From bd94fb8d313e2cace74dd1372f7d73aa4246fe1d Mon Sep 17 00:00:00 2001
From: NyanCatTW1 <17372086+NyanCatTW1@users.noreply.github.com>
Date: Tue, 23 Apr 2024 23:59:08 +0300
Subject: [PATCH] Handle more types of specifiers in vprintf

---
 ptx/lib/zluda_ptx_impl.bc  | Bin 232076 -> 232464 bytes
 ptx/lib/zluda_ptx_impl.cpp |  34 +++++++++++++++++++++++++++++++++-
 2 files changed, 33 insertions(+), 1 deletion(-)

diff --git a/ptx/lib/zluda_ptx_impl.bc b/ptx/lib/zluda_ptx_impl.bc
index 4e441511ffc37584dc5e8960d13debdc46e39a39..a6ddfdaaee3b32a33f7a5a9b50f6e9b4b947b2d5 100644
GIT binary patch
delta 1741
zcmXw(ZEzFE8OQhbq_t$9WLcOXV;Ro!O9(wBC&e(XtF~v^i5<*j&P*mH&<-cRsEjQu
zPT~(rK13_afK6QcXh5Th(?$j-qo#Q=#sxK`?a4OGr53hKNf9M!MF>+77)L2-#xx=6
z35T8M!+&>v&;Fm?d3J}7=*6&Jtc2>fk5|v=NF=LH+|cdX0k13un}h>R0&%KIfP~Gz
zhX7!W%55l60bm9I%yHVo-zg*SxW@qqg4d{9w{>pdsA8gs*RO3gsJPZ5qLbI|<7J|%
zjQ@<dR>o%ubt!+AIAi9AiPz2iLE@U3A0evDd_U2-flm`>%DBQ&Z~+wYNn&3a2M#NH
zkt$BB*!jU1EC5DDS(;?AHNTN;R0{Gu?fvC*d!K%`Pg!51O5ds0fN$P5ct2V9q^9<Z
zg3^IkCWmyh+p}=mmDK(QZ>4JD1<Cz^`+X1sP=VCu0ooM(*^8-$cLJ)q76bJ}Gh|g|
zpnNB31e#f0y&1=uph0yT8Ta~))Tioa4eO-9h>7ks^b73J;ICEJxn}@p@@lexdZYAH
zP}qMR{V+_rLV_tu`x;C-yD2I<#%p9D(*MlBc{-@abcWD@0hgYp|2ax!-wjvMcfHJ9
zB>mbUe^VZ}>823E)bb=5lk_ztI!e;p>(QQv{Ax<1Dw9vaj)>DZLR<%$zyso4@0|qY
zE-_YSfacJqk6^y2CZDf2uAA(?SLf~K*Hu-v77meXB@^rBsV5Aj6?td4ebY9tGGTKQ
zHjk_Mya|1Kst>4K_l<Ds7+AY%K2a~$7q<|Ezk)=c3_ppB$UME%K)t!8wb#iP{9!f1
zLr+0>q~CYIMu!6V4{SIq?LwQwJ`>qqPGGhrGSP^s=TIXU60*ZWRur;_g=|>JzA9uR
zLUuw(M}_oTLV8k2PYdaokbYlCOG0{1NXtSxp@h*vKH>*^YF}JhdF1N4U+t+`73UbD
zh9_c=RyfGD54nxH#~df9L&eGo-ZZ<lCtdoZNe>(Lve7!=?9>+ngPrE~4-Il_;&f@8
zu87k=h|>?p=^w@Eu{eIEO0EMibPxMu4wKf$B}-h|9GC2I$zK<jobmL*98I5&gRv^v
zt)0^2$s?p13{V=;Ww9h1S@}^`Zf50eth}9-A7^DZE2GUMw}E*f<_TIxhaA=a3Crzw
z#pS5y*px1@mAx!^$P=qVVI>E1E0ze_M?Y{>-wG$%S4GsHI5wqCRI_7}2cB3JZtZm7
zKKyd*;OyY=7}^EC5>v<83Ww!X`+BctnV@zEv~s2rrJ={Eo34<H+S(R@TySs+ZLWNH
zkm`xmXY$x^r-^(}4&d09s5-`hiI>V%k8AqEYHo(h3+lgheEs!uA!Qe6e;w#&RJmIe
zXt#EbF35Fe$|(}5mVnkmqP7r6?QkiEGY_gt8jlwn*zduOIS@C9@rx}4*`ZLr$Uu?A
z-QKi#V`N5{lLE>E3vT<l^U9WitY-9X6n+@4G6e&AShdwj@y^nW&-SH1m299A4W$`z
zMYu}ZfVw9VY#q`Skw=p55|iD8x-?6)&r+mos6ic+;4FzgZ(E_$6@kLpAEOl$xx%Ba
z$|OB!St=lNM2AASnuE!qfg0+(NwkUD&d@lu-}QY#UBBqR#25beUVJ3`F-XxgTCaRG
zyZ2kT1J>E{OLL*|^4_+2L7kcGlqSYf_70-XQ&M}q`Xtq6$jpVVP_Qp$p9plU*^Pdp
zdBHW$<>!Eo@?5%ArDSFHfwGg-SBA)JD6kR!Hf0}dOcv}$hc$i52UTtZfK>bnH03}>
zp11+Dqf<XU0$pbR)I9WciKiJHMbbQscJ;o`;NPQ#d6??r|HR;TP<kF}x_kf3;N$4a
zd05+>mKi*aewTp0?%s0@j-fjV8182O!Ql7M2j^g>JCb4WU(g#1(Bw&f#^70Waslr2
zByTYIESg<_gPzDu2LBcPeF4fI+P}=;bLiFrT=8T-XYdkwX%W`;2)7yhAsSzVdwZn+
zG58`ny$A<;_-`0|3H@ymMth_?48Dw(7oprEe#hXC(O?p0dRVO${|g;WLXB6-x8f`4
zy(HY^1$ryKiViHnW^Z!66<<Ramf&8mxY3ILjb@6@!?1UfvEm!(!g-kTPFhdBI|V<!
F_<tzPiEID>

