Fix reuseExistingVarDecl trigger condition (#463)

* reuseExistingVarDecl can inhibit other optimizations, it should only
happen in the second pass.
* reuseExistingVarDecl seems safe to use even when not at the root
level.

src/rewriter.fs

@@ -616,6 +616,7 @@ type private RewriterImpl(options: Options.Options, optimizationPass: Optimizati
     // Reuse an existing local variable declaration that won't be used anymore, instead of introducing a new one.
     // The reused identifier gets compressed better, and the declaration is sometimes removed.
     // float d1 = f(); float d2 = g();  ->  float d1 = f(); d1 = g();
+    // We only do this when the reused declaration is at the same level as the removed declaration.
     let reuseExistingVarDecl blockLevel b =
         let tryReplaceWithPrecedingAndFollowing f (xs : Stmt list) =
             let rec go f preceding = function
@@ -801,9 +802,7 @@ type private RewriterImpl(options: Options.Options, optimizationPass: Optimizati
             else removeUnusedAssignments blockLevel b
 
         let b =
-            if optimizationPass <> OptimizationPass.Second &&
-                // We ensure control flow analysis is trivial by only doing this at the root block level.
-                (match blockLevel with BlockLevel.FunctionRoot _ -> false | _ -> true) || hasPreprocessor
+            if optimizationPass <> OptimizationPass.Second || hasPreprocessor
             then b
             else reuseExistingVarDecl blockLevel b
 

tests/compression_results.log

@@ -1,24 +1,24 @@
-clod.frag (and others)                    8696 => 1484.471
-mouton/mouton.vert (and others)          16569 => 2405.061
-audio-flight-v2.frag                      4471 =>  874.321
+clod.frag (and others)                    8692 => 1483.915
+mouton/mouton.vert (and others)          16569 => 2403.743
+audio-flight-v2.frag                      4469 =>  873.544
 buoy.frag                                 3977 =>  597.317
 controllable-machinery.frag               7671 => 1220.647
-ed-209.frag                               7665 => 1345.543
+ed-209.frag                               7661 => 1340.206
 elevated.hlsl                             3405 =>  604.101
 endeavour.frag                            2567 =>  530.460
 from-the-seas-to-the-stars.frag          14173 => 2278.764
 frozen-wasteland.frag                     4455 =>  791.550
 kinder_painter.frag                       2832 =>  442.132
 leizex.frag                               2252 =>  506.309
-lunaquatic.frag                           5227 => 1044.673
-mandelbulb.frag                           2325 =>  539.673
+lunaquatic.frag                           5223 => 1047.141
+mandelbulb.frag                           2309 =>  533.408
 ohanami.frag                              3179 =>  703.261
-orchard.frag                              5384 => 1002.066
+orchard.frag                              5380 => 1000.024
 oscars_chair.frag                         4648 =>  986.069
-robin.frag                                6188 => 1038.630
-slisesix.frag                             4433 =>  884.925
+robin.frag                                6183 => 1038.390
+slisesix.frag                             4429 =>  890.365
 terrarium.frag                            3571 =>  744.738
 the_real_party_is_in_your_pocket.frag    11966 => 1765.100
 valley_ball.glsl                          4307 =>  881.820
-yx_long_way_from_home.frag                2901 =>  592.799
-Total: 132862 => 23264.432
+yx_long_way_from_home.frag                2885 =>  586.201
+Total: 132803 => 23249.205

