Merge branch 'main' of github.com:davidharrishmc/riscv-wally into main

.gitignore

@@ -70,3 +70,4 @@ synthDC/*.svf
 synthDC/runs/
 synthDC/hdl
 /pipelined/regression/power.saif
+tests/fp/vectors/*.tv

linux/bootmem.txt

@@ -0,0 +1,12 @@
+00001000: 00000297		auipc t0, 0		# t0 = 0x00001000
+00001004: 02828613		addi a2, t0,0x28	# a2 = 0x00001028
+00001008: f1402573		csrr a0, mhartid	# a0 = mhartid
+0000100c: 0202b583		ld a1, 32(t0) 		# a1 = 87000000 - device tree address
+00001010: 0182b283		ld t0, 24(t0)		# t0 = 80000000 - start of firmware
+00001014: 00028067		jr t0			# jump to firmware
+00001018: 0000000080000000			# firmware start address
+00001020: 0000000087000000			# flattened device tree load address
+00001028: 000000004942534f			# a2 points to this 8 dword data structure
+00001030: 0000000000000002
+00001038: 0000000080200000
+00001040: 0000000000000001

linux/buildroot-scripts/Makefile

@@ -28,7 +28,7 @@ ${DIS}/vmlinux.objdump.addr: ${DIS}/vmlinux.objdump
 	-cd ${DIS}; extractFunctionRadix.sh vmlinux.objdump
 
 ${DIS}/busybox.objdump: ${DIS}/rootfs/bin/busybox
-	riscv64-unknown-elf-objdump -D ${DIS}/rootfs/bin/busybox >> ${DIS}/busybox.objdump
+	riscv64-unknown-elf-objdump -S ${DIS}/rootfs/bin/busybox >> ${DIS}/busybox.objdump
 
 ${DIS}/rootfs/bin/busybox:
 	mkdir -p ${DIS}/rootfs

linux/testvector-generation/filterTrapsToInterrupts.py

@@ -0,0 +1,68 @@
+#! /usr/bin/python3
+import sys, os
+from functools import reduce
+
+################
+# Helper Funcs #
+################
+
+def tokenize(string):
+    tokens = []
+    token = ''
+    whitespace = 0
+    prevWhitespace = 0
+    for char in string:
+        prevWhitespace = whitespace
+        whitespace = char in ' \t\n'
+        if (whitespace):
+            if ((not prevWhitespace) and (token != '')): 
+                tokens.append(token)
+            token = ''
+        else:
+            token = token + char
+    return tokens
+
+def strip0x(num):
+    return num[2:]
+
+def stripZeroes(num):
+    num = num.strip('0')
+    if num=='':
+        return '0'
+    else:
+        return num
+
+#############
+# Main Code #
+#############
+print("Begin filtering traps down to just external interrupts.")
+
+# Parse Args
+if len(sys.argv) != 2:
+    sys.exit('Error filterTrapsToInterrupts.py expects 1 arg: <path_to_testvector_dir>')
+tvDir = sys.argv[1]+'/'
+trapsFilePath = tvDir+'traps.txt'
+if not os.path.exists(trapsFilePath):
+    sys.exit('Error input file '+trapsFilePath+'not found')
+
+with open(tvDir+'interrupts.txt', 'w') as interruptsFile:
+    with open(trapsFilePath, 'r') as trapsFile:
+        while True:
+            trap = trapsFile.readline()
+            if trap == '':
+                break
+            trapType = trap.split(' ')[-1]
+            if ('interrupt' in trap) and (('external' in trapType) or ('m_timer' in trapType)): # no s_timer because that is not controlled by CLINT
+                interruptsFile.write(trap) # overall line
+                interruptsFile.write(trapsFile.readline()) # attempted instr count
+                interruptsFile.write(trapsFile.readline()) # hart #
+                interruptsFile.write(trapsFile.readline()) # asynchronous
+                interruptsFile.write(trapsFile.readline()) # cause
+                interruptsFile.write(trapsFile.readline()) # epc
+                interruptsFile.write(trapsFile.readline()) # tval
+                interruptsFile.write(trapsFile.readline()) # description
+            else:
+                for i in range(7):
+                    trapsFile.readline()
+        
+print("Finished filtering traps down to just external interrupts.")

linux/testvector-generation/genInitMem.sh

@@ -53,7 +53,7 @@ then
     -ex "printf \"Warning - please verify that the second half of $rawUntrimmedBootmemFile is all 0s\n\"" \
     -ex "dump binary memory $rawUntrimmedBootmemFile 0x1000 0x2fff" \
     -ex "printf \"Creating $rawRamFile\n\"" \
-    -ex "dump binary memory $rawRamFile 0x80000000 0xffffffff" \
+    -ex "dump binary memory $rawRamFile 0x80000000 0x87ffffff" \
     -ex "kill" \
     -ex "q"
 

linux/testvector-generation/genTrace.sh

@@ -4,10 +4,12 @@ imageDir=$RISCV/buildroot/output/images
 tvDir=$RISCV/linux-testvectors
 recordFile="$tvDir/all.qemu"
 traceFile="$tvDir/all.txt"
+trapsFile="$tvDir/traps.txt"
 interruptsFile="$tvDir/interrupts.txt"
 
 read -p "Warning: running this script will overwrite the contents of:
   * $traceFile
+  * $trapsFile
   * $interruptsFile
 Would you like to proceed? (y/n) " -n 1 -r
 echo
@@ -28,6 +30,7 @@ then
     fi
 
