From patchwork Sat Sep 16 02:34:16 2017 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Richard Henderson X-Patchwork-Id: 9954079 Return-Path: Received: from mail.wl.linuxfoundation.org (pdx-wl-mail.web.codeaurora.org [172.30.200.125]) by pdx-korg-patchwork.web.codeaurora.org (Postfix) with ESMTP id 679CA601C0 for ; Sat, 16 Sep 2017 02:41:18 +0000 (UTC) Received: from mail.wl.linuxfoundation.org (localhost [127.0.0.1]) by mail.wl.linuxfoundation.org (Postfix) with ESMTP id 5F91429338 for ; Sat, 16 Sep 2017 02:41:18 +0000 (UTC) Received: by mail.wl.linuxfoundation.org (Postfix, from userid 486) id 540472933A; Sat, 16 Sep 2017 02:41:18 +0000 (UTC) X-Spam-Checker-Version: SpamAssassin 3.3.1 (2010-03-16) on pdx-wl-mail.web.codeaurora.org X-Spam-Level: X-Spam-Status: No, score=-6.8 required=2.0 tests=BAYES_00,DKIM_SIGNED, RCVD_IN_DNSWL_HI,T_DKIM_INVALID autolearn=ham version=3.3.1 Received: from lists.gnu.org (lists.gnu.org [208.118.235.17]) (using TLSv1 with cipher AES256-SHA (256/256 bits)) (No client certificate requested) by mail.wl.linuxfoundation.org (Postfix) with ESMTPS id 88A2C29338 for ; Sat, 16 Sep 2017 02:41:16 +0000 (UTC) Received: from localhost ([::1]:55718 helo=lists.gnu.org) by lists.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1dt32S-0005vb-Dy for patchwork-qemu-devel@patchwork.kernel.org; Fri, 15 Sep 2017 22:41:00 -0400 Received: from eggs.gnu.org ([2001:4830:134:3::10]:35751) by lists.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1dt2wB-0001Vd-GB for qemu-devel@nongnu.org; Fri, 15 Sep 2017 22:34:34 -0400 Received: from Debian-exim by eggs.gnu.org with spam-scanned (Exim 4.71) (envelope-from ) id 1dt2w8-0007jw-Jb for qemu-devel@nongnu.org; Fri, 15 Sep 2017 22:34:31 -0400 Received: from mail-pf0-x22d.google.com ([2607:f8b0:400e:c00::22d]:56173) by eggs.gnu.org with esmtps (TLS1.0:RSA_AES_128_CBC_SHA1:16) (Exim 4.71) (envelope-from ) id 1dt2w8-0007jl-9m for qemu-devel@nongnu.org; Fri, 15 Sep 2017 22:34:28 -0400 Received: by mail-pf0-x22d.google.com with SMTP id r71so2305607pfe.12 for ; Fri, 15 Sep 2017 19:34:28 -0700 (PDT) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=linaro.org; s=google; h=from:to:cc:subject:date:message-id:in-reply-to:references; bh=FPeApC4PMqN5hYDFGMlrW2WGw9xtSgDU3RhnoLsAY6M=; b=F45ig/JNRKTuWWLBkX0MtdRk+u2H11S7uo01F8dVAM9+95pVQCoj61OBhmQCn22pno cvQJDWwXG27NB7RRQbWsTfARHPKztKboq/Ozou2Yp3y/cTpAuqvMSG016WVwW86BxEyH gTPhdpJSKjw3HxkW2ZpNlXbPyIATEnIBoN2+U= X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20161025; h=x-gm-message-state:from:to:cc:subject:date:message-id:in-reply-to :references; bh=FPeApC4PMqN5hYDFGMlrW2WGw9xtSgDU3RhnoLsAY6M=; b=ZkCJcVTQctYLp+oG8sH4v1FcEPxaosA2X+0obGvkTz0RWMSHQvAW4Li8ez4vTzINil hLsVfLUjK645VshiaoaM8azkpAqJZAGIJvlVznmq9MfLCKfgfazWY/Rt4lRRgO9J0Kjh aE9PDSqsdgPeHwF1vh7Gm9TtSIZeIyg9tK2mudZpr3PegBMloV6j+WiKOJ32WZ6Nyh4A GZ+1oAYwCXTRU6HUJ5g0aLx9lAuWIP62g7S68xVkspX6/m3i/UP8hwUz5mvbT1Fq1P/m GwmcF2ppvXUnes7SlYZrLZgkrFqbfgmCmANEKvtxS6SpNp2I0AroQczdfUoB75yAi3J4 qEwQ== X-Gm-Message-State: AHPjjUjsaDzyilNJOWSp6JnR0BGmGP5t6IwiiLmS8wW6S5QJZ8x3AqG5 9VoRXX5sXkfzzjMLkl4SDg== X-Google-Smtp-Source: ADKCNb4rbiQ7i+P24XZLwPVDIrK2vvMdFETk/039pY0mRtaH/XOn8Dr2O2IWvJtptelGEzxeYEoryw== X-Received: by 10.101.70.203 with SMTP id n11mr25724463pgr.197.1505529266711; Fri, 15 Sep 2017 