Hexagon (target/hexagon) update

-----BEGIN PGP SIGNATURE-----
 
 iQEzBAABCgAdFiEENjXHiM5iuR/UxZq0ewJE+xLeRCIFAmQGzGYACgkQewJE+xLe
 RCJfxggAgAY9fqUxKk5FxziMX6sVxx2SiYGU1biABtHyduQvJ43U7In4cnbC65FL
 Pn+ZSixPRh51Ij2s9FTzDQ8IUNt/k1W7gCoEC3I0wDbQHdICEIlx30ry68QlUgOa
 T6ZDquJkanwFjb7kMnyM4LRBBxLG+OuYvKe1e+nzun0LDlfTS/sHitRlf4AJEBOK
 9h/Bdy81RcWlLzlc5tmD0f9rhtmkkFCTu/TGLt4G6sfn1xZbRdh5N1cFfUShPlQM
 qAgj+JgBOQoPKaBhQZnA3Ps9ZEM4+/8KPsr5oGweAcjqD8+kMGCmi2jv+60ES5Uq
 EDpn25COw7BVAo6vP/JfBEiZTg+YpA==
 =FDSG
 -----END PGP SIGNATURE-----

Merge tag 'pull-hex-20230306' of https://github.com/quic/qemu into staging

Hexagon (target/hexagon) update

# -----BEGIN PGP SIGNATURE-----
#
# iQEzBAABCgAdFiEENjXHiM5iuR/UxZq0ewJE+xLeRCIFAmQGzGYACgkQewJE+xLe
# RCJfxggAgAY9fqUxKk5FxziMX6sVxx2SiYGU1biABtHyduQvJ43U7In4cnbC65FL
# Pn+ZSixPRh51Ij2s9FTzDQ8IUNt/k1W7gCoEC3I0wDbQHdICEIlx30ry68QlUgOa
# T6ZDquJkanwFjb7kMnyM4LRBBxLG+OuYvKe1e+nzun0LDlfTS/sHitRlf4AJEBOK
# 9h/Bdy81RcWlLzlc5tmD0f9rhtmkkFCTu/TGLt4G6sfn1xZbRdh5N1cFfUShPlQM
# qAgj+JgBOQoPKaBhQZnA3Ps9ZEM4+/8KPsr5oGweAcjqD8+kMGCmi2jv+60ES5Uq
# EDpn25COw7BVAo6vP/JfBEiZTg+YpA==
# =FDSG
# -----END PGP SIGNATURE-----
# gpg: Signature made Tue 07 Mar 2023 05:32:22 GMT
# gpg:                using RSA key 3635C788CE62B91FD4C59AB47B0244FB12DE4422
# gpg: Good signature from "Taylor Simpson (Rock on) <tsimpson@quicinc.com>" [undefined]
# gpg: WARNING: This key is not certified with a trusted signature!
# gpg:          There is no indication that the signature belongs to the owner.
# Primary key fingerprint: 3635 C788 CE62 B91F D4C5  9AB4 7B02 44FB 12DE 4422

* tag 'pull-hex-20230306' of https://github.com/quic/qemu:
  Hexagon (target/hexagon) Improve code gen for predicated HVX instructions
  Hexagon (target/hexagon) Reduce manipulation of slot_cancelled
  Hexagon (target/hexagon) Remove gen_log_predicated_reg_write[_pair]
  Hexagon (target/hexagon) Change subtract from zero to change sign
  Hexagon (tests/tcg/hexagon) Enable HVX tests
  Hexagon (tests/tcg/hexagon) Remove __builtin from scatter_gather
  Hexagon (tests/tcg/hexagon) Update preg_alias.c
  Hexagon (target/hexagon) Analyze packet for HVX
  Hexagon (target/hexagon) Don't set pkt_has_store_s1 when not needed
  Hexagon (target/hexagon) Analyze packet before generating TCG
  Hexagon (target/hexagon) Add overrides for dealloc-return instructions
  Hexagon (target/hexagon) Add overrides for endloop1/endloop01
  Hexagon (target/hexagon) Add overrides for callr
  Hexagon (target/hexagon) Add overrides for jumpr31 instructions
  target/hexagon/idef-parser: Remove unused code paths
  target/hexagon/idef-parser: Elide mov in gen_pred_assign
  Hexagon (target/hexagon) Restore --disable-hexagon-idef-parser build

Signed-off-by: Peter Maydell <peter.maydell@linaro.org>

target/hexagon/README

@@ -52,6 +52,7 @@ header files in <BUILD_DIR>/target/hexagon
         gen_tcg_func_table.py           -> tcg_func_table_generated.c.inc
         gen_helper_funcs.py             -> helper_funcs_generated.c.inc
         gen_idef_parser_funcs.py        -> idef_parser_input.h
+        gen_analyze_funcs.py            -> analyze_funcs_generated.c.inc
 
 Qemu helper functions have 3 parts
     DEF_HELPER declaration indicates the signature of the helper
@@ -87,7 +88,6 @@ tcg_funcs_generated.c.inc
         TCGv RtV = hex_gpr[insn->regno[2]];
         gen_helper_A2_add(RdV, cpu_env, RsV, RtV);
         gen_log_reg_write(RdN, RdV);
-        ctx_log_reg_write(ctx, RdN);
     }
 
 helper_funcs_generated.c.inc
@@ -136,12 +136,9 @@ For HVX vectors, the generator behaves slightly differently.  The wide vectors
 won't fit in a TCGv or TCGv_i64, so we pass TCGv_ptr variables to pass the
 address to helper functions.  Here's an example for an HVX vector-add-word
 istruction.
-    static void generate_V6_vaddw(
+    static void generate_V6_vaddw(DisasContext *ctx)
-                    CPUHexagonState *env,
-                    DisasContext *ctx,
-                    Insn *insn,
-                    Packet *pkt)
     {
+        Insn *insn __attribute__((unused)) = ctx->insn;
         const int VdN = insn->regno[0];
         const intptr_t VdV_off =
             ctx_future_vreg_off(ctx, VdN, 1, true);
@@ -157,10 +154,7 @@ istruction.
         TCGv_ptr VvV = tcg_temp_new_ptr();
         tcg_gen_addi_ptr(VuV, cpu_env, VuV_off);
         tcg_gen_addi_ptr(VvV, cpu_env, VvV_off);
-        TCGv slot = tcg_constant_tl(insn->slot);
+        gen_helper_V6_vaddw(cpu_env, VdV, VuV, VvV);
-        gen_helper_V6_vaddw(cpu_env, VdV, VuV, VvV, slot);
-        gen_log_vreg_write(ctx, VdV_off, VdN, EXT_DFL, insn->slot, false);
-        ctx_log_vreg_write(ctx, VdN, EXT_DFL, false);
     }
 
 Notice that we also generate a variable named <operand>_off for each operand of
@@ -173,12 +167,9 @@ functions from tcg-op-gvec.h.  Here's the override for this instruction.
 Finally, we notice that the override doesn't use the TCGv_ptr variables, so
 we don't generate them when an override is present.  Here is what we generate
 when the override is present.
-    static void generate_V6_vaddw(
+    static void generate_V6_vaddw(DisasContext *ctx)
-                    CPUHexagonState *env,
-                    DisasContext *ctx,
-                    Insn *insn,
-                    Packet *pkt)
     {
+        Insn *insn __attribute__((unused)) = ctx->insn;
         const int VdN = insn->regno[0];
         const intptr_t VdV_off =
             ctx_future_vreg_off(ctx, VdN, 1, true);
@@ -189,10 +180,14 @@ when the override is present.
         const intptr_t VvV_off =
             vreg_src_off(ctx, VvN);
         fGEN_TCG_V6_vaddw({ fHIDE(int i;) fVFOREACH(32, i) { VdV.w[i] = VuV.w[i] + VvV.w[i] ; } });
-        gen_log_vreg_write(ctx, VdV_off, VdN, EXT_DFL, insn->slot, false);
-        ctx_log_vreg_write(ctx, VdN, EXT_DFL, false);
     }
 
+We also generate an analyze_<tag> function for each instruction.  Currently,
+these functions record the writes to registers by calling ctx_log_*.  During
+gen_start_packet, we invoke the analyze_<tag> function for each instruction in
+the packet, and we mark the implicit writes.  After the analysis is performed,
+we initialize hex_new_value for each of the predicated assignments.
+
 In addition to instruction semantics, we use a generator to create the decode
 tree.  This generation is also a two step process.  The first step is to run
 target/hexagon/gen_dectree_import.c to produce
@@ -277,10 +272,8 @@ For Hexagon Vector eXtensions (HVX), the following fields are used
     VRegs                       Vector registers
     future_VRegs                Registers to be stored during packet commit
     tmp_VRegs                   Temporary registers *not* stored during commit
-    VRegs_updated               Mask of predicated vector writes
     QRegs                       Q (vector predicate) registers
     future_QRegs                Registers to be stored during packet commit
-    QRegs_updated               Mask of predicated vector writes
 
 *** Debugging ***
 

target/hexagon/attribs_def.h.inc

@@ -44,6 +44,7 @@ DEF_ATTRIB(MEMSIZE_1B, "Memory width is 1 byte", "", "")
 DEF_ATTRIB(MEMSIZE_2B, "Memory width is 2 bytes", "", "")
 DEF_ATTRIB(MEMSIZE_4B, "Memory width is 4 bytes", "", "")
 DEF_ATTRIB(MEMSIZE_8B, "Memory width is 8 bytes", "", "")
+DEF_ATTRIB(SCALAR_LOAD, "Load is scalar", "", "")
 DEF_ATTRIB(SCALAR_STORE, "Store is scalar", "", "")
 DEF_ATTRIB(REGWRSIZE_1B, "Memory width is 1 byte", "", "")
 DEF_ATTRIB(REGWRSIZE_2B, "Memory width is 2 bytes", "", "")

target/hexagon/cpu.h

@@ -1,5 +1,5 @@
 /*
- *  Copyright(c) 2019-2022 Qualcomm Innovation Center, Inc. All Rights Reserved.
+ *  Copyright(c) 2019-2023 Qualcomm Innovation Center, Inc. All Rights Reserved.
  *
  *  This program is free software; you can redistribute it and/or modify
  *  it under the terms of the GNU General Public License as published by
@@ -111,11 +111,8 @@ typedef struct CPUArchState {
     MMVector future_VRegs[VECTOR_TEMPS_MAX] QEMU_ALIGNED(16);
     MMVector tmp_VRegs[VECTOR_TEMPS_MAX] QEMU_ALIGNED(16);
 
-    VRegMask VRegs_updated;
-
     MMQReg QRegs[NUM_QREGS] QEMU_ALIGNED(16);
     MMQReg future_QRegs[NUM_QREGS] QEMU_ALIGNED(16);
-    QRegMask QRegs_updated;
 
     /* Temporaries used within instructions */
     MMVectorPair VuuV QEMU_ALIGNED(16);

target/hexagon/gen_analyze_funcs.py

@@ -0,0 +1,252 @@
+#!/usr/bin/env python3
+
+##
+##  Copyright(c) 2022-2023 Qualcomm Innovation Center, Inc. All Rights Reserved.
+##
+##  This program is free software; you can redistribute it and/or modify
+##  it under the terms of the GNU General Public License as published by
+##  the Free Software Foundation; either version 2 of the License, or
+##  (at your option) any later version.
+##
+##  This program is distributed in the hope that it will be useful,
+##  but WITHOUT ANY WARRANTY; without even the implied warranty of
+##  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+##  GNU General Public License for more details.
+##
+##  You should have received a copy of the GNU General Public License
+##  along with this program; if not, see <http://www.gnu.org/licenses/>.
+##
+
+import sys
+import re
+import string
+import hex_common
+
+##
+## Helpers for gen_analyze_func
+##
+def is_predicated(tag):
+    return 'A_CONDEXEC' in hex_common.attribdict[tag]
+
+def analyze_opn_old(f, tag, regtype, regid, regno):
+    regN = "%s%sN" % (regtype, regid)
+    predicated = "true" if is_predicated(tag) else "false"
+    if (regtype == "R"):
+        if (regid in {"ss", "tt"}):
+            f.write("//    const int %s = insn->regno[%d];\n" % \
+                (regN, regno))
+        elif (regid in {"dd", "ee", "xx", "yy"}):
+            f.write("    const int %s = insn->regno[%d];\n" % (regN, regno))
+            f.write("    ctx_log_reg_write_pair(ctx, %s, %s);\n" % \
+                (regN, predicated))
+        elif (regid in {"s", "t", "u", "v"}):
+            f.write("//    const int %s = insn->regno[%d];\n" % \
+                (regN, regno))
+        elif (regid in {"d", "e", "x", "y"}):
+            f.write("    const int %s = insn->regno[%d];\n" % (regN, regno))
+            f.write("    ctx_log_reg_write(ctx, %s, %s);\n" % \
+                (regN, predicated))
+        else:
+            print("Bad register parse: ", regtype, regid)
+    elif (regtype == "P"):
+        if (regid in {"s", "t", "u", "v"}):
+            f.write("//    const int %s = insn->regno[%d];\n" % \
+                (regN, regno))
+        elif (regid in {"d", "e", "x"}):
+            f.write("    const int %s = insn->regno[%d];\n" % (regN, regno))
+            f.write("    ctx_log_pred_write(ctx, %s);\n" % (regN))
+        else:
+            print("Bad register parse: ", regtype, regid)
+    elif (regtype == "C"):
+        if (regid == "ss"):
+            f.write("//    const int %s = insn->regno[%d] + HEX_REG_SA0;\n" % \
+                (regN, regno))
+        elif (regid == "dd"):
+            f.write("    const int %s = insn->regno[%d] + HEX_REG_SA0;\n" % \
+                (regN, regno))
+            f.write("    ctx_log_reg_write_pair(ctx, %s, %s);\n" % \
+                (regN, predicated))
+        elif (regid == "s"):
+            f.write("//    const int %s = insn->regno[%d] + HEX_REG_SA0;\n" % \
+                (regN, regno))
+        elif (regid == "d"):
+            f.write("    const int %s = insn->regno[%d] + HEX_REG_SA0;\n" % \
+                (regN, regno))
+            f.write("    ctx_log_reg_write(ctx, %s, %s);\n" % \
+                (regN, predicated))
+        else:
+            print("Bad register parse: ", regtype, regid)
+    elif (regtype == "M"):
+        if (regid == "u"):
+            f.write("//    const int %s = insn->regno[%d];\n"% \
+                (regN, regno))
+        else:
+            print("Bad register parse: ", regtype, regid)
+    elif (regtype == "V"):
+        newv = "EXT_DFL"
+        if (hex_common.is_new_result(tag)):
+            newv = "EXT_NEW"
+        elif (hex_common.is_tmp_result(tag)):
+            newv = "EXT_TMP"
+        if (regid in {"dd", "xx"}):
+            f.write("    const int %s = insn->regno[%d];\n" %\
+                (regN, regno))
+            f.write("    ctx_log_vreg_write_pair(ctx, %s, %s, %s);\n" % \
+                (regN, newv, predicated))
+        elif (regid in {"uu", "vv"}):
+            f.write("//    const int %s = insn->regno[%d];\n" % \
+                (regN, regno))
+        elif (regid in {"s", "u", "v", "w"}):
+            f.write("//    const int %s = insn->regno[%d];\n" % \
+                (regN, regno))
+        elif (regid in {"d", "x", "y"}):
+            f.write("    const int %s = insn->regno[%d];\n" % \
+                (regN, regno))
+            f.write("    ctx_log_vreg_write(ctx, %s, %s, %s);\n" % \
+                (regN, newv, predicated))
+        else:
+            print("Bad register parse: ", regtype, regid)
+    elif (regtype == "Q"):
+        if (regid in {"d", "e", "x"}):
+            f.write("    const int %s = insn->regno[%d];\n" % \
+                (regN, regno))
+            f.write("    ctx_log_qreg_write(ctx, %s);\n" % (regN))
+        elif (regid in {"s", "t", "u", "v"}):
+            f.write("//    const int %s = insn->regno[%d];\n" % \
+                (regN, regno))
+        else:
+            print("Bad register parse: ", regtype, regid)
+    elif (regtype == "G"):
+        if (regid in {"dd"}):
+            f.write("//    const int %s = insn->regno[%d];\n" % \
+                (regN, regno))
+        elif (regid in {"d"}):
+            f.write("//    const int %s = insn->regno[%d];\n" % \
+                (regN, regno))
+        elif (regid in {"ss"}):
+            f.write("//    const int %s = insn->regno[%d];\n" % \
+                (regN, regno))
+        elif (regid in {"s"}):
+            f.write("//    const int %s = insn->regno[%d];\n" % \
+                (regN, regno))
+        else:
+            print("Bad register parse: ", regtype, regid)
+    elif (regtype == "S"):
+        if (regid in {"dd"}):
+            f.write("//    const int %s = insn->regno[%d];\n" % \
+                (regN, regno))
+        elif (regid in {"d"}):
+            f.write("//    const int %s = insn->regno[%d];\n" % \
+                (regN, regno))
+        elif (regid in {"ss"}):
+            f.write("//    const int %s = insn->regno[%d];\n" % \
+                (regN, regno))
+        elif (regid in {"s"}):
+            f.write("//    const int %s = insn->regno[%d];\n" % \
+                (regN, regno))
+        else:
+            print("Bad register parse: ", regtype, regid)
+    else:
+        print("Bad register parse: ", regtype, regid)
+
+def analyze_opn_new(f, tag, regtype, regid, regno):
+    regN = "%s%sN" % (regtype, regid)
+    if (regtype == "N"):
+        if (regid in {"s", "t"}):
+            f.write("//    const int %s = insn->regno[%d];\n" % \
+                (regN, regno))
+        else:
+            print("Bad register parse: ", regtype, regid)
+    elif (regtype == "P"):
+        if (regid in {"t", "u", "v"}):
+            f.write("//    const int %s = insn->regno[%d];\n" % \
+                (regN, regno))
+        else:
+            print("Bad register parse: ", regtype, regid)
+    elif (regtype == "O"):
+        if (regid == "s"):
+            f.write("//    const int %s = insn->regno[%d];\n" % \
+                (regN, regno))
+        else:
+            print("Bad register parse: ", regtype, regid)
+    else:
+        print("Bad register parse: ", regtype, regid)
+
+def analyze_opn(f, tag, regtype, regid, toss, numregs, i):
+    if (hex_common.is_pair(regid)):
+        analyze_opn_old(f, tag, regtype, regid, i)
+    elif (hex_common.is_single(regid)):
+        if hex_common.is_old_val(regtype, regid, tag):
+            analyze_opn_old(f,tag, regtype, regid, i)
+        elif hex_common.is_new_val(regtype, regid, tag):
+            analyze_opn_new(f, tag, regtype, regid, i)
+        else:
+            print("Bad register parse: ", regtype, regid, toss, numregs)
+    else:
+        print("Bad register parse: ", regtype, regid, toss, numregs)
+
+##
+## Generate the code to analyze the instruction
+##     For A2_add: Rd32=add(Rs32,Rt32), { RdV=RsV+RtV;}
+##     We produce:
+##     static void analyze_A2_add(DisasContext *ctx)
+##     {
+##         Insn *insn G_GNUC_UNUSED = ctx->insn;
+##         const int RdN = insn->regno[0];
+##         ctx_log_reg_write(ctx, RdN, false);
+##     //    const int RsN = insn->regno[1];
+##     //    const int RtN = insn->regno[2];
+##     }
+##
+def gen_analyze_func(f, tag, regs, imms):
+    f.write("static void analyze_%s(DisasContext *ctx)\n" %tag)
+    f.write('{\n')
+
+    f.write("    Insn *insn G_GNUC_UNUSED = ctx->insn;\n")
+
+    i=0
+    ## Analyze all the registers
+    for regtype, regid, toss, numregs in regs:
+        analyze_opn(f, tag, regtype, regid, toss, numregs, i)
+        i += 1
+
+    has_generated_helper = (not hex_common.skip_qemu_helper(tag) and
+                            not hex_common.is_idef_parser_enabled(tag))
+    if (has_generated_helper and
+        'A_SCALAR_LOAD' in hex_common.attribdict[tag]):
+        f.write("    ctx->need_pkt_has_store_s1 = true;\n")
+
+    f.write("}\n\n")
+
+def main():
+    hex_common.read_semantics_file(sys.argv[1])
+    hex_common.read_attribs_file(sys.argv[2])
+    hex_common.read_overrides_file(sys.argv[3])
+    hex_common.read_overrides_file(sys.argv[4])
+    ## Whether or not idef-parser is enabled is
+    ## determined by the number of arguments to
+    ## this script:
+    ##
+    ##   5 args. -> not enabled,
+    ##   6 args. -> idef-parser enabled.
+    ##
+    ## The 6:th arg. then holds a list of the successfully
+    ## parsed instructions.
+    is_idef_parser_enabled = len(sys.argv) > 6
+    if is_idef_parser_enabled:
+        hex_common.read_idef_parser_enabled_file(sys.argv[5])
+    hex_common.calculate_attribs()
+    tagregs = hex_common.get_tagregs()
+    tagimms = hex_common.get_tagimms()
+
+    with open(sys.argv[-1], 'w') as f:
+        f.write("#ifndef HEXAGON_TCG_FUNCS_H\n")
+        f.write("#define HEXAGON_TCG_FUNCS_H\n\n")
+
+        for tag in hex_common.tags:
+            gen_analyze_func(f, tag, tagregs[tag], tagimms[tag])
+
+        f.write("#endif    /* HEXAGON_TCG_FUNCS_H */\n")
+
+if __name__ == "__main__":
+    main()

target/hexagon/gen_helper_funcs.py

@@ -1,7 +1,7 @@
 #!/usr/bin/env python3
 
 ##
-##  Copyright(c) 2019-2022 Qualcomm Innovation Center, Inc. All Rights Reserved.
+##  Copyright(c) 2019-2023 Qualcomm Innovation Center, Inc. All Rights Reserved.
 ##
 ##  This program is free software; you can redistribute it and/or modify
 ##  it under the terms of the GNU General Public License as published by
@@ -226,6 +226,14 @@ def gen_helper_function(f, tag, tagregs, tagimms):
                     print("Bad register parse: ",regtype,regid,toss,numregs)
                 i += 1
 
