shithub: mc

Info  •  Files  •  Log  •  Branches

Correctly save/restore xmm registers.

    Our allocator wasn't acutally saving them, so we clobbered.

--- a/6/asm.h

+++ b/6/asm.h

@@ -3,7 +3,6 @@

 #define Maxdef (2*Maxarg)       /* maximum number of registers an insn can use or def */

 #define Wordsz 4                /* the size of a "natural int" */

 #define Ptrsz 8                 /* the size of a machine word (ie, pointer size) */

-#define Nsaved 13               /* number of registers saved in the ABI */

 typedef size_t regid;

@@ -131,7 +130,8 @@

     /* increased when we spill */

     Loc *stksz;

-    Loc *calleesave[Nsaved];

+    Loc *calleesave[Nreg];

+    size_t nsaved;

     /* register allocator state */

--- a/6/isel.c

+++ b/6/isel.c

@@ -852,7 +852,6 @@

                 return;

             break;

         case Imov:

-            assert(!isfloatmode(insn->args[1]->mode));

             if (insn->args[0]->type != Locreg || insn->args[1]->type != Locreg)

                 break;

             if (insn->args[0]->reg.colour == Rnone || insn->args[1]->reg.colour == Rnone)

@@ -924,12 +923,12 @@

     }

 }

+/* %rax is for int returns, %xmm0d is for floating returns */

 Reg savedregs[] = {

     Rrcx, Rrdx, Rrbx, Rrsi, Rrdi, Rr8, Rr9, Rr10, Rr11, Rr12, Rr13, Rr14, Rr15,

-    /*

-    Rxmm0d, Rxmm1d, Rxmm2d, Rxmm3d, Rxmm4d, Rxmm5d, Rxmm6d, Rxmm7d,

+    Rxmm1d, Rxmm2d, Rxmm3d, Rxmm4d, Rxmm5d, Rxmm6d, Rxmm7d,

     Rxmm8d, Rxmm9d, Rxmm10d, Rxmm11d, Rxmm12d, Rxmm13d, Rxmm14d, Rxmm15d,

-    */

+    Rnone

 };

 static void prologue(Isel *s, size_t sz)

@@ -937,6 +936,7 @@

     Loc *rsp;

     Loc *rbp;

     Loc *stksz;

+    Loc *phys;

     size_t i;

     rsp = locphysreg(Rrsp);

@@ -947,10 +947,12 @@

     g(s, Imov, rsp, rbp, NULL);

     g(s, Isub, stksz, rsp, NULL);

     /* save registers */

-    for (i = 0; i < sizeof(savedregs)/sizeof(savedregs[0]); i++) {

-        s->calleesave[i] = locreg(ModeQ);

-        g(s, Imov, locphysreg(savedregs[i]), s->calleesave[i], NULL);

+    for (i = 0; savedregs[i] != Rnone; i++) {

+        phys = locphysreg(savedregs[i]);

+        s->calleesave[i] = locreg(phys->mode);

+        g(s, Imov, phys, s->calleesave[i], NULL);

     }

+    s->nsaved = i;

     s->stksz = stksz; /* need to update if we spill */

 }

@@ -970,7 +972,7 @@

             g(s, Imov, ret, coreg(Rax, ret->mode), NULL);

     }

     /* restore registers */

-    for (i = 0; i < Nsaved; i++)

+    for (i = 0; savedregs[i] != Rnone; i++)

         g(s, Imov, s->calleesave[i], locphysreg(savedregs[i]), NULL);

     /* leave function */

     g(s, Imov, rbp, rsp, NULL);

