From 1f614a46b5c9ebceb5b33eb6335f187d3ac1f31c Mon Sep 17 00:00:00 2001
From: Stefan <stefan@pimaker.at>
Date: Tue, 31 Aug 2021 11:08:03 +0200
Subject: [PATCH] fixups and CSR memcpy test in bare_metal_test

---
 Makefile                         |  26 +++----
 bare_metal_test/Makefile         |   2 +-
 bare_metal_test/bare             | Bin 6336 -> 8468 bytes
 bare_metal_test/bare.c           | 113 ++++++++++++++++++++++++++++++-
 bare_metal_test/init.S           |   3 +-
 linux-_pi_-patches-for-rvc.patch |  33 ++++-----
 src/uart.h                       |   2 +-
 7 files changed, 145 insertions(+), 34 deletions(-)

diff --git a/Makefile b/Makefile
index f3f043da..dafb17fb 100644
--- a/Makefile
+++ b/Makefile
@@ -17,16 +17,6 @@ CFLAGS=-g
 $(TARGET): $(OBJS)
 	$(CC) $(LDFLAGS) -I./elfy/elfy.h $(OBJS) -o $@ $(LOADLIBES) $(LDLIBS) -L./elfy/target/release/ -Wl,--no-as-needed -ldl -lpthread -lelfy
 
-.PHONY: clean
-clean:
-	$(RM) $(TARGET) $(OBJS) $(DEPS)
-
-.PHONY: distclean
-distclean: clean
-	$(RM) $(TARGET) $(OBJS) $(DEPS) *_payload.*
-	$(MAKE) -C opensbi clean
-	$(MAKE) -C linux clean
-
 -include $(DEPS)
 
 # build device tree
@@ -39,9 +29,9 @@ LINUX_PAYLOAD=linux/arch/riscv/boot/Image
 
 BUILDROOT_MARKER=buildroot-2021.05/build.marker
 $(BUILDROOT_MARKER):
-	$(MAKE) -C buildroot-2021.05 make
+	$(MAKE) -C buildroot-2021.05
 	cd buildroot-2021.05 && cp init output/target/ && cp pi.js output/target/
-	$(MAKE) -C buildroot-2021.05 make
+	$(MAKE) -C buildroot-2021.05
 	touch $@
 
 $(PAYLOAD): $(shell find rust_payload/src -type f)
@@ -101,3 +91,15 @@ run: fw_payload.bin $(TARGET) dts.dtb
 
 run-v1: fw_payload.bin $(TARGET) dts.dtb
 	./rvc -b fw_payload.bin -d dts.dtb -v1
+
+
+.PHONY: clean
+clean:
+	$(RM) $(TARGET) $(OBJS) $(DEPS)
+
+.PHONY: distclean
+distclean: clean
+	$(RM) $(TARGET) $(OBJS) $(DEPS) *_payload.*
+	$(MAKE) -C opensbi clean
+	$(MAKE) -C linux clean
+	$(RM) $(BUILDROOT_MARKER)
diff --git a/bare_metal_test/Makefile b/bare_metal_test/Makefile
index 0b316554..35966e11 100644
--- a/bare_metal_test/Makefile
+++ b/bare_metal_test/Makefile
@@ -15,7 +15,7 @@ DEPS := $(OBJS:.o=.d)
 INC_DIRS := $(shell find $(SRC_DIRS) -type d)
 INC_FLAGS := $(addprefix -I,$(INC_DIRS))
 