+        ## For conditional instructions, we pass in the destination register
+        if 'A_CONDEXEC' in hex_common.attribdict[tag]:
+            for regtype, regid, toss, numregs in regs:
+                if (hex_common.is_writeonly(regid) and
+                    not hex_common.is_hvx_reg(regtype)):
+                    gen_helper_arg_opn(f, regtype, regid, i, tag)
+                    i += 1
+
         ## Arguments to the helper function are the source regs and immediates
         for regtype,regid,toss,numregs in regs:
             if (hex_common.is_read(regid)):
@@ -262,10 +270,11 @@ def gen_helper_function(f, tag, tagregs, tagimms):
         if hex_common.need_ea(tag): gen_decl_ea(f)
         ## Declare the return variable
         i=0
-        for regtype,regid,toss,numregs in regs:
+        if 'A_CONDEXEC' not in hex_common.attribdict[tag]:
-            if (hex_common.is_writeonly(regid)):
+            for regtype,regid,toss,numregs in regs:
-                gen_helper_dest_decl_opn(f,regtype,regid,i)
+                if (hex_common.is_writeonly(regid)):
-            i += 1
+                    gen_helper_dest_decl_opn(f,regtype,regid,i)
+                i += 1
 
         for regtype,regid,toss,numregs in regs:
             if (hex_common.is_read(regid)):

target/hexagon/gen_helper_protos.py

@@ -1,7 +1,7 @@
 #!/usr/bin/env python3
 
 ##
-##  Copyright(c) 2019-2022 Qualcomm Innovation Center, Inc. All Rights Reserved.
+##  Copyright(c) 2019-2023 Qualcomm Innovation Center, Inc. All Rights Reserved.
 ##
 ##  This program is free software; you can redistribute it and/or modify
 ##  it under the terms of the GNU General Public License as published by
@@ -87,6 +87,7 @@ def gen_helper_prototype(f, tag, tagregs, tagimms):
             if hex_common.need_slot(tag): def_helper_size += 1
             if hex_common.need_PC(tag): def_helper_size += 1
             if hex_common.helper_needs_next_PC(tag): def_helper_size += 1
+            if hex_common.need_condexec_reg(tag, regs): def_helper_size += 1
             f.write('DEF_HELPER_%s(%s' % (def_helper_size, tag))
             ## The return type is void
             f.write(', void' )
@@ -96,6 +97,7 @@ def gen_helper_prototype(f, tag, tagregs, tagimms):
             if hex_common.need_part1(tag): def_helper_size += 1
             if hex_common.need_slot(tag): def_helper_size += 1
             if hex_common.need_PC(tag): def_helper_size += 1
+            if hex_common.need_condexec_reg(tag, regs): def_helper_size += 1
             if hex_common.helper_needs_next_PC(tag): def_helper_size += 1
             f.write('DEF_HELPER_%s(%s' % (def_helper_size, tag))
 
@@ -121,6 +123,14 @@ def gen_helper_prototype(f, tag, tagregs, tagimms):
                     gen_def_helper_opn(f, tag, regtype, regid, toss, numregs, i)
                     i += 1
 
+        ## For conditional instructions, we pass in the destination register
+        if 'A_CONDEXEC' in hex_common.attribdict[tag]:
+            for regtype, regid, toss, numregs in regs:
+                if (hex_common.is_writeonly(regid) and
+                    not hex_common.is_hvx_reg(regtype)):
+                    gen_def_helper_opn(f, tag, regtype, regid, toss, numregs, i)
+                    i += 1
+
         ## Generate the qemu type for each input operand (regs and immediates)
         for regtype,regid,toss,numregs in regs:
             if (hex_common.is_read(regid)):

target/hexagon/gen_tcg.h

@@ -1,5 +1,5 @@
 /*
- *  Copyright(c) 2019-2022 Qualcomm Innovation Center, Inc. All Rights Reserved.
+ *  Copyright(c) 2019-2023 Qualcomm Innovation Center, Inc. All Rights Reserved.
  *
  *  This program is free software; you can redistribute it and/or modify
  *  it under the terms of the GNU General Public License as published by
@@ -332,8 +332,6 @@
         tcg_gen_movi_tl(EA, 0); \
         PRED;  \
         CHECK_NOSHUF_PRED(GET_EA, SIZE, LSB); \
-        PRED_LOAD_CANCEL(LSB, EA); \
-        tcg_gen_movi_tl(RdV, 0); \
         tcg_gen_brcondi_tl(TCG_COND_EQ, LSB, 0, label); \
         fLOAD(1, SIZE, SIGN, EA, RdV); \
         gen_set_label(label); \
@@ -391,8 +389,6 @@
         tcg_gen_movi_tl(EA, 0); \
         PRED;  \
         CHECK_NOSHUF_PRED(GET_EA, 8, LSB); \
-        PRED_LOAD_CANCEL(LSB, EA); \
-        tcg_gen_movi_i64(RddV, 0); \
         tcg_gen_brcondi_tl(TCG_COND_EQ, LSB, 0, label); \
         fLOAD(1, 8, u, EA, RddV); \
         gen_set_label(label); \
@@ -419,16 +415,16 @@
 
 #define fGEN_TCG_STORE(SHORTCODE) \
     do { \
-        TCGv HALF = tcg_temp_new(); \
+        TCGv HALF G_GNUC_UNUSED = tcg_temp_new(); \
-        TCGv BYTE = tcg_temp_new(); \
+        TCGv BYTE G_GNUC_UNUSED = tcg_temp_new(); \
         SHORTCODE; \
     } while (0)
 
 #define fGEN_TCG_STORE_pcr(SHIFT, STORE) \
     do { \
         TCGv ireg = tcg_temp_new(); \
-        TCGv HALF = tcg_temp_new(); \
+        TCGv HALF G_GNUC_UNUSED = tcg_temp_new(); \
-        TCGv BYTE = tcg_temp_new(); \
+        TCGv BYTE G_GNUC_UNUSED = tcg_temp_new(); \
         tcg_gen_mov_tl(EA, RxV); \
         gen_read_ireg(ireg, MuV, SHIFT); \
         gen_helper_fcircadd(RxV, RxV, ireg, MuV, hex_gpr[HEX_REG_CS0 + MuN]); \
@@ -491,6 +487,59 @@
 #define fGEN_TCG_S2_storerinew_pcr(SHORTCODE) \
     fGEN_TCG_STORE_pcr(2, fSTORE(1, 4, EA, NtN))
 
+/*
+ * dealloc_return
+ * Assembler mapped to
+ * r31:30 = dealloc_return(r30):raw
+ */
+#define fGEN_TCG_L4_return(SHORTCODE) \
+    gen_return(ctx, RddV, RsV)
+
+/*
+ * sub-instruction version (no RddV, so handle it manually)
+ */
+#define fGEN_TCG_SL2_return(SHORTCODE) \
+    do { \
+        TCGv_i64 RddV = get_result_gpr_pair(ctx, HEX_REG_FP); \
+        gen_return(ctx, RddV, hex_gpr[HEX_REG_FP]); \
+        gen_log_reg_write_pair(HEX_REG_FP, RddV); \
+    } while (0)
+
+/*
+ * Conditional returns follow this naming convention
+ *     _t                 predicate true
+ *     _f                 predicate false
+ *     _tnew_pt           predicate.new true predict taken
+ *     _fnew_pt           predicate.new false predict taken
+ *     _tnew_pnt          predicate.new true predict not taken
+ *     _fnew_pnt          predicate.new false predict not taken
+ * Predictions are not modelled in QEMU
+ *
+ * Example:
+ *     if (p1) r31:30 = dealloc_return(r30):raw
+ */
+#define fGEN_TCG_L4_return_t(SHORTCODE) \
+    gen_cond_return(ctx, RddV, RsV, PvV, TCG_COND_EQ);
+#define fGEN_TCG_L4_return_f(SHORTCODE) \
+    gen_cond_return(ctx, RddV, RsV, PvV, TCG_COND_NE)
+#define fGEN_TCG_L4_return_tnew_pt(SHORTCODE) \
+    gen_cond_return(ctx, RddV, RsV, PvN, TCG_COND_EQ)
+#define fGEN_TCG_L4_return_fnew_pt(SHORTCODE) \
+    gen_cond_return(ctx, RddV, RsV, PvN, TCG_COND_NE)
+#define fGEN_TCG_L4_return_tnew_pnt(SHORTCODE) \
+    gen_cond_return(ctx, RddV, RsV, PvN, TCG_COND_EQ)
+#define fGEN_TCG_L4_return_fnew_pnt(SHORTCODE) \
+    gen_cond_return(ctx, RddV, RsV, PvN, TCG_COND_NE)
+
+#define fGEN_TCG_SL2_return_t(SHORTCODE) \
+    gen_cond_return_subinsn(ctx, TCG_COND_EQ, hex_pred[0])
+#define fGEN_TCG_SL2_return_f(SHORTCODE) \
+    gen_cond_return_subinsn(ctx, TCG_COND_NE, hex_pred[0])
+#define fGEN_TCG_SL2_return_tnew(SHORTCODE) \
+    gen_cond_return_subinsn(ctx, TCG_COND_EQ, hex_new_pred_value[0])
+#define fGEN_TCG_SL2_return_fnew(SHORTCODE) \
+    gen_cond_return_subinsn(ctx, TCG_COND_NE, hex_new_pred_value[0])
+
 /*
  * Mathematical operations with more than one definition require
  * special handling
@@ -589,14 +638,24 @@
 
 #define fGEN_TCG_J2_call(SHORTCODE) \
     gen_call(ctx, riV)
+#define fGEN_TCG_J2_callr(SHORTCODE) \
+    gen_callr(ctx, RsV)
 
 #define fGEN_TCG_J2_callt(SHORTCODE) \
     gen_cond_call(ctx, PuV, TCG_COND_EQ, riV)
 #define fGEN_TCG_J2_callf(SHORTCODE) \
     gen_cond_call(ctx, PuV, TCG_COND_NE, riV)
+#define fGEN_TCG_J2_callrt(SHORTCODE) \
+    gen_cond_callr(ctx, TCG_COND_EQ, PuV, RsV)
+#define fGEN_TCG_J2_callrf(SHORTCODE) \
+    gen_cond_callr(ctx, TCG_COND_NE, PuV, RsV)
 
 #define fGEN_TCG_J2_endloop0(SHORTCODE) \
     gen_endloop0(ctx)
+#define fGEN_TCG_J2_endloop1(SHORTCODE) \
+    gen_endloop1(ctx)
+#define fGEN_TCG_J2_endloop01(SHORTCODE) \
+    gen_endloop01(ctx)
 
 /*
  * Compound compare and jump instructions
@@ -986,6 +1045,19 @@
 #define fGEN_TCG_S2_asl_r_r_sat(SHORTCODE) \
     gen_asl_r_r_sat(RdV, RsV, RtV)
 
+#define fGEN_TCG_SL2_jumpr31(SHORTCODE) \
+    gen_jumpr(ctx, hex_gpr[HEX_REG_LR])
+
+#define fGEN_TCG_SL2_jumpr31_t(SHORTCODE) \
+    gen_cond_jumpr31(ctx, TCG_COND_EQ, hex_pred[0])
+#define fGEN_TCG_SL2_jumpr31_f(SHORTCODE) \
+    gen_cond_jumpr31(ctx, TCG_COND_NE, hex_pred[0])
+
+#define fGEN_TCG_SL2_jumpr31_tnew(SHORTCODE) \
+    gen_cond_jumpr31(ctx, TCG_COND_EQ, hex_new_pred_value[0])
+#define fGEN_TCG_SL2_jumpr31_fnew(SHORTCODE) \
+    gen_cond_jumpr31(ctx, TCG_COND_NE, hex_new_pred_value[0])
+
 /* Floating point */
 #define fGEN_TCG_F2_conv_sf2df(SHORTCODE) \
     gen_helper_conv_sf2df(RddV, cpu_env, RsV)

target/hexagon/gen_tcg_funcs.py

@@ -1,7 +1,7 @@
 #!/usr/bin/env python3
 
 ##
-##  Copyright(c) 2019-2022 Qualcomm Innovation Center, Inc. All Rights Reserved.
+##  Copyright(c) 2019-2023 Qualcomm Innovation Center, Inc. All Rights Reserved.
 ##
 ##  This program is free software; you can redistribute it and/or modify
 ##  it under the terms of the GNU General Public License as published by
@@ -30,37 +30,33 @@ def gen_decl_ea_tcg(f, tag):
 
 def genptr_decl_pair_writable(f, tag, regtype, regid, regno):
     regN="%s%sN" % (regtype,regid)
-    f.write("    TCGv_i64 %s%sV = tcg_temp_new_i64();\n" % \
+    if (regtype == "R"):
-        (regtype, regid))
+        f.write("    const int %s = insn->regno[%d];\n" % (regN, regno))
-    if (regtype == "C"):
+    elif (regtype == "C"):
         f.write("    const int %s = insn->regno[%d] + HEX_REG_SA0;\n" % \
             (regN, regno))
     else:
-        f.write("    const int %s = insn->regno[%d];\n" % (regN, regno))
+        print("Bad register parse: ", regtype, regid)
-    if ('A_CONDEXEC' in hex_common.attribdict[tag]):
+    f.write("    TCGv_i64 %s%sV = get_result_gpr_pair(ctx, %s);\n" % \
-        f.write("    if (!is_preloaded(ctx, %s)) {\n" % regN)
+        (regtype, regid, regN))
-        f.write("        tcg_gen_mov_tl(hex_new_value[%s], hex_gpr[%s]);\n" % \
-                             (regN, regN))
-        f.write("    }\n")
-        f.write("    if (!is_preloaded(ctx, %s + 1)) {\n" % regN)
-        f.write("        tcg_gen_mov_tl(hex_new_value[%s + 1], hex_gpr[%s + 1]);\n" % \
-                             (regN, regN))
-        f.write("    }\n")
 
 def genptr_decl_writable(f, tag, regtype, regid, regno):
     regN="%s%sN" % (regtype,regid)
-    f.write("    TCGv %s%sV = tcg_temp_new();\n" % \
+    if (regtype == "R"):
-        (regtype, regid))
+        f.write("    const int %s = insn->regno[%d];\n" % (regN, regno))
-    if (regtype == "C"):
+        f.write("    TCGv %s%sV = get_result_gpr(ctx, %s);\n" % \
+            (regtype, regid, regN))
+    elif (regtype == "C"):
         f.write("    const int %s = insn->regno[%d] + HEX_REG_SA0;\n" % \
             (regN, regno))
-    else:
+        f.write("    TCGv %s%sV = get_result_gpr(ctx, %s);\n" % \
+            (regtype, regid, regN))
+    elif (regtype == "P"):
         f.write("    const int %s = insn->regno[%d];\n" % (regN, regno))
-    if ('A_CONDEXEC' in hex_common.attribdict[tag]):
+        f.write("    TCGv %s%sV = tcg_temp_new();\n" % \
-        f.write("    if (!is_preloaded(ctx, %s)) {\n" % regN)
+            (regtype, regid))
-        f.write("        tcg_gen_mov_tl(hex_new_value[%s], hex_gpr[%s]);\n" % \
+    else:
-                             (regN, regN))
+        print("Bad register parse: ", regtype, regid)
-        f.write("    }\n")
 
 def genptr_decl(f, tag, regtype, regid, regno):
     regN="%s%sN" % (regtype,regid)
@@ -166,17 +162,6 @@ def genptr_decl(f, tag, regtype, regid, regno):
                 f.write("        ctx_future_vreg_off(ctx, %s%sN," % \
                     (regtype, regid))
                 f.write(" 1, true);\n");
-            if 'A_CONDEXEC' in hex_common.attribdict[tag]:
-                f.write("    if (!is_vreg_preloaded(ctx, %s)) {\n" % (regN))
-                f.write("        intptr_t src_off =")
-                f.write(" offsetof(CPUHexagonState, VRegs[%s%sN]);\n"% \
-                                     (regtype, regid))
-                f.write("        tcg_gen_gvec_mov(MO_64, %s%sV_off,\n" % \
-                                     (regtype, regid))
-                f.write("                         src_off,\n")
-                f.write("                         sizeof(MMVector),\n")
-                f.write("                         sizeof(MMVector));\n")
-                f.write("    }\n")
 
             if (not hex_common.skip_qemu_helper(tag)):
                 f.write("    TCGv_ptr %s%sV = tcg_temp_new_ptr();\n" % \
@@ -191,8 +176,7 @@ def genptr_decl(f, tag, regtype, regid, regno):
                 (regtype, regid, regno))
             f.write("    const intptr_t %s%sV_off =\n" % \
                 (regtype, regid))
-            f.write("        offsetof(CPUHexagonState,\n")
+            f.write("        get_result_qreg(ctx, %s%sN);\n" % \
-            f.write("                 future_QRegs[%s%sN]);\n" % \
                 (regtype, regid))
             if (not hex_common.skip_qemu_helper(tag)):
                 f.write("    TCGv_ptr %s%sV = tcg_temp_new_ptr();\n" % \
@@ -274,8 +258,12 @@ def genptr_src_read(f, tag, regtype, regid):
             f.write("                                 hex_gpr[%s%sN + 1]);\n" % \
                 (regtype, regid))
         elif (regid in {"x", "y"}):
-            f.write("    tcg_gen_mov_tl(%s%sV, hex_gpr[%s%sN]);\n" % \
+            ## For read/write registers, we need to get the original value into
-                (regtype,regid,regtype,regid))
+            ## the result TCGv.  For conditional instructions, this is done in
+            ## gen_start_packet.  For unconditional instructions, we do it here.
+            if ('A_CONDEXEC' not in hex_common.attribdict[tag]):
+                f.write("    tcg_gen_mov_tl(%s%sV, hex_gpr[%s%sN]);\n" % \
+                    (regtype, regid, regtype, regid))
         elif (regid not in {"s", "t", "u", "v"}):
             print("Bad register parse: ", regtype, regid)
     elif (regtype == "P"):
@@ -385,37 +373,22 @@ def gen_helper_call_imm(f,immlett):
     f.write(", tcgv_%s" % hex_common.imm_name(immlett))
 
 def genptr_dst_write_pair(f, tag, regtype, regid):
-    if ('A_CONDEXEC' in hex_common.attribdict[tag]):
+    f.write("    gen_log_reg_write_pair(%s%sN, %s%sV);\n" % \
-        f.write("    gen_log_predicated_reg_write_pair(%s%sN, %s%sV, insn->slot);\n" % \
+        (regtype, regid, regtype, regid))
-            (regtype, regid, regtype, regid))
-    else:
-        f.write("    gen_log_reg_write_pair(%s%sN, %s%sV);\n" % \
-            (regtype, regid, regtype, regid))
-    f.write("    ctx_log_reg_write_pair(ctx, %s%sN);\n" % \
-        (regtype, regid))
 
 def genptr_dst_write(f, tag, regtype, regid):
     if (regtype == "R"):
         if (regid in {"dd", "xx", "yy"}):
             genptr_dst_write_pair(f, tag, regtype, regid)
         elif (regid in {"d", "e", "x", "y"}):
-            if ('A_CONDEXEC' in hex_common.attribdict[tag]):
+            f.write("    gen_log_reg_write(%s%sN, %s%sV);\n" % \
-                f.write("    gen_log_predicated_reg_write(%s%sN, %s%sV,\n" % \
+                (regtype, regid, regtype, regid))
-                    (regtype, regid, regtype, regid))
-                f.write("                                 insn->slot);\n")
-            else:
-                f.write("    gen_log_reg_write(%s%sN, %s%sV);\n" % \
-                    (regtype, regid, regtype, regid))
-            f.write("    ctx_log_reg_write(ctx, %s%sN);\n" % \
-                (regtype, regid))
         else:
             print("Bad register parse: ", regtype, regid)
     elif (regtype == "P"):
         if (regid in {"d", "e", "x"}):
             f.write("    gen_log_pred_write(ctx, %s%sN, %s%sV);\n" % \
                 (regtype, regid, regtype, regid))
-            f.write("    ctx_log_pred_write(ctx, %s%sN);\n" % \
-                (regtype, regid))
         else:
             print("Bad register parse: ", regtype, regid)
     elif (regtype == "C"):
@@ -432,43 +405,18 @@ def genptr_dst_write(f, tag, regtype, regid):
 
