Fix printvideo regression and minor improvements

build/definitions.mk

@@ -28,12 +28,9 @@
 #   Other flags will usually be the ones that provide us marginal gains
 #   in terms of performance and code size, but the world won't end when
 #   they aren't used. Flags that play a critical role in source working
-#   correctly will usually be specified on a target-by-target basis, in
+#   correctly will usually be specified on a object-by-object basis, in
 #   their respective packages.
 #
-#   NOTE might have gotten the per-target override flag flow wrong since
-#        flags defined that way might flow quadratically opposite blaze.
-#
 # VARIABLES
 #
 #     CCFLAGS      gcc frontend flags (.i, .c, .cc, .f, .S, .lds, etc.)

dsp/scale/gyarados.c

@@ -43,7 +43,7 @@
  * @see Magikarp
  */
 
-#define M      15
+#define M      14
 #define SQR(X) ((X) * (X))
 
 struct SamplingSolution {

libc/bits/bits.h

@@ -50,78 +50,104 @@ unsigned long hamming(unsigned long, unsigned long) pureconst;
   ((uint16_t)kReverseBits[(X)&0xff] << 010 | \
    kReverseBits[((uint16_t)(X) >> 010) & 0xff])
 
-#ifndef __GNUC__
+#define READ16LE(S)                             \
-#define READ16LE(P) ((unsigned)(P)[1] << 010 | (unsigned)(P)[0])
+  ((uint16_t)((unsigned char *)(S))[1] << 010 | \
-#define READ32LE(P)                                              \
+   (uint16_t)((unsigned char *)(S))[0] << 000)
-  ((unsigned long)(P)[3] << 030 | (unsigned long)(P)[2] << 020 | \
+#define READ32LE(S)                             \
-   (unsigned long)(P)[1] << 010 | (unsigned long)(P)[0])
+  ((uint32_t)((unsigned char *)(S))[3] << 030 | \
-#define READ64LE(P)                                                   \
+   (uint32_t)((unsigned char *)(S))[2] << 020 | \
-  ((unsigned long long)(P)[3] << 030 | (unsigned long)(P)[2] << 020 | \
+   (uint32_t)((unsigned char *)(S))[1] << 010 | \
-   (unsigned long long)(P)[1] << 010 | (unsigned long)(P)[0])
+   (uint32_t)((unsigned char *)(S))[0] << 000)
-#else
+#define READ64LE(S)                             \
-#define READ16LE(P) read16le(P)
+  ((uint64_t)((unsigned char *)(S))[7] << 070 | \
-#define READ32LE(P) read32le(P)
+   (uint64_t)((unsigned char *)(S))[6] << 060 | \
-#define READ64LE(P) read64le(P)
+   (uint64_t)((unsigned char *)(S))[5] << 050 | \
-#define read16le(P)                                       \
+   (uint64_t)((unsigned char *)(S))[4] << 040 | \
-  ({                                                      \
+   (uint64_t)((unsigned char *)(S))[3] << 030 | \
-    const unsigned char *Pu = (const unsigned char *)(P); \
+   (uint64_t)((unsigned char *)(S))[2] << 020 | \
-    (uint16_t) Pu[1] << 010 | (uint16_t)Pu[0];            \
+   (uint64_t)((unsigned char *)(S))[1] << 010 | \
+   (uint64_t)((unsigned char *)(S))[0] << 000)
+
+#define READ16BE(S)                             \
+  ((uint16_t)((unsigned char *)(S))[0] << 010 | \
+   (uint16_t)((unsigned char *)(S))[1] << 000)
+#define READ32BE(S)                             \
+  ((uint32_t)((unsigned char *)(S))[0] << 030 | \
+   (uint32_t)((unsigned char *)(S))[1] << 020 | \
+   (uint32_t)((unsigned char *)(S))[2] << 010 | \
+   (uint32_t)((unsigned char *)(S))[3] << 000)
+#define READ64BE(S)                             \
+  ((uint64_t)((unsigned char *)(S))[0] << 070 | \
+   (uint64_t)((unsigned char *)(S))[1] << 060 | \
+   (uint64_t)((unsigned char *)(S))[2] << 050 | \
+   (uint64_t)((unsigned char *)(S))[3] << 040 | \
+   (uint64_t)((unsigned char *)(S))[4] << 030 | \
+   (uint64_t)((unsigned char *)(S))[5] << 020 | \
+   (uint64_t)((unsigned char *)(S))[6] << 010 | \
+   (uint64_t)((unsigned char *)(S))[7] << 000)
+
+#define read16le(S)                            \
+  ({                                           \
+    unsigned char *Str = (unsigned char *)(S); \
+    READ16LE(Str);                             \
   })
-#define read32le(P)                                       \
+#define read32le(S)                            \
-  ({                                                      \
+  ({                                           \
-    const unsigned char *Pu = (const unsigned char *)(P); \
+    unsigned char *Str = (unsigned char *)(S); \
-    ((uint32_t)Pu[3] << 030 | (uint32_t)Pu[2] << 020 |    \
+    READ32LE(Str);                             \
-     (uint32_t)Pu[1] << 010 | (uint32_t)Pu[0] << 000);    \
   })
-#define read64le(P)                                       \
+#define read64le(S)                            \
-  ({                                                      \
+  ({                                           \
-    const unsigned char *Pu = (const unsigned char *)(P); \
+    unsigned char *Str = (unsigned char *)(S); \
-    ((uint64_t)Pu[7] << 070 | (uint64_t)Pu[6] << 060 |    \
+    READ64LE(Str);                             \
-     (uint64_t)Pu[5] << 050 | (uint64_t)Pu[4] << 040 |    \
-     (uint64_t)Pu[3] << 030 | (uint64_t)Pu[2] << 020 |    \
-     (uint64_t)Pu[1] << 010 | (uint64_t)Pu[0] << 000);    \
   })
-#endif
 