     touch $traceFile 
+    touch $trapsFile 
     touch $interruptsFile 
 
     # QEMU Simulation
@@ -38,7 +41,9 @@ then
     -bios $imageDir/fw_jump.elf -kernel $imageDir/Image -append "root=/dev/vda ro" -initrd $imageDir/rootfs.cpio \
     -singlestep -rtc clock=vm -icount shift=0,align=off,sleep=on,rr=replay,rrfile=$recordFile \
     -d nochain,cpu,in_asm,int \
-    2>&1 >./qemu-serial | ./parseQEMUtoGDB.py | ./parseGDBtoTrace.py $interruptsFile | ./remove_dup.awk > $traceFile)
+    2>&1 >./qemu-serial | ./parseQEMUtoGDB.py | ./parseGDBtoTrace.py $trapsFile | ./remove_dup.awk > $traceFile)
+
+    ./filterTrapsToInterrupts.py $tvDir
 
     echo "genTrace.sh completed!"
     echo "You may want to restrict write access to $tvDir now and give cad ownership of it."

linux/testvector-generation/parseQEMUtoGDB.py

@@ -29,10 +29,14 @@ def printPC(l):
 
 def printCSRs():
     global parseState, inPageFault, CSRs, pageFaultCSRs, regs, pageFaultCSRs, instrs
+    global interrupt_line
     if not inPageFault:
         for (csr,val) in CSRs.items():
             print('{}{}{:#x}  {}'.format(csr, ' '*(15-len(csr)), val, val))
-        print('-----')
+        print('-----') # end of current instruction
+        if len(interrupt_line)>0: # squish interrupts in between instructions
+            print(interrupt_line)
+            interrupt_line=""
 
 def parseCSRs(l):
     global parseState, inPageFault, CSRs, pageFaultCSRs, regs, pageFaultCSRs, instrs
@@ -119,15 +123,9 @@ for l in fileinput.input():
         break
     elif l.startswith('IN:'):
         # New disassembled instr
-        if len(interrupt_line)>0:
-            print(interrupt_line)
-            interrupt_line=""
         parseState = "instr"
     elif (parseState == "instr") and l.startswith('0x'):
         # New instruction
-        if len(interrupt_line)>0:
-            print(interrupt_line)
-            interrupt_line=""
         if "out of bounds" in l:
             sys.stderr.write("Detected QEMU page fault error\n")
             beginPageFault = not inPageFault

pipelined/config/rv32ia/wally-config.vh

@@ -43,6 +43,7 @@
 `define ZIFENCEI_SUPPORTED 1
 `define COUNTERS 32
 `define ZICOUNTERS_SUPPORTED 1
+`define ZFH_SUPPORTED 0
 
 // Microarchitectural Features
 `define UARCH_PIPELINED 1

pipelined/config/rv64ia/wally-config.vh

@@ -43,6 +43,7 @@
 `define ZIFENCEI_SUPPORTED 1
 `define COUNTERS 32
 `define ZICOUNTERS_SUPPORTED 1
+`define ZFH_SUPPORTED 0
 
 /// Microarchitectural Features
 `define UARCH_PIPELINED 1

pipelined/regression/wally-pipelined-batch.do

@@ -41,7 +41,6 @@ if {$2 eq "buildroot" || $2 eq "buildroot-checkpoint"} {
     run -all
     run -all
     exec ./slack-notifier/slack-notifier.py
-    quit 
 } else {
     vlog -lint -work wkdir/work_${1}_${2} +incdir+../config/$1 +incdir+../config/shared ../testbench/testbench.sv ../testbench/common/*.sv   ../src/*/*.sv ../src/*/*/*.sv -suppress 2583 -suppress 7063
     # start and run simulation
@@ -52,11 +51,15 @@ if {$2 eq "buildroot" || $2 eq "buildroot-checkpoint"} {
     #vopt work_$2.testbench -work work_$2 -o workopt_$2 +cover=sbectf
     #vsim -coverage -lib work_$2 workopt_$2
 
+    # power add generates the logging necessary for saif generation.
+    power add -r /dut/core/*
     run -all
-    quit
+    power off -r /dut/core/*
 } 
 
 #coverage report -file wally-pipelined-coverage.txt
 # These aren't doing anything helpful
 #coverage report -memory 
 #profile report -calltree -file wally-pipelined-calltree.rpt -cutoff 2
+power report -all -bsaif power.saif
+quit

pipelined/regression/wally-pipelined.do

@@ -57,12 +57,12 @@ if {$2 eq "buildroot" || $2 eq "buildroot-checkpoint"} {
     do wave.do
 
     # power add generates the logging necessary for saif generation.
-    #power add -r /dut/core/*
+    power add -r /dut/core/*
     #-- Run the Simulation 
 
     run -all
-    #power off -r /dut/core/*
+    power off -r /dut/core/*
-    #power report -all -bsaif power.saif
+    power report -all -bsaif power.saif
     noview ../testbench/testbench.sv
     view wave
 }

pipelined/src/fpu/fma.sv

@@ -193,7 +193,7 @@ module expadd(
             endcase
         end
 
-    end else begin
+    end else if (`FPSIZES == 4) begin
         always_comb begin
             case (FmtE)
                 2'h3: Denorm = 1;
@@ -619,7 +619,7 @@ module normalize(
             endcase
         end
 