-CFLAGS += -O3 -march=rv32ima -mabi=ilp32 -fno-stack-protector -fno-pie
+CFLAGS += -O2 -march=rv32ima -mabi=ilp32 -fno-stack-protector -fno-pie
 ASFLAGS += -march=rv32ima -mabi=ilp32
 LDFLAGS += -static -nostdlib -nostartfiles -nodefaultlibs -Triscv32.ld -Wl,--no-as-needed
 
diff --git a/bare_metal_test/bare b/bare_metal_test/bare
index 45102ac32cbe02cc335d7bb7c315b41666eca303..7008f5047ef478024c0e09aa22d32312fc2b2ec5 100755
GIT binary patch
literal 8468
zcmeHNe{2)i9sl0FyL(RfWw;P1Wz8ke4NeHgELp9hRZGfDhP0_LwowVO#9U0`E{TKB
zC@@vpWrve(>PFg&3fd4*O)XQmNifjdRhm|+ny|J&%aE9+3RZb7lSUU)endkw?E8FY
z2SVuJkNxK--M#Poe!rjZulId-cAOuCzrI=!1Skv<u7JBTn~=G)%8@jI1|s;O9LgNI
zH_tcmZU$!otiX$M7V_79l<eT-)S)25flZZ^)};!RDp0CGsRE@6lqyiFK&b+y3Y02P
zsz9j%r3#cP@c*oU;cX7$)wCPGoZEwgvN8OfphF+$wAWTNFosuq@d_Hlj|FYz5EuuP
z-~u5hsKUTaSNHqIaAi}vlG{*BE;m36#-SPF7`%G>B?y;T&0Ju$D}ec1fYq-9)(`;}
z*a}P=5LwOhBJ;f@vf9HU^B)sg{re(o_)uhl%OcYR8L=|+Rm+H%5icWNM!bx88SxbH
z6!8@C6!8@C6!8@C)bN4VVBU%=ELdR#v<l`!-`aqO)kHkZx7EXHkCmyi0Am18^H~*y
z3adu;5Ek|r!=l&lsuh+N2CF&)q(&C3ny^4vO0ZZ)AU;{Nd|}agG$WE)nOL=9LYAb6
zwIoA`UzRL?SR!>P$*Rjpq+XV-`mjuvreterMkWoiVl{*nvMi-o%Q6ZH$kYmiDOsMP
z*76J`kETSd=N*4K2~qG`trYC>@j2PObi!)gfHW1egLHzyLFhCA$Z8pEMF!oYh+N+u
ztb}`GMu8kxAp3}NlVdUW&sZ()hY$CorkXHq1z~{}fi<iXSbaob{;dMa1(XY{<|Tn?
zhXq?PA?<-&$OC2?{i(A|l9u5OduC8d6ZpG2bFF00XqVC=yr<4+meO<JfL-pDLUZ7d
zTCPdy*)VLE4U>=p&1zXQ3C)IfTWTjE0)48~M;iO*3wk*%)4Ts!q0b(jsXz1a?DQ<q
zbSXVUpZLUwbWop{CS;U?+K+WvOee_b2Bfhr3Y}1CkTCNi_UQFn>ijTd_b4}r?-=ge
zfJo|EM5|5^N&PXwsvi)@(iXv5DhQ<E7+4JhK$f+DwJhHgBwt(Cu3kyMSoDWJigYEN
zi03YRR`8E{fuij5^cH@;o(`k3zOZv~(bkk4nK^WOewX68s!xkCETX<5Zkiadj*@=$
zBabfiEBd?z4;OXs^3g98Yxtt_G3NQNqK^tX8QU~Iih7IaqN4q2^mG`P^~D$EqRw40
zv`)q_qMVQC_GgQ>7t(n}+%&ec-g6%o<#*W)P1@a~{1xkWUKgp9tsMX0UKo{n@9e9+
z%2|x(Z^d{jsHcdXCLZkDwb{z_xuVXLm{huUs{K=F|7>>ghBNoVFtc+qj}hfv)PG;4
zaHpL;%G;RVd(7N0sk=`(f%|!inY$->Bb>MYB=*)M@8`<#c&_bT)SK+LVa}_1N2S`G