19:34:26 -0700 (PDT) Received: from bigtime.twiddle.net (97-126-103-167.tukw.qwest.net. [97.126.103.167]) by smtp.gmail.com with ESMTPSA id o5sm4443621pfh.67.2017.09.15.19.34.25 (version=TLS1_2 cipher=ECDHE-RSA-CHACHA20-POLY1305 bits=256/256); Fri, 15 Sep 2017 19:34:25 -0700 (PDT) From: Richard Henderson To: qemu-devel@nongnu.org Date: Fri, 15 Sep 2017 19:34:16 -0700 Message-Id: <20170916023417.14599-6-richard.henderson@linaro.org> X-Mailer: git-send-email 2.13.5 In-Reply-To: <20170916023417.14599-1-richard.henderson@linaro.org> References: <20170916023417.14599-1-richard.henderson@linaro.org> X-detected-operating-system: by eggs.gnu.org: Genre and OS details not recognized. X-Received-From: 2607:f8b0:400e:c00::22d Subject: [Qemu-devel] [PATCH v3 5/6] tcg/i386: Add vector operations X-BeenThere: qemu-devel@nongnu.org X-Mailman-Version: 2.1.21 Precedence: list List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Cc: alex.bennee@linaro.org, f4bug@amsat.org Errors-To: qemu-devel-bounces+patchwork-qemu-devel=patchwork.kernel.org@nongnu.org Sender: "Qemu-devel" X-Virus-Scanned: ClamAV using ClamSMTP Signed-off-by: Richard Henderson --- tcg/i386/tcg-target.h | 36 +++- tcg/i386/tcg-target.inc.c | 423 +++++++++++++++++++++++++++++++++++++++++----- 2 files changed, 413 insertions(+), 46 deletions(-) diff --git a/tcg/i386/tcg-target.h b/tcg/i386/tcg-target.h index b89dababf4..df69f8db91 100644 --- a/tcg/i386/tcg-target.h +++ b/tcg/i386/tcg-target.h @@ -30,11 +30,10 @@ #ifdef __x86_64__ # define TCG_TARGET_REG_BITS 64 -# define TCG_TARGET_NB_REGS 16 #else # define TCG_TARGET_REG_BITS 32 -# define TCG_TARGET_NB_REGS 8 #endif +# define TCG_TARGET_NB_REGS 24 typedef enum { TCG_REG_EAX = 0, @@ -56,6 +55,19 @@ typedef enum { TCG_REG_R13, TCG_REG_R14, TCG_REG_R15, + + /* SSE registers; 64-bit has access to 8 more, but we won't + need more than a few and using only the first 8 minimizes + the need for a rex prefix on the sse instructions. */ + TCG_REG_XMM0, + TCG_REG_XMM1, + TCG_REG_XMM2, + TCG_REG_XMM3, + TCG_REG_XMM4, + TCG_REG_XMM5, + TCG_REG_XMM6, + TCG_REG_XMM7, + TCG_REG_RAX = TCG_REG_EAX, TCG_REG_RCX = TCG_REG_ECX, TCG_REG_RDX = TCG_REG_EDX, @@ -78,6 +90,17 @@ typedef enum { extern bool have_bmi1; extern bool have_popcnt; +#ifdef __SSE2__ +#define have_sse2 true +#else +extern bool have_sse2; +#endif +#ifdef __AVX2__ +#define have_avx2 true +#else +extern bool have_avx2; +#endif + /* optional instructions */ #define TCG_TARGET_HAS_div2_i32 1 #define TCG_TARGET_HAS_rot_i32 1 @@ -146,6 +169,15 @@ extern bool have_popcnt; #define TCG_TARGET_HAS_mulsh_i64 0 #endif +#define TCG_TARGET_HAS_v64 have_sse2 +#define TCG_TARGET_HAS_v128 have_sse2 +#define TCG_TARGET_HAS_v256 have_avx2 + +#define TCG_TARGET_HAS_andc_vec 1 +#define TCG_TARGET_HAS_orc_vec 0 +#define TCG_TARGET_HAS_not_vec 0 +#define TCG_TARGET_HAS_neg_vec 0 + #define TCG_TARGET_deposit_i32_valid(ofs, len) \ (((ofs) == 0 && (len) == 8) || ((ofs) == 8 && (len) == 8) || \ ((ofs) == 0 && (len) == 16)) diff --git a/tcg/i386/tcg-target.inc.c b/tcg/i386/tcg-target.inc.c index 69e49c9f58..df3be932d5 100644 --- a/tcg/i386/tcg-target.inc.c +++ b/tcg/i386/tcg-target.inc.c @@ -28,10 +28,11 @@ static const char * const tcg_target_reg_names[TCG_TARGET_NB_REGS] = { #if TCG_TARGET_REG_BITS == 64 "%rax", "%rcx", "%rdx", "%rbx", "%rsp", "%rbp", "%rsi", "%rdi", - "%r8", "%r9", "%r10", "%r11", "%r12", "%r13", "%r14", "%r15", #else "%eax", "%ecx", "%edx", "%ebx", "%esp", "%ebp", "%esi", "%edi", #endif + "%r8", "%r9", "%r10", "%r11", "%r12", "%r13", "%r14", "%r15", + "%xmm0", "%xmm1", "%xmm2", "%xmm3", "%xmm4", "%xmm5", "%xmm6", "%xmm7", }; #endif @@ -61,6 +62,14 @@ static const int tcg_target_reg_alloc_order[] = { TCG_REG_EDX, TCG_REG_EAX, #endif + TCG_REG_XMM0, + TCG_REG_XMM1, + TCG_REG_XMM2, + TCG_REG_XMM3, + TCG_REG_XMM4, + TCG_REG_XMM5, + TCG_REG_XMM6, + TCG_REG_XMM7, }; static const int tcg_target_call_iarg_regs[] = { @@ -94,7 +103,7 @@ static const int tcg_target_call_oarg_regs[] = { #define TCG_CT_CONST_I32 0x400 #define TCG_CT_CONST_WSZ 0x800 -/* Registers used with L constraint, which are the first argument +/* Registers used with L constraint, which are the first argument registers on x86_64, and two random call clobbered registers on i386. */ #if TCG_TARGET_REG_BITS == 64 @@ -126,6 +135,16 @@ static bool have_cmov; bool have_bmi1; bool have_popcnt; +#ifndef have_sse2 +bool have_sse2; +#endif +#ifdef have_avx2 +#define have_avx1 have_avx2 +#else +static bool have_avx1; +bool have_avx2; +#endif + #ifdef CONFIG_CPUID_H static bool have_movbe; static bool have_bmi2; @@ -192,14 +211,17 @@ static const char *target_parse_constraint(TCGArgConstraint *ct, tcg_regset_set_reg(ct->u.regs, TCG_REG_EDI); break; case 'q': + /* A register that can be used as a byte operand. */ ct->ct |= TCG_CT_REG; ct->u.regs = TCG_TARGET_REG_BITS == 64 ? 0xffff : 0xf; break; case 'Q': + /* A register with an addressable second byte (e.g. %ah). */ ct->ct |= TCG_CT_REG; ct->u.regs = 0xf; break; case 'r': + /* A general register. */ ct->ct |= TCG_CT_REG; ct->u.regs = TCG_TARGET_REG_BITS == 64 ? 0xffff : 0xff; break; @@ -207,6 +229,11 @@ static const char *target_parse_constraint(TCGArgConstraint *ct, /* With TZCNT/LZCNT, we can have operand-size as an input. */ ct->ct |= TCG_CT_CONST_WSZ; break; + case 'x': + /* A vector register. */ + ct->ct |= TCG_CT_REG; + ct->u.regs = 0xff0000; + break; /* qemu_ld/st address constraint */ case 'L': @@ -277,8 +304,9 @@ static inline int tcg_target_const_match(tcg_target_long val, TCGType type, # define P_REXB_RM 0 # define P_GS 0 #endif -#define P_SIMDF3 0x10000 /* 0xf3 opcode prefix */ -#define P_SIMDF2 0x20000 /* 0xf2 opcode prefix */ +#define P_SIMDF3 0x20000 /* 0xf3 opcode prefix */ +#define P_SIMDF2 0x40000 /* 0xf2 opcode prefix */ +#define P_VEXL 0x80000 /* Set VEX.L = 1 */ #define OPC_ARITH_EvIz (0x81) #define OPC_ARITH_EvIb (0x83) @@ -310,11 +338,30 @@ static inline int tcg_target_const_match(tcg_target_long val, TCGType type, #define OPC_MOVL_Iv (0xb8) #define OPC_MOVBE_GyMy (0xf0 | P_EXT38) #define OPC_MOVBE_MyGy (0xf1 | P_EXT38) +#define OPC_MOVDQA_GyMy (0x6f | P_EXT | P_DATA16) +#define OPC_MOVDQA_MyGy (0x7f | P_EXT | P_DATA16) +#define OPC_MOVDQU_GyMy (0x6f | P_EXT | P_SIMDF3) +#define OPC_MOVDQU_MyGy (0x7f | P_EXT | P_SIMDF3) +#define OPC_MOVQ_GyMy (0x7e | P_EXT | P_SIMDF3) +#define OPC_MOVQ_MyGy (0xd6 | P_EXT | P_DATA16) #define OPC_MOVSBL (0xbe | P_EXT) #define OPC_MOVSWL (0xbf | P_EXT) #define OPC_MOVSLQ (0x63 | P_REXW) #define OPC_MOVZBL (0xb6 | P_EXT) #define OPC_MOVZWL (0xb7 | P_EXT) +#define OPC_PADDB (0xfc | P_EXT | P_DATA16) +#define OPC_PADDW (0xfd | P_EXT | P_DATA16) +#define OPC_PADDD (0xfe | P_EXT | P_DATA16) +#define OPC_PADDQ (0xd4 | P_EXT | P_DATA16) +#define OPC_PAND (0xdb | P_EXT | P_DATA16) +#define OPC_PANDN (0xdf | P_EXT | P_DATA16) +#define OPC_PCMPEQB (0x74 | P_EXT | P_DATA16) +#define OPC_POR (0xeb | P_EXT | P_DATA16) +#define OPC_PSUBB (0xf8 | P_EXT | P_DATA16) +#define OPC_PSUBW (0xf9 | P_EXT | P_DATA16) +#define OPC_PSUBD (0xfa | P_EXT | P_DATA16) +#define OPC_PSUBQ (0xfb | P_EXT | P_DATA16) +#define OPC_PXOR (0xef | P_EXT | P_DATA16) #define OPC_POP_r32 (0x58) #define OPC_POPCNT (0xb8 | P_EXT | P_SIMDF3) #define OPC_PUSH_r32 (0x50) @@ -330,6 +377,7 @@ static inline int tcg_target_const_match(tcg_target_long val, TCGType type, #define OPC_SHRX (0xf7 | P_EXT38 | P_SIMDF2) #define OPC_TESTL (0x85) #define OPC_TZCNT (0xbc | P_EXT | P_SIMDF3) +#define OPC_VZEROUPPER (0x77 | P_EXT) #define OPC_XCHG_ax_r32 (0x90) #define OPC_GRP3_Ev (0xf7) @@ -479,11 +527,20 @@ static void tcg_out_modrm(TCGContext *s, int opc, int r, int rm) tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm)); } -static void tcg_out_vex_modrm(TCGContext *s, int opc, int r, int v, int rm) +static void tcg_out_vex_opc(TCGContext *s, int opc, int r, int v, + int rm, int index) { int tmp; - if ((opc & (P_REXW | P_EXT | P_EXT38)) || (rm & 8)) { + /* Use the two byte form if possible, which cannot encode + VEX.W, VEX.B, VEX.X, or an m-mmmm field other than P_EXT. */ + if ((opc & (P_EXT | P_EXT38 | P_REXW)) == P_EXT + && ((rm | index) & 8) == 0) { + /* Two byte VEX prefix. */ + tcg_out8(s, 0xc5); + + tmp = (r & 8 ? 0 : 0x80); /* VEX.R */ + } else { /* Three byte VEX prefix. */ tcg_out8(s, 0xc4); @@ -493,20 +550,17 @@ static void tcg_out_vex_modrm(TCGContext *s, int opc, int r, int v, int rm) } else if (opc & P_EXT) { tmp = 1; } else { - tcg_abort(); + g_assert_not_reached(); } - tmp |= 0x40; /* VEX.X */ - tmp |= (r & 8 ? 0 : 0x80); /* VEX.R */ - tmp |= (rm & 8 ? 0 : 0x20); /* VEX.B */ + tmp |= (r & 8 ? 0 : 0x80); /* VEX.R */ + tmp |= (index & 8 ? 0 : 0x40); /* VEX.X */ + tmp |= (rm & 8 ? 0 : 0x20); /* VEX.B */ tcg_out8(s, tmp); - tmp = (opc & P_REXW ? 0x80 : 0); /* VEX.W */ - } else { - /* Two byte VEX prefix. */ - tcg_out8(s, 0xc5); - - tmp = (r & 8 ? 0 : 0x80); /* VEX.R */ + tmp = (opc & P_REXW ? 0x80 : 0); /* VEX.W */ } + + tmp |= (opc & P_VEXL ? 0x04 : 0); /* VEX.L */ /* VEX.pp */ if (opc & P_DATA16) { tmp |= 1; /* 0x66 */ @@ -518,6 +572,11 @@ static void tcg_out_vex_modrm(TCGContext *s, int opc, int r, int v, int rm) tmp |= (~v & 15) << 3; /* VEX.vvvv */ tcg_out8(s, tmp); tcg_out8(s, opc); +} + +static void tcg_out_vex_modrm(TCGContext *s, int opc, int r, int v, int rm) +{ + tcg_out_vex_opc(s, opc, r, v, rm, 0); tcg_out8(s, 0xc0 | (LOWREGMASK(r) << 3) | LOWREGMASK(rm)); } @@ -526,8 +585,8 @@ static void tcg_out_vex_modrm(TCGContext *s, int opc, int r, int v, int rm) mode for absolute addresses, ~RM is the size of the immediate operand that will follow the instruction. */ -static void tcg_out_modrm_sib_offset(TCGContext *s, int opc, int r, int rm, - int index, int shift, intptr_t offset) +static void tcg_out_sib_offset(TCGContext *s, int r, int rm, int index, + int shift, intptr_t offset) { int mod, len; @@ -538,7 +597,6 @@ static void tcg_out_modrm_sib_offset(TCGContext *s, int opc, int r, int rm, intptr_t pc = (intptr_t)s->code_ptr + 5 + ~rm; intptr_t disp = offset - pc; if (disp == (int32_t)disp) { - tcg_out_opc(s, opc, r, 0, 0); tcg_out8(s, (LOWREGMASK(r) << 3) | 5); tcg_out32(s, disp); return; @@ -548,7 +606,6 @@ static void tcg_out_modrm_sib_offset(TCGContext *s, int opc, int r, int rm, use of the MODRM+SIB encoding and is therefore larger than rip-relative addressing. */ if (offset == (int32_t)offset) { - tcg_out_opc(s, opc, r, 0, 0); tcg_out8(s, (LOWREGMASK(r) << 3) | 4); tcg_out8(s, (4 << 3) | 5); tcg_out32(s, offset); @@ -556,10 +613,9 @@ static void tcg_out_modrm_sib_offset(TCGContext *s, int opc, int r, int rm, } /* ??? The memory isn't directly addressable. */ - tcg_abort(); + g_assert_not_reached(); } else { /* Absolute address. */ - tcg_out_opc(s, opc, r, 0, 0); tcg_out8(s, (r << 3) | 5); tcg_out32(s, offset); return; @@ -582,7 +638,6 @@ static void tcg_out_modrm_sib_offset(TCGContext *s, int opc, int r, int rm, that would be used for %esp is the escape to the two byte form. */ if (index < 0 && LOWREGMASK(rm) != TCG_REG_ESP) { /* Single byte MODRM format. */ - tcg_out_opc(s, opc, r, rm, 0); tcg_out8(s, mod | (LOWREGMASK(r) << 3) | LOWREGMASK(rm)); } else { /* Two byte MODRM+SIB format. */ @@ -596,7 +651,6 @@ static void tcg_out_modrm_sib_offset(TCGContext *s, int opc, int r, int rm, tcg_debug_assert(index != TCG_REG_ESP); } - tcg_out_opc(s, opc, r, rm, index); tcg_out8(s, mod | (LOWREGMASK(r) << 3) | 4); tcg_out8(s, (shift << 6) | (LOWREGMASK(index) << 3) | LOWREGMASK(rm)); } @@ -608,6 +662,21 @@ static void tcg_out_modrm_sib_offset(TCGContext *s, int opc, int r, int rm, } } +static void tcg_out_modrm_sib_offset(TCGContext *s, int opc, int r, int rm, + int index, int shift, intptr_t offset) +{ + tcg_out_opc(s, opc, r, rm < 0 ? 0 : rm, index < 0 ? 0 : index); + tcg_out_sib_offset(s, r, rm, index, shift, offset); +} + +static void tcg_out_vex_modrm_sib_offset(TCGContext *s, int opc, int r, int v, + int rm, int index, int shift, + intptr_t offset) +{ + tcg_out_vex_opc(s, opc, r, v, rm < 0 ? 0 : rm, index < 0 ? 0 : index); + tcg_out_sib_offset(s, r, rm, index, shift, offset); +} + /* A simplification of the above with no index or shift. */ static inline void tcg_out_modrm_offset(TCGContext *s, int opc, int r, int rm, intptr_t offset) @@ -615,6 +684,31 @@ static inline void tcg_out_modrm_offset(TCGContext *s, int opc, int r, tcg_out_modrm_sib_offset(s, opc, r, rm, -1, 0, offset); } +static inline void tcg_out_vex_modrm_offset(TCGContext *s, int opc, int r, + int v, int rm, intptr_t offset) +{ + tcg_out_vex_modrm_sib_offset(s, opc, r, v, rm, -1, 0, offset); +} + +static void tcg_out_maybe_vex_modrm(TCGContext *s, int opc, int r, int rm) +{ + if (have_avx1) { + tcg_out_vex_modrm(s, opc, r, 0, rm); + } else { + tcg_out_modrm(s, opc, r, rm); + } +} + +static void tcg_out_maybe_vex_modrm_offset(TCGContext *s, int opc, int r, + int rm, intptr_t offset) +{ + if (have_avx1) { + tcg_out_vex_modrm_offset(s, opc, r, 0, rm, offset); + } else { + tcg_out_modrm_offset(s, opc, r, rm, offset); + } +} + /* Generate dest op= src. Uses the same ARITH_* codes as tgen_arithi. */ static inline void tgen_arithr(TCGContext *s, int subop, int dest, int src) { @@ -625,12 +719,34 @@ static inline void