-    end else begin
+    end else if (`FPSIZES == 4) begin
         always_comb begin
             case (FmtM)
                 2'h3: SumExpTmp = SumExpTmpTmp;
@@ -664,7 +664,7 @@ module normalize(
             endcase
         end
 
-    end else begin
+    end else if (`FPSIZES == 4) begin
         logic Sum0LEZ, Sum0GEFL, Sum1LEZ, Sum1GEFL, Sum2LEZ, Sum2GEFL, Sum3LEZ, Sum3GEFL;
         assign Sum0LEZ  = SumExpTmpTmp[`NE+1] | ~|SumExpTmpTmp;
         assign Sum0GEFL = $signed(SumExpTmpTmp) >= $signed(-(`NE+2)'(`NF  )-(`NE+2)'(2));
@@ -719,7 +719,7 @@ module normalize(
         (|CorrSumShifted[3*`NF+2-`NF1:2*`NF+3]&((FmtM==`FMT1)|(FmtM==`FMT2))) | 
         (|CorrSumShifted[3*`NF+2-`NF2:3*`NF+3-`NF1]&(FmtM==`FMT2));
 
-    end else begin        
+    end else if (`FPSIZES == 4) begin        
         assign NormSumSticky = (|CorrSumShifted[2*`NF+2:0]) | 
         (|CorrSumShifted[3*`NF+2-`D_NF:2*`NF+3]&((FmtM==1)|(FmtM==0)|(FmtM==2))) | 
         (|CorrSumShifted[3*`NF+2-`S_NF:3*`NF+3-`D_NF]&((FmtM==0)|(FmtM==2))) |
@@ -876,7 +876,7 @@ module fmaround(
             endcase
         end
 
-    end else begin
+    end else if (`FPSIZES == 4) begin
         always_comb begin
             case (FmtM)
                 2'h3: begin
@@ -995,7 +995,7 @@ module fmaround(
             endcase
         end
 
-    end else begin        
+    end else if (`FPSIZES == 4) begin        
         always_comb begin
             case (FmtM)
                 2'h3: RoundAdd = Minus1 ? {`FLEN+1{1'b1}} : {{{`FLEN{1'b0}}}, Plus1};
@@ -1063,7 +1063,7 @@ module fmaflags(
             endcase
         end
 
-    end else begin        
+    end else if (`FPSIZES == 4) begin        
         always_comb begin
             case (FmtM)
                 2'h3: GtMaxExp =  &FullResultExp[`NE-1:0] | FullResultExp[`NE];
@@ -1223,7 +1223,7 @@ module resultselect(
             endcase
         end
 
-    end else begin 
+    end else if (`FPSIZES == 4) begin 
         always_comb begin
             case (FmtM)
                 2'h3: begin  

pipelined/src/fpu/unpack.sv

@@ -261,7 +261,7 @@ module unpack (
             endcase
         end
 
-    end else begin      // if all precsisons are supported - quad, double, single, and half
+    end else if (`FPSIZES == 4) begin      // if all precsisons are supported - quad, double, single, and half
     
         //    quad   |  double  |  single  |  half    
         //-------------------------------------------------------------------

pipelined/testbench/tests.vh

@@ -1450,30 +1450,30 @@ string imperas32f[] = '{
 
  string wally64priv[] = '{
     `WALLYTEST,
-    "rv64i_m/privilege/WALLY-CSR-permission-s-01", "0050a0",
+    "rv64i_m/privilege/WALLY-CSR-permission-s-01", "0060a0",
-    //"rv64i_m/privilege/WALLY-CSR-PERMISSIONS-M", "005070",
+    //"rv64i_m/privilege/WALLY-CSR-PERMISSIONS-M", "005090",
-    //"rv64i_m/privilege/WALLY-CSR-PERMISSIONS-S", "003070",
+    //"rv64i_m/privilege/WALLY-CSR-PERMISSIONS-S", "003090",
-    "rv64i_m/privilege/WALLY-CSR-permission-u-01", "0050a0",
+    "rv64i_m/privilege/WALLY-CSR-permission-u-01", "0060a0",
-   // "rv64i_m/privilege/WALLY-MARCHID", "003070",
+   // "rv64i_m/privilege/WALLY-MARCHID", "004090",
-/*    "rv64i_m/privilege/WALLY-MCAUSE", "003070",
+/*    "rv64i_m/privilege/WALLY-MCAUSE", "003090",
-    "rv64i_m/privilege/WALLY-MEDELEG", "003070",
+    "rv64i_m/privilege/WALLY-MEDELEG", "004090",
-    "rv64i_m/privilege/WALLY-MHARTID", "003070",
+    "rv64i_m/privilege/WALLY-MHARTID", "004090",
-    "rv64i_m/privilege/WALLY-MIMPID", "003070",*/
+    "rv64i_m/privilege/WALLY-MIMPID", "004090",*/
-    "rv64i_m/privilege/WALLY-minfo-01", "0040a0",
+    "rv64i_m/privilege/WALLY-minfo-01", "0050a0",
-    "rv64i_m/privilege/WALLY-misa-01", "0040a0",
+    "rv64i_m/privilege/WALLY-misa-01", "0050a0",
-    "rv64i_m/privilege/WALLY-MMU-SV39", "0040a0",
+    "rv64i_m/privilege/WALLY-MMU-SV39", "0050a0",
-    "rv64i_m/privilege/WALLY-MMU-SV48", "0040a0",
+    "rv64i_m/privilege/WALLY-MMU-SV48", "0050a0",
-/*    "rv64i_m/privilege/WALLY-MSTATUS", "002070",
+/*    "rv64i_m/privilege/WALLY-MSTATUS", "002090",
-    "rv64i_m/privilege/WALLY-MTVEC", "002070",
+    "rv64i_m/privilege/WALLY-MTVEC", "002090",
-    "rv64i_m/privilege/WALLY-MVENDORID", "003070", */
+    "rv64i_m/privilege/WALLY-MVENDORID", "004090", */
-    "rv64i_m/privilege/WALLY-PMA", "0040a0",
+    "rv64i_m/privilege/WALLY-PMA", "0050a0",
-    "rv64i_m/privilege/WALLY-PMP", "0040a0",
+    "rv64i_m/privilege/WALLY-PMP", "0050a0"
-//    "rv64i_m/privilege/WALLY-SCAUSE", "002070",
+//    "rv64i_m/privilege/WALLY-SCAUSE", "002090",
-    "rv64i_m/privilege/WALLY-scratch-01", "0040a0",
+//    "rv64i_m/privilege/WALLY-scratch-01", "0040a0",
-    "rv64i_m/privilege/WALLY-sscratch-s-01", "0040a0"
+//    "rv64i_m/privilege/WALLY-sscratch-s-01", "0040a0",
-//    "rv64i_m/privilege/WALLY-trap-01", "0040a0"
+//    "rv64i_m/privilege/WALLY-trap-01", "0050a0"
-//    "rv64i_m/privilege/WALLY-STVEC", "002070",
+//    "rv64i_m/privilege/WALLY-STVEC", "002090",
-//    "rv64i_m/privilege/WALLY-UCAUSE", "002070",
+//    "rv64i_m/privilege/WALLY-UCAUSE", "002090",
 