-#define WRITE16LE(P, V)                  \
+#define read16be(S)                            \
-  do {                                   \
+  ({                                           \
-    uint8_t *Ple = (unsigned char *)(P); \
+    unsigned char *Str = (unsigned char *)(S); \
-    uint16_t Vle = (V);                  \
+    READ16BE(Str);                             \
-    Ple[0] = (uint8_t)(Vle >> 000);      \
+  })
-    Ple[1] = (uint8_t)(Vle >> 010);      \
+#define read32be(S)                            \
+  ({                                           \
+    unsigned char *Str = (unsigned char *)(S); \
+    READ32BE(Str);                             \
+  })
+#define read64be(S)                            \
+  ({                                           \
+    unsigned char *Str = (unsigned char *)(S); \
+    READ64BE(Str);                             \
+  })
+
+#define WRITE16LE(P, V)             \
+  do {                              \
+    uint8_t *Ple = (uint8_t *)(P);  \
+    uint16_t Vle = (V);             \
+    Ple[0] = (uint8_t)(Vle >> 000); \
+    Ple[1] = (uint8_t)(Vle >> 010); \
   } while (0)
-
+#define WRITE32LE(P, V)             \
-#define WRITE32LE(P, V)                  \
+  do {                              \
-  do {                                   \
+    uint8_t *Ple = (uint8_t *)(P);  \
-    uint8_t *Ple = (unsigned char *)(P); \
+    uint32_t Vle = (V);             \
-    uint32_t Vle = (V);                  \
+    Ple[0] = (uint8_t)(Vle >> 000); \
-    Ple[0] = (uint8_t)(Vle >> 000);      \
+    Ple[1] = (uint8_t)(Vle >> 010); \
-    Ple[1] = (uint8_t)(Vle >> 010);      \
+    Ple[2] = (uint8_t)(Vle >> 020); \
-    Ple[2] = (uint8_t)(Vle >> 020);      \
+    Ple[3] = (uint8_t)(Vle >> 030); \
-    Ple[3] = (uint8_t)(Vle >> 030);      \
   } while (0)
-
+#define WRITE64LE(P, V)             \
-#define WRITE64LE(P, V)                  \
+  do {                              \
-  do {                                   \
+    uint8_t *Ple = (uint8_t *)(P);  \
-    uint8_t *Ple = (unsigned char *)(P); \
+    uint64_t Vle = (V);             \
-    uint64_t Vle = (V);                  \
+    Ple[0] = (uint8_t)(Vle >> 000); \
-    Ple[0] = (uint8_t)(Vle >> 000);      \
+    Ple[1] = (uint8_t)(Vle >> 010); \
-    Ple[1] = (uint8_t)(Vle >> 010);      \
+    Ple[2] = (uint8_t)(Vle >> 020); \
-    Ple[2] = (uint8_t)(Vle >> 020);      \
+    Ple[3] = (uint8_t)(Vle >> 030); \
-    Ple[3] = (uint8_t)(Vle >> 030);      \
+    Ple[4] = (uint8_t)(Vle >> 040); \
-    Ple[4] = (uint8_t)(Vle >> 040);      \
+    Ple[5] = (uint8_t)(Vle >> 050); \
-    Ple[5] = (uint8_t)(Vle >> 050);      \
+    Ple[6] = (uint8_t)(Vle >> 060); \
-    Ple[6] = (uint8_t)(Vle >> 060);      \
+    Ple[7] = (uint8_t)(Vle >> 070); \
-    Ple[7] = (uint8_t)(Vle >> 070);      \
   } while (0)
 