zenynH5dUXp9=sj@CgM*pbIBxcSV3OZD85HbRXHT%yX5eLJh##d5J$bhesu8+Kco70
z-+;!!3|OQ198ApzRh|#OoO5j=yC*Z@KI3k2U(7cE_};*GK{Y<p=7L(LfGQzorwmRU
zgTH%_IMkE`W4Wr|IC+3*53WA?KzKQi{}spF>v?U{1HfP$-^~>>46hHb$7k4>yD>~>
zSlTfln3Lse0aBNLW(hps(0r_qAna4+)E`)2ZaZrjXlM1$w=;ir1mC9-=4*+tnst$}
z-2T0JT{-jmSnk3vLi11lT9v~L>3#nzCXVBAq5dDS27CCPST7HqA+L;{u_;_U^82)i
zefUu0&QbBm&a~Jl9c^xu&Vp|KLvnnE|AhWWoEl!(?Z)xuSgy@h^L5$(qQgsJY*jB9
z_=^JnjFH}OZNeU+-^f04^@eNbEtC(dGCahw4o@D%9De55sd^z)f%TAgqMiLeb>$~e
z{+zM4r6M2qDE*5*Li6@r`P(S}i7g45S|&87zuh!1bo{sH?J@!1vt^(TNnlf?Q>ji>
zx&Ci<AdNjW3hcmfu!mYfrLA~YRzr4V_XOAD`s8H_*2}FxUKs-G71Y%ZVlN`jhji^Z
z#8E+|RLJh#ebeTjNo1i}RY1d>&W2Pld0adWwH)8sWK~*~Ge>Pt?pD6vcpM(^Sg%T8
z9mM$fnc|-ZMs9%q!`I92_w#_Yqzq##L)<ygpTqsj+aZ|C{e7OW>JdA<NLI@i;hFj>
z*gsx_XKW2*E6T^+x}F-#J$22YQ>VvryRPy*<$kk+*KWphRsV5hnRmu=r$1HYay(yi
zuBZ9_rKFRne_2g&-s!8Z4dgwQbM$syade&<Q>oByFQ|tjbez}y0P0_M_J7wWg+AK2
zIg*GjPbMOLo3ze&Qd_@%&Dv_Lt;6q$_UL^(+H@naNi!1q6T$5mCTJZ-l9P^3G?9p-
z*iqH|J(1pRk?uNeOC;7E?LbQfTT}E_t^eljN_ktGv{<j!6N&Zu>p0NmC8I_%*4wG|
zC1Smt2m}J<uE>dJie4lUOLp}{lQErZ7VIo{{qrRI6LGYch#K3vliEtZW38aj9q)JZ
zyp!J*>+Es_u0L)`k0%mQJ?Yx2F2ps}#wM+Ro)RlByw(-z?dXms^3hI_J5`-758Q)|
znEod135{2QZx3I)cq7`qh1X5%i1$W|`?*Q$jrZd=Yl&^Wni0nub@9UX#1m01x?PVI
zs*5$dh2zy%0a{nBTB-S;+!pKZNW?MWWt~l%w>;MLrO>jDSR%QjPFoQSGzNl*n0A<)
zOThVswRRS`Ymu|TUtE*?8h=+cIYB_0vX=Af!i|99T68!PT)yQ&vJZcobvOdxSF_^a
z@-N_u+vkx8AVV$A+&{+&=+Te!2LhO1#(5rc0yvK-V1CVAujF!EYwnu!+g*QN)c+Q)
zahUwZ*Bf!ZM8Wi7YTRBsu3tm@KAidb8Jri2U_Ohf@^uo|0J#Bv=lExFzKQ;FICD9~
zdSl5z9NLUF-ZO2{-eh72@WUI>0XwnHFk*cWf&3u>r=#+k;GSXUTp~UhNhI4)q0^;-
zA2`^?xqaJ`x?`cOE4m%puz@1__X-y9DF6^KcJw48n{iGioVlw&C8OJu5a@^`BPeyI
z&4z(={;tP+dN637-~En<QIYQ#{~aBNiG=~|q2j=0epKU~>+vq)JywgG?Q~dPoR@KG
z30?&~zW09A<NIG21)A?Yr@1}eGeO=ww+(KCk4C(v>J?;MUZ5X$=^$s~;P(6unZI+r
z!gYJxPorKt3b>D%cwwvX65xL%II-cBAFn+EFjQFJT<|*=X`Vx2d<Flv;mN`R=YrmU
E0g1la7XSbN