delta 1355
zcmaKreN0nV7{<>%w>`b!EnG^4TU5NT`MJwha5!-;J17e{7IWQL=+K#4f?A=i3|K_B
zqz76`|8TBq$5>`&UDS0SDg-x)Swbr~2^46P8A1F2=LZb7A@0kGcJ*Wb?&Ljre#!gf
zN#5jr-OX{kKZ<id4G(|Qy6}vePIOUi6Y7%Jq4A(SUvA8osZ;qfq%v&}004=Kw?!_h
zeDrtW7*8bnRITP!fTAQux&GBWoTl>B;4f949XO!Unz2r$t;U(iQ-zlxPc_a#o=O~p
zJUelnO1lHRv<gae7f#?P<veMS>cTpX%ErJGO$Xq%9)N_XrssFN*T?4BXe9+ewP|}*
zjd{CuYqXK57`eZY0W-%q>$T+NOtvv`j`NcdrK!`Mm_u23Q+={m-MMa3p@@2)Ldv-g
zf%g@^-@ax~onmDn#}*Ypy<#3n<7o|GI@P%;Bp}*3#ZII#Y1r$^axU3hw}%%hxq4YY
z-nh@E7{%KF=B>;KV2{rE5yaJxqih3hY?67OU|7@KXz{L|cI+BNOZj|X<1t!bsxS?V
z%`|iJui>c=1&X*SDl4et0())F2fBM(rw)c~&WJ9umB^AqmRz#5TbA6i<dLOjSvnv~
zURgRUOFmf&YIxfA3dqsk#LcuZ&XL1ChAuIXDN%!TbcWf_p;*U!zKt!=)BwlzPZzkI
zIj;1o0NxoO<A90fC4jpL&@D`_UmOwc+y9Ke_OdS`MC?U+!<HCt90~2Fl)%9<ZsW7*
zuIUrujxCXt2NnOExZQCxGFQtAw%M#Vnf>f7XlepK119u8fdnn3u{vdW<&4o7)389>
zBpFse0-HQjcf0Z2{8U@*%;*z{dYwrrC6y-QA!S37wnQ@JJ_gxT_fEsP*{Q{~8%LvH
zdq`ViKCeq6g)5xzF|IdI3IF`skRx}0BMIeg1I(NHLP1>wN(mpv4GP30s9#`?1JuS|
z)LzGU<1JKK2Mn4u%mHppKa0p&yfjrjbnsnUF=F#n*kC6RtVV2?nA@bJ!xVQ#=pLCk
z;#cgY=*&X*5!)GD0~9NnELZ6gK`!&-pr-7uo{KI}d_C*m6w3>33SGQMf98mx=*Ied
zqrP;+pUc$eW-qD+N%b4M8y?l!bwQ1Zo~L+`ofHy=aHAv6o?5WWt*+5e>Raq(Wo?RK
ztTdXY^=#4iyunnO8C_Esitofqo89!gKA6Q``Y-s~MFGQD!=@*tEGfgl%yIO?liA=*
z$Bh}CX+vF2htjO2y|QwwWwRgF3}$(9N!$jM&4i4-Kt-)I>t0KXI_YN=Ijx)B4c_GM
z*miF6;lY5kOaNe1XE}keEUF)mG{NLk!%cAQTwAyY)|QiFBJrWCJ<wM!9~a5j=$9Ti
zSdLGK<T#4-K&C=&7fCDH-wRh%gied(B<k*kwH0!QNZQeXAZ)G>f+9JM#)5FH!hKOB
z9q4Qy<jtW=BI!iu`=G(>zbuj;(6v73GP|#cqzC=c2fNLjZA>J6XkS14+dMKZk}&$J
zAFi-S6Cyc}{Qb~mX_*ws1$40=x-9rtkqn{fe(1Bb{3eoN^uYk^wvf9bxs1LTfMXWn
zo=C2uzyM^dEi)n+MMDE{v2`jUlAqAg5G=6z*$i?6-3dXHmFP0aEi{@ChRs%QY6h7=
RcfxSc>YaaT*-3c$!oLWH_T>No