-/* TODO(jart): these ones aren't coded correctly */
-#define read128le(P) ((uint128_t)read64le((P) + 8) << 0100 | read64le(P))
-#define read16be(P)  ((uint16_t)(*(P) << 010) | (uint16_t)(*((P) + 1)))
-#define read32be(P)  ((uint32_t)read16be(P) << 020 | (uint32_t)read16be((P) + 2))
-#define read64be(P)  ((uint64_t)read32be(P) << 040 | read32be((P) + 4))
-#define read128be(P) ((uint128_t)read64be(P) << 0100 | read64be((P) + 8))
-
 /*───────────────────────────────────────────────────────────────────────────│─╗
 │ cosmopolitan § bits » some assembly required                             ─╬─│┼
 ╚────────────────────────────────────────────────────────────────────────────│*/

libc/math.h

@@ -148,7 +148,7 @@ float atanf(float);
 float atanhf(float);
 float cbrtf(float);
 float ceilf(float);
-float copysignf(float, float y);
+float copysignf(float, float);
 float cosf(float);
 float coshf(float);
 float dremf(float, float);

libc/nexgen32e/memrchr.S

@@ -38,7 +38,7 @@ memrchr:.leafprologue
 	jz	5f
 	vmovd	%esi,%xmm0
 	vpbroadcastb %xmm0,%ymm0
-3:	vmovups	-32(%rdi,%rdx),%ymm1
+3:	vmovdqu	-32(%rdi,%rdx),%ymm1
 	vpcmpeqb %ymm1,%ymm0,%ymm1
 	vpmovmskb %ymm1,%eax
 	lzcnt	%eax,%eax

libc/nexgen32e/memrchr16.S

@@ -38,7 +38,7 @@ memrchr16:
 	jz	5f
 	vmovd	%esi,%xmm0
 	vpbroadcastw %xmm0,%ymm0
-3:	vmovups	-32(%rdi,%rdx,2),%ymm1
+3:	vmovdqu	-32(%rdi,%rdx,2),%ymm1
 	vpcmpeqw %ymm1,%ymm0,%ymm1
 	vpmovmskb %ymm1,%eax
 	lzcnt	%eax,%eax

libc/nexgen32e/memrchr32.S

@@ -38,7 +38,7 @@ wmemrchr:
 	jz	5f
 	vmovd	%esi,%xmm0
 	vpbroadcastd %xmm0,%ymm0
-3:	vmovups	-32(%rdi,%rdx,4),%ymm1
+3:	vmovdqu	-32(%rdi,%rdx,4),%ymm1
 	vpcmpeqd %ymm1,%ymm0,%ymm1
 	vpmovmskb %ymm1,%eax
 	lzcnt	%eax,%eax

libc/unicode/unicode.h

@@ -7,9 +7,9 @@ COSMOPOLITAN_C_START_
 │ cosmopolitan § characters » unicode                                      ─╬─│┼
 ╚────────────────────────────────────────────────────────────────────────────│*/
 
-extern const uint64_t kEastAsianWidth[];
+extern const uint8_t kEastAsianWidth[];
 extern const uint32_t kEastAsianWidthBits;
-extern const uint64_t kCombiningChars[];
+extern const uint8_t kCombiningChars[];
 extern const uint32_t kCombiningCharsBits;
 
 /*───────────────────────────────────────────────────────────────────────────│─╗

libc/unicode/wcwidth.c

@@ -17,7 +17,6 @@
 │ Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA                │
 │ 02110-1301 USA                                                               │
 ╚─────────────────────────────────────────────────────────────────────────────*/
-#include "libc/bits/bits.h"
 #include "libc/unicode/unicode.h"
 