tests/real/audio-flight-v2.frag.expected

@@ -115,9 +115,9 @@ vec2 map(vec3 p,float sg)
   float obj=fBox(o,vec3(.15*offWobble))-.015;
   if(obj<res.x)
     res=vec2(obj,11),g_hp=o;
-  float pid=mod(floor((q.z+3.625)/7.25),18.);
-  pid=pid<12.?
-    pid<6.?
+  float pid=floor((q.z+3.625)/7.25),trackmod=mod(pid,18.);
+  pid=trackmod<12.?
+    trackmod<6.?
       4.:
       6.:
     10.;
@@ -130,8 +130,8 @@ vec2 map(vec3 p,float sg)
   tun=floor(fs.xz)+.5;
   vec2 neighbour=tun+edge(fs.xz-tun);
   pid=floor(tun.y);
-  float height=smoothstep(.001,1.,texture(iChannel0,vec2(mod(pid,16.)*.0465,.1)).x);
-  ga=height;
+  trackmod=smoothstep(.001,1.,texture(iChannel0,vec2(mod(pid,16.)*.0465,.1)).x);
+  ga=trackmod;
   float ids=pMod(s.z,7.25);
   vec3 qid=pMod(q,7.25);
   ids=mod(ids,2.);
@@ -142,7 +142,7 @@ vec2 map(vec3 p,float sg)
         2.:
         2.:
       5.:
-    floor(height*1.45));
+    floor(trackmod*1.45));
   float d4a=length(o.xy-vec2(2.5,1.75))-.1,d4=length(o.xy-vec2(2.5,1.75))-.04+.027+.027*sin(o.z-time*4.5);
   if(d4<res.x)
     res=vec2(d4,12),g_hp=p;
@@ -154,16 +154,16 @@ vec2 map(vec3 p,float sg)
     d1.x=min(d4a,d1.x);
   if(d1.x<res.x)
     res=d1.xy,g_hp=p;
-  float me=fBox(fs-vec3(0,0,tun.y),vec3(.05,.15+height,.25)),next=fBox(fs-vec3(0,0,neighbour.y),vec3(.05,.001+height,.25));
-  height=min(me,next);
-  if(height<res.x)
-    res=vec2(height,4),g_hp=p;
+  float me=fBox(fs-vec3(0,0,tun.y),vec3(.05,.15+trackmod,.25)),next=fBox(fs-vec3(0,0,neighbour.y),vec3(.05,.001+trackmod,.25));
+  trackmod=min(me,next);
+  if(trackmod<res.x)
+    res=vec2(trackmod,4),g_hp=p;
   if(sg==1.)
     beams+=1e-4/(3e-6+d4*d4);
   if(sg==1.&&ids<1.)
     flight+=.00025/(1e-7+d4a*d4a);
   if(sg==1.&&pid<.1)
-    glow+=.00015/(2e-6+height*height);
+    glow+=.00015/(2e-6+trackmod*trackmod);
   if(sg==1.&&tm<13.25)
     objglow+=5e-4/(5e-4+obj*obj);
   return res;