 def genptr_dst_write_ext(f, tag, regtype, regid, newv="EXT_DFL"):
     if (regtype == "V"):
-        if (regid in {"dd", "xx", "yy"}):
+        if (regid in {"xx"}):
-            if ('A_CONDEXEC' in hex_common.attribdict[tag]):
-                is_predicated = "true"
-            else:
-                is_predicated = "false"
             f.write("    gen_log_vreg_write_pair(ctx, %s%sV_off, %s%sN, " % \
                 (regtype, regid, regtype, regid))
-            f.write("%s, insn->slot, %s);\n" % \
+            f.write("%s);\n" % \
-                (newv, is_predicated))
+                (newv))
-            f.write("    ctx_log_vreg_write_pair(ctx, %s%sN, %s,\n" % \
+        elif (regid in {"y"}):
-                (regtype, regid, newv))
+            f.write("    gen_log_vreg_write(ctx, %s%sV_off, %s%sN, %s);\n" % \
-            f.write("        %s);\n" % (is_predicated))
-        elif (regid in {"d", "x", "y"}):
-            if ('A_CONDEXEC' in hex_common.attribdict[tag]):
-                is_predicated = "true"
-            else:
-                is_predicated = "false"
-            f.write("    gen_log_vreg_write(ctx, %s%sV_off, %s%sN, %s, " % \
                 (regtype, regid, regtype, regid, newv))
-            f.write("insn->slot, %s);\n" % \
+        elif (regid not in {"dd", "d", "x"}):
-                (is_predicated))
-            f.write("    ctx_log_vreg_write(ctx, %s%sN, %s, %s);\n" % \
-                (regtype, regid, newv, is_predicated))
-        else:
             print("Bad register parse: ", regtype, regid)
     elif (regtype == "Q"):
-        if (regid in {"d", "e", "x"}):
+        if (regid not in {"d", "e", "x"}):
-            if ('A_CONDEXEC' in hex_common.attribdict[tag]):
-                is_predicated = "true"
-            else:
-                is_predicated = "false"
-            f.write("    gen_log_qreg_write(%s%sV_off, %s%sN, %s, " % \
-                (regtype, regid, regtype, regid, newv))
-            f.write("insn->slot, %s);\n" % (is_predicated))
-            f.write("    ctx_log_qreg_write(ctx, %s%sN, %s);\n" % \
-                (regtype, regid, is_predicated))
-        else:
             print("Bad register parse: ", regtype, regid)
     else:
         print("Bad register parse: ", regtype, regid)
@@ -500,15 +448,15 @@ def genptr_dst_write_opn(f,regtype, regid, tag):
 ##     For A2_add: Rd32=add(Rs32,Rt32), { RdV=RsV+RtV;}
 ##     We produce:
 ##    static void generate_A2_add(DisasContext *ctx)
-##       {
+##    {
-##           TCGv RdV = tcg_temp_new();
+##        Insn *insn __attribute__((unused)) = ctx->insn;
-##           const int RdN = insn->regno[0];
+##        const int RdN = insn->regno[0];
-##           TCGv RsV = hex_gpr[insn->regno[1]];
+##        TCGv RdV = get_result_gpr(ctx, RdN);
-##           TCGv RtV = hex_gpr[insn->regno[2]];
+##        TCGv RsV = hex_gpr[insn->regno[1]];
-##           <GEN>
+##        TCGv RtV = hex_gpr[insn->regno[2]];
-##           gen_log_reg_write(RdN, RdV);
+##        <GEN>
-##           ctx_log_reg_write(ctx, RdN);
+##        gen_log_reg_write(RdN, RdV);
-##       }
+##    }
 ##
 ##       where <GEN> depends on hex_common.skip_qemu_helper(tag)
 ##       if hex_common.skip_qemu_helper(tag) is True
@@ -592,6 +540,14 @@ def gen_tcg_func(f, tag, regs, imms):
         if (i > 0): f.write(", ")
         f.write("cpu_env")
         i=1
+        ## For conditional instructions, we pass in the destination register
+        if 'A_CONDEXEC' in hex_common.attribdict[tag]:
+            for regtype, regid, toss, numregs in regs:
+                if (hex_common.is_writeonly(regid) and
+                    not hex_common.is_hvx_reg(regtype)):
+                    gen_helper_call_opn(f, tag, regtype, regid, toss, \
+                                        numregs, i)
+                    i += 1
         for regtype,regid,toss,numregs in regs:
             if (hex_common.is_written(regid)):
                 if (not hex_common.is_hvx_reg(regtype)):

target/hexagon/gen_tcg_hvx.h

@@ -1,5 +1,5 @@
 /*
- *  Copyright(c) 2019-2022 Qualcomm Innovation Center, Inc. All Rights Reserved.
+ *  Copyright(c) 2019-2023 Qualcomm Innovation Center, Inc. All Rights Reserved.
  *
  *  This program is free software; you can redistribute it and/or modify
  *  it under the terms of the GNU General Public License as published by
@@ -133,16 +133,11 @@ static inline void assert_vhist_tmp(DisasContext *ctx)
     do { \
         TCGv lsb = tcg_temp_new(); \
         TCGLabel *false_label = gen_new_label(); \
-        TCGLabel *end_label = gen_new_label(); \
         tcg_gen_andi_tl(lsb, PsV, 1); \
         tcg_gen_brcondi_tl(TCG_COND_NE, lsb, PRED, false_label); \
         tcg_gen_gvec_mov(MO_64, VdV_off, VuV_off, \
                          sizeof(MMVector), sizeof(MMVector)); \
-        tcg_gen_br(end_label); \
         gen_set_label(false_label); \
-        tcg_gen_ori_tl(hex_slot_cancelled, hex_slot_cancelled, \
-                       1 << insn->slot); \
-        gen_set_label(end_label); \
     } while (0)
 
 
@@ -547,17 +542,12 @@ static inline void assert_vhist_tmp(DisasContext *ctx)
     do { \
         TCGv LSB = tcg_temp_new(); \
         TCGLabel *false_label = gen_new_label(); \
-        TCGLabel *end_label = gen_new_label(); \
         GET_EA; \
         PRED; \
         tcg_gen_brcondi_tl(TCG_COND_EQ, LSB, 0, false_label); \
         gen_vreg_load(ctx, DSTOFF, EA, true); \
         INC; \
-        tcg_gen_br(end_label); \
         gen_set_label(false_label); \
-        tcg_gen_ori_tl(hex_slot_cancelled, hex_slot_cancelled, \
-                       1 << insn->slot); \
-        gen_set_label(end_label); \
     } while (0)
 
 #define fGEN_TCG_PRED_VEC_LOAD_pred_pi \
@@ -717,17 +707,12 @@ static inline void assert_vhist_tmp(DisasContext *ctx)
     do { \
         TCGv LSB = tcg_temp_new(); \
         TCGLabel *false_label = gen_new_label(); \
-        TCGLabel *end_label = gen_new_label(); \
         GET_EA; \
         PRED; \
         tcg_gen_brcondi_tl(TCG_COND_EQ, LSB, 0, false_label); \
         gen_vreg_store(ctx, EA, SRCOFF, insn->slot, ALIGN); \
         INC; \
-        tcg_gen_br(end_label); \
         gen_set_label(false_label); \
-        tcg_gen_ori_tl(hex_slot_cancelled, hex_slot_cancelled, \
-                       1 << insn->slot); \
-        gen_set_label(end_label); \
     } while (0)
 
 #define fGEN_TCG_PRED_VEC_STORE_pred_pi(ALIGN) \

target/hexagon/genptr.c

@@ -1,5 +1,5 @@
 /*
- *  Copyright(c) 2019-2022 Qualcomm Innovation Center, Inc. All Rights Reserved.
+ *  Copyright(c) 2019-2023 Qualcomm Innovation Center, Inc. All Rights Reserved.
  *
  *  This program is free software; you can redistribute it and/or modify
  *  it under the terms of the GNU General Public License as published by
@@ -68,26 +68,17 @@ static inline void gen_masked_reg_write(TCGv new_val, TCGv cur_val,
     }
 }
 
-static inline void gen_log_predicated_reg_write(int rnum, TCGv val,
+static TCGv get_result_gpr(DisasContext *ctx, int rnum)
-                                                uint32_t slot)
 {
-    TCGv zero = tcg_constant_tl(0);
+    return hex_new_value[rnum];
-    TCGv slot_mask = tcg_temp_new();
+}
 
-    tcg_gen_andi_tl(slot_mask, hex_slot_cancelled, 1 << slot);
+static TCGv_i64 get_result_gpr_pair(DisasContext *ctx, int rnum)
-    tcg_gen_movcond_tl(TCG_COND_EQ, hex_new_value[rnum], slot_mask, zero,
+{
-                           val, hex_new_value[rnum]);
+    TCGv_i64 result = tcg_temp_new_i64();
-    if (HEX_DEBUG) {
+    tcg_gen_concat_i32_i64(result, hex_new_value[rnum],
-        /*
+                                   hex_new_value[rnum + 1]);
-         * Do this so HELPER(debug_commit_end) will know
+    return result;
-         *
-         * Note that slot_mask indicates the value is not written
-         * (i.e., slot was cancelled), so we create a true/false value before
-         * or'ing with hex_reg_written[rnum].
-         */
-        tcg_gen_setcond_tl(TCG_COND_EQ, slot_mask, slot_mask, zero);
-        tcg_gen_or_tl(hex_reg_written[rnum], hex_reg_written[rnum], slot_mask);
-    }
 }
 
 void gen_log_reg_write(int rnum, TCGv val)
@@ -102,39 +93,6 @@ void gen_log_reg_write(int rnum, TCGv val)
     }
 }
 
-static void gen_log_predicated_reg_write_pair(int rnum, TCGv_i64 val,
-                                              uint32_t slot)
-{
-    TCGv val32 = tcg_temp_new();
-    TCGv zero = tcg_constant_tl(0);
-    TCGv slot_mask = tcg_temp_new();
-
-    tcg_gen_andi_tl(slot_mask, hex_slot_cancelled, 1 << slot);
-    /* Low word */
-    tcg_gen_extrl_i64_i32(val32, val);
-    tcg_gen_movcond_tl(TCG_COND_EQ, hex_new_value[rnum],
-                       slot_mask, zero,
-                       val32, hex_new_value[rnum]);
-    /* High word */
-    tcg_gen_extrh_i64_i32(val32, val);
-    tcg_gen_movcond_tl(TCG_COND_EQ, hex_new_value[rnum + 1],
-                       slot_mask, zero,
-                       val32, hex_new_value[rnum + 1]);
-    if (HEX_DEBUG) {
-        /*
-         * Do this so HELPER(debug_commit_end) will know
-         *
-         * Note that slot_mask indicates the value is not written
-         * (i.e., slot was cancelled), so we create a true/false value before
-         * or'ing with hex_reg_written[rnum].
-         */
-        tcg_gen_setcond_tl(TCG_COND_EQ, slot_mask, slot_mask, zero);
-        tcg_gen_or_tl(hex_reg_written[rnum], hex_reg_written[rnum], slot_mask);
-        tcg_gen_or_tl(hex_reg_written[rnum + 1], hex_reg_written[rnum + 1],
-                      slot_mask);
-    }
-}
-
 static void gen_log_reg_write_pair(int rnum, TCGv_i64 val)
 {
     const target_ulong reg_mask_low = reg_immut_masks[rnum];
@@ -180,6 +138,7 @@ void gen_log_pred_write(DisasContext *ctx, int pnum, TCGv val)
                        hex_new_pred_value[pnum], base_val);
     }
     tcg_gen_ori_tl(hex_pred_written, hex_pred_written, 1 << pnum);
+    set_bit(pnum, ctx->pregs_written);
 }
 
 static inline void gen_read_p3_0(TCGv control_reg)
@@ -256,7 +215,6 @@ static void gen_write_p3_0(DisasContext *ctx, TCGv control_reg)
     for (int i = 0; i < NUM_PREGS; i++) {
         tcg_gen_extract_tl(hex_p8, control_reg, i * 8, 8);
         gen_log_pred_write(ctx, i, hex_p8);
-        ctx_log_pred_write(ctx, i);
     }
 }
 
@@ -274,7 +232,6 @@ static inline void gen_write_ctrl_reg(DisasContext *ctx, int reg_num,
         gen_write_p3_0(ctx, val);
     } else {
         gen_log_reg_write(reg_num, val);
-        ctx_log_reg_write(ctx, reg_num);
         if (reg_num == HEX_REG_QEMU_PKT_CNT) {
             ctx->num_packets = 0;
         }
@@ -291,15 +248,14 @@ static inline void gen_write_ctrl_reg_pair(DisasContext *ctx, int reg_num,
                                            TCGv_i64 val)
 {
     if (reg_num == HEX_REG_P3_0_ALIASED) {
+        TCGv result = get_result_gpr(ctx, reg_num + 1);
         TCGv val32 = tcg_temp_new();
         tcg_gen_extrl_i64_i32(val32, val);
         gen_write_p3_0(ctx, val32);
         tcg_gen_extrh_i64_i32(val32, val);
-        gen_log_reg_write(reg_num + 1, val32);
+        tcg_gen_mov_tl(result, val32);
-        ctx_log_reg_write(ctx, reg_num + 1);
     } else {
         gen_log_reg_write_pair(reg_num, val);
-        ctx_log_reg_write_pair(ctx, reg_num);
         if (reg_num == HEX_REG_QEMU_PKT_CNT) {
             ctx->num_packets = 0;
             ctx->num_insns = 0;
@@ -571,6 +527,13 @@ static void gen_cond_jumpr(DisasContext *ctx, TCGv dst_pc,
     gen_write_new_pc_addr(ctx, dst_pc, cond, pred);
 }
 
+static void gen_cond_jumpr31(DisasContext *ctx, TCGCond cond, TCGv pred)
+{
+    TCGv LSB = tcg_temp_new();
+    tcg_gen_andi_tl(LSB, pred, 1);
+    gen_cond_jumpr(ctx, hex_gpr[HEX_REG_LR], cond, LSB);
+}
+
 static void gen_cond_jump(DisasContext *ctx, TCGCond cond, TCGv pred,
                           int pc_off)
 {
@@ -669,27 +632,99 @@ static void gen_jumpr(DisasContext *ctx, TCGv new_pc)
 
 static void gen_call(DisasContext *ctx, int pc_off)
 {
-    TCGv next_PC =
+    TCGv lr = get_result_gpr(ctx, HEX_REG_LR);
-        tcg_constant_tl(ctx->pkt->pc + ctx->pkt->encod_pkt_size_in_bytes);
+    tcg_gen_movi_tl(lr, ctx->next_PC);
-    gen_log_reg_write(HEX_REG_LR, next_PC);
     gen_write_new_pc_pcrel(ctx, pc_off, TCG_COND_ALWAYS, NULL);
 }
 
+static void gen_callr(DisasContext *ctx, TCGv new_pc)
+{
+    TCGv lr = get_result_gpr(ctx, HEX_REG_LR);
+    tcg_gen_movi_tl(lr, ctx->next_PC);
+    gen_write_new_pc_addr(ctx, new_pc, TCG_COND_ALWAYS, NULL);
+}
+
 static void gen_cond_call(DisasContext *ctx, TCGv pred,
                           TCGCond cond, int pc_off)
 {
-    TCGv next_PC;
+    TCGv lr = get_result_gpr(ctx, HEX_REG_LR);
     TCGv lsb = tcg_temp_new();
     TCGLabel *skip = gen_new_label();
     tcg_gen_andi_tl(lsb, pred, 1);
     gen_write_new_pc_pcrel(ctx, pc_off, cond, lsb);
     tcg_gen_brcondi_tl(cond, lsb, 0, skip);
-    next_PC =
+    tcg_gen_movi_tl(lr, ctx->next_PC);
-        tcg_constant_tl(ctx->pkt->pc + ctx->pkt->encod_pkt_size_in_bytes);
-    gen_log_reg_write(HEX_REG_LR, next_PC);
     gen_set_label(skip);
 }
 
+static void gen_cond_callr(DisasContext *ctx,
+                           TCGCond cond, TCGv pred, TCGv new_pc)
+{
+    TCGv lsb = tcg_temp_new();
+    TCGLabel *skip = gen_new_label();
+    tcg_gen_andi_tl(lsb, pred, 1);
+    tcg_gen_brcondi_tl(cond, lsb, 0, skip);
+    gen_callr(ctx, new_pc);
+    gen_set_label(skip);
+}
+
+/* frame ^= (int64_t)FRAMEKEY << 32 */
+static void gen_frame_unscramble(TCGv_i64 frame)
+{
+    TCGv_i64 framekey = tcg_temp_new_i64();
+    tcg_gen_extu_i32_i64(framekey, hex_gpr[HEX_REG_FRAMEKEY]);
+    tcg_gen_shli_i64(framekey, framekey, 32);
+    tcg_gen_xor_i64(frame, frame, framekey);
+}
+
+static void gen_load_frame(DisasContext *ctx, TCGv_i64 frame, TCGv EA)
+{
+    Insn *insn = ctx->insn;  /* Needed for CHECK_NOSHUF */
+    CHECK_NOSHUF(EA, 8);
+    tcg_gen_qemu_ld64(frame, EA, ctx->mem_idx);
+}
+
+static void gen_return(DisasContext *ctx, TCGv_i64 dst, TCGv src)
+{
+    /*
+     * frame = *src
+     * dst = frame_unscramble(frame)
+     * SP = src + 8
+     * PC = dst.w[1]
+     */
+    TCGv_i64 frame = tcg_temp_new_i64();
+    TCGv r31 = tcg_temp_new();
+    TCGv r29 = get_result_gpr(ctx, HEX_REG_SP);
+
+    gen_load_frame(ctx, frame, src);
+    gen_frame_unscramble(frame);
+    tcg_gen_mov_i64(dst, frame);
+    tcg_gen_addi_tl(r29, src, 8);
+    tcg_gen_extrh_i64_i32(r31, dst);
+    gen_jumpr(ctx, r31);
+}
+
+/* if (pred) dst = dealloc_return(src):raw */
+static void gen_cond_return(DisasContext *ctx, TCGv_i64 dst, TCGv src,
+                            TCGv pred, TCGCond cond)
+{
+    TCGv LSB = tcg_temp_new();
+    TCGLabel *skip = gen_new_label();
+    tcg_gen_andi_tl(LSB, pred, 1);
+
+    tcg_gen_brcondi_tl(cond, LSB, 0, skip);
+    gen_return(ctx, dst, src);
+    gen_set_label(skip);
+}
+
+/* sub-instruction version (no RddV, so handle it manually) */
+static void gen_cond_return_subinsn(DisasContext *ctx, TCGCond cond, TCGv pred)
+{
+    TCGv_i64 RddV = get_result_gpr_pair(ctx, HEX_REG_FP);
+    gen_cond_return(ctx, RddV, hex_gpr[HEX_REG_FP], pred, cond);
+    gen_log_reg_write_pair(HEX_REG_FP, RddV);
+}
+
 static void gen_endloop0(DisasContext *ctx)
 {
     TCGv lpcfg = tcg_temp_new();
@@ -737,14 +772,95 @@ static void gen_endloop0(DisasContext *ctx)
         TCGLabel *label3 = gen_new_label();
         tcg_gen_brcondi_tl(TCG_COND_LEU, hex_gpr[HEX_REG_LC0], 1, label3);
         {
+            TCGv lc0 = get_result_gpr(ctx, HEX_REG_LC0);
             gen_jumpr(ctx, hex_gpr[HEX_REG_SA0]);
-            tcg_gen_subi_tl(hex_new_value[HEX_REG_LC0],
+            tcg_gen_subi_tl(lc0, hex_gpr[HEX_REG_LC0], 1);
-                            hex_gpr[HEX_REG_LC0], 1);
         }
         gen_set_label(label3);
     }
 }
 