 /**
@@ -28,10 +27,10 @@ int wcwidth(wchar_t ucs) {
   if (ucs < 32 || (ucs >= 0x7f && ucs < 0xa0)) {
     return -1;
   } else if (0 <= ucs && ucs < kCombiningCharsBits &&
-             bt(kCombiningChars, ucs)) {
+             !!(kCombiningChars[ucs >> 3] & (1 << (ucs & 7)))) {
     return 0;
   } else if (0 <= ucs && ucs < kEastAsianWidthBits) {
-    return 1 + bt(kEastAsianWidth, ucs);
+    return 1 + !!(kEastAsianWidth[ucs >> 3] & (1 << (ucs & 7)));
   } else {
     return 1;
   }

test/libc/unicode/wcwidth_test.c

@@ -19,6 +19,7 @@
 ╚─────────────────────────────────────────────────────────────────────────────*/
 #include "libc/bits/bits.h"
 #include "libc/str/str.h"
+#include "libc/testlib/ezbench.h"
 #include "libc/testlib/testlib.h"
 #include "libc/unicode/unicode.h"
 
@@ -73,3 +74,8 @@ TEST(strwidth, testTextDelimitingControlCodes_dontHaveSubstance) {
   EXPECT_EQ(0, strwidth("\0"));
   EXPECT_EQ(0, strwidth("\1"));
 }
+
+BENCH(wcwidth, bench) {
+  volatile int x;
+  EZBENCH2("wcwidth", donothing, x = wcwidth(VEIL("r", u'→')));
+}

test/tool/build/lib/pml4t_test.c

@@ -21,6 +21,7 @@
 #include "libc/macros.h"
 #include "libc/mem/mem.h"
 #include "libc/stdio/stdio.h"
+#include "libc/str/str.h"
 #include "libc/testlib/ezbench.h"
 #include "libc/testlib/testlib.h"
 #include "tool/build/lib/memory.h"

third_party/xed/x86.h

@@ -1,10 +1,5 @@
 #ifndef COSMOPOLITAN_THIRD_PARTY_XED_X86_H_
 #define COSMOPOLITAN_THIRD_PARTY_XED_X86_H_
-#include "libc/bits/bits.h"
-#include "libc/runtime/runtime.h"
-#include "libc/str/str.h"
-#if !(__ASSEMBLER__ + __LINKER__ + 0)
-COSMOPOLITAN_C_START_
 /*           ▓▓▓▓▓▓▓▓▓▓▓▓▓                      ▄▄▄▄
              ▓▓▓▓▓▓▓▓▓▓▓▓▓     ▄▓▓▓▓▓▓▄      ▄▓▓▓▓▓▓▓▓            ▄▓▓▓▀
              ▓▓▓▓    ▓▓▓▓▓    ▓     ▓▓▓▓    ▓▓       ▓▓▓      ▄▓▓▓▓
@@ -47,6 +42,9 @@ COSMOPOLITAN_C_START_
 #define xed_sib_scale(M)           (((M)&0xff) >> 6)
 #define xed_get_modrm_reg_field(M) (((M)&0x38) >> 3)
 
+#if !(__ASSEMBLER__ + __LINKER__ + 0)
+COSMOPOLITAN_C_START_
+
 enum XedMachineMode {
   XED_MACHINE_MODE_INVALID,
   XED_MACHINE_MODE_LONG_64,
@@ -321,44 +319,44 @@ struct XedChipFeatures {
 };
 