tgen_arithr(TCGContext *s, int subop, int dest, int src) tcg_out_modrm(s, OPC_ARITH_GvEv + (subop << 3) + ext, dest, src); } -static inline void tcg_out_mov(TCGContext *s, TCGType type, - TCGReg ret, TCGReg arg) +static void tcg_out_mov(TCGContext *s, TCGType type, TCGReg ret, TCGReg arg) { - if (arg != ret) { - int opc = OPC_MOVL_GvEv + (type == TCG_TYPE_I64 ? P_REXW : 0); - tcg_out_modrm(s, opc, ret, arg); + if (arg == ret) { + return; + } + switch (type) { + case TCG_TYPE_I32: + tcg_debug_assert(ret < 16 && arg < 16); + tcg_out_modrm(s, OPC_MOVL_GvEv, ret, arg); + break; + case TCG_TYPE_I64: + tcg_debug_assert(ret < 16 && arg < 16); + tcg_out_modrm(s, OPC_MOVL_GvEv | P_REXW, ret, arg); + break; + case TCG_TYPE_V64: + tcg_debug_assert(ret >= 16 && arg >= 16); + tcg_out_maybe_vex_modrm(s, OPC_MOVQ_GyMy, ret, arg); + break; + case TCG_TYPE_V128: + tcg_debug_assert(ret >= 16 && arg >= 16); + tcg_out_maybe_vex_modrm(s, OPC_MOVDQA_GyMy, ret, arg); + break; + case TCG_TYPE_V256: + tcg_debug_assert(ret >= 16 && arg >= 16); + tcg_out_vex_modrm(s, OPC_MOVDQA_GyMy | P_VEXL, ret, 0, arg); + break; + default: + g_assert_not_reached(); } } @@ -638,6 +754,36 @@ static void tcg_out_movi(TCGContext *s, TCGType type, TCGReg ret, tcg_target_long arg) { tcg_target_long diff; + int opc; + + switch (type) { + case TCG_TYPE_I32: + case TCG_TYPE_I64: + tcg_debug_assert(ret < 16); + break; + + case TCG_TYPE_V64: + case TCG_TYPE_V128: + case TCG_TYPE_V256: + tcg_debug_assert(ret >= 16); + /* ??? Revisit this as the implementation progresses. */ + if (arg == 0) { + opc = OPC_PXOR; + } else if (arg == -1) { + opc = OPC_PCMPEQB; + } else { + g_assert_not_reached(); + } + if (have_avx1) { + tcg_out_vex_modrm(s, opc, ret, ret, ret); + } else { + tcg_out_modrm(s, opc, ret, ret); + } + return; + + default: + g_assert_not_reached(); + } if (arg == 0) { tgen_arithr(s, ARITH_XOR, ret, ret); @@ -702,18 +848,64 @@ static inline void tcg_out_pop(TCGContext *s, int reg) tcg_out_opc(s, OPC_POP_r32 + LOWREGMASK(reg), 0, reg, 0); } -static inline void tcg_out_ld(TCGContext *s, TCGType type, TCGReg ret, - TCGReg arg1, intptr_t arg2) +static void tcg_out_ld(TCGContext *s, TCGType type, TCGReg ret, + TCGReg arg1, intptr_t arg2) { - int opc = OPC_MOVL_GvEv + (type == TCG_TYPE_I64 ? P_REXW : 0); - tcg_out_modrm_offset(s, opc, ret, arg1, arg2); + switch (type) { + case TCG_TYPE_I64: + tcg_debug_assert(ret < 16); + tcg_out_modrm_offset(s, OPC_MOVL_GvEv | P_REXW, ret, arg1, arg2); + break; + case TCG_TYPE_I32: + tcg_debug_assert(ret < 16); + tcg_out_modrm_offset(s, OPC_MOVL_GvEv, ret, arg1, arg2); + break; + case TCG_TYPE_V64: + tcg_debug_assert(ret >= 16); + tcg_out_maybe_vex_modrm_offset(s, OPC_MOVQ_GyMy, ret, arg1, arg2); + break; + case TCG_TYPE_V128: + tcg_debug_assert(ret >= 16); + tcg_out_maybe_vex_modrm_offset(s, OPC_MOVDQU_GyMy, ret, arg1, arg2); + break; + case TCG_TYPE_V256: + tcg_debug_assert(ret >= 16); + tcg_out_vex_modrm_offset(s, OPC_MOVDQU_GyMy | P_VEXL, + ret, 0, arg1, arg2); + break; + default: + g_assert_not_reached(); + } } -static inline void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg, - TCGReg arg1, intptr_t arg2) +static void tcg_out_st(TCGContext *s, TCGType type, TCGReg arg, + TCGReg arg1, intptr_t arg2) { - int opc = OPC_MOVL_EvGv + (type == TCG_TYPE_I64 ? P_REXW : 0); - tcg_out_modrm_offset(s, opc, arg, arg1, arg2); + switch (type) { + case TCG_TYPE_I64: + tcg_debug_assert(arg < 16); + tcg_out_modrm_offset(s, OPC_MOVL_EvGv | P_REXW, arg, arg1, arg2); + break; + case TCG_TYPE_I32: + tcg_debug_assert(arg < 16); + tcg_out_modrm_offset(s, OPC_MOVL_EvGv, arg, arg1, arg2); + break; + case TCG_TYPE_V64: + tcg_debug_assert(arg >= 16); + tcg_out_maybe_vex_modrm_offset(s, OPC_MOVQ_MyGy, arg, arg1, arg2); + break; + case TCG_TYPE_V128: + tcg_debug_assert(arg >= 16); + tcg_out_maybe_vex_modrm_offset(s, OPC_MOVDQU_MyGy, arg, arg1, arg2); + break; + case TCG_TYPE_V256: + tcg_debug_assert(arg >= 16); + tcg_out_vex_modrm_offset(s, OPC_MOVDQU_MyGy | P_VEXL, + arg, 0, arg1, arg2); + break; + default: + g_assert_not_reached(); + } } static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val, @@ -725,6 +917,8 @@ static bool tcg_out_sti(TCGContext *s, TCGType type, TCGArg val, return false; } rexw = P_REXW; + } else if (type != TCG_TYPE_I32) { + return false; } tcg_out_modrm_offset(s, OPC_MOVL_EvIz | rexw, 0, base, ofs); tcg_out32(s, val); @@ -2254,19 +2448,110 @@ static inline void tcg_out_op(TCGContext *s, TCGOpcode opc, } break; + case INDEX_op_add8_vec: + c = OPC_PADDB; + goto gen_simd; + case INDEX_op_add16_vec: + c = OPC_PADDW; + goto gen_simd; + case INDEX_op_add32_vec: + c = OPC_PADDD; + goto gen_simd; + case INDEX_op_add64_vec: + c = OPC_PADDQ; + goto gen_simd; + case INDEX_op_sub8_vec: + c = OPC_PSUBB; + goto gen_simd; + case INDEX_op_sub16_vec: + c = OPC_PSUBW; + goto gen_simd; + case INDEX_op_sub32_vec: + c = OPC_PSUBD; + goto gen_simd; + case INDEX_op_sub64_vec: + c = OPC_PSUBQ; + goto gen_simd; + case INDEX_op_and_vec: + c = OPC_PAND; + goto gen_simd; + case INDEX_op_or_vec: + c = OPC_POR; + goto gen_simd; + case INDEX_op_xor_vec: + c = OPC_PXOR; + gen_simd: + if (args[3] == 2) { + c |= P_VEXL; + } + if (have_avx1) { + tcg_out_vex_modrm(s, c, a0, a1, a2); + } else { + tcg_out_modrm(s, c, a0, a2); + } + break; + case INDEX_op_andc_vec: + c = OPC_PANDN; + if (args[3] == 2) { + c |= P_VEXL; + } + if (have_avx1) { + tcg_out_vex_modrm(s, c, a0, a2, a1); + } else { + tcg_out_modrm(s, c, a0, a1); + } + break; + + case INDEX_op_ld_vec: + case INDEX_op_ldz_vec: + switch (args[3]) { + case 0: + tcg_out_ld(s, TCG_TYPE_V64, a0, a1, a2); + break; + case 1: + tcg_out_ld(s, TCG_TYPE_V128, a0, a1, a2); + break; + case 2: + tcg_out_ld(s, TCG_TYPE_V256, a0, a1, a2); + break; + default: + g_assert_not_reached(); + } + break; + + case INDEX_op_st_vec: + switch (args[3]) { + case 0: + tcg_out_st(s, TCG_TYPE_V64, a0, a1, a2); + break; + case 1: + tcg_out_st(s, TCG_TYPE_V128, a0, a1, a2); + break; + case 2: + tcg_out_st(s, TCG_TYPE_V256, a0, a1, a2); + break; + default: + g_assert_not_reached(); + } + break; + case INDEX_op_mb: tcg_out_mb(s, a0); break; case INDEX_op_mov_i32: /* Always emitted via tcg_out_mov. */ case INDEX_op_mov_i64: + case INDEX_op_mov_vec: case INDEX_op_movi_i32: /* Always emitted via tcg_out_movi. */ case INDEX_op_movi_i64: + case INDEX_op_movi_vec: case INDEX_op_call: /* Always emitted via tcg_out_call. */ default: tcg_abort(); } #undef OP_32_64 +#undef OP_128_256 +#undef OP_64_128_256 } static const TCGTargetOpDef *tcg_target_op_def(TCGOpcode op) @@ -2292,6 +2577,10 @@ static const