+static void gen_endloop1(DisasContext *ctx)
+{
+    /*
+     *    if (hex_gpr[HEX_REG_LC1] > 1) {
+     *        PC = hex_gpr[HEX_REG_SA1];
+     *        hex_new_value[HEX_REG_LC1] = hex_gpr[HEX_REG_LC1] - 1;
+     *    }
+     */
+    TCGLabel *label = gen_new_label();
+    tcg_gen_brcondi_tl(TCG_COND_LEU, hex_gpr[HEX_REG_LC1], 1, label);
+    {
+        TCGv lc1 = get_result_gpr(ctx, HEX_REG_LC1);
+        gen_jumpr(ctx, hex_gpr[HEX_REG_SA1]);
+        tcg_gen_subi_tl(lc1, hex_gpr[HEX_REG_LC1], 1);
+    }
+    gen_set_label(label);
+}
+
+static void gen_endloop01(DisasContext *ctx)
+{
+    TCGv lpcfg = tcg_temp_new();
+    TCGLabel *label1 = gen_new_label();
+    TCGLabel *label2 = gen_new_label();
+    TCGLabel *label3 = gen_new_label();
+    TCGLabel *done = gen_new_label();
+
+    GET_USR_FIELD(USR_LPCFG, lpcfg);
+
+    /*
+     *    if (lpcfg == 1) {
+     *        hex_new_pred_value[3] = 0xff;
+     *        hex_pred_written |= 1 << 3;
+     *    }
+     */
+    tcg_gen_brcondi_tl(TCG_COND_NE, lpcfg, 1, label1);
+    {
+        tcg_gen_movi_tl(hex_new_pred_value[3], 0xff);
+        tcg_gen_ori_tl(hex_pred_written, hex_pred_written, 1 << 3);
+    }
+    gen_set_label(label1);
+
+    /*
+     *    if (lpcfg) {
+     *        SET_USR_FIELD(USR_LPCFG, lpcfg - 1);
+     *    }
+     */
+    tcg_gen_brcondi_tl(TCG_COND_EQ, lpcfg, 0, label2);
+    {
+        tcg_gen_subi_tl(lpcfg, lpcfg, 1);
+        SET_USR_FIELD(USR_LPCFG, lpcfg);
+    }
+    gen_set_label(label2);
+
+    /*
+     *    if (hex_gpr[HEX_REG_LC0] > 1) {
+     *        PC = hex_gpr[HEX_REG_SA0];
+     *        hex_new_value[HEX_REG_LC0] = hex_gpr[HEX_REG_LC0] - 1;
+     *    } else {
+     *        if (hex_gpr[HEX_REG_LC1] > 1) {
+     *            hex_next_pc = hex_gpr[HEX_REG_SA1];
+     *            hex_new_value[HEX_REG_LC1] = hex_gpr[HEX_REG_LC1] - 1;
+     *        }
+     *    }
+     */
+    tcg_gen_brcondi_tl(TCG_COND_LEU, hex_gpr[HEX_REG_LC0], 1, label3);
+    {
+        TCGv lc0 = get_result_gpr(ctx, HEX_REG_LC0);
+        gen_jumpr(ctx, hex_gpr[HEX_REG_SA0]);
+        tcg_gen_subi_tl(lc0, hex_gpr[HEX_REG_LC0], 1);
+        tcg_gen_br(done);
+    }
+    gen_set_label(label3);
+    tcg_gen_brcondi_tl(TCG_COND_LEU, hex_gpr[HEX_REG_LC1], 1, done);
+    {
+        TCGv lc1 = get_result_gpr(ctx, HEX_REG_LC1);
+        gen_jumpr(ctx, hex_gpr[HEX_REG_SA1]);
+        tcg_gen_subi_tl(lc1, hex_gpr[HEX_REG_LC1], 1);
+    }
+    gen_set_label(done);
+}
+
 static void gen_cmp_jumpnv(DisasContext *ctx,
                            TCGCond cond, TCGv val, TCGv src, int pc_off)
 {
@@ -869,68 +985,32 @@ static intptr_t vreg_src_off(DisasContext *ctx, int num)
 }
 
 static void gen_log_vreg_write(DisasContext *ctx, intptr_t srcoff, int num,
-                               VRegWriteType type, int slot_num,
+                               VRegWriteType type)
-                               bool is_predicated)
 {
-    TCGLabel *label_end = NULL;
     intptr_t dstoff;
 
-    if (is_predicated) {
-        TCGv cancelled = tcg_temp_new();
-        label_end = gen_new_label();
-
-        /* Don't do anything if the slot was cancelled */
-        tcg_gen_extract_tl(cancelled, hex_slot_cancelled, slot_num, 1);
-        tcg_gen_brcondi_tl(TCG_COND_NE, cancelled, 0, label_end);
-    }
-
     if (type != EXT_TMP) {
         dstoff = ctx_future_vreg_off(ctx, num, 1, true);
         tcg_gen_gvec_mov(MO_64, dstoff, srcoff,
                          sizeof(MMVector), sizeof(MMVector));
-        tcg_gen_ori_tl(hex_VRegs_updated, hex_VRegs_updated, 1 << num);
     } else {
         dstoff = ctx_tmp_vreg_off(ctx, num, 1, false);
         tcg_gen_gvec_mov(MO_64, dstoff, srcoff,
                          sizeof(MMVector), sizeof(MMVector));
     }
-
-    if (is_predicated) {
-        gen_set_label(label_end);
-    }
 }
 
 static void gen_log_vreg_write_pair(DisasContext *ctx, intptr_t srcoff, int num,
-                                    VRegWriteType type, int slot_num,
+                                    VRegWriteType type)
-                                    bool is_predicated)
 {
-    gen_log_vreg_write(ctx, srcoff, num ^ 0, type, slot_num, is_predicated);
+    gen_log_vreg_write(ctx, srcoff, num ^ 0, type);
     srcoff += sizeof(MMVector);
-    gen_log_vreg_write(ctx, srcoff, num ^ 1, type, slot_num, is_predicated);
+    gen_log_vreg_write(ctx, srcoff, num ^ 1, type);
 }
 
-static void gen_log_qreg_write(intptr_t srcoff, int num, int vnew,
+static intptr_t get_result_qreg(DisasContext *ctx, int qnum)
-                               int slot_num, bool is_predicated)
 {
-    TCGLabel *label_end = NULL;
+    return  offsetof(CPUHexagonState, future_QRegs[qnum]);
-    intptr_t dstoff;
-
-    if (is_predicated) {
-        TCGv cancelled = tcg_temp_new();
-        label_end = gen_new_label();
-
-        /* Don't do anything if the slot was cancelled */
-        tcg_gen_extract_tl(cancelled, hex_slot_cancelled, slot_num, 1);
-        tcg_gen_brcondi_tl(TCG_COND_NE, cancelled, 0, label_end);
-    }
-
-    dstoff = offsetof(CPUHexagonState, future_QRegs[num]);
-    tcg_gen_gvec_mov(MO_64, dstoff, srcoff, sizeof(MMQReg), sizeof(MMQReg));
-
-    if (is_predicated) {
-        tcg_gen_ori_tl(hex_QRegs_updated, hex_QRegs_updated, 1 << num);
-        gen_set_label(label_end);
-    }
 }
 
 static void gen_vreg_load(DisasContext *ctx, intptr_t dstoff, TCGv src,

target/hexagon/hex_common.py

@@ -1,7 +1,7 @@
 #!/usr/bin/env python3
 
 ##
-##  Copyright(c) 2019-2022 Qualcomm Innovation Center, Inc. All Rights Reserved.
+##  Copyright(c) 2019-2023 Qualcomm Innovation Center, Inc. All Rights Reserved.
 ##
 ##  This program is free software; you can redistribute it and/or modify
 ##  it under the terms of the GNU General Public License as published by
@@ -89,6 +89,7 @@ def calculate_attribs():
     add_qemu_macro_attrib('fWRITE_P3', 'A_WRITES_PRED_REG')
     add_qemu_macro_attrib('fSET_OVERFLOW', 'A_IMPLICIT_WRITES_USR')
     add_qemu_macro_attrib('fSET_LPCFG', 'A_IMPLICIT_WRITES_USR')
+    add_qemu_macro_attrib('fLOAD', 'A_SCALAR_LOAD')
     add_qemu_macro_attrib('fSTORE', 'A_SCALAR_STORE')
 
     # Recurse down macros, find attributes from sub-macros
@@ -236,6 +237,13 @@ def helper_needs_next_PC(tag):
 def need_pkt_has_multi_cof(tag):
     return 'A_COF' in attribdict[tag]
 
+def need_condexec_reg(tag, regs):
+    if 'A_CONDEXEC' in attribdict[tag]:
+        for regtype, regid, toss, numregs in regs:
+            if is_writeonly(regid) and not is_hvx_reg(regtype):
+                return True
+    return False
+
 def skip_qemu_helper(tag):
     return tag in overrides.keys()
 

target/hexagon/idef-parser/idef-parser.h

@@ -82,7 +82,6 @@ enum ImmUnionTag {
     VALUE,
     QEMU_TMP,
     IMM_PC,
-    IMM_NPC,
     IMM_CONSTEXT,
 };
 

target/hexagon/idef-parser/idef-parser.lex

@@ -5,7 +5,7 @@
 
 %{
 /*
- *  Copyright(c) 2019-2022 rev.ng Labs Srl. All Rights Reserved.
+ *  Copyright(c) 2019-2023 rev.ng Labs Srl. All Rights Reserved.
  *
  *  This program is free software; you can redistribute it and/or modify
  *  it under the terms of the GNU General Public License as published by
@@ -140,8 +140,6 @@ STRING_LIT               \"(\\.|[^"\\])*\"
                            yylval->rvalue.is_dotnew = true;
                            yylval->rvalue.signedness = SIGNED;
                            return PRED; }
-"IV1DEAD()"              |
-"fPAUSE(uiV);"           { return ';'; }
 "+="                     { return INC; }
 "-="                     { return DEC; }
 "++"                     { return PLUSPLUS; }
@@ -159,9 +157,8 @@ STRING_LIT               \"(\\.|[^"\\])*\"
 "else"                   { return ELSE; }
 "for"                    { return FOR; }
 "fREAD_IREG"             { return ICIRC; }
-"fPART1"                 { return PART1; }
 "if"                     { return IF; }
-"fFRAME_SCRAMBLE"        { return FSCR; }
+"fFRAME_SCRAMBLE"        |
 "fFRAME_UNSCRAMBLE"      { return FSCR; }
 "fFRAMECHECK"            { return FCHK; }
 "Constant_extended"      { return CONSTEXT; }
@@ -312,14 +309,10 @@ STRING_LIT               \"(\\.|[^"\\])*\"
 "(unsigned int)"         { yylval->cast.bit_width = 32;
                            yylval->cast.signedness = UNSIGNED;
                            return CAST; }
-"fREAD_PC()"             |
+"fREAD_PC()"             { return PC; }
-"PC"                     { return PC; }
-"fREAD_NPC()"            |
-"NPC"                    { return NPC; }
-"fGET_LPCFG"             |
 "USR.LPCFG"              { return LPCFG; }
 "LOAD_CANCEL(EA)"        { return LOAD_CANCEL; }
-"STORE_CANCEL(EA)"       |
+"STORE_CANCEL(EA)"       { return STORE_CANCEL; }
 "CANCEL"                 { return CANCEL; }
 "N"{LOWER_ID}"N"         { yylval->rvalue.type = REGISTER_ARG;
                            yylval->rvalue.reg.type = DOTNEW;
@@ -360,14 +353,6 @@ STRING_LIT               \"(\\.|[^"\\])*\"
                            yylval->rvalue.bit_width = 32;
                            yylval->rvalue.signedness = UNSIGNED;
                            return REG; }
-"fREAD_LC"[01]           { yylval->rvalue.type = REGISTER;
-                           yylval->rvalue.reg.type = CONTROL;
-                           yylval->rvalue.reg.id = HEX_REG_LC0
-                                                 + (yytext[8] - '0') * 2;
-                           yylval->rvalue.reg.bit_width = 32;
-                           yylval->rvalue.bit_width = 32;
-                           yylval->rvalue.signedness = UNSIGNED;
-                           return REG; }
 "LC"[01]                 { yylval->rvalue.type = REGISTER;
                            yylval->rvalue.reg.type = CONTROL;
                            yylval->rvalue.reg.id = HEX_REG_LC0
@@ -376,14 +361,6 @@ STRING_LIT               \"(\\.|[^"\\])*\"
                            yylval->rvalue.bit_width = 32;
                            yylval->rvalue.signedness = UNSIGNED;
                            return REG; }
-"fREAD_SA"[01]           { yylval->rvalue.type = REGISTER;
-                           yylval->rvalue.reg.type = CONTROL;
-                           yylval->rvalue.reg.id = HEX_REG_SA0
-                                                 + (yytext[8] - '0') * 2;
-                           yylval->rvalue.reg.bit_width = 32;
-                           yylval->rvalue.bit_width = 32;
-                           yylval->rvalue.signedness = UNSIGNED;
-                           return REG; }
 "SA"[01]                 { yylval->rvalue.type = REGISTER;
                            yylval->rvalue.reg.type = CONTROL;
                            yylval->rvalue.reg.id = HEX_REG_SA0

target/hexagon/idef-parser/idef-parser.y

@@ -1,6 +1,6 @@
 %{
 /*
- *  Copyright(c) 2019-2022 rev.ng Labs Srl. All Rights Reserved.
+ *  Copyright(c) 2019-2023 rev.ng Labs Srl. All Rights Reserved.
  *
  *  This program is distributed in the hope that it will be useful,
  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
@@ -52,8 +52,8 @@
 %token IN INAME VAR
 %token ABS CROUND ROUND CIRCADD COUNTONES INC DEC ANDA ORA XORA PLUSPLUS ASL
 %token ASR LSR EQ NEQ LTE GTE MIN MAX ANDL FOR ICIRC IF MUN FSCR FCHK SXT
-%token ZXT CONSTEXT LOCNT BREV SIGN LOAD STORE PC NPC LPCFG
+%token ZXT CONSTEXT LOCNT BREV SIGN LOAD STORE PC LPCFG
-%token LOAD_CANCEL CANCEL IDENTITY PART1 ROTL INSBITS SETBITS EXTRANGE
+%token LOAD_CANCEL STORE_CANCEL CANCEL IDENTITY ROTL INSBITS SETBITS EXTRANGE
 %token CAST4_8U FAIL CARRY_FROM_ADD ADDSAT64 LSBNEW
 %token TYPE_SIZE_T TYPE_INT TYPE_SIGNED TYPE_UNSIGNED TYPE_LONG
 
@@ -336,15 +336,6 @@ assign_statement : lvalue '=' rvalue
                        OUT(c, &@1, &$1, " = ", &$3, ";\n");
                        $$ = $1;
                    }
-                 | PC '=' rvalue
-                   {
-                       @1.last_column = @3.last_column;
-                       yyassert(c, &@1, !is_inside_ternary(c),
-                                "Assignment side-effect not modeled!");
-                       $3 = gen_rvalue_truncate(c, &@1, &$3);
-                       $3 = rvalue_materialize(c, &@1, &$3);
-                       OUT(c, &@1, "gen_write_new_pc(", &$3, ");\n");
-                   }
                  | LOAD '(' IMM ',' IMM ',' SIGN ',' var ',' lvalue ')'
                    {
                        @1.last_column = @12.last_column;
@@ -412,7 +403,6 @@ control_statement : frame_check
                   | cancel_statement
                   | if_statement
                   | for_statement
-                  | fpart1_statement
                   ;
 
 frame_check : FCHK '(' rvalue ',' rvalue ')' ';'
@@ -422,10 +412,11 @@ cancel_statement : LOAD_CANCEL
                    {
                        gen_load_cancel(c, &@1);
                    }
-                 | CANCEL
+                 | STORE_CANCEL
                    {
                        gen_cancel(c, &@1);
                    }
+                 | CANCEL
                  ;
 
 if_statement : if_stmt
@@ -462,17 +453,6 @@ for_statement : FOR '(' IMM '=' IMM ';' IMM '<' IMM ';' IMM PLUSPLUS ')'
                 }
               ;
 
-fpart1_statement : PART1
-                   {
-                       OUT(c, &@1, "if (insn->part1) {\n");
-                   }
-                   '(' statements ')'
-                   {
-                       @1.last_column = @3.last_column;
-                       OUT(c, &@1, "return; }\n");
-                   }
-                 ;
-
 if_stmt : IF '(' rvalue ')'
           {
               @1.last_column = @3.last_column;
@@ -512,20 +492,6 @@ rvalue : FAIL
              rvalue.signedness = UNSIGNED;
              $$ = rvalue;
          }
-       | NPC
-         {
-             /*
-              * NPC is only read from CALLs, so we can hardcode it
-              * at translation time
-              */
-             HexValue rvalue;
-             memset(&rvalue, 0, sizeof(HexValue));
-             rvalue.type = IMMEDIATE;
-             rvalue.imm.type = IMM_NPC;
-             rvalue.bit_width = 32;
-             rvalue.signedness = UNSIGNED;
-             $$ = rvalue;
-         }
        | CONSTEXT
          {
              HexValue rvalue;
@@ -781,11 +747,6 @@ rvalue : FAIL
              /* Ones count */
              $$ = gen_ctpop_op(c, &@1, &$3);
          }
-       | LPCFG
-         {
-             $$ = gen_tmp(c, &@1, 32, UNSIGNED);
-             OUT(c, &@1, "GET_USR_FIELD(USR_LPCFG, ", &$$, ");\n");
-         }
        | EXTRACT '(' rvalue ',' rvalue ')'
          {
              @1.last_column = @6.last_column;

target/hexagon/idef-parser/macros.inc

@@ -97,16 +97,8 @@
 #define fWRITE_LR(A) (LR = A)
 #define fWRITE_FP(A) (FP = A)
 #define fWRITE_SP(A) (SP = A)
-/*
- * Note: There is a rule in the parser that matches `PC = ...` and emits
- * a call to `gen_write_new_pc`. We need to call `gen_write_new_pc` to
- * get the correct semantics when there are multiple stores in a packet.
- */
-#define fBRANCH(LOC, TYPE) (PC = LOC)
-#define fJUMPR(REGNO, TARGET, TYPE) (PC = TARGET)
 #define fWRITE_LOOP_REGS0(START, COUNT) SA0 = START; (LC0 = COUNT)
 #define fWRITE_LOOP_REGS1(START, COUNT) SA1 = START; (LC1 = COUNT)
-#define fWRITE_LC0(VAL) (LC0 = VAL)
 #define fWRITE_LC1(VAL) (LC1 = VAL)
 #define fSET_LPCFG(VAL) (USR.LPCFG = VAL)
 #define fWRITE_P0(VAL) P0 = VAL;
@@ -121,7 +113,6 @@
 #define fEA_GPI(IMM) (EA = fREAD_GP() + IMM)
 #define fPM_I(REG, IMM) (REG = REG + IMM)
 #define fPM_M(REG, MVAL) (REG = REG + MVAL)
-#define fWRITE_NPC(VAL) (PC = VAL)
 
 /* Unary operators */
 #define fROUND(A) (A + 0x8000)

target/hexagon/idef-parser/parser-helpers.c

@@ -1,5 +1,5 @@
 /*
- *  Copyright(c) 2019-2022 rev.ng Labs Srl. All Rights Reserved.
+ *  Copyright(c) 2019-2023 rev.ng Labs Srl. All Rights Reserved.
  *
  *  This program is free software; you can redistribute it and/or modify
  *  it under the terms of the GNU General Public License as published by
@@ -185,9 +185,6 @@ void imm_print(Context *c, YYLTYPE *locp, HexImm *imm)
     case IMM_PC:
         EMIT(c, "ctx->base.pc_next");
         break;
-    case IMM_NPC:
-        EMIT(c, "ctx->npc");
-        break;
     case IMM_CONSTEXT:
         EMIT(c, "insn->extension_valid");
         break;
@@ -1323,10 +1320,6 @@ void gen_write_reg(Context *c, YYLTYPE *locp, HexValue *reg, HexValue *value)
         locp,
         "gen_log_reg_write(", &reg->reg.id, ", ",
         &value_m, ");\n");
-    OUT(c,
-        locp,
-        "ctx_log_reg_write(ctx, ", &reg->reg.id,
-        ");\n");
 }
 
 void gen_assign(Context *c,
@@ -1675,9 +1668,7 @@ void gen_inst_init_args(Context *c, YYLTYPE *locp)
     for (unsigned i = 0; i < c->inst.init_list->len; i++) {
         HexValue *val = &g_array_index(c->inst.init_list, HexValue, i);
         if (val->type == REGISTER_ARG) {
-            char reg_id[5];
+            /* Nothing to do here */
-            reg_compose(c, locp, &val->reg, reg_id);
-            EMIT_HEAD(c, "tcg_gen_movi_i%u(%s, 0);\n", val->bit_width, reg_id);
         } else if (val->type == PREDICATE) {
             char suffix = val->is_dotnew ? 'N' : 'V';
             EMIT_HEAD(c, "tcg_gen_movi_i%u(P%c%c, 0);\n", val->bit_width,
@@ -1722,13 +1713,10 @@ void gen_pred_assign(Context *c, YYLTYPE *locp, HexValue *left_pred,
         *left_pred = gen_tmp(c, locp, 32, UNSIGNED);
     }
     /* Extract first 8 bits, and store new predicate value */
-    OUT(c, locp, "tcg_gen_mov_i32(", left_pred, ", ", &r, ");\n");
+    OUT(c, locp, "tcg_gen_andi_i32(", left_pred, ", ", &r, ", 0xff);\n");
-    OUT(c, locp, "tcg_gen_andi_i32(", left_pred, ", ", left_pred,
-        ", 0xff);\n");
     if (is_direct) {
         OUT(c, locp, "gen_log_pred_write(ctx, ", pred_id, ", ", left_pred,
             ");\n");
-        OUT(c, locp, "ctx_log_pred_write(ctx, ", pred_id, ");\n");
     }
 }
 