  };
 

tests/fp/sample/Makefile

@@ -1,19 +0,0 @@
-# Makefile
-
-CC     = gcc
-CFLAGS = -O3
-LIBS   = -lm
-LFLAGS = -L. 
-IFLAGS = -I../../../addins/SoftFloat-3e/source/include/
-LIBS   = ../../../addins/SoftFloat-3e/build/Linux-x86_64-GCC/softfloat.a
-SRCS   = $(wildcard *.c)
-
-PROGS = $(patsubst %.c,%,$(SRCS))
-
-all:	$(PROGS)
-
-%: %.c
-	$(CC) $(CFLAGS) $(IFLAGS) $(LFLAGS) -o $@ $< $(LIBS)
-
-clean: 
-	rm -f $(PROGS)

tests/fp/sample/div.c

@@ -1,52 +0,0 @@
-#include <stdio.h>
-#include <stdint.h>
-#include "softfloat.h"
-#include "softfloat_types.h"
-
-int float_rounding_mode = 0;
-
-union dp {
-  unsigned short x[4];
-  double y;
-} X;
-
-
-int main()
-{
-    uint8_t rounding_mode;
-    uint8_t exceptions;
-
-    uint64_t n, d, result;
-    float64_t   d_n, d_d, d_result;
-
-    n = 0x3feffffffefffff6;
-    d = 0xffeffffffffffffe;
-    //n = 0x00000000400001ff;
-    //d = 0x3ffffdfffffffbfe;    
-
-    d_n.v = n;
-    d_d.v = d;
-
-    softfloat_roundingMode = rounding_mode;
-    softfloat_exceptionFlags = 0;
-    softfloat_detectTininess = softfloat_tininess_beforeRounding;
-
-    d_result = f64_div(d_n, d_d);
-
-    //result = d_result.v;    
-    //exceptions = softfloat_exceptionFlags & 0x1f;
-
-    X.x[3] = (d_result.v & 0xffff000000000000) >> 48;
-    X.x[2] = (d_result.v & 0x0000ffff00000000) >> 32;    
-    X.x[1] = (d_result.v & 0x00000000ffff0000) >> 16;
-    X.x[0] = (d_result.v & 0x000000000000ffff);
-
-    printf("Number = %.4x\n", X.x[3]);
-    printf("Number = %.4x\n", X.x[2]);
-    printf("Number = %.4x\n", X.x[1]);
-    printf("Number = %.4x\n", X.x[0]);
-    printf("Number = %1.25lg\n", X.y);    
-    
-
-    return 0;
-}

tests/fp/sample/fma.c

@@ -1,47 +0,0 @@
-#include <stdio.h>
-#include <stdint.h>
-#include "softfloat.h"
-#include "softfloat_types.h"
-
-int float_rounding_mode = 0;
-
-union sp {
-  unsigned short x[2];
-  float y;
-} X;
-
-
-int main()
-{
-    uint8_t rounding_mode;
-    uint8_t exceptions;
-
-    uint32_t multiplier, multiplicand, addend, result;
-    float32_t f_multiplier, f_multiplicand, f_addend, f_result;
-
-    multiplier = 0xbf800000;
-    multiplicand = 0xbf800000;
-    addend = 0xffaaaaaa;
-
-    f_multiplier.v = multiplier;
-    f_multiplicand.v = multiplicand;
-    f_addend.v = addend;
-
-    softfloat_roundingMode = rounding_mode;
-    softfloat_exceptionFlags = 0;
-    softfloat_detectTininess = softfloat_tininess_beforeRounding;
-
-    f_result = f32_mulAdd(f_multiplier, f_multiplicand, f_addend);
-
-    result = f_result.v;    
-    exceptions = softfloat_exceptionFlags & 0x1f;
-
-    printf("%x\n", f_result.v);
-
-    // Print out SP number
-    X.x[1] = (f_result.v & 0xffff0000) >> 16;
-    X.x[0] = (f_result.v & 0x0000ffff);
-    printf("Number = %f\n", X.y);
-
-    return 0;
-}

tests/wally-riscv-arch-test/riscv-test-suite/rv64i_m/privilege/references/WALLY-PMA.reference_output