 struct XedOperands { /*
-     ┌prefix66
+    ┌rep
-     │┌log₂𝑏
+    │ ┌log₂𝑏
-     ││ ┌rexx
+    │ │ ┌rexx
-     ││ │┌index
+    │ │ │┌index
-     ││ ││  ┌mod
+    │ │ ││  ┌mod
-     ││ ││  │ ┌rexb
+    │ │ ││  │ ┌rexb
-     ││ ││  │ │┌base
+    │ │ ││  │ │┌base
-     ││ ││  │ ││  ┌asz
+    │ │ ││  │ ││  ┌asz
-     ││ ││  │ ││  │┌rex         REGISTER
+    │ │ ││  │ ││  │┌rex         REGISTER
-     ││ ││  │ ││  ││┌rexb       DISPATCH
+    │ │ ││  │ ││  ││┌rexb       DISPATCH
-     ││ ││  │ ││  │││┌srm       ENCODING
+    │ │ ││  │ ││  │││┌srm       ENCODING
-     ││ ││  │ ││  ││││  ┌rexw
+    │ │ ││  │ ││  ││││  ┌rexw
-     ││ ││  │ ││  ││││  │┌rex
+    │ │ ││  │ ││  ││││  │┌rex
-     ││ ││  │ ││  ││││  ││┌rexb
+    │ │ ││  │ ││  ││││  ││┌rexb
-     ││ ││  │ ││  ││││  │││┌rm
+    │ │ ││  │ ││  ││││  │││┌rm
-     ││ ││  │ ││  ││││  ││││  ┌osz
+    │ │ ││  │ ││  ││││  ││││  ┌osz
-     ││ ││  │ ││  ││││  ││││  │┌rex
+    │ │ ││  │ ││  ││││  ││││  │┌rex
-     ││ ││  │ ││  ││││  ││││  ││┌rexr
+    │ │ ││  │ ││  ││││  ││││  ││┌rexr
-     ││ ││  │ ││  ││││  ││││  │││┌reg
+    │ │ ││  │ ││  ││││  ││││  │││┌reg
-     ││2││  │2││  ││││  ││││  ││││
+    │ │2││  │2││  ││││  ││││  ││││
-     ││8││24│2││18││││12││││ 6││││ 0
+    │ │8││24│2││18││││12││││ 6││││ 0
-     │├┐│├─┐├┐│├─┐│││├─┐│││├─┐│││├─┐
+    ├┐├┐│├─┐├┐│├─┐│││├─┐│││├─┐│││├─┐
   0b00000000000000000000000000000000*/
   uint32_t rde;
-  bool osz : 1;   // operand size override prefix
+  bool rexw : 1;   // rex.w or rex.wb or etc. 64-bit override
-  bool rexw : 1;  // rex.w or rex.wb or etc. 64-bit override
+  bool rexb : 1;   // rex.b or rex.wb or etc. see modrm table
-  bool rexb : 1;  // rex.b or rex.wb or etc. see modrm table
+  bool rexr : 1;   // rex.r or rex.wr or etc. see modrm table
-  bool rexr : 1;  // rex.r or rex.wr or etc. see modrm table
+  bool rex : 1;    // any rex prefix including rex
-  bool rex : 1;   // any rex prefix including rex
+  bool rexx : 1;   // rex.x or rex.wx or etc. see sib table
-  bool asz : 1;   // address size override
+  bool rexrr : 1;  // evex
-  bool rexx : 1;  // rex.x or rex.wx or etc. see sib table
+  bool asz : 1;    // address size override
+  bool osz : 1;    // operand size override prefix
   bool out_of_bytes : 1;
   bool is_intel_specific : 1;
   bool ild_f2 : 1;
   bool ild_f3 : 1;
   bool has_modrm : 1;
   bool has_sib : 1;
-  bool prefix66 : 1;  // rexw except for xmm ops
   bool realmode : 1;
   bool amd3dnow : 1;
   uint8_t max_bytes;
@@ -385,8 +383,6 @@ struct XedOperands { /*
   uint8_t map : 4;            // enum XedIldMap
   uint8_t hint : 2;           // static branch prediction
   uint8_t seg_ovd : 3;        // XED_SEG_xx
-  uint8_t first_f2f3 : 2;     // internal see rep, ild_f2, ild_f3
-  uint8_t last_f2f3 : 2;      // internal see rep, ild_f2, ild_f3
   uint8_t error : 5;          // enum XedError
   uint8_t mode : 3;           // real,legacy,long
   uint8_t rep : 2;            // 0, 2 (0xf2 repnz), 3 (0xf3 rep/repe)
@@ -410,7 +406,6 @@ struct XedOperands { /*
   uint8_t vex_prefix;  // vex
   uint8_t zeroing;     // evex
   uint8_t bcrc;        // evex
-  uint8_t rexrr;       // evex
   uint8_t llrc;        // evex
   uint8_t vl;          // evex
   uint8_t mask;        // evex
@@ -453,7 +448,7 @@ forceinline void xed_operands_set_mode(struct XedOperands *p,
       p->mode = 0;
       break;
     default:
-      abort();
+      unreachable;
   }
 }
 