@@ -1739,7 +1727,6 @@ void gen_cancel(Context *c, YYLTYPE *locp)
 
 void gen_load_cancel(Context *c, YYLTYPE *locp)
 {
-    gen_cancel(c, locp);
     OUT(c, locp, "if (insn->slot == 0 && pkt->pkt_has_store_s1) {\n");
     OUT(c, locp, "ctx->s1_store_processed = false;\n");
     OUT(c, locp, "process_store(ctx, 1);\n");

target/hexagon/macros.h

@@ -1,5 +1,5 @@
 /*
- *  Copyright(c) 2019-2022 Qualcomm Innovation Center, Inc. All Rights Reserved.
+ *  Copyright(c) 2019-2023 Qualcomm Innovation Center, Inc. All Rights Reserved.
  *
  *  This program is free software; you can redistribute it and/or modify
  *  it under the terms of the GNU General Public License as published by
@@ -205,26 +205,11 @@ static inline void gen_cancel(uint32_t slot)
 
 #define CANCEL gen_cancel(slot);
 #else
-#define CANCEL cancel_slot(env, slot)
+#define CANCEL do { } while (0)
 #endif
 
 #define LOAD_CANCEL(EA) do { CANCEL; } while (0)
 
-#ifdef QEMU_GENERATE
-static inline void gen_pred_cancel(TCGv pred, uint32_t slot_num)
- {
-    TCGv slot_mask = tcg_temp_new();
-    TCGv tmp = tcg_temp_new();
-    TCGv zero = tcg_constant_tl(0);
-    tcg_gen_ori_tl(slot_mask, hex_slot_cancelled, 1 << slot_num);
-    tcg_gen_andi_tl(tmp, pred, 1);
-    tcg_gen_movcond_tl(TCG_COND_EQ, hex_slot_cancelled, tmp, zero,
-                       slot_mask, hex_slot_cancelled);
-}
-#define PRED_LOAD_CANCEL(PRED, EA) \
-    gen_pred_cancel(PRED, insn->is_endloop ? 4 : insn->slot)
-#endif
-
 #define STORE_CANCEL(EA) { env->slot_cancelled |= (1 << slot); }
 
 #define fMAX(A, B) (((A) > (B)) ? (A) : (B))
@@ -415,16 +400,6 @@ static inline TCGv gen_read_ireg(TCGv result, TCGv val, int shift)
 #define fBRANCH(LOC, TYPE)          fWRITE_NPC(LOC)
 #define fJUMPR(REGNO, TARGET, TYPE) fBRANCH(TARGET, COF_TYPE_JUMPR)
 #define fHINTJR(TARGET) { /* Not modelled in qemu */}
-#define fCALL(A) \
-    do { \
-        fWRITE_LR(fREAD_NPC()); \
-        fBRANCH(A, COF_TYPE_CALL); \
-    } while (0)
-#define fCALLR(A) \
-    do { \
-        fWRITE_LR(fREAD_NPC()); \
-        fBRANCH(A, COF_TYPE_CALLR); \
-    } while (0)
 #define fWRITE_LOOP_REGS0(START, COUNT) \
     do { \
         WRITE_RREG(HEX_REG_LC0, COUNT);  \

target/hexagon/meson.build

@@ -1,5 +1,5 @@
 ##
-##  Copyright(c) 2020-2021 Qualcomm Innovation Center, Inc. All Rights Reserved.
+##  Copyright(c) 2020-2023 Qualcomm Innovation Center, Inc. All Rights Reserved.
 ##
 ##  This program is free software; you can redistribute it and/or modify
 ##  it under the terms of the GNU General Public License as published by
@@ -276,4 +276,13 @@ tcg_funcs_generated = custom_target(
 )
 hexagon_ss.add(tcg_funcs_generated)
 
+analyze_funcs_generated = custom_target(
+    'analyze_funcs_generated.c.inc',
+    output: 'analyze_funcs_generated.c.inc',
+    depends: helper_dep,
+    depend_files: [hex_common_py, attribs_def, gen_tcg_h, gen_tcg_hvx_h],
+    command: [python, files('gen_analyze_funcs.py'), helper_in, '@OUTPUT@'],
+)
+hexagon_ss.add(analyze_funcs_generated)
+
 target_arch += {'hexagon': hexagon_ss}

target/hexagon/op_helper.c

@@ -1,5 +1,5 @@
 /*
- *  Copyright(c) 2019-2022 Qualcomm Innovation Center, Inc. All Rights Reserved.
+ *  Copyright(c) 2019-2023 Qualcomm Innovation Center, Inc. All Rights Reserved.
  *
  *  This program is free software; you can redistribute it and/or modify
  *  it under the terms of the GNU General Public License as published by
@@ -30,6 +30,7 @@
 #include "mmvec/mmvec.h"
 #include "mmvec/macros.h"
 #include "op_helper.h"
+#include "translate.h"
 
 #define SF_BIAS        127
 #define SF_MANTBITS    23
@@ -105,30 +106,6 @@ void log_store64(CPUHexagonState *env, target_ulong addr,
     env->mem_log_stores[slot].data64 = val;
 }
 
-void write_new_pc(CPUHexagonState *env, bool pkt_has_multi_cof,
-                         target_ulong addr)
-{
-    HEX_DEBUG_LOG("write_new_pc(0x" TARGET_FMT_lx ")\n", addr);
-
-    if (pkt_has_multi_cof) {
-        /*
-         * If more than one branch is taken in a packet, only the first one
-         * is actually done.
-         */
-        if (env->branch_taken) {
-            HEX_DEBUG_LOG("INFO: multiple branches taken in same packet, "
-                          "ignoring the second one\n");
-        } else {
-            fCHECK_PCALIGN(addr);
-            env->gpr[HEX_REG_PC] = addr;
-            env->branch_taken = 1;
-        }
-    } else {
-        fCHECK_PCALIGN(addr);
-        env->gpr[HEX_REG_PC] = addr;
-    }
-}
-
 /* Handy place to set a breakpoint */
 void HELPER(debug_start_packet)(CPUHexagonState *env)
 {
@@ -439,9 +416,10 @@ int32_t HELPER(vacsh_pred)(CPUHexagonState *env,
     return PeV;
 }
 
-static void probe_store(CPUHexagonState *env, int slot, int mmu_idx)
+static void probe_store(CPUHexagonState *env, int slot, int mmu_idx,
+                        bool is_predicated)
 {
-    if (!(env->slot_cancelled & (1 << slot))) {
+    if (!is_predicated || !(env->slot_cancelled & (1 << slot))) {
         size1u_t width = env->mem_log_stores[slot].width;
         target_ulong va = env->mem_log_stores[slot].va;
         uintptr_t ra = GETPC();
@@ -461,9 +439,12 @@ void HELPER(probe_noshuf_load)(CPUHexagonState *env, target_ulong va,
 }
 
 /* Called during packet commit when there are two scalar stores */
-void HELPER(probe_pkt_scalar_store_s0)(CPUHexagonState *env, int mmu_idx)
+void HELPER(probe_pkt_scalar_store_s0)(CPUHexagonState *env, int args)
 {
-    probe_store(env, 0, mmu_idx);
+    int mmu_idx = FIELD_EX32(args, PROBE_PKT_SCALAR_STORE_S0, MMU_IDX);
+    bool is_predicated =
+        FIELD_EX32(args, PROBE_PKT_SCALAR_STORE_S0, IS_PREDICATED);
+    probe_store(env, 0, mmu_idx, is_predicated);
 }
 
 void HELPER(probe_hvx_stores)(CPUHexagonState *env, int mmu_idx)
@@ -510,15 +491,18 @@ void HELPER(probe_hvx_stores)(CPUHexagonState *env, int mmu_idx)
 void HELPER(probe_pkt_scalar_hvx_stores)(CPUHexagonState *env, int mask,
                                          int mmu_idx)
 {
-    bool has_st0        = (mask >> 0) & 1;
+    bool has_st0 = FIELD_EX32(mask, PROBE_PKT_SCALAR_HVX_STORES, HAS_ST0);
-    bool has_st1        = (mask >> 1) & 1;
+    bool has_st1 = FIELD_EX32(mask, PROBE_PKT_SCALAR_HVX_STORES, HAS_ST1);
-    bool has_hvx_stores = (mask >> 2) & 1;
+    bool has_hvx_stores =
+        FIELD_EX32(mask, PROBE_PKT_SCALAR_HVX_STORES, HAS_HVX_STORES);
+    bool s0_is_pred = FIELD_EX32(mask, PROBE_PKT_SCALAR_HVX_STORES, S0_IS_PRED);
+    bool s1_is_pred = FIELD_EX32(mask, PROBE_PKT_SCALAR_HVX_STORES, S1_IS_PRED);
 
     if (has_st0) {
-        probe_store(env, 0, mmu_idx);
+        probe_store(env, 0, mmu_idx, s0_is_pred);
     }
     if (has_st1) {
-        probe_store(env, 1, mmu_idx);
+        probe_store(env, 1, mmu_idx, s1_is_pred);
     }
     if (has_hvx_stores) {
         HELPER(probe_hvx_stores)(env, mmu_idx);
@@ -1193,7 +1177,7 @@ float32 HELPER(sffms)(CPUHexagonState *env, float32 RxV,
 {
     float32 neg_RsV;
     arch_fpop_start(env);
-    neg_RsV = float32_sub(float32_zero, RsV, &env->fp_status);
+    neg_RsV = float32_set_sign(RsV, float32_is_neg(RsV) ? 0 : 1);
     RxV = internal_fmafx(neg_RsV, RtV, RxV, 0, &env->fp_status);
     arch_fpop_end(env);
     return RxV;
@@ -1468,12 +1452,6 @@ void HELPER(vwhist128qm)(CPUHexagonState *env, int32_t uiV)
     }
 }
 
-void cancel_slot(CPUHexagonState *env, uint32_t slot)
-{
-    HEX_DEBUG_LOG("Slot %d cancelled\n", slot);
-    env->slot_cancelled |= (1 << slot);
-}
-
 /* These macros can be referenced in the generated helper functions */
 #define warn(...) /* Nothing */
 #define fatal(...) g_assert_not_reached();

target/hexagon/op_helper.h

@@ -1,5 +1,5 @@
 /*
- *  Copyright(c) 2019-2021 Qualcomm Innovation Center, Inc. All Rights Reserved.
+ *  Copyright(c) 2019-2023 Qualcomm Innovation Center, Inc. All Rights Reserved.
  *
  *  This program is free software; you can redistribute it and/or modify
  *  it under the terms of the GNU General Public License as published by
@@ -19,7 +19,6 @@
 #define HEXAGON_OP_HELPER_H
 
 /* Misc functions */
-void cancel_slot(CPUHexagonState *env, uint32_t slot);
 void write_new_pc(CPUHexagonState *env, bool pkt_has_multi_cof, target_ulong addr);
 
 uint8_t mem_load1(CPUHexagonState *env, uint32_t slot, target_ulong vaddr);

target/hexagon/translate.c

@@ -1,5 +1,5 @@
 /*
- *  Copyright(c) 2019-2022 Qualcomm Innovation Center, Inc. All Rights Reserved.
+ *  Copyright(c) 2019-2023 Qualcomm Innovation Center, Inc. All Rights Reserved.
  *
  *  This program is free software; you can redistribute it and/or modify
  *  it under the terms of the GNU General Public License as published by
@@ -29,6 +29,15 @@
 #include "translate.h"
 #include "printinsn.h"
 
+#include "analyze_funcs_generated.c.inc"
+
+typedef void (*AnalyzeInsn)(DisasContext *ctx);
+static const AnalyzeInsn opcode_analyze[XX_LAST_OPCODE] = {
+#define OPCODE(X)    [X] = analyze_##X
+#include "opcodes_def_generated.h.inc"
+#undef OPCODE
+};
+
 TCGv hex_gpr[TOTAL_PER_THREAD_REGS];
 TCGv hex_pred[NUM_PREGS];
 TCGv hex_this_PC;
@@ -47,8 +56,6 @@ TCGv hex_dczero_addr;
 TCGv hex_llsc_addr;
 TCGv hex_llsc_val;
 TCGv_i64 hex_llsc_val_i64;
-TCGv hex_VRegs_updated;
-TCGv hex_QRegs_updated;
 TCGv hex_vstore_addr[VSTORES_MAX];
 TCGv hex_vstore_size[VSTORES_MAX];
 TCGv hex_vstore_pending[VSTORES_MAX];
@@ -239,7 +246,15 @@ static bool check_for_attrib(Packet *pkt, int attrib)
 
 static bool need_slot_cancelled(Packet *pkt)
 {
-    return check_for_attrib(pkt, A_CONDEXEC);
+    /* We only need slot_cancelled for conditional store instructions */
+    for (int i = 0; i < pkt->num_insns; i++) {
+        uint16_t opcode = pkt->insn[i].opcode;
+        if (GET_ATTRIB(opcode, A_CONDEXEC) &&
+            GET_ATTRIB(opcode, A_SCALAR_STORE)) {
+            return true;
+        }
+    }
+    return false;
 }
 
 static bool need_pred_written(Packet *pkt)
@@ -265,6 +280,77 @@ static bool need_next_PC(DisasContext *ctx)
     return false;
 }
 
+/*
+ * The opcode_analyze functions mark most of the writes in a packet
+ * However, there are some implicit writes marked as attributes
+ * of the applicable instructions.
+ */
+static void mark_implicit_reg_write(DisasContext *ctx, int attrib, int rnum)
+{
+    uint16_t opcode = ctx->insn->opcode;
+    if (GET_ATTRIB(opcode, attrib)) {
+        /*
+         * USR is used to set overflow and FP exceptions,
+         * so treat it as conditional
+         */
+        bool is_predicated = GET_ATTRIB(opcode, A_CONDEXEC) ||
+                             rnum == HEX_REG_USR;
+
+        /* LC0/LC1 is conditionally written by endloop instructions */
+        if ((rnum == HEX_REG_LC0 || rnum == HEX_REG_LC1) &&
+            (opcode == J2_endloop0 ||
+             opcode == J2_endloop1 ||
+             opcode == J2_endloop01)) {
+            is_predicated = true;
+        }
+
+        ctx_log_reg_write(ctx, rnum, is_predicated);
+    }
+}
+
+static void mark_implicit_reg_writes(DisasContext *ctx)
+{
+    mark_implicit_reg_write(ctx, A_IMPLICIT_WRITES_FP,  HEX_REG_FP);
+    mark_implicit_reg_write(ctx, A_IMPLICIT_WRITES_SP,  HEX_REG_SP);
+    mark_implicit_reg_write(ctx, A_IMPLICIT_WRITES_LR,  HEX_REG_LR);
+    mark_implicit_reg_write(ctx, A_IMPLICIT_WRITES_LC0, HEX_REG_LC0);
+    mark_implicit_reg_write(ctx, A_IMPLICIT_WRITES_SA0, HEX_REG_SA0);
+    mark_implicit_reg_write(ctx, A_IMPLICIT_WRITES_LC1, HEX_REG_LC1);
+    mark_implicit_reg_write(ctx, A_IMPLICIT_WRITES_SA1, HEX_REG_SA1);
+    mark_implicit_reg_write(ctx, A_IMPLICIT_WRITES_USR, HEX_REG_USR);
+    mark_implicit_reg_write(ctx, A_FPOP, HEX_REG_USR);
+}
+
+static void mark_implicit_pred_write(DisasContext *ctx, int attrib, int pnum)
+{
+    if (GET_ATTRIB(ctx->insn->opcode, attrib)) {
+        ctx_log_pred_write(ctx, pnum);
+    }
+}
+
+static void mark_implicit_pred_writes(DisasContext *ctx)
+{
+    mark_implicit_pred_write(ctx, A_IMPLICIT_WRITES_P0, 0);
+    mark_implicit_pred_write(ctx, A_IMPLICIT_WRITES_P1, 1);
+    mark_implicit_pred_write(ctx, A_IMPLICIT_WRITES_P2, 2);
+    mark_implicit_pred_write(ctx, A_IMPLICIT_WRITES_P3, 3);
+}
+
+static void analyze_packet(DisasContext *ctx)
+{
+    Packet *pkt = ctx->pkt;
+    ctx->need_pkt_has_store_s1 = false;
+    for (int i = 0; i < pkt->num_insns; i++) {
+        Insn *insn = &pkt->insn[i];
+        ctx->insn = insn;
+        if (opcode_analyze[insn->opcode]) {
+            opcode_analyze[insn->opcode](ctx);
+        }
+        mark_implicit_reg_writes(ctx);
+        mark_implicit_pred_writes(ctx);
+    }
+}
+
 static void gen_start_packet(DisasContext *ctx)
 {
     Packet *pkt = ctx->pkt;
@@ -275,6 +361,7 @@ static void gen_start_packet(DisasContext *ctx)
     ctx->next_PC = next_PC;
     ctx->reg_log_idx = 0;
     bitmap_zero(ctx->regs_written, TOTAL_PER_THREAD_REGS);
+    bitmap_zero(ctx->predicated_regs, TOTAL_PER_THREAD_REGS);
     ctx->preg_log_idx = 0;
     bitmap_zero(ctx->pregs_written, NUM_PREGS);
     ctx->future_vregs_idx = 0;
@@ -283,14 +370,27 @@ static void gen_start_packet(DisasContext *ctx)
     bitmap_zero(ctx->vregs_updated_tmp, NUM_VREGS);
     bitmap_zero(ctx->vregs_updated, NUM_VREGS);
     bitmap_zero(ctx->vregs_select, NUM_VREGS);
+    bitmap_zero(ctx->predicated_future_vregs, NUM_VREGS);
+    bitmap_zero(ctx->predicated_tmp_vregs, NUM_VREGS);
     ctx->qreg_log_idx = 0;
     for (i = 0; i < STORES_MAX; i++) {
         ctx->store_width[i] = 0;
     }
-    tcg_gen_movi_tl(hex_pkt_has_store_s1, pkt->pkt_has_store_s1);
     ctx->s1_store_processed = false;
     ctx->pre_commit = true;
 
+    analyze_packet(ctx);
+
+    if (ctx->need_pkt_has_store_s1) {
+        tcg_gen_movi_tl(hex_pkt_has_store_s1, pkt->pkt_has_store_s1);
+    }
+
+    /*
+     * pregs_written is used both in the analyze phase as well as the code
+     * gen phase, so clear it again.
+     */
+    bitmap_zero(ctx->pregs_written, NUM_PREGS);
+
     if (HEX_DEBUG) {
         /* Handy place to set a breakpoint before the packet executes */
         gen_helper_debug_start_packet(cpu_env);
@@ -313,9 +413,42 @@ static void gen_start_packet(DisasContext *ctx)
         tcg_gen_movi_tl(hex_pred_written, 0);
     }
 
-    if (pkt->pkt_has_hvx) {
+    /* Preload the predicated registers into hex_new_value[i] */
-        tcg_gen_movi_tl(hex_VRegs_updated, 0);
+    if (!bitmap_empty(ctx->predicated_regs, TOTAL_PER_THREAD_REGS)) {
-        tcg_gen_movi_tl(hex_QRegs_updated, 0);
+        int i = find_first_bit(ctx->predicated_regs, TOTAL_PER_THREAD_REGS);
+        while (i < TOTAL_PER_THREAD_REGS) {
+            tcg_gen_mov_tl(hex_new_value[i], hex_gpr[i]);
+            i = find_next_bit(ctx->predicated_regs, TOTAL_PER_THREAD_REGS,
+                              i + 1);
+        }
+    }
+
+    /* Preload the predicated HVX registers into future_VRegs and tmp_VRegs */
+    if (!bitmap_empty(ctx->predicated_future_vregs, NUM_VREGS)) {
+        int i = find_first_bit(ctx->predicated_future_vregs, NUM_VREGS);
+        while (i < NUM_VREGS) {
+            const intptr_t VdV_off =
+                ctx_future_vreg_off(ctx, i, 1, true);
+            intptr_t src_off = offsetof(CPUHexagonState, VRegs[i]);
+            tcg_gen_gvec_mov(MO_64, VdV_off,
+                             src_off,
+                             sizeof(MMVector),
+                             sizeof(MMVector));
+            i = find_next_bit(ctx->predicated_future_vregs, NUM_VREGS, i + 1);
+        }
+    }
+    if (!bitmap_empty(ctx->predicated_tmp_vregs, NUM_VREGS)) {
+        int i = find_first_bit(ctx->predicated_tmp_vregs, NUM_VREGS);
+        while (i < NUM_VREGS) {
+            const intptr_t VdV_off =
+                ctx_tmp_vreg_off(ctx, i, 1, true);
+            intptr_t src_off = offsetof(CPUHexagonState, VRegs[i]);
+            tcg_gen_gvec_mov(MO_64, VdV_off,
+                             src_off,
+                             sizeof(MMVector),
+                             sizeof(MMVector));
+            i = find_next_bit(ctx->predicated_tmp_vregs, NUM_VREGS, i + 1);
+        }
     }
 }
 
