From b655e89af08e30eeef1085b35bb4dc46349101d2 Mon Sep 17 00:00:00 2001
From: ptitSeb <sebastien.chev@gmail.com>
Date: Wed, 22 Dec 2021 11:44:11 +0100
Subject: [PATCH] [DYNAREC] Fixed idiv optimisation, and added a new test about
 idiv opcode

---
 CMakeLists.txt               |   7 ++++-
 src/dynarec/arm_emitter.h    |   8 +++++-
 src/dynarec/dynarec_arm_00.c |  20 +++++++--------
 tests/ref20.txt              |  11 ++++++++
 tests/test20                 | Bin 0 -> 15544 bytes
 tests/test20.c               |  48 +++++++++++++++++++++++++++++++++++
 6 files changed, 81 insertions(+), 13 deletions(-)
 create mode 100644 tests/ref20.txt
 create mode 100755 tests/test20
 create mode 100644 tests/test20.c

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 60292e75..ff08d1e2 100755
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -738,11 +738,16 @@ add_test(NAME longjumpInSignals COMMAND ${CMAKE_COMMAND} -D TEST_PROGRAM=${CMAKE
     -D TEST_REFERENCE=${CMAKE_SOURCE_DIR}/tests/ref18.txt
     -P ${CMAKE_SOURCE_DIR}/runTest.cmake )
 
-    add_test(NAME x87 COMMAND ${CMAKE_COMMAND} -D TEST_PROGRAM=${CMAKE_BINARY_DIR}/${BOX86} 
+add_test(NAME x87 COMMAND ${CMAKE_COMMAND} -D TEST_PROGRAM=${CMAKE_BINARY_DIR}/${BOX86} 
     -D TEST_ARGS=${CMAKE_SOURCE_DIR}/tests/test19 -D TEST_OUTPUT=tmpfile.txt 
     -D TEST_REFERENCE=${CMAKE_SOURCE_DIR}/tests/ref19.txt
     -P ${CMAKE_SOURCE_DIR}/runTest.cmake )
 
+add_test(NAME idiv COMMAND ${CMAKE_COMMAND} -D TEST_PROGRAM=${CMAKE_BINARY_DIR}/${BOX86} 
+    -D TEST_ARGS=${CMAKE_SOURCE_DIR}/tests/test20 -D TEST_OUTPUT=tmpfile.txt 
+    -D TEST_REFERENCE=${CMAKE_SOURCE_DIR}/tests/ref20.txt
+    -P ${CMAKE_SOURCE_DIR}/runTest.cmake )
+
 file(GLOB extension_tests "${CMAKE_SOURCE_DIR}/tests/extensions/*.c")
 foreach(file ${extension_tests})
     get_filename_component(testname "${file}" NAME_WE)
diff --git a/src/dynarec/arm_emitter.h b/src/dynarec/arm_emitter.h
index dbd93429..cd01b17e 100755
--- a/src/dynarec/arm_emitter.h
+++ b/src/dynarec/arm_emitter.h
@@ -245,9 +245,12 @@ Op is 20-27
 // cmp.s dst, src, #imm
 #define CMPS_IMM8(src, imm8) \
     EMIT(0xe3500000 | ((0) << 12) | ((src) << 16) | brIMM(imm8) )
-// cmn.s dst, src, #imm
+// cmn.s.cond dst, src, #imm
 #define CMNS_IMM8_COND(cond, src, imm8) \
     EMIT((cond) | 0x03700000 | ((0) << 12) | ((src) << 16) | brIMM(imm8) )
+// cmn.s dst, src, #imm
+#define CMNS_IMM8(src, imm8) \
+    EMIT(c__ | 0x03700000 | ((0) << 12) | ((src) << 16) | brIMM(imm8) )