third_party/xed/x86ild.greg.c

@@ -517,17 +517,17 @@ privileged static int xed_consume_byte(struct XedDecodedInst *d) {
 }
 
 privileged static void xed_prefix_scanner(struct XedDecodedInst *d) {
+  xed_bits_t first_f2f3, last_f2f3;
   xed_bits_t b, max_bytes, length, nprefixes, nseg_prefixes, nrexes, rex;
   length = d->length;
   max_bytes = d->op.max_bytes;
-  rex = nrexes = nprefixes = nseg_prefixes = 0;
+  first_f2f3 = last_f2f3 = rex = nrexes = nprefixes = nseg_prefixes = 0;
   while (length < max_bytes) {
     b = d->bytes[length];
     if (xed_get_prefix_table_bit(b) == 0) goto out;
     switch (b) {
       case 0x66:
         d->op.osz = true;
-        d->op.prefix66 = true;
         rex = 0;
         break;
       case 0x67:
@@ -558,17 +558,17 @@ privileged static void xed_prefix_scanner(struct XedDecodedInst *d) {
         break;
       case 0xF3:
         d->op.ild_f3 = true;
-        d->op.last_f2f3 = 3;
+        last_f2f3 = 3;
-        if (!d->op.first_f2f3) {
+        if (!first_f2f3) {
-          d->op.first_f2f3 = 3;
+          first_f2f3 = 3;
         }
         rex = 0;
         break;
       case 0xF2:
         d->op.ild_f2 = true;
-        d->op.last_f2f3 = 2;
+        last_f2f3 = 2;
-        if (!d->op.first_f2f3) {
+        if (!first_f2f3) {
-          d->op.first_f2f3 = 2;
+          first_f2f3 = 2;
         }
         rex = 0;
         break;
@@ -597,9 +597,9 @@ out:
     d->op.rex = true;
   }
   if (d->op.mode_first_prefix) {
-    d->op.rep = d->op.first_f2f3;
+    d->op.rep = first_f2f3;
   } else {
-    d->op.rep = d->op.last_f2f3;
+    d->op.rep = last_f2f3;
   }
   switch (d->op.ild_seg) {
     case 0x2e:
@@ -752,8 +752,8 @@ privileged static void xed_evex_scanner(struct XedDecodedInst *d) {
       d->op.rexb = ~evex1.s.b_inv & 1;
       d->op.rexrr = ~evex1.s.rr_inv & 1;
     }
-    d->op.map = evex1.s.map;
     d->op.rexw = evex2.s.rexw & 1;
+    d->op.map = evex1.s.map;
     d->op.vexdest3 = evex2.s.vexdest3;
     d->op.vexdest210 = evex2.s.vexdest210;
     d->op.ubit = evex2.s.ubit;
@@ -1199,7 +1199,7 @@ privileged static void xed_encode_rde(struct XedDecodedInst *x) {
   /* fragile examples:        addb, cmpxchgb */
   /* fragile counterexamples: btc, bsr, etc. */
   const uint8_t kWordLog2[2][2][2] = {{{2, 3}, {1, 3}}, {{0, 0}, {0, 0}}};
-  x->op.rde = x->op.prefix66 << 30 |
+  x->op.rde = x->op.rep << 30 |
               kWordLog2[~x->op.opcode & 1][x->op.osz][x->op.rexw] << 28 |
               x->op.mod << 22 | x->op.rexw << 11 | x->op.osz << 5 |
               x->op.asz << 17 | (x->op.rexx << 3 | x->op.index) << 24 |

tool/build/lib/cvt.c

@@ -241,9 +241,7 @@ static void OpVdqWpdCvtpd2dq(struct Machine *m) {
 }
 
 void OpCvt(struct Machine *m, unsigned long op) {
-  op |= m->xedd->op.rep;
+  switch (op | Rep(m->xedd) | Osz(m->xedd)) {
-  op |= Prefix66(m->xedd);
-  switch (op) {
     case kOpCvt0f2a + 0:
       OpVpsQpiCvtpi2ps(m);
       break;

tool/build/lib/disarg.c

@@ -23,8 +23,17 @@
 #include "libc/macros.h"
 #include "libc/str/str.h"
 #include "tool/build/lib/dis.h"
+#include "tool/build/lib/endian.h"
 #include "tool/build/lib/modrm.h"
 
