Double-quads can be loaded and stored (more) correctly.

--HG--
branch : dtrg-videocore

diff --git a/mach/vc4/ncg/table b/mach/vc4/ncg/table
index 3c96d71..e5cc0e1 100644 (file)
--- a/mach/vc4/ncg/table
+++ b/mach/vc4/ncg/table
@@ -386,8 +386,8 @@ PATTERNS
 
        pat ldl                            /* Load double-word from local */
                leaving
-                       lol $1 + QUAD*0
                        lol $1 + QUAD*1
+                       lol $1 + QUAD*0
 
        pat stl inreg($1)>0                /* Store to local */
                with CONST+GPRI
@@ -402,8 +402,8 @@ PATTERNS
 
        pat sdl                            /* Store double-word to local */
                leaving
-                       stl $1 + QUAD*1
                        stl $1 + QUAD*0
+                       stl $1 + QUAD*1
                        
        pat lil inreg($1)>0                /* Load from indirected local */
                uses REG
@@ -496,14 +496,15 @@ PATTERNS
             dec
             sti QUAD
 
-       pat lde                             /* Load external */
-               uses REG, REG
-               gen
-                       lea %a, {LABEL, $1}
-                       ld %b, {GPROFFSET, %a, 4}
-                       ld %a, {GPROFFSET, %a, 0}
-               yields %b %a
+       pat lde                             /* Load double external */
+               leaving
+                       lae $1
+                       loi QUAD*2
 
+       pat sde                             /* Store double external */
+               leaving
+                       lae $1
+                       sti QUAD*2
 
 
 /* Structures */
@@ -570,7 +571,7 @@ PATTERNS
                                ldhs %a, {GPROFFSET, %a, 0}
                        yields %a
                
-       pat loi $1==WORD                  /* Load short indirect */
+       pat loi $1==WORD                   /* Load short indirect */
                with LABEL
                        uses REG
                        gen
@@ -583,7 +584,7 @@ PATTERNS
                                ldh %a, {GPROFFSET, %a, 0}
                        yields %a
 
-       pat loi $1==QUAD                  /* Load quad indirect */
+       pat loi $1==QUAD                   /* Load quad indirect */
                with LABEL
                        uses REG
                        gen
@@ -596,6 +597,22 @@ PATTERNS
                                ld %a, {GPROFFSET, %a, 0}
                        yields %a
 
+       pat loi $1==2*QUAD                 /* Load double-quad indirect */
+               with LABEL
+                       uses REG, REG
+                       gen
+                               lea %b, %1
+                               ld %a, {GPROFFSET, %b, 0}
+                               ld %b, {GPROFFSET, %b, 4}
+                       yields %b %a
+               with GPRI
+                       uses reusing %1, REG, REG
+                       gen
+                               add %b, %1, GP
+                               ld %a, {GPROFFSET, %b, 0}
+                               ld %b, {GPROFFSET, %b, 4}
+                       yields %b %a
+
        pat loi !nicesize($1)              /* Load arbitrary size */
                leaving
                        loc $1
@@ -638,6 +655,20 @@ PATTERNS
                                add %a, %1, GP
                                st %2, {GPROFFSET, %a, 0}
 
+       pat sti $1==2*QUAD                 /* Load double-quad indirect */
+               with LABEL GPRI GPRI
+                       uses REG
+                       gen
+                               lea %a, %1
+                               st %2, {GPROFFSET, %a, 0}
+                               st %3, {GPROFFSET, %a, 4}
+               with GPRI GPRI GPRI
+                       uses reusing %1, REG=%1
+                       gen
+                               add %a, GP
+                               st %2, {GPROFFSET, %a, 0}
+                               st %3, {GPROFFSET, %a, 4}
+
        pat sti                            /* Store arbitrary size */
                leaving
                        loc $1