From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from eggs.gnu.org ([2001:4830:134:3::10]:59830) by lists.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1ddVPh-00042f-Vb for qemu-devel@nongnu.org; Fri, 04 Aug 2017 01:44:47 -0400 Received: from Debian-exim by eggs.gnu.org with spam-scanned (Exim 4.71) (envelope-from ) id 1ddVPg-0003YI-6k for qemu-devel@nongnu.org; Fri, 04 Aug 2017 01:44:46 -0400 Received: from mail-pg0-x242.google.com ([2607:f8b0:400e:c05::242]:37030) by eggs.gnu.org with esmtps (TLS1.0:RSA_AES_128_CBC_SHA1:16) (Exim 4.71) (envelope-from ) id 1ddVPf-0003Vw-Tb for qemu-devel@nongnu.org; Fri, 04 Aug 2017 01:44:44 -0400 Received: by mail-pg0-x242.google.com with SMTP id 83so822317pgb.4 for ; Thu, 03 Aug 2017 22:44:43 -0700 (PDT) Received: from bigtime.twiddle.net (97-126-108-236.tukw.qwest.net. [97.126.108.236]) by smtp.gmail.com with ESMTPSA id o14sm1061063pfi.158.2017.08.03.22.44.41 for (version=TLS1_2 cipher=ECDHE-RSA-CHACHA20-POLY1305 bits=256/256); Thu, 03 Aug 2017 22:44:41 -0700 (PDT) Sender: Richard Henderson From: Richard Henderson Date: Thu, 3 Aug 2017 22:44:18 -0700 Message-Id: <20170804054426.10590-16-rth@twiddle.net> In-Reply-To: <20170804054426.10590-1-rth@twiddle.net> References: <20170804054426.10590-1-rth@twiddle.net> Subject: [Qemu-devel] [PATCH for-2.11 15/23] tcg/arm: Improve tlb load for armv7 List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , To: qemu-devel@nongnu.org Use UBFX to avoid limitation on CPU_TLB_BITS. Since we're dropping the initial shift, we need to replace the page masking. We can use MOVW+BIC to do this without shifting. The result is the same size as the armv6 path with one less conditional instruction. Signed-off-by: Richard Henderson --- tcg/arm/tcg-target.inc.c | 72 ++++++++++++++++++++++++++++++++++-------------- 1 file changed, 52 insertions(+), 20 deletions(-) diff --git a/tcg/arm/tcg-target.inc.c b/tcg/arm/tcg-target.inc.c index 81ea900852..66c369c239 100644 --- a/tcg/arm/tcg-target.inc.c +++ b/tcg/arm/tcg-target.inc.c @@ -1173,18 +1173,33 @@ static TCGReg tcg_out_tlb_read(TCGContext *s, TCGReg addrlo, TCGReg addrhi, unsigned s_bits = opc & MO_SIZE; unsigned a_bits = get_alignment_bits(opc); - /* Should generate something like the following: - * shr tmp, addrlo, #TARGET_PAGE_BITS (1) + /* V7 generates the following: + * ubfx r0, addrlo, #TARGET_PAGE_BITS, #CPU_TLB_BITS * add r2, env, #high - * and r0, tmp, #(CPU_TLB_SIZE - 1) (2) - * add r2, r2, r0, lsl #CPU_TLB_ENTRY_BITS (3) - * ldr r0, [r2, #cmp] (4) + * add r2, r2, r0, lsl #CPU_TLB_ENTRY_BITS + * ldr r0, [r2, #cmp] + * ldr r2, [r2, #add] + * movw tmp, #page_align_mask + * bic tmp, addrlo, tmp + * cmp r0, tmp + * + * Otherwise we generate: + * shr tmp, addrlo, #TARGET_PAGE_BITS + * add