 // tst.s dst, src1, src2, lsl #imm
 #define TSTS_REG_LSL_IMM5(src1, src2, imm5) \
     EMIT(0xe1100000 | ((0) << 12) | ((src1) << 16) | brLSL(imm5, src2) )
@@ -257,6 +260,9 @@ Op is 20-27
 // tst.s dst, src1, #imm ror rot*2
 #define TSTS_IMM8_ROR(src, imm8, rot) \
     EMIT(0xe3100000 | ((0) << 12) | ((src) << 16) | ((rot)<<8) | brIMM(imm8) )
+// tst.s.cond dst, src1, #imm ror rot*2
+#define TSTS_IMM8_ROR_COND(cond, src, imm8, rot) \
+    EMIT((cond) | 0x03100000 | ((0) << 12) | ((src) << 16) | ((rot)<<8) | brIMM(imm8) )
 // orr dst, src1, src2, lsl #imm
 #define ORR_REG_LSL_IMM5(dst, src1, src2, imm5) \
     EMIT(0xe1800000 | ((dst) << 12) | ((src1) << 16) | brLSL(imm5, src2) )
diff --git a/src/dynarec/dynarec_arm_00.c b/src/dynarec/dynarec_arm_00.c
index 96084ab4..fdbaa8bc 100755
--- a/src/dynarec/dynarec_arm_00.c
+++ b/src/dynarec/dynarec_arm_00.c
@@ -2683,31 +2683,29 @@ uintptr_t dynarec00(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int ninst,
                         GETED;
                         SDIV(x2, xEAX, ed); // x1 = xEAX / ed
                         MLS(xEDX, x2, ed, xEAX);  // x14 = xEAX mod ed (i.e. xEAX - x1*ed)
-                        TSTS_IMM8_ROR(xEDX, 0b10, 1);   // test if reminder is negative
-                        ADD_REG_LSL_IMM5_COND(cNE, xEDX, xEDX, ed, 0); // add ed if negative
                         MOV_REG(xEAX, x2);
                     } else {
                         GETEDH(x1);
-                        // disabling the use of SDIV for now. It breaks X3Reunion
-                        // seems to be an issue with large number
-                        // for example EDX=0xffffffff EAX=0xe7f30000 and div ecx=0x186a0
-                        // gives wrong modulo
-                        if(arm_div && 0) {
+                        if(arm_div) {
+                            // check if a 32bits division is enough
                             CMPS_IMM8(xEDX, 0); // compare to 0
-                            CMNS_IMM8_COND(cNE, xEDX, 1); // compare to FFFFFFFF if not 0
+                            TSTS_IMM8_ROR_COND(cEQ, xEAX, 0b10, 1);   // also test that xEAX is not signed!
                             B_MARK(cEQ);
+                            CMNS_IMM8(xEDX, 1); // compare to FFFFFFFF if not 0
+                            B_MARK2(cNE);
+                            TSTS_IMM8_ROR(xEAX, 0b10, 1);   // also test that xEAX is signed!
+                            B_MARK(cNE);
+                            MARK2;