TCGTargetOpDef *tcg_target_op_def(TCGOpcode op) = { .args_ct_str = { "r", "r", "L", "L" } }; static const TCGTargetOpDef L_L_L_L = { .args_ct_str = { "L", "L", "L", "L" } }; + static const TCGTargetOpDef x_0_x = { .args_ct_str = { "x", "0", "x" } }; + static const TCGTargetOpDef x_x_0 = { .args_ct_str = { "x", "x", "0" } }; + static const TCGTargetOpDef x_x_x = { .args_ct_str = { "x", "x", "x" } }; + static const TCGTargetOpDef x_r = { .args_ct_str = { "x", "r" } }; switch (op) { case INDEX_op_goto_ptr: @@ -2493,6 +2782,26 @@ static const TCGTargetOpDef *tcg_target_op_def(TCGOpcode op) return &s2; } + case INDEX_op_ld_vec: + case INDEX_op_ldz_vec: + case INDEX_op_st_vec: + return &x_r; + + case INDEX_op_add8_vec: + case INDEX_op_add16_vec: + case INDEX_op_add32_vec: + case INDEX_op_add64_vec: + case INDEX_op_sub8_vec: + case INDEX_op_sub16_vec: + case INDEX_op_sub32_vec: + case INDEX_op_sub64_vec: + case INDEX_op_and_vec: + case INDEX_op_or_vec: + case INDEX_op_xor_vec: + return have_avx1 ? &x_x_x : &x_0_x; + case INDEX_op_andc_vec: + return have_avx1 ? &x_x_x : &x_x_0; + default: break; } @@ -2577,6 +2886,9 @@ static void tcg_target_qemu_prologue(TCGContext *s) tcg_out_addi(s, TCG_REG_CALL_STACK, stack_addend); + if (have_avx2) { + tcg_out_vex_opc(s, OPC_VZEROUPPER, 0, 0, 0, 0); + } for (i = ARRAY_SIZE(tcg_target_callee_save_regs) - 1; i >= 0; i--) { tcg_out_pop(s, tcg_target_callee_save_regs[i]); } @@ -2598,9 +2910,16 @@ static void tcg_out_nop_fill(tcg_insn_unit *p, int count) static void tcg_target_init(TCGContext *s) { #ifdef CONFIG_CPUID_H - unsigned a, b, c, d; + unsigned a, b, c, d, b7 = 0; int max = __get_cpuid_max(0, 0); + if (max >= 7) { + /* BMI1 is available on AMD Piledriver and Intel Haswell CPUs. */ + __cpuid_count(7, 0, a, b7, c, d); + have_bmi1 = (b7 & bit_BMI) != 0; + have_bmi2 = (b7 & bit_BMI2) != 0; + } + if (max >= 1) { __cpuid(1, a, b, c, d); #ifndef have_cmov @@ -2609,17 +2928,26 @@ static void tcg_target_init(TCGContext *s) available, we'll use a small forward branch. */ have_cmov = (d & bit_CMOV) != 0; #endif +#ifndef have_sse2 + have_sse2 = (d & bit_SSE2) != 0; +#endif /* MOVBE is only available on Intel Atom and Haswell CPUs, so we need to probe for it. */ have_movbe = (c & bit_MOVBE) != 0; have_popcnt = (c & bit_POPCNT) != 0; - } - if (max >= 7) { - /* BMI1 is available on AMD Piledriver and Intel Haswell CPUs. */ - __cpuid_count(7, 0, a, b, c, d); - have_bmi1 = (b & bit_BMI) != 0; - have_bmi2 = (b & bit_BMI2) != 0; +#ifndef have_avx2 + /* There are a number of things we must check before we can be + sure of not hitting invalid opcode. */ + if (c & bit_OSXSAVE) { + unsigned xcrl, xcrh; + asm ("xgetbv" : "=a" (xcrl), "=d" (xcrh) : "c" (0)); + if ((xcrl & 6) == 6) { + have_avx1 = (c & bit_AVX) != 0; + have_avx2 = (b7 & bit_AVX2) != 0; + } + } +#endif } max = __get_cpuid_max(0x8000000, 0); @@ -2636,6 +2964,13 @@ static void tcg_target_init(TCGContext *s) } else { tcg_target_available_regs[TCG_TYPE_I32] = 0xff; } + if (have_sse2) { + tcg_target_available_regs[TCG_TYPE_V64] = 0xff0000; + tcg_target_available_regs[TCG_TYPE_V128] = 0xff0000; + } + if (have_avx2) { + tcg_target_available_regs[TCG_TYPE_V256] = 0xff0000; + } tcg_target_call_clobber_regs = 0; tcg_regset_set_reg(tcg_target_call_clobber_regs, TCG_REG_EAX);