@@ -1,23 +1,11 @@
 beef00b4 # Test 12.3.2.1: read 64 bits success in CLINT 
 0000dead # all of these read successes are also confirming successful writes 
-00000007 # write 32 bits with access fault in CLINT
+beef00b5 # read 32 bits success in CLINT (bottom 32 bits sign extended)
-00000000
+ffffffff
-00000005 # read 32 bits with access fault in CLINT
+000000b6 # read 16 bits success in CLINT (bottom 16 bits sign extended)
-00000000
-00000bad
-00000000
-00000007 # write 16 bits with access fault in CLINT
-00000000
-00000005 # read 16 bits with access fault in CLINT
-00000000
-00000bad
-00000000
-00000007 # write 8 bits with access fault in CLINT
-00000000
-00000005 # read 8 bits with access fault in CLINT
-00000000
-00000bad
 00000000
+ffffffb7 # read 8 bits success in CLINT (bottom 8 bits sign extended)
+ffffffff
 00000001 # execute test with access fault in CLINT
 00000000
 00000bad

tests/wally-riscv-arch-test/riscv-test-suite/rv64i_m/privilege/src/WALLY-PMA.S

@@ -85,12 +85,12 @@ test_cases:
 # Use timecmp register as readable and writable section of the CLINT
 .8byte CLINT_BASE + 0x4000, 0x0000DEADBEEF00B4, write64_test  # 64-bit write:  success
 .8byte CLINT_BASE + 0x4000, 0x0000DEADBEEF00B4, read64_test # 64-bit read:   success
-.8byte CLINT_BASE + 0x4000, 0x0000DEADBEEF00B5, write32_test # 32-bit write:  failure *** due to non-native length access
+.8byte CLINT_BASE + 0x4000, 0x0000DEADBEEF00B5, write32_test # 32-bit write: success
-.8byte CLINT_BASE + 0x4000, 0x0000DEADBEEF00B5, read32_test # 32-bit read:   failure
+.8byte CLINT_BASE + 0x4000, 0x0000DEADBEEF00B5, read32_test # 32-bit read:   success
-.8byte CLINT_BASE + 0x4000, 0x0000DEADBEEF00B6, write16_test # 16-bit write:  failure
+.8byte CLINT_BASE + 0x4000, 0x0000DEADBEEF00B6, write16_test # 16-bit write: success
-.8byte CLINT_BASE + 0x4000, 0x0000DEADBEEF00B6, read16_test # 16-bit read:   failure
+.8byte CLINT_BASE + 0x4000, 0x0000DEADBEEF00B6, read16_test # 16-bit read:   success
-.8byte CLINT_BASE + 0x4000, 0x0000DEADBEEF00B7, write08_test # 08-bit write:  failure
+.8byte CLINT_BASE + 0x4000, 0x0000DEADBEEF00B7, write08_test # 08-bit write: success
-.8byte CLINT_BASE + 0x4000, 0x0000DEADBEEF00B7, read08_test # 08-bit read:   failure
+.8byte CLINT_BASE + 0x4000, 0x0000DEADBEEF00B7, read08_test # 08-bit read:   success
 
 .8byte CLINT_BASE, 0xbad, executable_test# execute: instruction access fault
 

tests/wally-riscv-arch-test/riscv-test-suite/rv64i_m/privilege/src/WALLY-TEST-LIB-64.h

@@ -41,7 +41,7 @@ RVTEST_CODE_BEGIN
     //   
     //   Initialize x6 as a virtual pointer to the test results
     //   Initialize x16 as a physical pointer to the test results
-    //   Set up stack pointer (sp = x2)
+    //   Set up stack pointer, mscratch, sscratch
     //   
 	// ---------------------------------------------------------------------------------------------
 
@@ -50,8 +50,12 @@ RVTEST_CODE_BEGIN
     la x16, test_1_res // x16 reserved for the physical address equivalent of x6 to be used in trap handlers
                         // any time either is used, both must be updated.
 
-    // address for stack
+    // address for normal user stack, mscratch stack, and sscratch stack
-    la sp, top_of_stack
+    la sp, mscratch_top
+    csrw mscratch, sp
+    la sp, sscratch_top
+    csrw sscratch, sp
+    la sp, stack_top
 
 .endm
 
@@ -72,7 +76,7 @@ cause_instr_addr_misaligned:
 
 cause_instr_access:
     la x28, 0x0 // address zero is an address with no memory
-    sd x1, -8(sp) // push the return adress ontot the stack
+    sd x1, -8(sp) // push the return adress onto the stack
     addi sp, sp, -8
     jalr x28 // cause instruction access trap
     ld x1, 0(sp) // pop return adress back from the stack
@@ -83,7 +87,7 @@ cause_illegal_instr:
     .word 0x00000000 // a 32 bit zros is an illegal instruction
     ret
 