                         }
                         if(ed!=x1) {MOV_REG(x1, ed);}
                         STM(xEmu, (1<<xEAX) | (1<<xECX) | (1<<xEDX));
                         CALL(idiv32, -1, 0);
                         LDM(xEmu, (1<<xEAX) | (1<<xECX) | (1<<xEDX));
-                        if(arm_div && 0) {
+                        if(arm_div) {
                             B_NEXT(c__);
                             MARK;
                             SDIV(x2, xEAX, ed); // x2 = xEAX / ed
                             MLS(xEDX, x2, ed, xEAX);  // x14 = xEAX mod ed (i.e. xEAX - x2*ed)
-                            TSTS_IMM8_ROR(xEDX, 0b10, 1);   // test if reminder is negative
-                            ADD_REG_LSL_IMM5_COND(cNE, xEDX, xEDX, ed, 0); // add ed if negative
                             MOV_REG(xEAX, x2);
                             SET_DFNONE(x2); // flags are undefined
                         }
diff --git a/tests/ref20.txt b/tests/ref20.txt
new file mode 100644
index 00000000..2646fa37
--- /dev/null
+++ b/tests/ref20.txt
@@ -0,0 +1,11 @@
+SDHLD/SAR/IDIV 0x2d0 0x500 => 0x0 / 0x9000 (47185920/1280 => 36864 + 0)
+SDHLD/SAR/IDIV 0xfffffa5d 0x186a0 => 0xfffef4a0 / 0xfffffc4f (281474882142208/100000 => -945 + -68448)
+SDHLD/SAR/IDIV 0x1701 0x186a0 => 0xa220 / 0xf13 (385941504/100000 => 3859 + 41504)
+SDHLD/SAR/IDIV 0xffff9a19 0x186a0 => 0xffff6d00 / 0xffffbd38 (281473267073024/100000 => -17096 + -37632)
+SDHLD/SAR/IDIV 0xffffe7f3 0x186a0 => 0xffffebe0 / 0xfffff03d (281474573205504/100000 => -4035 + -5152)
+SDHLD/SAR/IDIV 0x15840 0x186a0 => 0xdd20 / 0xe19b (5775556608/100000 => 57755 + 56608)
+SDHLD/SAR/IDIV 0xff451330 0x186a0 => 0xfffee8a0 / 0xff857f2f (280672139739136/100000 => -8028369 + -71520)
+SDHLD/SAR/IDIV 0xffff626a 0x186a0 => 0xffff2fc0 / 0xffff98ba (281472332857344/100000 => -26438 + -53312)
+SDHLD/SAR/IDIV 0x9120 0x186a0 => 0x16d20 / 0x5f1b (2434793472/100000 => 24347 + 93472)
+
+Done
diff --git a/tests/test20 b/tests/test20
new file mode 100755
index 0000000000000000000000000000000000000000..d329fa9d6fa8a025ef13eb5a1b0762440adbcfa1
GIT binary patch
literal 15544
zcmeHOe{ft?6~6m}HqyA+mWcdl<xNSUg?6*4Sb+-cHoIw}+ma?}19r;mCc8<NZg%7D
z+qwmmhNaNB-I>UA{6mnbRK_}kqhry*am2I>r810Ds?Ol(jCFuvD^XIZ5FzsXzW3e@
zFE$P{{-^U!=Dc&yJ?EZ#?|1Lp-22|U_eW!GE|*J4@(7Qh7F;aE4Uly=nwA@cA{L32
z;!|RoSb!|rB=1B4vKM(|5BMN-$(KMU(}rw=A!g|n6d=<s{FWV(X>vVm<csWf1&YNX
zw-97|`SSrG&fjYCYE&8ov)zmbS1|+o4D947=w#X%gM0-3k|WT`($@L9mN=roZgC7Y
zvb1l39qI+d_^4iV!ESF~J?bK}JcMXYr}~=Hi8bj|rceyzvw^jytr0fBAU1B^Dt7<&
