From 68f9627a24b22c06673cce110df43b2744adaa29 Mon Sep 17 00:00:00 2001
From: Luke Kenneth Casson Leighton <lkcl@lkcl.net>
Date: Sun, 30 Sep 2018 07:07:43 +0100
Subject: [PATCH] add in predication to sv instruction execution

this relies on setting the value of the destination register
(and source registers) to zero.  a bad hack but it will do
---
 riscv/insn_template_sv.cc | 11 ++++++---
 riscv/sv.cc               | 20 ++++++++++++++--
 riscv/sv_decode.h         | 50 ++++++++++++++++++++++++++-------------
 3 files changed, 59 insertions(+), 22 deletions(-)

diff --git a/riscv/insn_template_sv.cc b/riscv/insn_template_sv.cc
index a83d7b0..97ca8cb 100644
--- a/riscv/insn_template_sv.cc
+++ b/riscv/insn_template_sv.cc
@@ -18,9 +18,13 @@ reg_t FN(processor_t* p, insn_t s_insn, reg_t pc)
   // need to know if register is used as float or int.
   // REGS_PATTERN is generated by id_regs.py (per opcode)
   unsigned int floatintmap = REGS_PATTERN;
-  reg_t predicate = 0;
+  reg_t predicate = ~0x0;
   sv_insn_t insn(p, bits, floatintmap,
                  predicate, predicate, predicate, predicate);
+  bool zeroing;
+#if defined(USING_REG_RD) || defined(USING_REG_FRD)
+  predicate = insn.predicate(s_insn.rd(), floatintmap & REG_RD, zeroing);
+#endif
   // identify which regs have had their CSR entries set as vectorised.
   // really could do with a macro for-loop here... oh well...
   // integer ops, RD, RS1, RS2, RS3 (use sv_int_tb)
@@ -42,8 +46,9 @@ reg_t FN(processor_t* p, insn_t s_insn, reg_t pc)
       #include INCLUDEFILE
       if (vlen > 1)
       {
-        fprintf(stderr, "reg %s %x vloop %d vlen %d stop %d\n",
-                xstr(INSN), INSNCODE, voffs, vlen, insn.stop_vloop());
+        fprintf(stderr, "reg %s %x vloop %d vlen %d stop %d pred %lx\n",
+                xstr(INSN), INSNCODE, voffs, vlen, insn.stop_vloop(),
+                predicate & (1<<voffs));
       }
       insn.reset_caches(); // ready to increment offsets in next iteration
       if (insn.stop_vloop())
diff --git a/riscv/sv.cc b/riscv/sv.cc
index b180b59..726dc5d 100644
--- a/riscv/sv.cc
+++ b/riscv/sv.cc
@@ -66,7 +66,7 @@ bool sv_insn_t::sv_check_reg(bool intreg, uint64_t reg)
  * of SV.  it's "supposed" to "just" be a vectorisation API. it isn't:
  * it's quite a bit more.
  */
-uint64_t sv_insn_t::remap(uint64_t reg, bool intreg, int &voffs)
+uint64_t sv_insn_t::remap(uint64_t reg, bool intreg, int &voffs, uint64_t &pred)
 {
   // okaay so first determine which map to use.  intreg is passed
   // in (ultimately) from id_regs.py's examination of the use of
@@ -92,8 +92,9 @@ uint64_t sv_insn_t::remap(uint64_t reg, bool intreg, int &voffs)
   // we return the re-mapped register...
   if (!r->isvec) // scalar
   {
+    // reg remapped even as scalar
     vloop_continue = false;
-    return reg; // ... remapped at this point...
+    return predicated(reg, voffs, pred); // returns x0 if pred bit false
   }
 
   // aaand now, as it's a "vector", FINALLY we can add on the loop-offset
@@ -101,6 +102,10 @@ uint64_t sv_insn_t::remap(uint64_t reg, bool intreg, int &voffs)
   // and, at last, we have "parallelism" a la contiguous registers.
   reg += voffs; // wheww :)
 
