_register_operand" "")))] + "TARGET_NEON && reload_completed && IS_VFP_REGNUM (REGNO (operands[0]))" + [(set (match_dup 2) (vec_duplicate:V2SI (match_dup 1))) + (set (match_dup 0) (ashiftrt:DI (match_dup 0) (const_int 32)))] + { + operands[2] = gen_rtx_REG (V2SImode, REGNO (operands[0])); + }) + +(define_split + [(set (match_operand:DI 0 "s_register_operand" "") + (sign_extend:DI (match_operand:HI 1 "s_register_operand" "")))] + "TARGET_NEON && reload_completed && IS_VFP_REGNUM (REGNO (operands[0]))" + [(set (match_dup 2) (vec_duplicate:V4HI (match_dup 1))) + (set (match_dup 0) (ashiftrt:DI (match_dup 0) (const_int 48)))] + { + operands[2] = gen_rtx_REG (V4HImode, REGNO (operands[0])); + }) + +(define_split + [(set (match_operand:DI 0 "s_register_operand" "") + (sign_extend:DI (match_operand:QI 1 "s_register_operand" "")))] + "TARGET_NEON && reload_completed && IS_VFP_REGNUM (REGNO (operands[0]))" + [(set (match_dup 2) (vec_duplicate:V8QI (match_dup 1))) + (set (match_dup 0) (ashiftrt:DI (match_dup 0) (const_int 56)))] + { + operands[2] = gen_rtx_REG (V8QImode, REGNO (operands[0])); + }) + +(define_split + [(set (match_operand:DI 0 "s_register_operand" "") + (zero_extend:DI (match_operand:SI 1 "s_register_operand" "")))] + "TARGET_NEON && reload_completed && IS_VFP_REGNUM (REGNO (operands[0]))" + [(set (match_dup 2) (vec_duplicate:V2SI (match_dup 1))) + (set (match_dup 0) (lshiftrt:DI (match_dup 0) (const_int 32)))] + { + operands[2] = gen_rtx_REG (V2SImode, REGNO (operands[0])); + }) + +(define_split + [(set (match_operand:DI 0 "s_register_operand" "") + (zero_extend:DI (match_operand:HI 1 "s_register_operand" "")))] + "TARGET_NEON && reload_completed && IS_VFP_REGNUM (REGNO (operands[0]))" + [(set (match_dup 2) (vec_duplicate:V4HI (match_dup 1))) + (set (match_dup 0) (lshiftrt:DI (match_dup 0) (const_int 48)))] + { + operands[2] = gen_rtx_REG (V4HImode, REGNO (operands[0])); + }) + +(define_split + [(set (match_operand:DI 0 "s_register_operand" "") + (zero_extend:DI (match_operand:QI 1 "s_register_operand" "")))] + "TARGET_NEON && reload_completed && IS_VFP_REGNUM (REGNO (operands[0]))" + [(set (match_dup 2) (vec_duplicate:V8QI (match_dup 1))) + (set (match_dup 0) (lshiftrt:DI (match_dup 0) (const_int 56)))] + { + operands[2] = gen_rtx_REG (V8QImode, REGNO (operands[0])); + }) === added file 'gcc/testsuite/gcc.target/arm/neon-extend-1.c' --- gcc/testsuite/gcc.target/arm/neon-extend-1.c 1970-01-01 00:00:00 +0000 +++ gcc/testsuite/gcc.target/arm/neon-extend-1.c 2012-09-19 13:19:31 +0000 @@ -0,0 +1,13 @@ +/* { dg-require-effective-target arm_neon_hw } */ +/* { dg-options "-O2" } */ +/* { dg-add-options arm_neon } */ + +void +f (unsigned int a) +{ + unsigned long long b = a; + asm volatile ("@ extended to %0" : : "w" (b)); +} + +/* { dg-final { scan-assembler "vdup.32" } } */ +/* { dg-final { scan-assembler "vshr.u64" } } */ === added file 'gcc/testsuite/gcc.target/arm/neon-extend-2.c' --- gcc/testsuite/gcc.target/arm/neon-extend-2.c 1970-01-01 00:00:00 +0000 +++ gcc/testsuite/gcc.target/arm/neon-extend-2.c 2012-09-19 13:19:31 +0000 @@ -0,0 +1,13 @@ +/* { dg-require-effective-target arm_neon_hw } */ +/* { dg-options "-O2" } */ +/* { dg-add-options arm_neon } */ + +void +f (int a) +{ + long long b = a; + asm volatile ("@ extended to %0" : : "w" (b)); +} + +/* { dg-final { scan-assembler "vdup.32" } } */ +/* { dg-final { scan-assembler "vshr.s64" } } */