tests/real/buoy.frag.expected

@@ -131,11 +131,10 @@ vec3 ShadeBoat(vec3 pos,vec3 ray)
   pos=vec3(dot(pos,boatRight),dot(pos,boatUp),dot(pos,boatForward));
   vec3 lightDir=normalize(vec3(-2,3,1));
   float aa=4./iResolution.x;
-  pos=mix(vec3(1,.8,.08),mix(vec3(.04),mix(mix(vec3(1),vec3(.04),smoothstep(-aa*4.,aa*4.,cos(atan(pos.x,pos.z)*6.))),mix(vec3(.04),vec3(1,.01,0),smoothstep(.25-aa,.25,abs(pos.y))),smoothstep(.2-aa*1.5,.2,abs(pos.y))),smoothstep(.05-aa,.05,abs(abs(pos.y)-.6))),smoothstep(.05-aa,.05,abs(abs(pos.y)-.65)))*(smoothstep(-.1,1.,dot(norm,lightDir))*vec3(1,.9,.8)+vec3(.06,.1,.1));
-  lightDir=pow(max(0.,dot(norm,normalize(lightDir-ray))),1e2)*1e2/32.*vec3(1);
-  vec3 rr=reflect(ray,norm);
-  lightDir+=mix(vec3(0,.04,.04),Sky(rr),smoothstep(-.1,.1,rr.y));
-  return mix(pos,lightDir,mix(.001,1.,pow(1.-abs(dot(norm,ray)),5.)));
+  pos=mix(vec3(1,.8,.08),mix(vec3(.04),mix(mix(vec3(1),vec3(.04),smoothstep(-aa*4.,aa*4.,cos(atan(pos.x,pos.z)*6.))),mix(vec3(.04),vec3(1,.01,0),smoothstep(.25-aa,.25,abs(pos.y))),smoothstep(.2-aa*1.5,.2,abs(pos.y))),smoothstep(.05-aa,.05,abs(abs(pos.y)-.6))),smoothstep(.05-aa,.05,abs(abs(pos.y)-.65)));
+  vec3 specular=pow(max(0.,dot(norm,normalize(lightDir-ray))),1e2)*1e2/32.*vec3(1),rr=reflect(ray,norm);
+  specular+=mix(vec3(0,.04,.04),Sky(rr),smoothstep(-.1,.1,rr.y));
+  return mix(pos*(smoothstep(-.1,1.,dot(norm,lightDir))*vec3(1,.9,.8)+vec3(.06,.1,.1)),specular,mix(.001,1.,pow(1.-abs(dot(norm,ray)),5.)));
 }
 float OceanDistanceField(vec3 pos)
 {
@@ -173,16 +172,16 @@ vec3 ShadeOcean(vec3 pos,vec3 ray,vec2 fragCoord)
   vec3 norm=OceanNormal(pos);
   float ndotr=dot(ray,norm);
   vec3 reflectedRay=ray-2.*norm*ndotr;
-  norm=normalize(ray+(-cos(1.33*acos(-ndotr))-ndotr)*norm);
-  ray=Sky(reflectedRay);
+  ray=normalize(ray+(-cos(1.33*acos(-ndotr))-ndotr)*norm);
+  norm=Sky(reflectedRay);
   float t=TraceBoat(pos-0.*reflectedRay,reflectedRay);
   if(t>0.)
-    ray=ShadeBoat(pos+t*reflectedRay,reflectedRay);
-  t=TraceBoat(pos-0.*norm,norm);
+    norm=ShadeBoat(pos+t*reflectedRay,reflectedRay);
+  t=TraceBoat(pos-0.*ray,ray);
   reflectedRay=vec3(0,.04,.04);
   if(t>0.)
-    reflectedRay=mix(reflectedRay,ShadeBoat(pos+t*norm,norm),exp(-t));
-  reflectedRay=mix(reflectedRay,ray,pow(1.-abs(ndotr),5.));
+    reflectedRay=mix(reflectedRay,ShadeBoat(pos+t*ray,ray),exp(-t));
+  reflectedRay=mix(reflectedRay,norm,pow(1.-abs(ndotr),5.));
   return mix(reflectedRay,vec3(1),WaveCrests(pos,fragCoord));
 }
 void mainImage(out vec4 fragColor,vec2 fragCoord)

tests/real/ed-209.frag.expected

@@ -39,14 +39,14 @@ float sdTriPrism(vec3 p)
 float sdCappedCone(vec3 p)
 {
   vec3 b=vec3(0,0,-.1),a=vec3(0);
-  float rba=.1,baba=dot(b-a,b-a),paba=dot(p-a,b-a)/baba,x=sqrt(dot(p-a,p-a)-paba*paba*baba),cax=max(0.,x-(paba<.5?
+  float baba=dot(b-a,b-a),paba=dot(p-a,b-a)/baba,x=sqrt(dot(p-a,p-a)-paba*paba*baba),cax=max(0.,x-(paba<.5?
     .25:
-    .35)),cay=abs(paba-.5)-.5,f=clamp((rba*(x-.25)+paba*baba)/(rba*rba+baba),0.,1.);
-  x=x-.25-f*rba;
-  rba=paba-f;
+    .35)),cay=abs(paba-.5)-.5,f=clamp((.1*(x-.25)+paba*baba)/(.01+baba),0.,1.);
+  x=x-.25-f*.1;
+  paba-=f;
   return(x<0.&&cay<0.?
     -1.:
-    1.)*sqrt(min(cax*cax+cay*cay*baba,x*x+rba*rba*baba));
+    1.)*sqrt(min(cax*cax+cay*cay*baba,x*x+paba*paba*baba));
 }
 float sdCappedCylinder(vec3 p,float h,float r)
 {