-cause_breakpnt: // ****
+cause_breakpnt:
     ebreak
     ret
 
@@ -128,7 +132,9 @@ cause_time_interrupt:
     sw x31,4(x29)          // store into most significant word of MTIMECMP
 nowrap:
     sw x28, 0(x29)         // store into least significant word of MTIMECMP
-    loop: j loop         // wait until interrupt occurs
+loop:
+    wfi 
+    j loop         // wait until interrupt occurs
     ret
 
 cause_soft_interrupt:
@@ -150,7 +156,7 @@ end_trap_triggers:
 
 .macro TRAP_HANDLER MODE, VECTORED=1, DEBUG=0
     // MODE decides which mode this trap handler will be taken in (M or S mode)
-    // Vectored decides whether interrumpts are handled with the vector table at trap_handler_MODE (1)
+    // Vectored decides whether interrupts are handled with the vector table at trap_handler_MODE (1)
     //      vs Using the non-vector approach the rest of the trap handler takes (0)
     // DEBUG decides whether we will print mtval a string with status.mpie, status.mie, and status.mpp to the signature (1)
     //      vs not saving that info to the signature (0)
@@ -216,23 +222,28 @@ trap_handler_\MODE\():
     // *** ASSUMES that a cause value of 0 for an interrupt is unimplemented
     // otherwise, a vectored interrupt handler should jump to trap_handler_\MODE\() + 4 * Interrupt cause code
     // No matter the value of VECTORED, exceptions (not interrupts) are handled in an unvecotred way
-    j soft_interrupt_\MODE\()    // 1: instruction access fault // the zero spot is taken up by the instruction to skip this table.
+    j s_soft_vector_\MODE\()    // 1: instruction access fault // the zero spot is taken up by the instruction to skip this table.
     j segfault_\MODE\()            // 2: reserved
-    j soft_interrupt_\MODE\()    // 3: breakpoint
+    j m_soft_vector_\MODE\()    // 3: breakpoint
     j segfault_\MODE\()            // 4: reserved
-    j time_interrupt_\MODE\()    // 5: load access fault
+    j s_time_vector_\MODE\()    // 5: load access fault
     j segfault_\MODE\()            // 6: reserved
-    j time_interrupt_\MODE\()    // 7: store access fault
+    j m_time_vector_\MODE\()    // 7: store access fault
     j segfault_\MODE\()            // 8: reserved
-    j ext_interrupt_\MODE\()     // 9: ecall from S-mode
+    j s_ext_vector_\MODE\()     // 9: ecall from S-mode
     j segfault_\MODE\()            // 10: reserved
-    j ext_interrupt_\MODE\()     // 11: ecall from M-mode
+    j m_ext_vector_\MODE\()     // 11: ecall from M-mode
     // 12 through >=16 are reserved or designated for platform use
 
 trap_unvectored_\MODE\():
+    csrrw sp, \MODE\()scratch, sp // swap sp and scratch so we can use the scratch stack in the trap hanler without messing up sp's value or the stack itself.
+    // *** NOTE: this means that nested traps will be screwed up but they shouldn't happen in any of these tests
+
+trap_stack_saved_\MODE\(): // jump here after handling vectored interupt since we already switch sp and scratch there
     // save registers on stack before using
     sd x1, -8(sp)       
     sd x5, -16(sp)
+    sd x7, -24(sp) 
 
     // Record trap
     csrr x1, \MODE\()cause     // record the mcause
@@ -262,49 +273,35 @@ trap_unvectored_\MODE\():
     // Respond to trap based on cause
     // All interrupts should return after being logged
     csrr x1, \MODE\()cause
+    slli x1, x1, 3          // multiply cause by 8 to get offset in vector Table
     li x5, 0x8000000000000000   // if msb is set, it is an interrupt
     and x5, x5, x1
-    bnez x5, trapreturn_\MODE\()   // return from interrupt
+    bnez x5, interrupt_handler_\MODE\()  // return from interrupt
     // Other trap handling is specified in the vector Table
-    slli x1, x1, 3      // multiply cause by 8 to get offset in vector Table
     la x5, exception_vector_table_\MODE\()
     add x5, x5, x1      // compute address of vector in Table
     ld x5, 0(x5)        // fectch address of handler from vector Table
     jr x5               // and jump to the handler
 
+interrupt_handler_\MODE\():
+    la x5, interrupt_vector_table_\MODE\() // NOTE THIS IS NOT THE SAME AS VECTORED INTERRUPTS!!!
+    add x5, x5, x1      // compute address of vector in Table
+    ld x5, 0(x5)        // fectch address of handler from vector Table
+    jr x5               // and jump to the handler
+
 segfault_\MODE\():
-    ld x5, -16(sp)      // restore registers from stack before faulting
+    sd x7, -24(sp)  // restore registers from stack before faulting 
+    ld x5, -16(sp)
     ld x1, -8(sp)       
     j terminate_test          // halt program.
 