r2, env, #high + * and r0, tmp, #(CPU_TLB_SIZE - 1) + * add r2, r2, r0, lsl #CPU_TLB_ENTRY_BITS + * ldr r0, [r2, #cmp] + * ldr r2, [r2, #add] * tst addrlo, #s_mask - * ldr r2, [r2, #add] (5) * cmpeq r0, tmp, lsl #TARGET_PAGE_BITS */ - tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, - 0, addrlo, SHIFT_IMM_LSR(TARGET_PAGE_BITS)); + if (use_armv7_instructions) { + tcg_out_extract(s, COND_AL, TCG_REG_R0, addrlo, + TARGET_PAGE_BITS, CPU_TLB_BITS); + } else { + tcg_out_dat_reg(s, COND_AL, ARITH_MOV, TCG_REG_TMP, + 0, addrlo, SHIFT_IMM_LSR(TARGET_PAGE_BITS)); + } /* We checked that the offset is contained within 16 bits above. */ if (add_off > 0xfff || (use_armv6_instructions && cmp_off > 0xff)) { @@ -1194,9 +1209,10 @@ static TCGReg tcg_out_tlb_read(TCGContext *s, TCGReg addrlo, TCGReg addrhi, add_off -= cmp_off & 0xff00; cmp_off &= 0xff; } - - tcg_out_dat_imm(s, COND_AL, ARITH_AND, - TCG_REG_R0, TCG_REG_TMP, CPU_TLB_SIZE - 1); + if (!use_armv7_instructions) { + tcg_out_dat_imm(s, COND_AL, ARITH_AND, + TCG_REG_R0, TCG_REG_TMP, CPU_TLB_SIZE - 1); + } tcg_out_dat_reg(s, COND_AL, ARITH_ADD, TCG_REG_R2, base, TCG_REG_R0, SHIFT_IMM_LSL(CPU_TLB_ENTRY_BITS)); @@ -1212,24 +1228,40 @@ static TCGReg tcg_out_tlb_read(TCGContext *s, TCGReg addrlo, TCGReg addrhi, } } + /* Load the tlb addend. */ + tcg_out_ld32_12(s, COND_AL, TCG_REG_R2, TCG_REG_R2, add_off); + /* Check alignment. We don't support inline unaligned acceses, but we can easily support overalignment checks. */ if (a_bits < s_bits) { a_bits = s_bits; } - if (a_bits) { - tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addrlo, (1 << a_bits) - 1); - } - /* Load the tlb addend. */ - tcg_out_ld32_12(s, COND_AL, TCG_REG_R2, TCG_REG_R2, add_off); + if (use_armv7_instructions) { + tcg_target_ulong mask = ~(TARGET_PAGE_MASK | ((1 << a_bits) - 1)); + int rot = encode_imm(mask); - tcg_out_dat_reg(s, (a_bits ? COND_EQ : COND_AL), ARITH_CMP, 0, - TCG_REG_R0, TCG_REG_TMP, SHIFT_IMM_LSL(TARGET_PAGE_BITS)); + if (rot >= 0) { + tcg_out_dat_imm(s, COND_AL, ARITH_BIC, TCG_REG_TMP, addrlo, + rotl(mask, rot) | (rot << 7)); + } else { + tcg_out_movi32(s, COND_AL, TCG_REG_TMP, mask); + tcg_out_dat_reg(s, COND_AL, ARITH_BIC, TCG_REG_TMP, + addrlo, TCG_REG_TMP, 0); + } + tcg_out_dat_reg(s, COND_AL, ARITH_CMP, 0, TCG_REG_R0, TCG_REG_TMP, 0); + } else { + if (a_bits) { + tcg_out_dat_imm(s, COND_AL, ARITH_TST, 0, addrlo, + (1 << a_bits) - 1); + } + tcg_out_dat_reg(s, (a_bits ? COND_EQ : COND_AL), ARITH_CMP, + 0, TCG_REG_R0, TCG_REG_TMP, + SHIFT_IMM_LSL(TARGET_PAGE_BITS)); + } if (TARGET_LONG_BITS == 64) { - tcg_out_dat_reg(s, COND_EQ, ARITH_CMP, 0, - TCG_REG_R1, addrhi, SHIFT_IMM_LSL(0)); + tcg_out_dat_reg(s, COND_EQ, ARITH_CMP, 0, TCG_REG_R1, addrhi, 0); } return TCG_REG_R2; -- 2.13.3