tests/real/heart.expected

@@ -10,12 +10,12 @@ pub const HEART_FRAG: &'static [u8] = b"\
  void main()\
  {\
    vec2 f=(2.*gl_FragCoord.xy-y)/y.y;\
-   float u=mod(v,2.)/2.,a=pow(u,.2)*.5+.5;\
-   a-=a*.2*sin(u*6.2831*5.)*exp(-u*6.);\
+   float g=mod(v,2.)/2.,a=pow(g,.2)*.5+.5;\
+   a-=a*.2*sin(g*6.2831*5.)*exp(-g*6.);\
    f*=vec2(.5,1.5)+a*vec2(.5,-.5);\
-   a=atan(f.x,f.y)/3.141593;\
-   u=length(f);\
-   a=abs(a);\
-   a=(13.*a-22.*a*a+10.*a*a*a)/(6.-5.*a);\
-   gl_FragColor=vec4(step(u,a)*pow(1.-u/a,.25),0,0,1);\
+   g=atan(f.x,f.y)/3.141593;\
+   a=length(f);\
+   g=abs(g);\
+   g=(13.*g-22.*g*g+10.*g*g*g)/(6.-5.*g);\
+   gl_FragColor=vec4(step(a,g)*pow(1.-a/g,.25),0,0,1);\
  }\0";

tests/real/heart.frag.expected

@@ -12,14 +12,14 @@ const char *heart_frag =
  "void main()"
  "{"
    "vec2 f=(2.*gl_FragCoord.xy-y)/y.y;"
-   "float u=mod(v,2.)/2.,a=pow(u,.2)*.5+.5;"
-   "a-=a*.2*sin(u*6.2831*5.)*exp(-u*6.);"
+   "float g=mod(v,2.)/2.,a=pow(g,.2)*.5+.5;"
+   "a-=a*.2*sin(g*6.2831*5.)*exp(-g*6.);"
    "f*=vec2(.5,1.5)+a*vec2(.5,-.5);"
-   "a=atan(f.x,f.y)/3.141593;"
-   "u=length(f);"
-   "a=abs(a);"
-   "a=(13.*a-22.*a*a+10.*a*a*a)/(6.-5.*a);"
-   "gl_FragColor=vec4(step(u,a)*pow(1.-u/a,.25),0,0,1);"
+   "g=atan(f.x,f.y)/3.141593;"
+   "a=length(f);"
+   "g=abs(g);"
+   "g=(13.*g-22.*g*g+10.*g*g*g)/(6.-5.*g);"
+   "gl_FragColor=vec4(step(a,g)*pow(1.-a/g,.25),0,0,1);"
  "}";
 
 #endif // HEART_FRAG_EXPECTED_

tests/real/kinder_painter.expected

@@ -170,23 +170,23 @@
    "b[3]=vec4(.5,.5,.7,3);"
    "b[4]=vec4(1,.9,.9,2);"
    "b[5]=vec4(1,.9,.9,2);"
-   "float x=.15*z-6.2831*s.x/v.x,u=2.+3.*s.y/v.y;"
-   "vec2 a=vec2(cos(x),sin(x));"
-   "vec3 n=normalize(vec3(y.x*a.x-a.y,y.y,a.x+y.x*a.y)),e=vec3(u*a.y,0,-u*a.x);"
-   "u=m(e,n,i,w);"
-   "e+=n*u;"
-   "f.xyz=vec3(0,1.5,-3)-e;"
+   "float x=.15*z-6.2831*s.x/v.x,a=2.+3.*s.y/v.y;"
+   "vec2 g=vec2(cos(x),sin(x));"
+   "vec3 n=normalize(vec3(y.x*g.x-g.y,y.y,g.x+y.x*g.y)),u=vec3(a*g.y,0,-a*g.x);"
+   "x=m(u,n,i,w);"
+   "u+=n*x;"
+   "f.xyz=vec3(0,1.5,-3)-u;"
    "f.w=length(f.xyz);"
    "f.xyz=f.xyz/f.w;"
-   "w=d(e,i,w,n,f,o);"
-   "vec4 g,C;"
-   "u=m(e,o.xyz,i,g);"
-   "e+=o.xyz*u;"
-   "f.xyz=vec3(0,1.5,-1)-e;"
+   "w=d(u,i,w,n,f,o);"
+   "vec4 t,e;"
+   "x=m(u,o.xyz,i,t);"
+   "u+=o.xyz*x;"
+   "f.xyz=vec3(0,1.5,-1)-u;"
    "f.w=length(f.xyz);"
    "f.xyz=f.xyz/f.w;"
-   "g=d(e,i,g,o.xyz,f,C);"
-   "w=mix(w,g,.5-.5*o.w);"
+   "t=d(u,i,t,o.xyz,f,e);"
+   "w=mix(w,t,.5-.5*o.w);"
    "gl_FragColor=w;"
  "}",
 