delta 682
zcmY+CPfQa*6vkh7yE{Y?ww4ymmUh~ztcfJW8a5%(*fy>eXfX9)jDKidYGcG~vMpzq
z5L=GwGoCc@NQg0##+0KcxOvq(hh8{vQ7=S%Q%DJOm^Z)uzW2@B&A!WipLors)bN``
zQc(aVl#S{AK}k(uUi%^4&~PLWk>&U+Ob3>6Gf?VptKbFIuL?FAIs6nJ*BLR>Ai3Hp
zRy}}YUA-yv(>=ds2{y?uuQiJ+Yum*jgHCIb$*90Nk=bIya+)!@)nu7g^C#F+yHDg{
z!^)q`+L|H0Y}D>#hT0Ukvd!+=@0s5L0Q27Oe?i6yJ0CiK4x80!Vb+|z`M<~0Z52)?
zT64mL{dycf>QUU$!+4;FBk<r!*{wX{Rj1Bv$8{@qox7FV(qcV+3SUMH>_jpb=vFwO
zwZ0B;E0>$%Mt14h;)3f?(|FZ<deyYeNxVIpzLHRYz7c>7k|<#GSnhJF$NnS6FQan-
zdg>$gfdUu1T{sYp-XD>`%M+d@VTx)A4#PEKi<p)tHo*m|CG!8G9N{k#Z~DMHPgU?k
zssMY-V%P9n^xE54g3VWI^|JjOug6~C=h)CCn!dNRqf?x!m`%(t-YyXWkg&{yxF0)1
zf0Sp0aDXTK=#(M}y3#D~N~sV&=3~APHu#vI#1wAuq_P?YSm@ezxu)zAWd;2Q)3cwt

diff --git a/bare_metal_test/bare.c b/bare_metal_test/bare.c
index 79c66384..5adea961 100644
--- a/bare_metal_test/bare.c
+++ b/bare_metal_test/bare.c
@@ -16,26 +16,89 @@ void print(const char* str) {
     }
 }
 
+void print_hex(unsigned int number) {
+    for (int i = 0; i < 8; i++) {
+        switch (number & 0xF0000000) {
+            case 0x00000000: uart_putc('0'); break;
+            case 0x10000000: uart_putc('1'); break;
+            case 0x20000000: uart_putc('2'); break;
+            case 0x30000000: uart_putc('3'); break;
+            case 0x40000000: uart_putc('4'); break;
+            case 0x50000000: uart_putc('5'); break;
+            case 0x60000000: uart_putc('6'); break;
+            case 0x70000000: uart_putc('7'); break;
+            case 0x80000000: uart_putc('8'); break;
+            case 0x90000000: uart_putc('9'); break;
+            case 0xA0000000: uart_putc('A'); break;
+            case 0xB0000000: uart_putc('B'); break;
+            case 0xC0000000: uart_putc('C'); break;
+            case 0xD0000000: uart_putc('D'); break;
+            case 0xE0000000: uart_putc('E'); break;
+            case 0xF0000000: uart_putc('F'); break;
+        }
+        number <<= 4;
+    }
+}
+
 static bool got_ssip = false;
 volatile static long long a = 0xaaaaffffffff;
 
+static int pattern_good = 0x12345678;
+static int pattern_bad = 0xdeadbeef;
+
+static char print_buf[256];
+
 // runs in supervisor mode via mideleg (machine-mode trap delegation)
 void handle_trap(void) {
     print("[bare/trap] got SSIP!\n");
     got_ssip = true;
 }
 