+  // before returning, put the register through the predication wringer.
+  // this will return x0 if predication is false
+  reg = predicated(reg, voffs, pred);
+
   // however... before returning, we increment the loop-offset for
   // this particular register, so that on the next loop the next
   // contiguous register will be used.
@@ -149,3 +154,14 @@ reg_t sv_insn_t::predicate(uint64_t reg, bool intreg, bool &zeroing)
   }
   return predicate;
 }
+
+uint64_t sv_insn_t::predicated(uint64_t reg, int offs, uint64_t pred)
+{
+    if (pred & (1<<offs))
+    {
+        return reg;
+    }
+    fprintf(stderr, "predication %ld %d %lx\n", reg, offs, pred);
+    return 0;
+}
+
diff --git a/riscv/sv_decode.h b/riscv/sv_decode.h
index 0858b45..039c439 100644
--- a/riscv/sv_decode.h
+++ b/riscv/sv_decode.h
@@ -29,25 +29,29 @@ public:
             offs_rs2(0), offs_rs3(0),
             prd(p_rd), prs1(p_rs1), prs2(p_rs2), prs3(p_rs3) {}
   uint64_t rd ()
-    { return _remap(insn_t::rd (), fimap & REG_RD , offs_rd , cached_rd); }
+    { return _remap(insn_t::rd (), fimap & REG_RD ,
+                    offs_rd , cached_rd, prd); }
   uint64_t rs1()
-    { return _remap(insn_t::rs1(), fimap & REG_RS1, offs_rs1, cached_rs1); }
+    { return _remap(insn_t::rs1(), fimap & REG_RS1,
+                    offs_rs1, cached_rs1, prs1); }
   uint64_t rs2()
-    { return _remap(insn_t::rs2(), fimap & REG_RS2, offs_rs2, cached_rs2); }
+    { return _remap(insn_t::rs2(), fimap & REG_RS2,
+                    offs_rs2, cached_rs2, prs2); }
   uint64_t rs3()
-    { return _remap(insn_t::rs3(), fimap & REG_RS3, offs_rs3, cached_rs3); }
+    { return _remap(insn_t::rs3(), fimap & REG_RS3,
+                    offs_rs3, cached_rs3, prs3); }
   uint64_t rvc_rs1 ()
     { return _remap(insn_t::rvc_rs1(), fimap & REG_RVC_RS1,
-                    offs_rs1, cached_rs1); }
+                    offs_rs1, cached_rs1, prs1); }
   uint64_t rvc_rs1s ()
     { return _remap(insn_t::rvc_rs1s(), fimap & REG_RVC_RS1S,
-                    offs_rs1, cached_rs1); }
+                    offs_rs1, cached_rs1, prs1); }
   uint64_t rvc_rs2 ()
     { return _remap(insn_t::rvc_rs2(), fimap & REG_RVC_RS2,
-                    offs_rs2, cached_rs2); }
+                    offs_rs2, cached_rs2, prs2); }
   uint64_t rvc_rs2s ()
     { return _remap(insn_t::rvc_rs2s(), fimap & REG_RVC_RS2S,
-                    offs_rs2, cached_rs2); }
+                    offs_rs2, cached_rs2, prs2); }
 
   void reset_caches(void)
   {
@@ -77,29 +81,41 @@ private:
   int offs_rs1;
   int offs_rs2;
   int offs_rs3;
-  uint64_t prd;
-  uint64_t prs1;
-  uint64_t prs2;
-  uint64_t prs3;
+  uint64_t &prd;
+  uint64_t &prs1;
+  uint64_t &prs2;
+  uint64_t &prs3;
+
   // remaps the register through the lookup table.
   // will need to take the current loop index/offset somehow
-  uint64_t remap(uint64_t reg, bool isint, int &offs);
+  uint64_t remap(uint64_t reg, bool isint, int &offs, uint64_t &pred);
 
   // cached version of remap: if remap is called multiple times
   // by an emulated instruction it would increment the loop offset
   // before it's supposed to.
-  uint64_t _remap(uint64_t reg, bool isint, int &offs, uint64_t &cached)
+  uint64_t _remap(uint64_t reg, bool isint, int &offs,
+                  uint64_t &cached, uint64_t &pred)
   {
     if (cached == 0xff)
     {
-        cached = remap(reg, isint, offs);
+        cached = remap(reg, isint, offs, pred);
     }
-    else if (!sv_check_reg(isint, reg))
+    else
     {
-        vloop_continue = false;
+        if (!sv_check_reg(isint, reg))
+        {
+            vloop_continue = false;
+        }
+        reg = predicated(reg, offs, pred);
+        if (reg == 0)
+        {
+            return 0;
+        }
     }
     return cached;
   }
+
+  uint64_t predicated(uint64_t reg, int offs, uint64_t pred);
 };
 
 #endif
-- 
2.30.2