tests/real/lunaquatic.frag.expected

@@ -109,11 +109,11 @@ vec3 calculateSky(vec3 ro,vec3 rd,int addPlanet)
 {
   vec3 color=max(vec3(0),(max(vec3(0),pow(dot(lightDir,rd),6.))*.7-rd.y)*mix(vec3(1,.5,0),vec3(1),min(1.,lightDir.y*1.5))+lightDir.y*3.);
   float phi=atan(rd.x,rd.z),theta=acos(rd.y/length(rd));
-  theta=smoothstep(0.,.5,norm(.5*smoothrnd(3e2*vec2(phi,theta)))+norm(.75*smoothrnd(5e2*vec2(phi,theta)))-1.25)*saturate(1.-lightDir.y*5.);
+  phi=smoothstep(0.,.5,norm(.5*smoothrnd(3e2*vec2(phi,theta)))+norm(.75*smoothrnd(5e2*vec2(phi,theta)))-1.25)*saturate(1.-lightDir.y*5.);
   if(addPlanet>0)
     {
       vec4 p=calcPlanet(ro,rd);
-      color+=theta*saturate(1.-p.w)+p.xyz;
+      color+=phi*saturate(1.-p.w)+p.xyz;
     }
   rd.xy+=ro.xy*1e-4;
   return color+(calculateSkySub(normalize(rd+vec3(sin(Y.z*.1),0,cos(Y.z*.1))*.1)*3.)+calculateSkySub(normalize(rd+vec3(sin(Y.z*.1),0,cos(Y.z*.1))*.2)*5.)*.1+calculateSkySub(normalize(rd+vec3(sin(Y.z*.1),0,cos(Y.z*.1))*.4)*7.)*.1-calculateSkySub(normalize(rd+vec3(sin(Y.z*.2),0,0)*.5))*1.5)*saturate(rd.y+.5);
@@ -128,9 +128,8 @@ float isoSurface(vec3 p)
 }
 float traceIso(vec3 ro,vec3 rd,float mint,float maxt)
 {
-  float s=99.,lt,liso,exact;
-  s=(maxt-mint)/s;
-  for(float t=mint;t<maxt;t+=s)
+  float lt,liso,exact,delt=(maxt-mint)/99.;
+  for(float t=mint;t<maxt;t+=delt)
     {
       vec3 p=ro+t*rd;
       float iso=isoSurface(p);