+// byte-for-byte memcpy
+void memcpy_manual(unsigned char *dst, const unsigned char* src, int n) {
+    for (int i = 0; i < n; i++) {
+        dst[i] = src[i];
+    }
+}
+
+// CSR-based memcpy
+void memcpy_csr(unsigned char *dst, const unsigned char* src, int n) {
+    print("[bare/memcpy_csr] src=0x");
+    print_hex((unsigned int)src);
+    print(" dst=0x");
+    print_hex((unsigned int)dst);
+    print("\n");
+    __asm volatile ("csrw 0x0b1, %0" :: "r" (src));
+    __asm volatile ("csrw 0x0b2, %0" :: "r" (dst));
+    __asm volatile ("csrw 0x0b3, %0" :: "r" (n));
+    __asm volatile ("csrw 0x0b0, %0" :: "r" (1) : "memory");
+}
+
 // called from assembly, runs in supervisor mode (via sret)
 int main(void) {
     print("[bare] in main()\n");
 
+    print("[bare] testing printing...\n[bare] 0x");
+    print_hex(0xabcd1234);
+    uart_putc('\n');
+
     print("[bare] testing arithmetic...\n");
 
     long long b = a * 400LLU * -1;
     if (b != -0x10AAB2FFFFFFE70LL) {
-        print("[bare] got wrong result :(\n");
+        print("[bare/error] got wrong result :(\n[bare/error] low=0x");
+        print_hex((unsigned int)b);
+        print("\n[bare/error] high=0x");
+        print_hex((unsigned int)(b >> 32));
+        print("\n");
+        __asm volatile ("ebreak");
+        while (true) {}
     } else {
-        print("[bare] got correct result!\n");
+        print("[bare] got correct result!\n[bare] low=0x");
+        print_hex((unsigned int)b);
+        print("\n[bare] high=0x");
+        print_hex((unsigned int)(b >> 32));
+        print("\n");
     }
 
     print("[bare] testing trap handler...\n");
@@ -43,6 +106,52 @@ int main(void) {
 
     while (!got_ssip) {}
 
+    print("[bare] testing memcpy...\n");
+    int src[128];
+    int dst_man[128];
+    int dst_csr[128];
+
+    for (int i = 0; i < 128; i++) {
+        src[i] = pattern_good;
+        dst_man[i] = pattern_bad;
+        dst_csr[i] = pattern_bad;
+    }
+
+    memcpy_manual((unsigned char*)dst_man,
+                  (const unsigned char*)src,
+                  sizeof(src));
+
+    memcpy_csr((unsigned char*)dst_csr,
+               (const unsigned char*)src,
+               sizeof(src));
+
+    for (int i = 0; i < 128; i++) {
+        int bad = ~0;
+
+        if (dst_man[i] != src[i] || dst_man[i] != pattern_good) {
+            print("[bare/error] memcpy (manual) failed\n");
+            bad = dst_man[i];
+        }
+        if (dst_csr[i] != src[i] || dst_csr[i] != pattern_good) {
+            print("[bare/error] memcpy (CSR) failed\n");
+            bad = dst_csr[i];
+        }
+
+        if (bad != (~0)) {
+            print("[bare/error] dst[0x");
+            print_hex(i);
+            print("] = 0x");
+            print_hex(bad);
+            print(" src[0x");
+            print_hex(i);
+            print("] = 0x");
+            print_hex(src[i]);
+            print("\n");
+            __asm volatile ("ebreak");
+            while (true) {}
+        }
+    }
+
     print("[bare] selftest done!\n");
     print("[bare] now go run something more exciting...\n");
 
diff --git a/bare_metal_test/init.S b/bare_metal_test/init.S
index 16d2a6b8..9d37a06e 100644
--- a/bare_metal_test/init.S
+++ b/bare_metal_test/init.S
@@ -88,7 +88,7 @@ _s_trap_entry:
     STORE  t2, 34*REGBYTES(x2)
     STORE  t3, 35*REGBYTES(x2)
 
-    move  a0, sp
+    mv   a0, sp
     jal handle_trap
 
     mv   a0, sp
@@ -101,7 +101,6 @@ _s_trap_entry:
 