+#define SSTRCMP8(S1, S2)        \
+  ({                            \
+    uint64_t x, y;              \
+    x = READ64BE(S1);           \
+    y = READ64BE(S2);           \
+    x > y ? 1 : x < y ? -1 : 0; \
+  })
+
 static const char kScale[4][4] = {"", ",2", ",4", ",8"};
 static const char kSegName[8][3] = {"es", "cs", "ss", "ds", "fs", "gs"};
 static const char kSegOverride[8][3] = {"", "cs", "ds", "es", "fs", "gs", "ss"};
@@ -541,12 +550,14 @@ static const struct DisArg {
 };
 
 char *DisArg(struct DisBuilder b, char *p, const char *s) {
+  char key[8];
   int c, m, l, r;
   l = 0;
   r = ARRAYLEN(kDisArgs) - 1;
+  strncpy(key, s, 8);
   while (l <= r) {
     m = (l + r) >> 1;
-    c = strcmp(kDisArgs[m].s, s);
+    c = SSTRCMP8(kDisArgs[m].s, key);
     if (c < 0) {
       l = m + 1;
     } else if (c > 0) {

tool/build/lib/disinst.c

@@ -45,7 +45,7 @@ static int IsRepOpcode(struct DisBuilder b) {
 
 static char *DisRepPrefix(struct DisBuilder b, char *p) {
   const char *s;
-  if (b.xedd->op.rep && b.xedd->op.map == XED_ILD_MAP0) {
+  if (Rep(b.xedd) && b.xedd->op.map == XED_ILD_MAP0) {
     switch (IsRepOpcode(b)) {
       case 0:
         break;
@@ -53,7 +53,7 @@ static char *DisRepPrefix(struct DisBuilder b, char *p) {
         p = stpcpy(p, "rep ");
         break;
       case 2:
-        p = stpcpy(p, b.xedd->op.rep == 2 ? "repnz " : "repz ");
+        p = stpcpy(p, Rep(b.xedd) == 2 ? "repnz " : "repz ");
         break;
       default:
         break;
@@ -135,7 +135,7 @@ static char *DisName(struct DisBuilder b, char *bp, const char *name,
       p = DisBranchTaken(b, p);
     }
     if (wantsuffixsd) {
-      if (b.xedd->op.prefix66) {
+      if (Osz(b.xedd)) {
         *p++ = 'd';
       } else {
         *p++ = 's';

tool/build/lib/disspec.c

@@ -49,7 +49,7 @@ char *DisOpFpu1(struct XedDecodedInst *x, char *p, const char *extra) {
 
 char *DisOp66(struct XedDecodedInst *x, char *p, const char *s, const char *a,
               const char *b) {
-  stpcpy(stpcpy(p, s), !Prefix66(x) ? a : b);
+  stpcpy(stpcpy(p, s), !Osz(x) ? a : b);
   return p;
 }
 