tests/real/mandelbulb.expected

@@ -49,7 +49,6 @@ const char *mandelbulb_frag =
  "}"
  "bool f(vec3 i,vec3 s,out float y,out vec3 t,out vec4 o)"
  "{"
-   "float p=1.;"
    "vec2 e;"
    "if(!f(i,s,e))"
      "return false;"
@@ -62,24 +61,24 @@ const char *mandelbulb_frag =
    "float c;"
    "vec3 x;"
    "vec4 n;"
-   "p=1./sqrt(1.+p*p);"
+   "float w=1./sqrt(2.);"
    "for(float f=e.x;f<e.y;)"
      "{"
        "vec3 z=i+s*f;"
-       "float d=clamp(.001*f*p,1e-6,.005),w=d*.1;"
+       "float p=clamp(.001*f*w,1e-6,.005),d=p*.1;"
        "vec4 r;"
        "float l;"
        "if(v(z,l,n))"
          "return y=f,t=normalize(x),o=n,true;"
        "float m;"
-       "v(z+vec3(w,0,0),m,r);"
+       "v(z+vec3(d,0,0),m,r);"
        "float g;"
-       "v(z+vec3(0,w,0),g,r);"
+       "v(z+vec3(0,d,0),g,r);"
        "float C;"
-       "v(z+vec3(0,0,w),C,r);"
+       "v(z+vec3(0,0,d),C,r);"
        "x=vec3(m-l,g-l,C-l);"
-       "c=.5*l*w/length(x);"
-       "if(c<d)"
+       "c=.5*l*d/length(x);"
+       "if(c<p)"
          "return o=n,t=normalize(x),y=f,true;"
        "f+=c;"
      "}"
@@ -90,31 +89,29 @@ const char *mandelbulb_frag =
    "vec2 v=-1.+2.*gl_FragCoord.xy/resolution.xy,i=v*vec2(1.33,1);"
    "vec3 y=vec3(.577);"
    "float t=1.4+.2*cos(6.28318*time/20.);"
-   "vec3 c=vec3(t*sin(6.28318*time/20.),.3-.4*sin(6.28318*time/20.),t*cos(6.28318*time/20.)),e=normalize(vec3(0,.1,0)-c),s=normalize(cross(e,vec3(0,1,0)));"
-   "s=normalize(i.x*s+i.y*normalize(cross(s,e))+1.5*e);"
-   "vec3 x;"
+   "vec3 c=vec3(t*sin(6.28318*time/20.),.3-.4*sin(6.28318*time/20.),t*cos(6.28318*time/20.)),s=normalize(vec3(0,.1,0)-c),e=normalize(cross(s,vec3(0,1,0))),d=normalize(i.x*e+i.y*normalize(cross(e,s))+1.5*s);"
    "vec4 o;"
-   "if(f(c,s,t,e,o))"
+   "if(f(c,d,t,e,o))"
      "{"
-       "vec3 i=c+t*s;"
+       "vec3 i=c+t*d;"
        "float v=clamp(.2+.8*dot(y,e),0.,1.);"
        "v*=v;"
-       "float p=clamp(.3+.7*dot(vec3(-.707,0,.707),e),0.,1.),w=clamp(1.25*o.w-.4,0.,1.);"
-       "w=w*w*.5+.5*w;"
+       "float p=clamp(.3+.7*dot(vec3(-.707,0,.707),e),0.,1.),n=clamp(1.25*o.w-.4,0.,1.);"
+       "n=n*n*.5+.5*n;"
        "float r;"
-       "vec3 n;"
-       "vec4 g;"
+       "vec3 C;"
+       "vec4 x;"
        "if(v>.001)"
-         "if(f(i,y,r,n,g))"
+         "if(f(i,y,r,C,x))"
            "v=.1;"
-       "x=mix(mix(mix(vec3(1),vec3(.8,.6,.2),sqrt(o.x)*1.25),vec3(.8,.3,.3),sqrt(o.y)*1.25),vec3(.7,.4,.3),sqrt(o.z)*1.25)*((.5+.5*e.y)*vec3(.14,.15,.16)*.8+v*vec3(1,.85,.4)+.5*p*vec3(.08,.1,.14))*vec3(pow(w,.8),w,pow(w,1.1));"
-       "x=1.5*(x*.15+.85*sqrt(x));"
+       "s=mix(mix(mix(vec3(1),vec3(.8,.6,.2),sqrt(o.x)*1.25),vec3(.8,.3,.3),sqrt(o.y)*1.25),vec3(.7,.4,.3),sqrt(o.z)*1.25)*((.5+.5*e.y)*vec3(.14,.15,.16)*.8+v*vec3(1,.85,.4)+.5*p*vec3(.08,.1,.14))*vec3(pow(n,.8),n,pow(n,1.1));"
+       "s=1.5*(s*.15+.85*sqrt(s));"
      "}"
    "else"
-     " x=1.3*vec3(1,.98,.9)*(.7+.3*s.y);"
+     " s=1.3*vec3(1,.98,.9)*(.7+.3*d.y);"
    "v=v*.5+.5;"
-   "x=clamp(x*(.7+4.8*v.x*v.y*(1.-v.x)*(1.-v.y)),0.,1.);"
-   "gl_FragColor=vec4(x,1);"
+   "s=clamp(s*(.7+4.8*v.x*v.y*(1.-v.x)*(1.-v.y)),0.,1.);"
+   "gl_FragColor=vec4(s,1);"
  "}";
 
 #endif // MANDELBULB_EXPECTED_