@@ -336,66 +469,6 @@ bool is_gather_store_insn(DisasContext *ctx)
     return false;
 }
 
-/*
- * The LOG_*_WRITE macros mark most of the writes in a packet
- * However, there are some implicit writes marked as attributes
- * of the applicable instructions.
- */
-static void mark_implicit_reg_write(DisasContext *ctx, int attrib, int rnum)
-{
-    uint16_t opcode = ctx->insn->opcode;
-    if (GET_ATTRIB(opcode, attrib)) {
-        /*
-         * USR is used to set overflow and FP exceptions,
-         * so treat it as conditional
-         */
-        bool is_predicated = GET_ATTRIB(opcode, A_CONDEXEC) ||
-                             rnum == HEX_REG_USR;
-
-        /* LC0/LC1 is conditionally written by endloop instructions */
-        if ((rnum == HEX_REG_LC0 || rnum == HEX_REG_LC1) &&
-            (opcode == J2_endloop0 ||
-             opcode == J2_endloop1 ||
-             opcode == J2_endloop01)) {
-            is_predicated = true;
-        }
-
-        if (is_predicated && !is_preloaded(ctx, rnum)) {
-            tcg_gen_mov_tl(hex_new_value[rnum], hex_gpr[rnum]);
-        }
-
-        ctx_log_reg_write(ctx, rnum);
-    }
-}
-
-static void mark_implicit_pred_write(DisasContext *ctx, int attrib, int pnum)
-{
-    if (GET_ATTRIB(ctx->insn->opcode, attrib)) {
-        ctx_log_pred_write(ctx, pnum);
-    }
-}
-
-static void mark_implicit_reg_writes(DisasContext *ctx)
-{
-    mark_implicit_reg_write(ctx, A_IMPLICIT_WRITES_FP,  HEX_REG_FP);
-    mark_implicit_reg_write(ctx, A_IMPLICIT_WRITES_SP,  HEX_REG_SP);
-    mark_implicit_reg_write(ctx, A_IMPLICIT_WRITES_LR,  HEX_REG_LR);
-    mark_implicit_reg_write(ctx, A_IMPLICIT_WRITES_LC0, HEX_REG_LC0);
-    mark_implicit_reg_write(ctx, A_IMPLICIT_WRITES_SA0, HEX_REG_SA0);
-    mark_implicit_reg_write(ctx, A_IMPLICIT_WRITES_LC1, HEX_REG_LC1);
-    mark_implicit_reg_write(ctx, A_IMPLICIT_WRITES_SA1, HEX_REG_SA1);
-    mark_implicit_reg_write(ctx, A_IMPLICIT_WRITES_USR, HEX_REG_USR);
-    mark_implicit_reg_write(ctx, A_FPOP, HEX_REG_USR);
-}
-
-static void mark_implicit_pred_writes(DisasContext *ctx)
-{
-    mark_implicit_pred_write(ctx, A_IMPLICIT_WRITES_P0, 0);
-    mark_implicit_pred_write(ctx, A_IMPLICIT_WRITES_P1, 1);
-    mark_implicit_pred_write(ctx, A_IMPLICIT_WRITES_P2, 2);
-    mark_implicit_pred_write(ctx, A_IMPLICIT_WRITES_P3, 3);
-}
-
 static void mark_store_width(DisasContext *ctx)
 {
     uint16_t opcode = ctx->insn->opcode;
@@ -423,9 +496,7 @@ static void mark_store_width(DisasContext *ctx)
 static void gen_insn(DisasContext *ctx)
 {
     if (ctx->insn->generate) {
-        mark_implicit_reg_writes(ctx);
         ctx->insn->generate(ctx);
-        mark_implicit_pred_writes(ctx);
         mark_store_width(ctx);
     } else {
         gen_exception_end_tb(ctx, HEX_EXCP_INVALID_OPCODE);
@@ -646,65 +717,31 @@ static void gen_commit_hvx(DisasContext *ctx)
     /*
      *    for (i = 0; i < ctx->vreg_log_idx; i++) {
      *        int rnum = ctx->vreg_log[i];
-     *        if (ctx->vreg_is_predicated[i]) {
+     *        env->VRegs[rnum] = env->future_VRegs[rnum];
-     *            if (env->VRegs_updated & (1 << rnum)) {
-     *                env->VRegs[rnum] = env->future_VRegs[rnum];
-     *            }
-     *        } else {
-     *            env->VRegs[rnum] = env->future_VRegs[rnum];
-     *        }
      *    }
      */
     for (i = 0; i < ctx->vreg_log_idx; i++) {
         int rnum = ctx->vreg_log[i];
-        bool is_predicated = ctx->vreg_is_predicated[i];
         intptr_t dstoff = offsetof(CPUHexagonState, VRegs[rnum]);
         intptr_t srcoff = ctx_future_vreg_off(ctx, rnum, 1, false);
         size_t size = sizeof(MMVector);
 
-        if (is_predicated) {
+        tcg_gen_gvec_mov(MO_64, dstoff, srcoff, size, size);
-            TCGv cmp = tcg_temp_new();
-            TCGLabel *label_skip = gen_new_label();
-
-            tcg_gen_andi_tl(cmp, hex_VRegs_updated, 1 << rnum);
-            tcg_gen_brcondi_tl(TCG_COND_EQ, cmp, 0, label_skip);
-            tcg_gen_gvec_mov(MO_64, dstoff, srcoff, size, size);
-            gen_set_label(label_skip);
-        } else {
-            tcg_gen_gvec_mov(MO_64, dstoff, srcoff, size, size);
-        }
     }
 
     /*
      *    for (i = 0; i < ctx->qreg_log_idx; i++) {
      *        int rnum = ctx->qreg_log[i];
-     *        if (ctx->qreg_is_predicated[i]) {
+     *        env->QRegs[rnum] = env->future_QRegs[rnum];
-     *            if (env->QRegs_updated) & (1 << rnum)) {
-     *                env->QRegs[rnum] = env->future_QRegs[rnum];
-     *            }
-     *        } else {
-     *            env->QRegs[rnum] = env->future_QRegs[rnum];
-     *        }
      *    }
      */
     for (i = 0; i < ctx->qreg_log_idx; i++) {
         int rnum = ctx->qreg_log[i];
-        bool is_predicated = ctx->qreg_is_predicated[i];
         intptr_t dstoff = offsetof(CPUHexagonState, QRegs[rnum]);
         intptr_t srcoff = offsetof(CPUHexagonState, future_QRegs[rnum]);
         size_t size = sizeof(MMQReg);
 
-        if (is_predicated) {
+        tcg_gen_gvec_mov(MO_64, dstoff, srcoff, size, size);
-            TCGv cmp = tcg_temp_new();
-            TCGLabel *label_skip = gen_new_label();
-
-            tcg_gen_andi_tl(cmp, hex_QRegs_updated, 1 << rnum);
-            tcg_gen_brcondi_tl(TCG_COND_EQ, cmp, 0, label_skip);
-            tcg_gen_gvec_mov(MO_64, dstoff, srcoff, size, size);
-            gen_set_label(label_skip);
-        } else {
-            tcg_gen_gvec_mov(MO_64, dstoff, srcoff, size, size);
-        }
     }
 
     if (pkt_has_hvx_store(ctx->pkt)) {
@@ -775,13 +812,27 @@ static void gen_commit_packet(DisasContext *ctx)
             TCGv mask_tcgv;
 
             if (has_store_s0) {
-                mask |= (1 << 0);
+                mask =
+                    FIELD_DP32(mask, PROBE_PKT_SCALAR_HVX_STORES, HAS_ST0, 1);
             }
             if (has_store_s1) {
-                mask |= (1 << 1);
+                mask =
+                    FIELD_DP32(mask, PROBE_PKT_SCALAR_HVX_STORES, HAS_ST1, 1);
             }
             if (has_hvx_store) {
-                mask |= (1 << 2);
+                mask =
+                    FIELD_DP32(mask, PROBE_PKT_SCALAR_HVX_STORES,
+                               HAS_HVX_STORES, 1);
+            }
+            if (has_store_s0 && slot_is_predicated(pkt, 0)) {
+                mask =
+                    FIELD_DP32(mask, PROBE_PKT_SCALAR_HVX_STORES,
+                               S0_IS_PRED, 1);
+            }
+            if (has_store_s1 && slot_is_predicated(pkt, 1)) {
+                mask =
+                    FIELD_DP32(mask, PROBE_PKT_SCALAR_HVX_STORES,
+                               S1_IS_PRED, 1);
             }
             mask_tcgv = tcg_constant_tl(mask);
             gen_helper_probe_pkt_scalar_hvx_stores(cpu_env, mask_tcgv, mem_idx);
@@ -791,8 +842,15 @@ static void gen_commit_packet(DisasContext *ctx)
          * process_store_log will execute the slot 1 store first,
          * so we only have to probe the store in slot 0
          */
-        TCGv mem_idx = tcg_constant_tl(ctx->mem_idx);
+        int args = 0;
-        gen_helper_probe_pkt_scalar_store_s0(cpu_env, mem_idx);
+        args =
+            FIELD_DP32(args, PROBE_PKT_SCALAR_STORE_S0, MMU_IDX, ctx->mem_idx);
+        if (slot_is_predicated(pkt, 0)) {
+            args =
+                FIELD_DP32(args, PROBE_PKT_SCALAR_STORE_S0, IS_PREDICATED, 1);
+        }
+        TCGv args_tcgv = tcg_constant_tl(args);
+        gen_helper_probe_pkt_scalar_store_s0(cpu_env, args_tcgv);
     }
 
     process_store_log(ctx);
@@ -1029,10 +1087,6 @@ void hexagon_translate_init(void)
         offsetof(CPUHexagonState, llsc_val), "llsc_val");
     hex_llsc_val_i64 = tcg_global_mem_new_i64(cpu_env,
         offsetof(CPUHexagonState, llsc_val_i64), "llsc_val_i64");
-    hex_VRegs_updated = tcg_global_mem_new(cpu_env,
-        offsetof(CPUHexagonState, VRegs_updated), "VRegs_updated");
-    hex_QRegs_updated = tcg_global_mem_new(cpu_env,
-        offsetof(CPUHexagonState, QRegs_updated), "QRegs_updated");
     for (i = 0; i < STORES_MAX; i++) {
         snprintf(store_addr_names[i], NAME_LEN, "store_addr_%d", i);
         hex_store_addr[i] = tcg_global_mem_new(cpu_env,

target/hexagon/translate.h

@@ -1,5 +1,5 @@
 /*
- *  Copyright(c) 2019-2022 Qualcomm Innovation Center, Inc. All Rights Reserved.
+ *  Copyright(c) 2019-2023 Qualcomm Innovation Center, Inc. All Rights Reserved.
  *
  *  This program is free software; you can redistribute it and/or modify
  *  it under the terms of the GNU General Public License as published by
@@ -38,6 +38,7 @@ typedef struct DisasContext {
     int reg_log[REG_WRITES_MAX];
     int reg_log_idx;
     DECLARE_BITMAP(regs_written, TOTAL_PER_THREAD_REGS);
+    DECLARE_BITMAP(predicated_regs, TOTAL_PER_THREAD_REGS);
     int preg_log[PRED_WRITES_MAX];
     int preg_log_idx;
     DECLARE_BITMAP(pregs_written, NUM_PREGS);
@@ -48,52 +49,54 @@ typedef struct DisasContext {
     int tmp_vregs_idx;
     int tmp_vregs_num[VECTOR_TEMPS_MAX];
     int vreg_log[NUM_VREGS];
-    bool vreg_is_predicated[NUM_VREGS];
     int vreg_log_idx;
     DECLARE_BITMAP(vregs_updated_tmp, NUM_VREGS);
     DECLARE_BITMAP(vregs_updated, NUM_VREGS);
     DECLARE_BITMAP(vregs_select, NUM_VREGS);
+    DECLARE_BITMAP(predicated_future_vregs, NUM_VREGS);
+    DECLARE_BITMAP(predicated_tmp_vregs, NUM_VREGS);
     int qreg_log[NUM_QREGS];
-    bool qreg_is_predicated[NUM_QREGS];
     int qreg_log_idx;
     bool pre_commit;
     TCGCond branch_cond;
     target_ulong branch_dest;
     bool is_tight_loop;
+    bool need_pkt_has_store_s1;
 } DisasContext;
 
-static inline void ctx_log_reg_write(DisasContext *ctx, int rnum)
-{
-    if (test_bit(rnum, ctx->regs_written)) {
-        HEX_DEBUG_LOG("WARNING: Multiple writes to r%d\n", rnum);
-    }
-    ctx->reg_log[ctx->reg_log_idx] = rnum;
-    ctx->reg_log_idx++;
-    set_bit(rnum, ctx->regs_written);
-}
-
-static inline void ctx_log_reg_write_pair(DisasContext *ctx, int rnum)
-{
-    ctx_log_reg_write(ctx, rnum);
-    ctx_log_reg_write(ctx, rnum + 1);
-}
-
 static inline void ctx_log_pred_write(DisasContext *ctx, int pnum)
 {
-    ctx->preg_log[ctx->preg_log_idx] = pnum;
+    if (!test_bit(pnum, ctx->pregs_written)) {
-    ctx->preg_log_idx++;
+        ctx->preg_log[ctx->preg_log_idx] = pnum;
-    set_bit(pnum, ctx->pregs_written);
+        ctx->preg_log_idx++;
+        set_bit(pnum, ctx->pregs_written);
+    }
 }
 
-static inline bool is_preloaded(DisasContext *ctx, int num)
+static inline void ctx_log_reg_write(DisasContext *ctx, int rnum,
+                                     bool is_predicated)
 {
-    return test_bit(num, ctx->regs_written);
+    if (rnum == HEX_REG_P3_0_ALIASED) {
+        for (int i = 0; i < NUM_PREGS; i++) {
+            ctx_log_pred_write(ctx, i);
+        }
+    } else {
+        if (!test_bit(rnum, ctx->regs_written)) {
+            ctx->reg_log[ctx->reg_log_idx] = rnum;
+            ctx->reg_log_idx++;
+            set_bit(rnum, ctx->regs_written);
+        }
+        if (is_predicated) {
+            set_bit(rnum, ctx->predicated_regs);
+        }
+    }
 }
 
-static inline bool is_vreg_preloaded(DisasContext *ctx, int num)
+static inline void ctx_log_reg_write_pair(DisasContext *ctx, int rnum,
+                                          bool is_predicated)
 {
-    return test_bit(num, ctx->vregs_updated) ||
+    ctx_log_reg_write(ctx, rnum, is_predicated);
-           test_bit(num, ctx->vregs_updated_tmp);
+    ctx_log_reg_write(ctx, rnum + 1, is_predicated);
 }
 
 intptr_t ctx_future_vreg_off(DisasContext *ctx, int regnum,
@@ -106,17 +109,25 @@ static inline void ctx_log_vreg_write(DisasContext *ctx,
                                       bool is_predicated)
 {
     if (type != EXT_TMP) {
-        ctx->vreg_log[ctx->vreg_log_idx] = rnum;
+        if (!test_bit(rnum, ctx->vregs_updated)) {
-        ctx->vreg_is_predicated[ctx->vreg_log_idx] = is_predicated;
+            ctx->vreg_log[ctx->vreg_log_idx] = rnum;
-        ctx->vreg_log_idx++;
+            ctx->vreg_log_idx++;
+            set_bit(rnum, ctx->vregs_updated);
+        }
 
         set_bit(rnum, ctx->vregs_updated);
+        if (is_predicated) {
+            set_bit(rnum, ctx->predicated_future_vregs);
+        }
     }
     if (type == EXT_NEW) {
         set_bit(rnum, ctx->vregs_select);
     }
     if (type == EXT_TMP) {
         set_bit(rnum, ctx->vregs_updated_tmp);
+        if (is_predicated) {
+            set_bit(rnum, ctx->predicated_tmp_vregs);
+        }
     }
 }
 
@@ -129,10 +140,9 @@ static inline void ctx_log_vreg_write_pair(DisasContext *ctx,
 }
 
 static inline void ctx_log_qreg_write(DisasContext *ctx,
-                                      int rnum, bool is_predicated)
+                                      int rnum)
 {
     ctx->qreg_log[ctx->qreg_log_idx] = rnum;
-    ctx->qreg_is_predicated[ctx->qreg_log_idx] = is_predicated;
     ctx->qreg_log_idx++;
 }
 
@@ -153,12 +163,20 @@ extern TCGv hex_dczero_addr;
 extern TCGv hex_llsc_addr;
 extern TCGv hex_llsc_val;
 extern TCGv_i64 hex_llsc_val_i64;
-extern TCGv hex_VRegs_updated;
-extern TCGv hex_QRegs_updated;
 extern TCGv hex_vstore_addr[VSTORES_MAX];
 extern TCGv hex_vstore_size[VSTORES_MAX];
 extern TCGv hex_vstore_pending[VSTORES_MAX];
 
 bool is_gather_store_insn(DisasContext *ctx);
 void process_store(DisasContext *ctx, int slot_num);
+
+FIELD(PROBE_PKT_SCALAR_STORE_S0, MMU_IDX,       0, 2)
+FIELD(PROBE_PKT_SCALAR_STORE_S0, IS_PREDICATED, 2, 1)
+
+FIELD(PROBE_PKT_SCALAR_HVX_STORES, HAS_ST0,        0, 1)
+FIELD(PROBE_PKT_SCALAR_HVX_STORES, HAS_ST1,        1, 1)
+FIELD(PROBE_PKT_SCALAR_HVX_STORES, HAS_HVX_STORES, 2, 1)
+FIELD(PROBE_PKT_SCALAR_HVX_STORES, S0_IS_PRED,     3, 1)
+FIELD(PROBE_PKT_SCALAR_HVX_STORES, S1_IS_PRED,     4, 1)
+
 #endif

tests/tcg/hexagon/Makefile.target

@@ -1,5 +1,5 @@
 ##
-##  Copyright(c) 2019-2022 Qualcomm Innovation Center, Inc. All Rights Reserved.
+##  Copyright(c) 2019-2023 Qualcomm Innovation Center, Inc. All Rights Reserved.
 ##
 ##  This program is free software; you can redistribute it and/or modify
 ##  it under the terms of the GNU General Public License as published by
@@ -45,6 +45,10 @@ HEX_TESTS += fpstuff
 HEX_TESTS += overflow
 HEX_TESTS += signal_context
 HEX_TESTS += reg_mut
+HEX_TESTS += vector_add_int
+HEX_TESTS += scatter_gather
+HEX_TESTS += hvx_misc
+HEX_TESTS += hvx_histogram
 
 HEX_TESTS += test_abs
 HEX_TESTS += test_bitcnt
@@ -78,3 +82,10 @@ TESTS += $(HEX_TESTS)
 usr: usr.c
 	$(CC) $(CFLAGS) -mv67t -O2 -Wno-inline-asm -Wno-expansion-to-defined $< -o $@ $(LDFLAGS)
 
+scatter_gather: CFLAGS += -mhvx
+vector_add_int: CFLAGS += -mhvx -fvectorize
+hvx_misc: CFLAGS += -mhvx
+hvx_histogram: CFLAGS += -mhvx -Wno-gnu-folding-constant
+
+hvx_histogram: hvx_histogram.c hvx_histogram_row.S
+	$(CC) $(CFLAGS) $(CROSS_CC_GUEST_CFLAGS) $^ -o $@ $(LDFLAGS)

tests/tcg/hexagon/fpstuff.c

@@ -1,5 +1,5 @@
 /*
- *  Copyright(c) 2020-2022 Qualcomm Innovation Center, Inc. All Rights Reserved.
+ *  Copyright(c) 2020-2023 Qualcomm Innovation Center, Inc. All Rights Reserved.
  *
  *  This program is free software; you can redistribute it and/or modify
  *  it under the terms of the GNU General Public License as published by
@@ -40,6 +40,7 @@ const int SF_HEX_NAN =                    0xffffffff;
 const int SF_small_neg =                  0xab98fba8;
 const int SF_denorm =                     0x00000001;
 const int SF_random =                     0x346001d6;
+const int SF_neg_zero =                   0x80000000;
 
 const long long DF_QNaN =                 0x7ff8000000000000ULL;
 const long long DF_SNaN =                 0x7ff7000000000000ULL;
@@ -536,6 +537,33 @@ static void check_sffixupd(void)
     check32(result, 0x146001d6);
 }
 