--- a/6/ra.c

+++ b/6/ra.c

@@ -42,22 +42,23 @@

 /* A map of which registers interfere */

 #define Northogonal 32

 Reg regmap[Northogonal][Nmode] = {

-    [0]  = {Rnone, Ral, Rax, Reax, Rrax},

-    [1]  = {Rnone, Rcl, Rcx, Recx, Rrcx},

-    [2]  = {Rnone, Rdl, Rdx, Redx, Rrdx},

-    [3]  = {Rnone, Rbl, Rbx, Rebx, Rrbx},

-    [4]  = {Rnone, Rsil, Rsi, Resi, Rrsi},

-    [5]  = {Rnone, Rdil, Rdi, Redi, Rrdi},

-    [6]  = {Rnone, Rr8b, Rr8w, Rr8d, Rr8},

-    [7]  = {Rnone, Rr9b, Rr9w, Rr9d, Rr9},

-    [8]  = {Rnone, Rr10b, Rr10w, Rr10d, Rr10},

-    [9]  = {Rnone, Rr11b, Rr11w, Rr11d, Rr11},

-    [10]  = {Rnone, Rr12b, Rr12w, Rr12d, Rr12},

-    [11]  = {Rnone, Rr13b, Rr13w, Rr13d, Rr13},

-    [12]  = {Rnone, Rr14b, Rr14w, Rr14d, Rr14},

-    [13]  = {Rnone, Rr15b, Rr15w, Rr15d, Rr15},

-    [14]  = {Rnone, Rnone, Rnone, Rnone, Rnone},

-    [15]  = {Rnone, Rnone, Rnone, Rnone, Rnone},

+        /* None,   ModeB, ModeW, ModeL, ModeQ, ModeF, ModeD */

+    [0]  = {Rnone, Ral,   Rax,   Reax,  Rrax,  Rnone,  Rnone},

+    [1]  = {Rnone, Rcl,   Rcx,   Recx,  Rrcx,  Rnone,  Rnone},

+    [2]  = {Rnone, Rdl,   Rdx,   Redx,  Rrdx,  Rnone,  Rnone},

+    [3]  = {Rnone, Rbl,   Rbx,   Rebx,  Rrbx,  Rnone,  Rnone},

+    [4]  = {Rnone, Rsil,  Rsi,   Resi,  Rrsi,  Rnone,  Rnone},

+    [5]  = {Rnone, Rdil,  Rdi,   Redi,  Rrdi,  Rnone,  Rnone},

+    [6]  = {Rnone, Rr8b,  Rr8w,  Rr8d,  Rr8,   Rnone,  Rnone},

+    [7]  = {Rnone, Rr9b,  Rr9w,  Rr9d,  Rr9,   Rnone,  Rnone},

+    [8]  = {Rnone, Rr10b, Rr10w, Rr10d, Rr10,  Rnone,  Rnone},

+    [9]  = {Rnone, Rr11b, Rr11w, Rr11d, Rr11,  Rnone,  Rnone},

+    [10] = {Rnone, Rr12b, Rr12w, Rr12d, Rr12,  Rnone,  Rnone},

+    [11] = {Rnone, Rr13b, Rr13w, Rr13d, Rr13,  Rnone,  Rnone},

+    [12] = {Rnone, Rr14b, Rr14w, Rr14d, Rr14,  Rnone,  Rnone},

+    [13] = {Rnone, Rr15b, Rr15w, Rr15d, Rr15,  Rnone,  Rnone},

+    [14] = {Rnone, Rnone, Rnone, Rnone, Rnone, Rnone,  Rnone},

+    [15] = {Rnone, Rnone, Rnone, Rnone, Rnone, Rnone,  Rnone},

     [16] = {Rnone, Rnone, Rnone, Rnone, Rnone, Rxmm0f, Rxmm0d},

     [17] = {Rnone, Rnone, Rnone, Rnone, Rnone, Rxmm1f, Rxmm1d},

     [18] = {Rnone, Rnone, Rnone, Rnone, Rnone, Rxmm2f, Rxmm2d},

@@ -383,6 +384,8 @@

         return;

     if (v == Rrbp || v == Rrsp || v == Rrip)

         return;

+    if (rclass(locmap[u]) != rclass(locmap[v]))

+        return;

     gbputedge(s, u, v);

     gbputedge(s, v, u);

@@ -1202,7 +1205,7 @@

     s->shouldspill = mkbs();

     s->neverspill = mkbs();

     s->initial = mkbs();

-    for (i = 0; i < Nsaved; i++)

+    for (i = 0; i < s->nsaved; i++)

         bsput(s->shouldspill, s->calleesave[i]->reg.id);

     do {

         setup(s);