z-WOkbYt^H>o?iCin-h;Ox$3b;TUeg)8G}9~`m?_z$~rLZ29}CYkTmkX|Akj4n6}dn
zo1!5737ftI>7_P(Inu0K6774W0;zznx9PO+x9JMfY?l`a_G2-lsU5egc`cF3sD*qo
zAw~*XUX0{Y8ErtQYJV}V4x}>ibn2cYRAAN*=Xg$2hvTV?pib=8iH)(24Xx_hKuh4$
zv*{pW_iR!V$y{<UmDiHFo=vUkY$n+g?@Ke<;BYo$MN>`Jc^%PEpz1M@CAu#Bxy3C=
zNi6bLr9Q?{*l6YXQ>l-$OgD5A=P1WyTFTo@eQC&)lh79cm?=sO=6OmC-VEhJ%p3SE
zMAd#u0M|%qOmyLpgMfp8gMfp8gMfp8gMfp8gMfp8gTNdFx=WLFwLxq!Myu~O3sEke
z*L8j0gyz;ym7c5HUe%9sH>Mxv9)+YIuE%wV$+KtRsvqV~XNcyrrwpCD)FEb`oiOw%
zTnEOX8&rukc}uTN>0g@JI`(>RC0g57`O2a0+UqcmRtspsmY;HKSeh)3MGsU;b*N)3
zy1!CdiA3e-PD6e0?I$aRLzSN}VEL=w(xkU^qOROF-P@@b_K!sm)_PIp9bfF#3kPcl
z(6z3yms?JEjz!1I(SyAW-#P9-cE;ZY-|;Pa;lbK<a4tO;Twgd)_)%%HzEWC_N;+#-
z8`=JriAr+n&h2-pJ1X&+>cDu*U+x<F6FSh@(^cc1ceFZzCkd6()M*r|l*TExY}1cU
zqj>GD4|M%}d1({M$-lYUz<BKpgB-t?LB6ISU3q}ln@(!W$`4E;2*%<PKvkM-+`=L8
z|FT-qt`Xza%G1}tPt*?KGJ1SH%EPPM9s{_nIyqK7R=euh`)=1OFHdWY<ws^1^0am_
zhP`u(ei{MGJ7y5mj5aXNAO?P;)i>5~9K3sfrl1N@xVBv4P@+AH0h4}|yZ_oU^bj#>
zpMm8>iH@$h@y>3CrQ9eX7;l*<_cVFSv8K9)r%w5gz2|@TAJv9M5*^fE+m0Zk&)<#8
zd$(bB){nhezOQLzt^1#t0N<KMeU*LZaa~nKue22{@0jxd4eiynrqu=?{4qUBCtPUS
zO>Y-oEj?Gy5v*2ri)wjwQx!EB^F9l&(W<^0{p4InwnGjA4gwAW4gwAW4gwAW4gwAW
z4gwAW|F;O>m7?*^vOCfqi!^tKyP7*99ov*(aYfO*h>Ih5FDh=i)l{2Jv2sN^ooHT>
zP>_%7^|-9Mcs`I$@wF)4SMGRE*AG2}_oJT_g7Wfjn)=&sm@;&S%g7I25;gqp?lbe1
zubcj^D?RJm5sD8Qif`#UA1FKn`84E7$Ui{t$Nn@y{hY3!gsed-*#EcxP1n<q9guq=
z`GkVxx~EIH?yGZMS?~4kcP*^rLksR|mm<zOI)LW)6+(fHzGwcb>mf+tt84Ssf4*Un
z<}Hfbu2_HF+NR}<N81d1*lrT@dcf1D<5FgTLf|3rYw&v(Qbh1Iz()T??gFk0-G%%7
z3*3cb#P4wzv{-|uFtWgX@EBae)$ob%3*qNm(f)Q6dkQ>(xW;??u&-{PJLao@*c0~o
zAM%EMjirlPeM-vL6!xu*_*O@J!A(A8qpvYs_4&gmeD&cMe0AaHeR$l%Hco?ogIIve
zAm*V;aSVBUBIJ;RfP;X8fP;X8fP;X8fP;X8fP;X8!2bdQJMo6u59Z#6-|F&>Ha9)^
zFjq_?|8huvmwP27_c~WFA3FDY)c^j0o}ENtY-`11*x&MV+j>ZDfS&*_0&~A4znA77
zo0~kih|9p-Q{wY_fvvEZfRX9FkbLnS0`psW?ib(J^(>QkOPTG29dA;zY%{iC*;}Dw
zOP&31_hKV#Ac)_;OInA#Ht;sfwlswA$Xi=)R#rxmeW`dxX$c1K8TFcVtE`l=Hn_GW
z*wPY|5#~yM|4%Yc6qw)Bn>vvY;-y(TIs+o6B#(iziK(d*X%lNxH$QMD2BwY=p+Urm
z<T21QF*S93f(@FOwOfxWtbTfo5yPH-RySsZ*g30nW(XlM^9Y}zbFPp)qQOikrtlb}
zZ=Mfej>i~b%&A$sS&(y#<S|CvoMWbLK6Q{|<FRHEzXdn#){}#f)H#zdw?c+MJkB!o
zCu9idoY5bW2Q%7&+Q{^ZbGrTE4Ry}ly_}dJH~MSLRhxbt@_7%DXb(ZMKE}n<9pgL%
z?-`N{Gd>T=`;J8a`ypljUFiS6*Z(-|pIc6EmbVI;DrUzyozFUCeBLoXhc5AQ0k0F#
zWqjIy2c73NOj;}^nFyJh3%&EZk+#O$jrrh4N2X9M`?Cb?U19@FL0$m*2=L(jx5Vue
zPePaYE5JWu^JnX~LO*A9gZ{gq6EBh<{(QC~=hG5*gUCTo+WhxJmqb_&Lznfj{10t*
z7W;)wSD^n2I`3!FV)td@_t3qH*?yk2__NKQ<re^taU}RGiT!^DvMGcM$dB>;9{hPm
zh|kGrzm)hXAI4h-T?ru%ME_OLdH*~Mo&Euvo$a|9x+DdO9{Qtx&UE?<Et%KW1_S+~