+static void check_sffms(void)
+{
+    int result;
+
+    /* Check that sffms properly deals with -0 */
+    result = SF_neg_zero;
+    asm ("%0 -= sfmpy(%1 , %2)\n\t"
+        : "+r"(result)
+        : "r"(SF_ZERO), "r"(SF_ZERO)
+        : "r12", "r8");
+    check32(result, SF_neg_zero);
+
+    result = SF_ZERO;
+    asm ("%0 -= sfmpy(%1 , %2)\n\t"
+        : "+r"(result)
+        : "r"(SF_neg_zero), "r"(SF_ZERO)
+        : "r12", "r8");
+    check32(result, SF_ZERO);
+
+    result = SF_ZERO;
+    asm ("%0 -= sfmpy(%1 , %2)\n\t"
+        : "+r"(result)
+        : "r"(SF_ZERO), "r"(SF_neg_zero)
+        : "r12", "r8");
+    check32(result, SF_ZERO);
+}
+
 static void check_float2int_convs()
 {
     int res32;
@@ -688,6 +716,7 @@ int main()
     check_invsqrta();
     check_sffixupn();
     check_sffixupd();
+    check_sffms();
     check_float2int_convs();
 
     puts(err ? "FAIL" : "PASS");

tests/tcg/hexagon/preg_alias.c

@@ -1,5 +1,5 @@
 /*
- *  Copyright(c) 2019-2022 Qualcomm Innovation Center, Inc. All Rights Reserved.
+ *  Copyright(c) 2019-2023 Qualcomm Innovation Center, Inc. All Rights Reserved.
  *
  *  This program is free software; you can redistribute it and/or modify
  *  it under the terms of the GNU General Public License as published by
@@ -65,7 +65,7 @@ static inline void creg_alias(int cval, PRegs *pregs)
       : "=r"(pregs->pregs.p0), "=r"(pregs->pregs.p1),
         "=r"(pregs->pregs.p2), "=r"(pregs->pregs.p3)
       : "r"(cval)
-      : "p0", "p1", "p2", "p3");
+      : "c4", "p0", "p1", "p2", "p3");
 }
 
 int err;
@@ -92,7 +92,7 @@ static inline void creg_alias_pair(unsigned int cval, PRegs *pregs)
        : "=r"(pregs->pregs.p0), "=r"(pregs->pregs.p1),
          "=r"(pregs->pregs.p2), "=r"(pregs->pregs.p3), "=r"(c5)
        : "r"(cval_pair)
-       : "p0", "p1", "p2", "p3");
+       : "c4", "c5", "p0", "p1", "p2", "p3");
 
   check(c5, 0xdeadbeef);
 }
@@ -117,7 +117,7 @@ static void test_packet(void)
          "}\n\t"
          : "+r"(result)
          : "r"(0xffffffff), "r"(0xff00ffff), "r"(0x837ed653)
-         : "p0", "p1", "p2", "p3");
+         : "c4", "p0", "p1", "p2", "p3");
     check(result, old_val);
 
     /* Test a predicated store */
@@ -129,7 +129,7 @@ static void test_packet(void)
          "}\n\t"
          :
          : "r"(0), "r"(0xffffffff), "r"(&result)
-         : "p0", "p1", "p2", "p3", "memory");
+         : "c4", "p0", "p1", "p2", "p3", "memory");
     check(result, 0x0);
 }
 

tests/tcg/hexagon/scatter_gather.c

@@ -1,5 +1,5 @@
 /*
- *  Copyright(c) 2019-2021 Qualcomm Innovation Center, Inc. All Rights Reserved.
+ *  Copyright(c) 2019-2023 Qualcomm Innovation Center, Inc. All Rights Reserved.
  *
  *  This program is free software; you can redistribute it and/or modify
  *  it under the terms of the GNU General Public License as published by
@@ -40,47 +40,6 @@ typedef long HVX_VectorPair   __attribute__((__vector_size__(256)))
 typedef long HVX_VectorPred   __attribute__((__vector_size__(128)))
                               __attribute__((aligned(128)));
 
-#define VSCATTER_16(BASE, RGN, OFF, VALS) \
-    __builtin_HEXAGON_V6_vscattermh_128B((int)BASE, RGN, OFF, VALS)
-#define VSCATTER_16_MASKED(MASK, BASE, RGN, OFF, VALS) \
-    __builtin_HEXAGON_V6_vscattermhq_128B(MASK, (int)BASE, RGN, OFF, VALS)
-#define VSCATTER_32(BASE, RGN, OFF, VALS) \
-    __builtin_HEXAGON_V6_vscattermw_128B((int)BASE, RGN, OFF, VALS)
-#define VSCATTER_32_MASKED(MASK, BASE, RGN, OFF, VALS) \
-    __builtin_HEXAGON_V6_vscattermwq_128B(MASK, (int)BASE, RGN, OFF, VALS)
-#define VSCATTER_16_32(BASE, RGN, OFF, VALS) \
-    __builtin_HEXAGON_V6_vscattermhw_128B((int)BASE, RGN, OFF, VALS)
-#define VSCATTER_16_32_MASKED(MASK, BASE, RGN, OFF, VALS) \
-    __builtin_HEXAGON_V6_vscattermhwq_128B(MASK, (int)BASE, RGN, OFF, VALS)
-#define VSCATTER_16_ACC(BASE, RGN, OFF, VALS) \
-    __builtin_HEXAGON_V6_vscattermh_add_128B((int)BASE, RGN, OFF, VALS)
-#define VSCATTER_32_ACC(BASE, RGN, OFF, VALS) \
-    __builtin_HEXAGON_V6_vscattermw_add_128B((int)BASE, RGN, OFF, VALS)
-#define VSCATTER_16_32_ACC(BASE, RGN, OFF, VALS) \
-    __builtin_HEXAGON_V6_vscattermhw_add_128B((int)BASE, RGN, OFF, VALS)
-
-#define VGATHER_16(DSTADDR, BASE, RGN, OFF) \
-    __builtin_HEXAGON_V6_vgathermh_128B(DSTADDR, (int)BASE, RGN, OFF)
-#define VGATHER_16_MASKED(DSTADDR, MASK, BASE, RGN, OFF) \
-    __builtin_HEXAGON_V6_vgathermhq_128B(DSTADDR, MASK, (int)BASE, RGN, OFF)
-#define VGATHER_32(DSTADDR, BASE, RGN, OFF) \
-    __builtin_HEXAGON_V6_vgathermw_128B(DSTADDR, (int)BASE, RGN, OFF)
-#define VGATHER_32_MASKED(DSTADDR, MASK, BASE, RGN, OFF) \
-    __builtin_HEXAGON_V6_vgathermwq_128B(DSTADDR, MASK, (int)BASE, RGN, OFF)
-#define VGATHER_16_32(DSTADDR, BASE, RGN, OFF) \
-    __builtin_HEXAGON_V6_vgathermhw_128B(DSTADDR, (int)BASE, RGN, OFF)
-#define VGATHER_16_32_MASKED(DSTADDR, MASK, BASE, RGN, OFF) \
-    __builtin_HEXAGON_V6_vgathermhwq_128B(DSTADDR, MASK, (int)BASE, RGN, OFF)
-
-#define VSHUFF_H(V) \
-    __builtin_HEXAGON_V6_vshuffh_128B(V)
-#define VSPLAT_H(X) \
-    __builtin_HEXAGON_V6_lvsplath_128B(X)
-#define VAND_VAL(PRED, VAL) \
-    __builtin_HEXAGON_V6_vandvrt_128B(PRED, VAL)
-#define VDEAL_H(V) \
-    __builtin_HEXAGON_V6_vdealh_128B(V)
-
 int err;
 
 /* define the number of rows/cols in a square matrix */
@@ -108,22 +67,22 @@ unsigned short vscatter16_32_ref[SCATTER_BUFFER_SIZE];
 unsigned short vgather16_32_ref[MATRIX_SIZE];
 
 /* declare the arrays of offsets */
-unsigned short half_offsets[MATRIX_SIZE];
+unsigned short half_offsets[MATRIX_SIZE] __attribute__((aligned(128)));
-unsigned int   word_offsets[MATRIX_SIZE];
+unsigned int   word_offsets[MATRIX_SIZE] __attribute__((aligned(128)));
 
 /* declare the arrays of values */
-unsigned short half_values[MATRIX_SIZE];
+unsigned short half_values[MATRIX_SIZE] __attribute__((aligned(128)));
-unsigned short half_values_acc[MATRIX_SIZE];
+unsigned short half_values_acc[MATRIX_SIZE] __attribute__((aligned(128)));
-unsigned short half_values_masked[MATRIX_SIZE];
+unsigned short half_values_masked[MATRIX_SIZE] __attribute__((aligned(128)));
-unsigned int   word_values[MATRIX_SIZE];
+unsigned int   word_values[MATRIX_SIZE] __attribute__((aligned(128)));
-unsigned int   word_values_acc[MATRIX_SIZE];
+unsigned int   word_values_acc[MATRIX_SIZE] __attribute__((aligned(128)));
-unsigned int   word_values_masked[MATRIX_SIZE];
+unsigned int   word_values_masked[MATRIX_SIZE] __attribute__((aligned(128)));
 
 /* declare the arrays of predicates */
-unsigned short half_predicates[MATRIX_SIZE];
+unsigned short half_predicates[MATRIX_SIZE] __attribute__((aligned(128)));
-unsigned int   word_predicates[MATRIX_SIZE];
+unsigned int   word_predicates[MATRIX_SIZE] __attribute__((aligned(128)));
 
-/* make this big enough for all the intrinsics */
+/* make this big enough for all the operations */
 const size_t region_len = sizeof(vtcm);
 
 /* optionally add sync instructions */
@@ -261,164 +220,201 @@ void create_offsets_values_preds_16_32(void)
     }
 }
 
-/* scatter the 16 bit elements using intrinsics */
+/* scatter the 16 bit elements using HVX */
 void vector_scatter_16(void)
 {
-    /* copy the offsets and values to vectors */
+    asm ("m0 = %1\n\t"
-    HVX_Vector offsets = *(HVX_Vector *)half_offsets;
+         "v0 = vmem(%2 + #0)\n\t"
-    HVX_Vector values = *(HVX_Vector *)half_values;
+         "v1 = vmem(%3 + #0)\n\t"
-
+         "vscatter(%0, m0, v0.h).h = v1\n\t"
-    VSCATTER_16(&vtcm.vscatter16, region_len, offsets, values);
+         : : "r"(vtcm.vscatter16), "r"(region_len),
+             "r"(half_offsets), "r"(half_values)
+         : "m0", "v0", "v1", "memory");
 
     sync_scatter(vtcm.vscatter16);
 }
 
-/* scatter-accumulate the 16 bit elements using intrinsics */
+/* scatter-accumulate the 16 bit elements using HVX */
 void vector_scatter_16_acc(void)
 {
-    /* copy the offsets and values to vectors */
+    asm ("m0 = %1\n\t"
-    HVX_Vector offsets = *(HVX_Vector *)half_offsets;
+         "v0 = vmem(%2 + #0)\n\t"
-    HVX_Vector values = *(HVX_Vector *)half_values_acc;
+         "v1 = vmem(%3 + #0)\n\t"
-
+         "vscatter(%0, m0, v0.h).h += v1\n\t"
-    VSCATTER_16_ACC(&vtcm.vscatter16, region_len, offsets, values);
+         : : "r"(vtcm.vscatter16), "r"(region_len),
+             "r"(half_offsets), "r"(half_values_acc)
+         : "m0", "v0", "v1", "memory");
 
     sync_scatter(vtcm.vscatter16);
 }
 
-/* scatter the 16 bit elements using intrinsics */
+/* masked scatter the 16 bit elements using HVX */
 void vector_scatter_16_masked(void)
 {
-    /* copy the offsets and values to vectors */
+    asm ("r1 = #-1\n\t"
-    HVX_Vector offsets = *(HVX_Vector *)half_offsets;
+         "v0 = vmem(%0 + #0)\n\t"
-    HVX_Vector values = *(HVX_Vector *)half_values_masked;
+         "q0 = vand(v0, r1)\n\t"
-    HVX_Vector pred_reg = *(HVX_Vector *)half_predicates;
+         "m0 = %2\n\t"
-    HVX_VectorPred preds = VAND_VAL(pred_reg, ~0);
+         "v0 = vmem(%3 + #0)\n\t"
-
+         "v1 = vmem(%4 + #0)\n\t"
-    VSCATTER_16_MASKED(preds, &vtcm.vscatter16, region_len, offsets, values);
+         "if (q0) vscatter(%1, m0, v0.h).h = v1\n\t"
+         : : "r"(half_predicates), "r"(vtcm.vscatter16), "r"(region_len),
+             "r"(half_offsets), "r"(half_values_masked)
+         : "r1", "q0", "m0", "q0", "v0", "v1", "memory");
 
     sync_scatter(vtcm.vscatter16);
 }
 
-/* scatter the 32 bit elements using intrinsics */
+/* scatter the 32 bit elements using HVX */
 void vector_scatter_32(void)
 {
-    /* copy the offsets and values to vectors */
+    HVX_Vector *offsetslo = (HVX_Vector *)word_offsets;
-    HVX_Vector offsetslo = *(HVX_Vector *)word_offsets;
+    HVX_Vector *offsetshi = (HVX_Vector *)&word_offsets[MATRIX_SIZE / 2];
-    HVX_Vector offsetshi = *(HVX_Vector *)&word_offsets[MATRIX_SIZE / 2];
+    HVX_Vector *valueslo = (HVX_Vector *)word_values;
-    HVX_Vector valueslo = *(HVX_Vector *)word_values;
+    HVX_Vector *valueshi = (HVX_Vector *)&word_values[MATRIX_SIZE / 2];
-    HVX_Vector valueshi = *(HVX_Vector *)&word_values[MATRIX_SIZE / 2];
 
-    VSCATTER_32(&vtcm.vscatter32, region_len, offsetslo, valueslo);
+    asm ("m0 = %1\n\t"
-    VSCATTER_32(&vtcm.vscatter32, region_len, offsetshi, valueshi);
+         "v0 = vmem(%2 + #0)\n\t"
+         "v1 = vmem(%3 + #0)\n\t"
+         "vscatter(%0, m0, v0.w).w = v1\n\t"
+         : : "r"(vtcm.vscatter32), "r"(region_len),
+             "r"(offsetslo), "r"(valueslo)
+         : "m0", "v0", "v1", "memory");
+    asm ("m0 = %1\n\t"
+         "v0 = vmem(%2 + #0)\n\t"
+         "v1 = vmem(%3 + #0)\n\t"
+         "vscatter(%0, m0, v0.w).w = v1\n\t"
+         : : "r"(vtcm.vscatter32), "r"(region_len),
+             "r"(offsetshi), "r"(valueshi)
+         : "m0", "v0", "v1", "memory");
 
     sync_scatter(vtcm.vscatter32);
 }
 
-/* scatter-acc the 32 bit elements using intrinsics */
+/* scatter-accumulate the 32 bit elements using HVX */
 void vector_scatter_32_acc(void)
 {
-    /* copy the offsets and values to vectors */
+    HVX_Vector *offsetslo = (HVX_Vector *)word_offsets;
-    HVX_Vector offsetslo = *(HVX_Vector *)word_offsets;
+    HVX_Vector *offsetshi = (HVX_Vector *)&word_offsets[MATRIX_SIZE / 2];
-    HVX_Vector offsetshi = *(HVX_Vector *)&word_offsets[MATRIX_SIZE / 2];
+    HVX_Vector *valueslo = (HVX_Vector *)word_values_acc;
-    HVX_Vector valueslo = *(HVX_Vector *)word_values_acc;
+    HVX_Vector *valueshi = (HVX_Vector *)&word_values_acc[MATRIX_SIZE / 2];
-    HVX_Vector valueshi = *(HVX_Vector *)&word_values_acc[MATRIX_SIZE / 2];
 
-    VSCATTER_32_ACC(&vtcm.vscatter32, region_len, offsetslo, valueslo);
+    asm ("m0 = %1\n\t"
-    VSCATTER_32_ACC(&vtcm.vscatter32, region_len, offsetshi, valueshi);
+         "v0 = vmem(%2 + #0)\n\t"
+         "v1 = vmem(%3 + #0)\n\t"
+         "vscatter(%0, m0, v0.w).w += v1\n\t"
+         : : "r"(vtcm.vscatter32), "r"(region_len),
+             "r"(offsetslo), "r"(valueslo)
+         : "m0", "v0", "v1", "memory");
+    asm ("m0 = %1\n\t"
+         "v0 = vmem(%2 + #0)\n\t"
+         "v1 = vmem(%3 + #0)\n\t"
+         "vscatter(%0, m0, v0.w).w += v1\n\t"
+         : : "r"(vtcm.vscatter32), "r"(region_len),
+             "r"(offsetshi), "r"(valueshi)
+         : "m0", "v0", "v1", "memory");
 
     sync_scatter(vtcm.vscatter32);
 }
 
-/* scatter the 32 bit elements using intrinsics */
+/* masked scatter the 32 bit elements using HVX */
 void vector_scatter_32_masked(void)
 {
-    /* copy the offsets and values to vectors */
+    HVX_Vector *offsetslo = (HVX_Vector *)word_offsets;
-    HVX_Vector offsetslo = *(HVX_Vector *)word_offsets;
+    HVX_Vector *offsetshi = (HVX_Vector *)&word_offsets[MATRIX_SIZE / 2];
-    HVX_Vector offsetshi = *(HVX_Vector *)&word_offsets[MATRIX_SIZE / 2];
+    HVX_Vector *valueslo = (HVX_Vector *)word_values_masked;
-    HVX_Vector valueslo = *(HVX_Vector *)word_values_masked;
+    HVX_Vector *valueshi = (HVX_Vector *)&word_values_masked[MATRIX_SIZE / 2];
-    HVX_Vector valueshi = *(HVX_Vector *)&word_values_masked[MATRIX_SIZE / 2];
+    HVX_Vector *predslo = (HVX_Vector *)word_predicates;
-    HVX_Vector pred_reglo = *(HVX_Vector *)word_predicates;
+    HVX_Vector *predshi = (HVX_Vector *)&word_predicates[MATRIX_SIZE / 2];
-    HVX_Vector pred_reghi = *(HVX_Vector *)&word_predicates[MATRIX_SIZE / 2];
-    HVX_VectorPred predslo = VAND_VAL(pred_reglo, ~0);
-    HVX_VectorPred predshi = VAND_VAL(pred_reghi, ~0);
 
-    VSCATTER_32_MASKED(predslo, &vtcm.vscatter32, region_len, offsetslo,
+    asm ("r1 = #-1\n\t"
-                       valueslo);
+         "v0 = vmem(%0 + #0)\n\t"
-    VSCATTER_32_MASKED(predshi, &vtcm.vscatter32, region_len, offsetshi,
+         "q0 = vand(v0, r1)\n\t"
-                       valueshi);
+         "m0 = %2\n\t"
+         "v0 = vmem(%3 + #0)\n\t"
+         "v1 = vmem(%4 + #0)\n\t"
+         "if (q0) vscatter(%1, m0, v0.w).w = v1\n\t"
+         : : "r"(predslo), "r"(vtcm.vscatter32), "r"(region_len),
+             "r"(offsetslo), "r"(valueslo)
+         : "r1", "q0", "m0", "q0", "v0", "v1", "memory");
+    asm ("r1 = #-1\n\t"
+         "v0 = vmem(%0 + #0)\n\t"
+         "q0 = vand(v0, r1)\n\t"
+         "m0 = %2\n\t"
+         "v0 = vmem(%3 + #0)\n\t"
+         "v1 = vmem(%4 + #0)\n\t"
+         "if (q0) vscatter(%1, m0, v0.w).w = v1\n\t"
+         : : "r"(predshi), "r"(vtcm.vscatter32), "r"(region_len),
+             "r"(offsetshi), "r"(valueshi)
+         : "r1", "q0", "m0", "q0", "v0", "v1", "memory");
 
