ARM: Fix the implementation of the VFP ldm and stm macroops.

There were four bugs in these instructions. First, the loaded value was being
stored into a floating point register as floating point, changing the value as
it was transfered. Second, the meaning of the "up" bit had been reversed.
Third, the statically sized microop array wasn't bit enough for all possible
inputs. It's now dynamically sized and should always be big enough. Fourth,
the offset was stored as an unsigned 8 bit value. Negative offsets would look
like moderately large positive offsets.

2 files changed, 22 insertions, 12 deletions

diff --git a/src/arch/arm/insts/macromem.cc b/src/arch/arm/insts/macromem.cc
index e5c374089..8e9e60f30 100644
--- a/src/arch/arm/insts/macromem.cc
+++ b/src/arch/arm/insts/macromem.cc
@@ -136,8 +136,6 @@ MacroVFPMemOp::MacroVFPMemOp(const char *mnem, ExtMachInst machInst,
                              bool writeback, bool load, uint32_t offset) :
     PredMacroOp(mnem, machInst, __opClass)
 {
-    const int maxMicroops = 17;
-    microOps = new StaticInstPtr[maxMicroops];
     int i = 0;
 
     // The lowest order bit selects fldmx (set) or fldmd (clear). These seem
@@ -153,26 +151,39 @@ MacroVFPMemOp::MacroVFPMemOp(const char *mnem, ExtMachInst machInst,
     if (count > NumFloatArchRegs)
         count = NumFloatArchRegs;
 
+    numMicroops = count * (single ? 1 : 2) + (writeback ? 1 : 0);
+    microOps = new StaticInstPtr[numMicroops];
+
     uint32_t addr = 0;
 
-    if (up)
-        addr = -4 * offset;
+    if (!up)
+        addr = 4 * offset;
 
+    bool tempUp = up;
     for (int j = 0; j < count; j++) {
         if (load) {
             microOps[i++] = new MicroLdrFpUop(machInst, vd++, rn,
-                                              true, addr);
+                                              tempUp, addr);
             if (!single)
                 microOps[i++] = new MicroLdrFpUop(machInst, vd++, rn,
-                                                  true, addr + 4);
+                                                  tempUp, addr + 4);
         } else {
             microOps[i++] = new MicroStrFpUop(machInst, vd++, rn,
-                                              true, addr);
+                                              tempUp, addr);
             if (!single)
                 microOps[i++] = new MicroStrFpUop(machInst, vd++, rn,
-                                                  true, addr + 4);
+                                                  tempUp, addr + 4);
+        }
+        if (!tempUp) {
+            addr -= (single ? 4 : 8);
+            // The microops don't handle negative displacement, so turn if we
+            // hit zero, flip polarity and start adding.
+            if (addr == 0) {
+                tempUp = true;
+            }
+        } else {
+            addr += (single ? 4 : 8);
         }
-        addr += (single ? 4 : 8);
     }
 
     if (writeback) {
@@ -185,8 +196,7 @@ MacroVFPMemOp::MacroVFPMemOp(const char *mnem, ExtMachInst machInst,
         }
     }
 
-    numMicroops = i;
-    assert(numMicroops <= maxMicroops);
+    assert(numMicroops == i);
     microOps[numMicroops - 1]->setLastMicroop();
 }
 
diff --git a/src/arch/arm/isa/insts/macromem.isa b/src/arch/arm/isa/insts/macromem.isa
index 781cbca7a..0870a966f 100644
--- a/src/arch/arm/isa/insts/macromem.isa
+++ b/src/arch/arm/isa/insts/macromem.isa
@@ -59,7 +59,7 @@ let {{
                                     'predicate_test': predicateTest},
                                    ['IsMicroop'])
 
-    microLdrFpUopCode = "Fa = cSwap(Mem.uw, ((CPSR)Cpsr).e);"
+    microLdrFpUopCode = "Fa.uw = cSwap(Mem.uw, ((CPSR)Cpsr).e);"
     microLdrFpUopIop = InstObjParams('ldrfp_uop', 'MicroLdrFpUop',
                                      'MicroMemOp',
                                      {'memacc_code': microLdrFpUopCode,