zKd0rj!oUEMA0FROwPCfN2RQOLyphPNgXwHvJgp|QY%Z_H3q{eN9Ue(1wPYfIqaX8m
z@GMD6jpuUlJ!&$e<@ShyTzojGCJMvDdr-utsqob1xT<PfS9nuYjc$&pI4?5SAtBVr
zotwj(I$GyijKeC(!ciSH+HOT?k8}xjW9;?~;h1`RTU&RuN9_r3h(#IIIO5WuFBnbz
zhyyIe8`lK}lbSlxuWCbu%&tIR(K<R4lILLv#D#7nLQN*(T6{iZqKI7kxtrNoJX-@;
zQO-OyW1Q8QwX&wUCu-D0KC2GJGYOuynS09Tf}=ch!pvg8F%as&h&uFTMDNSzt%{Aq
zKdOrJJaRO6&}ZJ$J8CMC+O6VV&hHu4;(d@>&Xhwkg|kM<+=vKdvRX1Qm?;GM3Mrh-
zN+qnU@P>{xT6|C#wxM`_NCXmlGKgnNEoWMGCv*8!HZw<2Vap}c^fwbDX-x!-ZUvAG
z3}&Hf$s(?tn()aQ;}S>?S(9-nfe=zNi<-02EFsgVAU>SxN99?Kk<m1cZvdEy0Pc%n
zOm}(feWc`?h(8s)UG3-lWI_2eay2seUzvXL99@%if|TnP{(eY5t_R4<Jbn@6y@T}*
z<8hrprk}YoFyk@iW-!-Z^y8Y0JdQM13@BpM%b)peU@iyf$2A=J38W>_mvk4H>n{3n
zy+P*hqpX+t<XvE{)0pNugd9ZLUN2N}516YF`t5}eIfS%~$3AfFR|3;-19$>HWd2vD
zAO9CizXM3Kt*i$>v<53oTSPz7!(i#Bz>gfF!6Fu=?;~L8w-cFUwnq}v<VV5MkH72W
zee9G)P!0MvuF{YH-^k|b#<a`!eb@3cLMTwcM<66AR^kVcj7LBIKPShKmL%hGJs|zK
z?jpySVUhHE0XkQ6fY(?r@LZ;}(UocPufbg9(U0qn>HpyON7y+Q^y7Mj=S-)pJP0Ee
z@%sjt&7mLHb(75(;=+1sHa}y)g62Fxwn4Ie=b*`WTo+shKh`6Oex!H7%xAqy2n;`I
zLzWSPa>h!-Y!LmfwbB%<OR_O$maQFrI}wn3C49nZ#P&mSF0)QaM{W>eyR9H&k+{UP
P`*}W#IWLW{$>hHP)s$x|

literal 0
HcmV?d00001

diff --git a/tests/test20.c b/tests/test20.c
new file mode 100644
index 00000000..017e925c
--- /dev/null
+++ b/tests/test20.c
@@ -0,0 +1,48 @@
+#include <string.h>
+#include <stdio.h>
+#include <stddef.h>
+#include <stdlib.h>
+#include <stdint.h>
+#include <math.h>
+
+#if defined(__x86_64__)
+#error Nope!
+#else
+__attribute__((naked)) uint64_t _idiv_(uint32_t a, uint32_t b)
+{
+    asm volatile (
+    "xor %%eax, %%eax\n"
+    "mov 4(%%esp), %%edx\n"
+    "mov 8(%%esp), %%ecx\n"
+    "shrd $0x10, %%edx, %%eax\n"
+    "sar $0x10, %%edx\n"
+    "idiv %%ecx\n"
+    "ret"
+    :::);
+}
+#endif
+
+int main(int argc, const char** argv)
+{
+ uint32_t tests[][2] = {
+  {0x000002d0, 0x00000500},
+  {0xfffffa5d, 0x000186a0},
+  {0x00001701, 0x000186a0},
+  {0xffff9a19, 0x000186a0},
+  {0xffffe7f3, 0x000186a0},
+  {0x00015840, 0x000186a0},
+  {0xff451330, 0x000186a0},
+  {0xffff626a, 0x000186a0},
+  {0x00009120, 0x000186a0},
+ };
+ int n = sizeof(tests)/sizeof(tests[0]); 
+ uint64_t res;
+ for(int i=0; i<n; ++i) {
+   printf("SDHLD/SAR/IDIV 0x%x 0x%x ", tests[i][0], tests[i][1]);
+   res = _idiv_(tests[i][0], tests[i][1]);
+   uint32_t divi = res & 0xffffffffLL;
+   uint32_t modo = res >> 32;
+   printf("=> 0x%x / 0x%x (%lld/%d => %d + %d)\n", modo, divi, ((int64_t)tests[i][0])<<16, tests[i][1], divi, modo);
+ }
+ printf("\nDone\n");
+}