-    sync_scatter(vtcm.vscatter16);
+    sync_scatter(vtcm.vscatter32);
 }
 
-/* scatter the 16 bit elements with 32 bit offsets using intrinsics */
+/* scatter the 16 bit elements with 32 bit offsets using HVX */
 void vector_scatter_16_32(void)
 {
-    HVX_VectorPair offsets;
+    asm ("m0 = %1\n\t"
-    HVX_Vector values;
+         "v0 = vmem(%2 + #0)\n\t"
-
+         "v1 = vmem(%2 + #1)\n\t"
-    /* get the word offsets in a vector pair */
+         "v2 = vmem(%3 + #0)\n\t"
-    offsets = *(HVX_VectorPair *)word_offsets;
+         "v2.h = vshuff(v2.h)\n\t"  /* shuffle the values for the scatter */
-
+         "vscatter(%0, m0, v1:0.w).h = v2\n\t"
-    /* these values need to be shuffled for the scatter */
+         : : "r"(vtcm.vscatter16_32), "r"(region_len),
-    values = *(HVX_Vector *)half_values;
+             "r"(word_offsets), "r"(half_values)
-    values = VSHUFF_H(values);
+         : "m0", "v0", "v1", "v2", "memory");
-
-    VSCATTER_16_32(&vtcm.vscatter16_32, region_len, offsets, values);
 
     sync_scatter(vtcm.vscatter16_32);
 }
 
-/* scatter-acc the 16 bit elements with 32 bit offsets using intrinsics */
+/* scatter-accumulate the 16 bit elements with 32 bit offsets using HVX */
 void vector_scatter_16_32_acc(void)
 {
-    HVX_VectorPair offsets;
+    asm ("m0 = %1\n\t"
-    HVX_Vector values;
+         "v0 = vmem(%2 + #0)\n\t"
-
+         "v1 = vmem(%2 + #1)\n\t"
-    /* get the word offsets in a vector pair */
+         "v2 = vmem(%3 + #0)\n\t" \
-    offsets = *(HVX_VectorPair *)word_offsets;
+         "v2.h = vshuff(v2.h)\n\t"  /* shuffle the values for the scatter */
-
+         "vscatter(%0, m0, v1:0.w).h += v2\n\t"
-    /* these values need to be shuffled for the scatter */
+         : : "r"(vtcm.vscatter16_32), "r"(region_len),
-    values = *(HVX_Vector *)half_values_acc;
+             "r"(word_offsets), "r"(half_values_acc)
-    values = VSHUFF_H(values);
+         : "m0", "v0", "v1", "v2", "memory");
-
-    VSCATTER_16_32_ACC(&vtcm.vscatter16_32, region_len, offsets, values);
 
     sync_scatter(vtcm.vscatter16_32);
 }
 
-/* masked scatter the 16 bit elements with 32 bit offsets using intrinsics */
+/* masked scatter the 16 bit elements with 32 bit offsets using HVX */
 void vector_scatter_16_32_masked(void)
 {
-    HVX_VectorPair offsets;
+    asm ("r1 = #-1\n\t"
-    HVX_Vector values;
+         "v0 = vmem(%0 + #0)\n\t"
-    HVX_Vector pred_reg;
+         "v0.h = vshuff(v0.h)\n\t"  /* shuffle the predicates */
-
+         "q0 = vand(v0, r1)\n\t"
-    /* get the word offsets in a vector pair */
+         "m0 = %2\n\t"
-    offsets = *(HVX_VectorPair *)word_offsets;
+         "v0 = vmem(%3 + #0)\n\t"
-
+         "v1 = vmem(%3 + #1)\n\t"
-    /* these values need to be shuffled for the scatter */
+         "v2 = vmem(%4 + #0)\n\t" \
-    values = *(HVX_Vector *)half_values_masked;
+         "v2.h = vshuff(v2.h)\n\t"  /* shuffle the values for the scatter */
-    values = VSHUFF_H(values);
+         "if (q0) vscatter(%1, m0, v1:0.w).h = v2\n\t"
-
+         : : "r"(half_predicates), "r"(vtcm.vscatter16_32), "r"(region_len),
-    pred_reg = *(HVX_Vector *)half_predicates;
+             "r"(word_offsets), "r"(half_values_masked)
-    pred_reg = VSHUFF_H(pred_reg);
+         : "r1", "q0", "m0", "v0", "v1", "v2", "memory");
-    HVX_VectorPred preds = VAND_VAL(pred_reg, ~0);
-
-    VSCATTER_16_32_MASKED(preds, &vtcm.vscatter16_32, region_len, offsets,
-                          values);
 
     sync_scatter(vtcm.vscatter16_32);
 }
 
-/* gather the elements from the scatter16 buffer */
+/* gather the elements from the scatter16 buffer using HVX */
 void vector_gather_16(void)
 {
-    HVX_Vector *vgather = (HVX_Vector *)&vtcm.vgather16;
+    asm ("m0 = %1\n\t"
-    HVX_Vector offsets = *(HVX_Vector *)half_offsets;
+         "v0 = vmem(%2 + #0)\n\t"
+         "{ vtmp.h = vgather(%0, m0, v0.h).h\n\t"
+         "  vmem(%3 + #0) = vtmp.new }\n\t"
+         : : "r"(vtcm.vscatter16), "r"(region_len),
+             "r"(half_offsets), "r"(vtcm.vgather16)
+         : "m0", "v0", "memory");
 
-    VGATHER_16(vgather, &vtcm.vscatter16, region_len, offsets);
+    sync_gather(vtcm.vgather16);
-
-    sync_gather(vgather);
 }
 
 static unsigned short gather_16_masked_init(void)
@@ -427,31 +423,51 @@ static unsigned short gather_16_masked_init(void)
     return letter | (letter << 8);
 }
 
+/* masked gather the elements from the scatter16 buffer using HVX */
 void vector_gather_16_masked(void)
 {
-    HVX_Vector *vgather = (HVX_Vector *)&vtcm.vgather16;
+    unsigned short init = gather_16_masked_init();
-    HVX_Vector offsets = *(HVX_Vector *)half_offsets;
-    HVX_Vector pred_reg = *(HVX_Vector *)half_predicates;
-    HVX_VectorPred preds = VAND_VAL(pred_reg, ~0);
 
-    *vgather = VSPLAT_H(gather_16_masked_init());
+    asm ("v0.h = vsplat(%5)\n\t"
-    VGATHER_16_MASKED(vgather, preds, &vtcm.vscatter16, region_len, offsets);
+         "vmem(%4 + #0) = v0\n\t"  /* initialize the write area */
+         "r1 = #-1\n\t"
+         "v0 = vmem(%0 + #0)\n\t"
+         "q0 = vand(v0, r1)\n\t"
+         "m0 = %2\n\t"
+         "v0 = vmem(%3 + #0)\n\t"
+         "{ if (q0) vtmp.h = vgather(%1, m0, v0.h).h\n\t"
+         "  vmem(%4 + #0) = vtmp.new }\n\t"
+         : : "r"(half_predicates), "r"(vtcm.vscatter16), "r"(region_len),
+             "r"(half_offsets), "r"(vtcm.vgather16), "r"(init)
+         : "r1", "q0", "m0", "v0", "memory");
 
-    sync_gather(vgather);
+    sync_gather(vtcm.vgather16);
 }
 
-/* gather the elements from the scatter32 buffer */
+/* gather the elements from the scatter32 buffer using HVX */
 void vector_gather_32(void)
 {
-    HVX_Vector *vgatherlo = (HVX_Vector *)&vtcm.vgather32;
+    HVX_Vector *vgatherlo = (HVX_Vector *)vtcm.vgather32;
-    HVX_Vector *vgatherhi =
+    HVX_Vector *vgatherhi = (HVX_Vector *)&vtcm.vgather32[MATRIX_SIZE / 2];
-        (HVX_Vector *)((int)&vtcm.vgather32 + (MATRIX_SIZE * 2));
+    HVX_Vector *offsetslo = (HVX_Vector *)word_offsets;
-    HVX_Vector offsetslo = *(HVX_Vector *)word_offsets;
+    HVX_Vector *offsetshi = (HVX_Vector *)&word_offsets[MATRIX_SIZE / 2];
-    HVX_Vector offsetshi = *(HVX_Vector *)&word_offsets[MATRIX_SIZE / 2];
 
-    VGATHER_32(vgatherlo, &vtcm.vscatter32, region_len, offsetslo);
+    asm ("m0 = %1\n\t"
-    VGATHER_32(vgatherhi, &vtcm.vscatter32, region_len, offsetshi);
+         "v0 = vmem(%2 + #0)\n\t"
+         "{ vtmp.w = vgather(%0, m0, v0.w).w\n\t"
+         "  vmem(%3 + #0) = vtmp.new }\n\t"
+         : : "r"(vtcm.vscatter32), "r"(region_len),
+             "r"(offsetslo), "r"(vgatherlo)
+         : "m0", "v0", "memory");
+    asm ("m0 = %1\n\t"
+         "v0 = vmem(%2 + #0)\n\t"
+         "{ vtmp.w = vgather(%0, m0, v0.w).w\n\t"
+         "  vmem(%3 + #0) = vtmp.new }\n\t"
+         : : "r"(vtcm.vscatter32), "r"(region_len),
+             "r"(offsetshi), "r"(vgatherhi)
+         : "m0", "v0", "memory");
 
+    sync_gather(vgatherlo);
     sync_gather(vgatherhi);
 }
 
@@ -461,79 +477,88 @@ static unsigned int gather_32_masked_init(void)
     return letter | (letter << 8) | (letter << 16) | (letter << 24);
 }
 
+/* masked gather the elements from the scatter32 buffer using HVX */
 void vector_gather_32_masked(void)
 {
-    HVX_Vector *vgatherlo = (HVX_Vector *)&vtcm.vgather32;
+    unsigned int init = gather_32_masked_init();
-    HVX_Vector *vgatherhi =
+    HVX_Vector *vgatherlo = (HVX_Vector *)vtcm.vgather32;
-        (HVX_Vector *)((int)&vtcm.vgather32 + (MATRIX_SIZE * 2));
+    HVX_Vector *vgatherhi = (HVX_Vector *)&vtcm.vgather32[MATRIX_SIZE / 2];
-    HVX_Vector offsetslo = *(HVX_Vector *)word_offsets;
+    HVX_Vector *offsetslo = (HVX_Vector *)word_offsets;
-    HVX_Vector offsetshi = *(HVX_Vector *)&word_offsets[MATRIX_SIZE / 2];
+    HVX_Vector *offsetshi = (HVX_Vector *)&word_offsets[MATRIX_SIZE / 2];
-    HVX_Vector pred_reglo = *(HVX_Vector *)word_predicates;
+    HVX_Vector *predslo = (HVX_Vector *)word_predicates;
-    HVX_VectorPred predslo = VAND_VAL(pred_reglo, ~0);
+    HVX_Vector *predshi = (HVX_Vector *)&word_predicates[MATRIX_SIZE / 2];
-    HVX_Vector pred_reghi = *(HVX_Vector *)&word_predicates[MATRIX_SIZE / 2];
-    HVX_VectorPred predshi = VAND_VAL(pred_reghi, ~0);
 
-    *vgatherlo = VSPLAT_H(gather_32_masked_init());
+    asm ("v0.h = vsplat(%5)\n\t"
-    *vgatherhi = VSPLAT_H(gather_32_masked_init());
+         "vmem(%4 + #0) = v0\n\t"  /* initialize the write area */
-    VGATHER_32_MASKED(vgatherlo, predslo, &vtcm.vscatter32, region_len,
+         "r1 = #-1\n\t"
-                      offsetslo);
+         "v0 = vmem(%0 + #0)\n\t"
-    VGATHER_32_MASKED(vgatherhi, predshi, &vtcm.vscatter32, region_len,
+         "q0 = vand(v0, r1)\n\t"
-                      offsetshi);
+         "m0 = %2\n\t"
+         "v0 = vmem(%3 + #0)\n\t"
+         "{ if (q0) vtmp.w = vgather(%1, m0, v0.w).w\n\t"
+         "  vmem(%4 + #0) = vtmp.new }\n\t"
+         : : "r"(predslo), "r"(vtcm.vscatter32), "r"(region_len),
+             "r"(offsetslo), "r"(vgatherlo), "r"(init)
+         : "r1", "q0", "m0", "v0", "memory");
+    asm ("v0.h = vsplat(%5)\n\t"
+         "vmem(%4 + #0) = v0\n\t"  /* initialize the write area */
+         "r1 = #-1\n\t"
+         "v0 = vmem(%0 + #0)\n\t"
+         "q0 = vand(v0, r1)\n\t"
+         "m0 = %2\n\t"
+         "v0 = vmem(%3 + #0)\n\t"
+         "{ if (q0) vtmp.w = vgather(%1, m0, v0.w).w\n\t"
+         "  vmem(%4 + #0) = vtmp.new }\n\t"
+         : : "r"(predshi), "r"(vtcm.vscatter32), "r"(region_len),
+             "r"(offsetshi), "r"(vgatherhi), "r"(init)
+         : "r1", "q0", "m0", "v0", "memory");
 
     sync_gather(vgatherlo);
     sync_gather(vgatherhi);
 }
 
-/* gather the elements from the scatter16_32 buffer */
+/* gather the elements from the scatter16_32 buffer using HVX */
 void vector_gather_16_32(void)
 {
-    HVX_Vector *vgather;
+    asm ("m0 = %1\n\t"
-    HVX_VectorPair offsets;
+         "v0 = vmem(%2 + #0)\n\t"
-    HVX_Vector values;
+         "v1 = vmem(%2 + #1)\n\t"
+         "{ vtmp.h = vgather(%0, m0, v1:0.w).h\n\t"
+         "  vmem(%3 + #0) = vtmp.new }\n\t"
+         "v0 = vmem(%3 + #0)\n\t"
+         "v0.h = vdeal(v0.h)\n\t"  /* deal the elements to get the order back */
+         "vmem(%3 + #0) = v0\n\t"
+         : : "r"(vtcm.vscatter16_32), "r"(region_len),
+             "r"(word_offsets), "r"(vtcm.vgather16_32)
+         : "m0", "v0", "v1", "memory");
 
-    /* get the vtcm address to gather from */
+    sync_gather(vtcm.vgather16_32);
-    vgather = (HVX_Vector *)&vtcm.vgather16_32;
-
-    /* get the word offsets in a vector pair */
-    offsets = *(HVX_VectorPair *)word_offsets;
-
-    VGATHER_16_32(vgather, &vtcm.vscatter16_32, region_len, offsets);
-
-    /* deal the elements to get the order back */
-    values = *(HVX_Vector *)vgather;
-    values = VDEAL_H(values);
-
-    /* write it back to vtcm address */
-    *(HVX_Vector *)vgather = values;
 }
 
+/* masked gather the elements from the scatter16_32 buffer using HVX */
 void vector_gather_16_32_masked(void)
 {
-    HVX_Vector *vgather;
+    unsigned short init = gather_16_masked_init();
-    HVX_VectorPair offsets;
-    HVX_Vector pred_reg;
-    HVX_VectorPred preds;
-    HVX_Vector values;
 
-    /* get the vtcm address to gather from */
+    asm ("v0.h = vsplat(%5)\n\t"
-    vgather = (HVX_Vector *)&vtcm.vgather16_32;
+         "vmem(%4 + #0) = v0\n\t"  /* initialize the write area */
+         "r1 = #-1\n\t"
+         "v0 = vmem(%0 + #0)\n\t"
+         "v0.h = vshuff(v0.h)\n\t"  /* shuffle the predicates */
+         "q0 = vand(v0, r1)\n\t"
+         "m0 = %2\n\t"
+         "v0 = vmem(%3 + #0)\n\t"
+         "v1 = vmem(%3 + #1)\n\t"
+         "{ if (q0) vtmp.h = vgather(%1, m0, v1:0.w).h\n\t"
+         "  vmem(%4 + #0) = vtmp.new }\n\t"
+         "v0 = vmem(%4 + #0)\n\t"
+         "v0.h = vdeal(v0.h)\n\t"  /* deal the elements to get the order back */
+         "vmem(%4 + #0) = v0\n\t"
+         : : "r"(half_predicates), "r"(vtcm.vscatter16_32), "r"(region_len),
+             "r"(word_offsets), "r"(vtcm.vgather16_32), "r"(init)
+         : "r1", "q0", "m0", "v0", "v1", "memory");
 
-    /* get the word offsets in a vector pair */
+    sync_gather(vtcm.vgather16_32);
-    offsets = *(HVX_VectorPair *)word_offsets;
-    pred_reg = *(HVX_Vector *)half_predicates;
-    pred_reg = VSHUFF_H(pred_reg);
-    preds = VAND_VAL(pred_reg, ~0);
-
-   *vgather = VSPLAT_H(gather_16_masked_init());
-   VGATHER_16_32_MASKED(vgather, preds, &vtcm.vscatter16_32, region_len,
-                        offsets);
-
-    /* deal the elements to get the order back */
-    values = *(HVX_Vector *)vgather;
-    values = VDEAL_H(values);
-
-    /* write it back to vtcm address */
-    *(HVX_Vector *)vgather = values;
 }
 
 static void check_buffer(const char *name, void *c, void *r, size_t size)
@@ -579,6 +604,7 @@ void scalar_scatter_16_acc(unsigned short *vscatter16)
     }
 }
 
+/* scatter-accumulate the 16 bit elements using C */
 void check_scatter_16_acc()
 {
     memset(vscatter16_ref, FILL_CHAR,
@@ -589,7 +615,7 @@ void check_scatter_16_acc()
                  SCATTER_BUFFER_SIZE * sizeof(unsigned short));
 }
 
-/* scatter the 16 bit elements using C */
+/* masked scatter the 16 bit elements using C */
 void scalar_scatter_16_masked(unsigned short *vscatter16)
 {
     for (int i = 0; i < MATRIX_SIZE; i++) {
@@ -628,7 +654,7 @@ void check_scatter_32()
                  SCATTER_BUFFER_SIZE * sizeof(unsigned int));
 }
 
-/* scatter the 32 bit elements using C */
+/* scatter-accumulate the 32 bit elements using C */
 void scalar_scatter_32_acc(unsigned int *vscatter32)
 {
     for (int i = 0; i < MATRIX_SIZE; ++i) {
@@ -646,7 +672,7 @@ void check_scatter_32_acc()
                  SCATTER_BUFFER_SIZE * sizeof(unsigned int));
 }
 
-/* scatter the 32 bit elements using C */
+/* masked scatter the 32 bit elements using C */
 void scalar_scatter_32_masked(unsigned int *vscatter32)
 {
     for (int i = 0; i < MATRIX_SIZE; i++) {
@@ -667,7 +693,7 @@ void check_scatter_32_masked()
                   SCATTER_BUFFER_SIZE * sizeof(unsigned int));
 }
 
-/* scatter the 32 bit elements using C */
+/* scatter the 16 bit elements with 32 bit offsets using C */
 void scalar_scatter_16_32(unsigned short *vscatter16_32)
 {
     for (int i = 0; i < MATRIX_SIZE; ++i) {
@@ -684,7 +710,7 @@ void check_scatter_16_32()
                  SCATTER_BUFFER_SIZE * sizeof(unsigned short));
 }
 
-/* scatter the 32 bit elements using C */
+/* scatter-accumulate the 16 bit elements with 32 bit offsets using C */
 void scalar_scatter_16_32_acc(unsigned short *vscatter16_32)
 {
     for (int i = 0; i < MATRIX_SIZE; ++i) {
@@ -702,6 +728,7 @@ void check_scatter_16_32_acc()
                  SCATTER_BUFFER_SIZE * sizeof(unsigned short));
 }
 
+/* masked scatter the 16 bit elements with 32 bit offsets using C */
 void scalar_scatter_16_32_masked(unsigned short *vscatter16_32)
 {
     for (int i = 0; i < MATRIX_SIZE; i++) {
@@ -738,6 +765,7 @@ void check_gather_16()
                    MATRIX_SIZE * sizeof(unsigned short));
 }
 
+/* masked gather the elements from the scatter buffer using C */
 void scalar_gather_16_masked(unsigned short *vgather16)
 {
     for (int i = 0; i < MATRIX_SIZE; ++i) {
@@ -756,7 +784,7 @@ void check_gather_16_masked()
                  MATRIX_SIZE * sizeof(unsigned short));
 }
 
-/* gather the elements from the scatter buffer using C */
+/* gather the elements from the scatter32 buffer using C */
 void scalar_gather_32(unsigned int *vgather32)
 {
     for (int i = 0; i < MATRIX_SIZE; ++i) {
@@ -772,6 +800,7 @@ void check_gather_32(void)
                  MATRIX_SIZE * sizeof(unsigned int));
 }
 
+/* masked gather the elements from the scatter32 buffer using C */
 void scalar_gather_32_masked(unsigned int *vgather32)
 {
     for (int i = 0; i < MATRIX_SIZE; ++i) {
@@ -781,7 +810,6 @@ void scalar_gather_32_masked(unsigned int *vgather32)
     }
 }
 
-
 void check_gather_32_masked(void)
 {
     memset(vgather32_ref, gather_32_masked_init(),
@@ -791,7 +819,7 @@ void check_gather_32_masked(void)
                  vgather32_ref, MATRIX_SIZE * sizeof(unsigned int));
 }
 
-/* gather the elements from the scatter buffer using C */
+/* gather the elements from the scatter16_32 buffer using C */
 void scalar_gather_16_32(unsigned short *vgather16_32)
 {
     for (int i = 0; i < MATRIX_SIZE; ++i) {
@@ -807,6 +835,7 @@ void check_gather_16_32(void)
                  MATRIX_SIZE * sizeof(unsigned short));
 }
 
+/* masked gather the elements from the scatter16_32 buffer using C */
 void scalar_gather_16_32_masked(unsigned short *vgather16_32)
 {
     for (int i = 0; i < MATRIX_SIZE; ++i) {