diff --git a/ptx/lib/zluda_ptx_impl.cpp b/ptx/lib/zluda_ptx_impl.cpp
index a6417f6..7c6600f 100644
--- a/ptx/lib/zluda_ptx_impl.cpp
+++ b/ptx/lib/zluda_ptx_impl.cpp
@@ -1395,8 +1395,36 @@ extern "C"
             char c = *(s++);
             if (c == 0)
                 break;
-            if (c == '%')
+            if (c != '%')
+                continue;
+
+            // %% requires no additional handling
+            if (*s == '%')
             {
+                s++;
+                continue;
+            }
+
+            // %s uses __ockl_printf_append_string_n
+            // https://github.com/ROCm/clr/blob/5a715ed160c69a2deebac187a0368900a774f3e8/hipamd/include/hip/amd_detail/hip_assert.h#L73
+            if (*s == 's')
+            {
+                s++;
+                const char *value = (const char *)read_valist(valist_ptr, valist_offset, 8);
+                handle = __ockl_printf_append_string_n(handle, value, strlen_plus_one(value), 0);
+                continue;
+            }
+
+            // Keep scanning until we figure out the length of this specifier or if we reach the end of the string
+            while (*s != 0) {
+                // "The width is not specified in the format string, but as an additional integer value argument preceding the argument that has to be formatted."
+                if (*s == '*') {
+                    s++;
+                    uint64_t value = read_valist(valist_ptr, valist_offset, 4);
+                    handle = __ockl_printf_append_args(handle, 1, value, 0, 0, 0, 0, 0, 0, 0);
+                    continue;
+                }
+
                 uint8_t len = 0;
                 if (parse_printf_specifier(s, len))
                 {
@@ -1408,6 +1436,9 @@ extern "C"
                     if (specifier_with_length)
                     {
                         s += specifier_with_length;
+                    } else {
+                        s++;
+                        continue;
                     }
                 }
 
@@ -1416,6 +1447,7 @@ extern "C"
                     uint64_t value = read_valist(valist_ptr, valist_offset, len);
                     handle = __ockl_printf_append_args(handle, 1, value, 0, 0, 0, 0, 0, 0, 0);
                 }
+                break;
             }
         }
         return (uint32_t)__ockl_printf_append_args(handle, 0, 0, 0, 0, 0, 0, 0, 0, 1);