     # restore x registers
     LOAD  x1,1*REGBYTES(a0)
-    LOAD  x2,2*REGBYTES(a0)
     LOAD  x3,3*REGBYTES(a0)
     LOAD  x4,4*REGBYTES(a0)
     LOAD  x5,5*REGBYTES(a0)
diff --git a/linux-_pi_-patches-for-rvc.patch b/linux-_pi_-patches-for-rvc.patch
index 225a4ba7..a940012a 100644
--- a/linux-_pi_-patches-for-rvc.patch
+++ b/linux-_pi_-patches-for-rvc.patch
@@ -1,44 +1,45 @@
-From fbba62aff7a727723aa75715e1d1838b484ba842 Mon Sep 17 00:00:00 2001
+From 30022eebb78abc5a8a9289aecbd30a155fafcff0 Mon Sep 17 00:00:00 2001
 From: Stefan <stefan@pimaker.at>
 Date: Sun, 1 Aug 2021 20:58:20 +0200
 Subject: [PATCH] _pi_ patches for rvc
 
-* use MEMOP extension for faster memcpy (disabled atm)
+* use MEMOP extension for faster memcpy
 * don't poison init kernel area, we don't care about security
 * print more messages about initramfs loading, since otherwise it just
   looks like it got stuck
 
 Signed-off-by: Stefan <stefan@pimaker.at>
 ---
- arch/riscv/lib/memcpy.S | 13 +++++++++++++
+ arch/riscv/lib/memcpy.S | 14 ++++++++++++++
  include/linux/mm.h      |  2 +-
  init/initramfs.c        | 19 +++++++++++++++++--
- 3 files changed, 31 insertions(+), 3 deletions(-)
+ 3 files changed, 32 insertions(+), 3 deletions(-)
 
 diff --git a/arch/riscv/lib/memcpy.S b/arch/riscv/lib/memcpy.S
-index 51ab716253fa..67636579413d 100644
+index 51ab716253fa..67e3cf24d9d0 100644
 --- a/arch/riscv/lib/memcpy.S
 +++ b/arch/riscv/lib/memcpy.S
-@@ -41,6 +41,18 @@ WEAK(memcpy)
+@@ -41,6 +41,19 @@ WEAK(memcpy)
  	beqz a4, 4f
  	add a3, a1, a4
  3:
-+	#lla t0, 10f
-+	#li t2, 1025
-+	#bltu a2, t2, 11f
++	# fast memcpy for page-sized chunks
++	lla t0, 10f
++	li t2, 4096
++	bne a2, t2, 11f
 +
-+	#csrw 0x0b1, a1 # src
-+	#csrw 0x0b2, t6 # dst
-+	#csrw 0x0b3, a2 # n
-+	#li t2, 0x1
-+	#csrw 0x0b0, t2 # memcpy
-+	#j 10f
++	csrw 0x0b1, a1 # src
++	csrw 0x0b2, t6 # dst
++	csrw 0x0b3, a2 # n
++	li t2, 0x1
++	csrw 0x0b0, t2 # memcpy
++	j 10f
 +
 +11:
  	REG_L a4,       0(a1)
  	REG_L a5,   SZREG(a1)
  	REG_L a6, 2*SZREG(a1)
-@@ -76,6 +88,7 @@ WEAK(memcpy)
+@@ -76,6 +89,7 @@ WEAK(memcpy)
  	REG_S t1, 15*SZREG(t6)
  	addi t6, t6, 16*SZREG
  	bltu a1, a3, 3b
diff --git a/src/uart.h b/src/uart.h
index 530a779f..eb366489 100644
--- a/src/uart.h
+++ b/src/uart.h
@@ -57,7 +57,7 @@ void uart_tick(cpu_t *cpu) {
 
     uint thr = UART_GET1(THR);
     if ((cpu->clock % 0x16) == 0 && thr != 0) {
-        printf(" %d", (unsigned char)thr);
+        printf("%c", (unsigned char)thr);
         fflush(stdout);
         UART_SET1(THR, 0);
         UART_SET2(LSR, (UART_GET2(LSR) | LSR_THR_EMPTY));