 trapreturn_\MODE\():
-    // look at the instruction to figure out whether to add 2 or 4 bytes to PC, or go to address specified in a1
+    csrr x1, \MODE\()epc
-    csrr x1, \MODE\()epc       // get the mepc
+    addi x1, x1, 4 // return to the address AFTER the trapping instruction
-    addi x1, x1, 4 // *** should be 2 for compressed instructions, see note.
-
-
-// ****** KMG: the following is no longer as easy to determine. mepc gets the virtual address of the trapped instruction, 
-// ********     but in the handler, we work in M mode with physical addresses
-//              This means the address in mepc is suddenly pointing somewhere else.
-//              to get this to work, We could either retranslate the vaddr back into a paddr (probably on the scale of difficult to intractible)
-//              or we could come up with some other ingenious way to stay in M mode and see if the instruction was compressed.
-
-//     lw x5, 0(x1)        // read the faulting instruction
-//     li x1, 3            // check bottom 2 bits of instruction to see if compressed
-//     and x5, x5, x1      // mask the other bits
-//     beq x5, x1, trapreturn_uncompressed_\MODE\()  // if 11, the instruction is return_uncompressed
-
-// trapreturn_compressed_\MODE\():
-//     csrr x1, mepc       // get the mepc again
-//     addi x1, x1, 2      // add 2 to find the next instruction
-//     j trapreturn_specified_\MODE\() // and return
-
-// trapreturn_uncompressed_\MODE\():
-//      csrr x1, mepc       // get the mepc again    
-//      addi x1, x1, 4      // add 4 to find the next instruction
 
 trapreturn_specified_\MODE\():
     // reset the necessary pointers and registers (x1, x5, x6, and the return address going to mepc)
+    // note that we don't need to change x7 since it was a temporary register with no important address in it.
     // so that when we return to a new virtual address, they're all in the right spot as well.
 
     beqz a1, trapreturn_finished_\MODE\() // either update values, of go to default return address.
@@ -318,13 +315,13 @@ trapreturn_specified_\MODE\():
     sll x5, x5, a2
     addi x5, x5, -1 // x5 = mask bits for offset into current pagetype
 
-    // reset the top of the stack, x1
+    // reset the top of the stack, which will be put into ra
     ld x7, -8(sp) 
     and x7, x5, x7 // x7 = offset for x1
     add x7, x7, a1 // x7 = new address for x1
     sd x7, -8(sp)
 
-    // reset the second spot in the stack, x5
+    // reset the second spot in the stack, which will be put into x5
     ld x7, -16(sp)
     and x7, x5, x7 // x7 = offset for x5
     add x7, x7, a1 // x7 = new address for x5
@@ -334,7 +331,7 @@ trapreturn_specified_\MODE\():
     and x7, x5, x6 // x7 = offset for x6
     add x6, x7, a1 // x6 = new address for the result pointer
     
-    // set return address, stored temporarily in x1, to the next instruction, but in the new virtual page.
+    // reset x1, which temporarily holds the return address that will be written to mepc.
     and x1, x5, x1 // x1 = offset for the return address
     add x1, x1, a1 // x1 = new return address.
 
@@ -343,9 +340,9 @@ trapreturn_specified_\MODE\():
 
 trapreturn_finished_\MODE\():
     csrw \MODE\()epc, x1            // update the epc with address of next instruction
-    ld x5, -16(sp)      // restore registers from stack before returning
+    ld x7, -24(sp)     // restore registers from stack before returning
+    ld x5, -16(sp)
     ld x1, -8(sp)
-    // *** this should be handled by indirectly clearing this bit csrw \MODE\()ip, 0x0 // clear interrupt pending register to indicate interrupt has been handled
     \MODE\()ret  // return from trap
 
 ecallhandler_\MODE\():
@@ -380,18 +377,17 @@ ecallhandler_changetousermode_\MODE\():
     j trapreturn_\MODE\()
 
 instrpagefault_\MODE\():
-    ld x1, -8(sp) // load return address int x1 (the address AFTER the jal to the faulting address)
+    ld x1, -8(sp) // load return address from stack into ra (the address AFTER the jal to the faulting address)
     j trapreturn_finished_\MODE\() // puts x1 into mepc, restores stack and returns to program (outside of faulting page)
 
 instrfault_\MODE\():
-    ld x1, -8(sp) // load return address int x1 (the address AFTER the jal to the faulting address)
+    ld x1, -8(sp) // load return address from stack into ra (the address AFTER the jal to the faulting address)
-    j trapreturn_finished_\MODE\() // return to the code after recording the mcause
+    j trapreturn_finished_\MODE\() // return to the code at ra value from before trap
 
 illegalinstr_\MODE\():
     j trapreturn_\MODE\() // return to the code after recording the mcause
 
 accessfault_\MODE\():
-    // *** What do I have to do here?
     j trapreturn_\MODE\()
 
 addr_misaligned_\MODE\():
@@ -400,36 +396,67 @@ addr_misaligned_\MODE\():
 breakpt_\MODE\():
     j trapreturn_\MODE\()
 
-soft_interrupt_\MODE\():
+s_soft_vector_\MODE\():
-    li x5, 0x7EC // write 0x7EC (looks like VEC) to the output before the mcause and extras to indicate that this trap was handled with a vector table. 
+    csrrw sp, \MODE\()scratch, sp // swap sp and scratch so we can use the scratch stack in the trap hanler without messing up sp's value or the stack itself.
-    sd x5, 0(x16)
+    sd x5, -8(sp) // put x5 on the scratch stack before messing with it
+    li x5, 0x7EC01 // write 0x7ec01 (for "VEC"tored and 01 for the interrupt code)
+    j vectored_int_end_\MODE\()
+
+m_soft_vector_\MODE\():
+    csrrw sp, \MODE\()scratch, sp // swap sp and scratch so we can use the scratch stack in the trap hanler without messing up sp's value or the stack itself.
+    sd x5, -8(sp) // put x5 on the scratch stack before messing with it
+    li x5, 0x7EC03 // write 0x7ec03 (for "VEC"tored and 03 for the interrupt code)
+    j vectored_int_end_\MODE\()
+
+s_time_vector_\MODE\():
+    csrrw sp, \MODE\()scratch, sp // swap sp and scratch so we can use the scratch stack in the trap hanler without messing up sp's value or the stack itself.
+    sd x5, -8(sp) // put x5 on the scratch stack before messing with it
+    li x5, 0x7EC05 // write 0x7ec05 (for "VEC"tored and 05 for the interrupt code)
+    j vectored_int_end_\MODE\()
+
+m_time_vector_\MODE\():
+    csrrw sp, \MODE\()scratch, sp // swap sp and scratch so we can use the scratch stack in the trap hanler without messing up sp's value or the stack itself.
+    sd x5, -8(sp) // put x5 on the scratch stack before messing with it
+    li x5, 0x7EC07 // write 0x7ec07 (for "VEC"tored and 07 for the interrupt code)
+    j vectored_int_end_\MODE\()
+
+s_ext_vector_\MODE\():
+    csrrw sp, \MODE\()scratch, sp // swap sp and scratch so we can use the scratch stack in the trap hanler without messing up sp's value or the stack itself.
+    sd x5, -8(sp) // put x5 on the scratch stack before messing with it
+    li x5, 0x7EC09 // write 0x7ec09 (for "VEC"tored and 08 for the interrupt code)
+    j vectored_int_end_\MODE\()
+
+m_ext_vector_\MODE\():
+    csrrw sp, \MODE\()scratch, sp // swap sp and scratch so we can use the scratch stack in the trap hanler without messing up sp's value or the stack itself.
+    sd x5, -8(sp) // put x5 on the scratch stack before messing with it
+    li x5, 0x7EC0B // write 0x7ec0B (for "VEC"tored and 0B for the interrupt code)
+    j vectored_int_end_\MODE\()
+
+vectored_int_end_\MODE\():
+    sd x5, 0(x16) // store to signature to show vectored interrupts succeeded. 
     addi x6, x6, 8
     addi x16, x16, 8
+    ld x5, -8(sp) // restore x5 before continuing to handle trap in case its needed.
+    j trap_stack_saved_\MODE\()
+
+soft_interrupt_\MODE\():
     la x28, 0x02000000 // Reset by clearing MSIP interrupt from CLINT
     sw x0, 0(x28)
-    j trap_unvectored_\MODE\()
+    j trapreturn_\MODE\()
 
 time_interrupt_\MODE\():
-    li x5, 0x7EC
-    sd x5, 0(x16)
-    addi x6, x6, 8
-    addi x16, x16, 8
-
     la x29, 0x02004000    // MTIMECMP register in CLINT
     li x30, 0xFFFFFFFF
     sd x30, 0(x29) // reset interrupt by setting mtimecmp to 0xFFFFFFFF
     
-    j trap_unvectored_\MODE\()
+    ld x1, -8(sp) // load return address from stack into ra (the address AFTER the jal to the faulting address)
+    j trapreturn_finished_\MODE\() // return to the code at ra value from before trap
 
 ext_interrupt_\MODE\():
-    li x5, 0x7EC
-    sd x5, 0(x16)
-    addi x6, x6, 8
-    addi x16, x16, 8
     li x28, 0x10060000 // reset interrupt by clearing all the GPIO bits
     sw x0, 8(x28) // disable the first pin as an output
     sw x0, 40(x28) // write a 0 to the first output pin (reset interrupt)
-    j trap_unvectored_\MODE\()
+    j trapreturn_\MODE\()
 
     // Table of trap behavior
     // lists what to do on each exception (not interrupts)
@@ -455,6 +482,21 @@ exception_vector_table_\MODE\():
     .8byte segfault_\MODE\()      // 14: reserved
     .8byte trapreturn_\MODE\()    // 15: store page fault
 
+    .align 3 // aligns this data table to an 8 byte boundary
+interrupt_vector_table_\MODE\():
+    .8byte segfault_\MODE\()            // 0: reserved
+    .8byte s_soft_vector_\MODE\()    // 1: instruction access fault // the zero spot is taken up by the instruction to skip this table.
+    .8byte segfault_\MODE\()            // 2: reserved
+    .8byte m_soft_vector_\MODE\()    // 3: breakpoint
+    .8byte segfault_\MODE\()            // 4: reserved
+    .8byte s_time_vector_\MODE\()    // 5: load access fault
+    .8byte segfault_\MODE\()            // 6: reserved
+    .8byte m_time_vector_\MODE\()    // 7: store access fault
+    .8byte segfault_\MODE\()            // 8: reserved
+    .8byte s_ext_vector_\MODE\()     // 9: ecall from S-mode
+    .8byte segfault_\MODE\()            // 10: reserved
+    .8byte m_ext_vector_\MODE\()     // 11: ecall from M-mode
+
 .align 3
 trap_return_pagetype_table_\MODE\():
     .8byte 0xC  // 0: kilopage has 12 offset bits
@@ -1051,9 +1093,20 @@ rvtest_data:
 RVTEST_DATA_END
 
 .align 3 // align stack to 8 byte boundary
-bottom_of_stack:
+stack_bottom:
     .fill 1024, 4, 0xdeadbeef 
-top_of_stack:
+stack_top:
+
+.align 3
+mscratch_bottom:
+    .fill 512, 4, 0xdeadbeef
+mscratch_top:
+
+.align 3
+sscratch_bottom:
+    .fill 512, 4, 0xdeadbeef
+sscratch_top:
+
 
 
 RVMODEL_DATA_BEGIN