@@ -80,7 +80,7 @@ char *DisOpVpsWpsVssWssVpdWpdVsdWsd(struct XedDecodedInst *x, char *p,
     stpcpy(q, "ss %Vss Wss");
   } else if (x->op.ild_f2) {
     stpcpy(q, "sd %Vsd Wsd");
-  } else if (Prefix66(x)) {
+  } else if (Osz(x)) {
     stpcpy(q, "pd %Vpd Wpd");
   } else {
     stpcpy(q, "ps %Vps Wps");
@@ -580,8 +580,8 @@ const char *DisSpecMap1(struct XedDecodedInst *x, char *p) {
     RCASE(0x28, "movapSD %Vps Wps");
     RCASE(0x29, "movapSD Wps %Vps");
     RCASE(0x2B, "movntpSD Mps %Vps");
-    RCASE(0x2E, Prefix66(x) ? "ucomisd %Vsd Wsd" : "ucomiss %Vss Wss");
+    RCASE(0x2E, Osz(x) ? "ucomisd %Vsd Wsd" : "ucomiss %Vss Wss");
-    RCASE(0x2F, Prefix66(x) ? "comisd %Vsd Wsd" : "comiss %Vss Wss");
+    RCASE(0x2F, Osz(x) ? "comisd %Vsd Wsd" : "comiss %Vss Wss");
     RCASE(0x31, "rdtsc");
     RCASE(0x40, "cmovo %Gvqp Evqp");
     RCASE(0x41, "cmovno %Gvqp Evqp");
@@ -693,7 +693,7 @@ const char *DisSpecMap1(struct XedDecodedInst *x, char *p) {
     RCASE(0xD3, DisOpPqQqVdqWdq(x, p, "psrlq"));
     RCASE(0xD4, DisOpPqQqVdqWdq(x, p, "paddq"));
     RCASE(0xD5, DisOpPqQqVdqWdq(x, p, "pmullw"));
-    RCASE(0xD7, Prefix66(x) ? "pmovmskb %Gdqp %Udq" : "pmovmskb %Gdqp %Nq");
+    RCASE(0xD7, Osz(x) ? "pmovmskb %Gdqp %Udq" : "pmovmskb %Gdqp %Nq");
     RCASE(0xD8, DisOpPqQqVdqWdq(x, p, "psubusb"));
     RCASE(0xD9, DisOpPqQqVdqWdq(x, p, "psubusw"));
     RCASE(0xDA, DisOpPqQqVdqWdq(x, p, "pminub"));
@@ -708,7 +708,7 @@ const char *DisSpecMap1(struct XedDecodedInst *x, char *p) {
     RCASE(0xE3, DisOpPqQqVdqWdq(x, p, "pavgw"));
     RCASE(0xE4, DisOpPqQqVdqWdq(x, p, "pmulhuw"));
     RCASE(0xE5, DisOpPqQqVdqWdq(x, p, "pmulhw"));
-    RCASE(0xE7, Prefix66(x) ? "movntdq Mdq %Vdq" : "movntq Mq %Pq");
+    RCASE(0xE7, Osz(x) ? "movntdq Mdq %Vdq" : "movntq Mq %Pq");
     RCASE(0xE8, DisOpPqQqVdqWdq(x, p, "psubsb"));
     RCASE(0xE9, DisOpPqQqVdqWdq(x, p, "psubsw"));
     RCASE(0xEA, DisOpPqQqVdqWdq(x, p, "pminsw"));
@@ -740,7 +740,7 @@ const char *DisSpecMap1(struct XedDecodedInst *x, char *p) {
       }
       break;
     case 0x70:
-      switch (x->op.rep | Prefix66(x)) {
+      switch (Rep(x) | Osz(x)) {
         RCASE(0, "pshufw %Pq Qq Ib");
         RCASE(1, "pshufd %Vdq Wdq Ib");
         RCASE(2, "pshuflw %Vdq Wdq Ib");
@@ -799,7 +799,7 @@ const char *DisSpecMap1(struct XedDecodedInst *x, char *p) {
         return "movss %Vss Wss";
       } else if (x->op.ild_f2) {
         return "movsd %Vsd Wsd";
-      } else if (Prefix66(x)) {
+      } else if (Osz(x)) {
         return "movupd %Vpd Wpd";
       } else {
         return "movups %Vps Wps";
@@ -810,14 +810,14 @@ const char *DisSpecMap1(struct XedDecodedInst *x, char *p) {
         return "movss Wss %Vss";
       } else if (x->op.ild_f2) {
         return "movsd Wsd %Vsd";
-      } else if (Prefix66(x)) {
+      } else if (Osz(x)) {
         return "movupd Wpd %Vpd";
       } else {
         return "movups Wps %Vps";
       }
       break;
     case 0xC4:
-      if (!Prefix66(x)) {
+      if (!Osz(x)) {
         if (IsModrmRegister(x)) {
           return "pinsrw %Pq %Rdqp Ib";
         } else {
@@ -832,14 +832,14 @@ const char *DisSpecMap1(struct XedDecodedInst *x, char *p) {
       }
       break;
     case 0xC5:
-      if (!Prefix66(x)) {
+      if (!Osz(x)) {
         return "pextrw %Gdqp %Nq Ib";
       } else {
         return "pextrw %Gdqp %Udq Ib";
       }
       break;
     case 0xC6:
-      if (!Prefix66(x)) {
+      if (!Osz(x)) {
         return "shufps %Vps Wps Ib";
       } else {
         return "shufpd %Vpd Wpd Ib";
@@ -853,7 +853,7 @@ const char *DisSpecMap1(struct XedDecodedInst *x, char *p) {
       }
       break;
     case 0xD6:
-      if (Prefix66(x)) {
+      if (Osz(x)) {
         return "movq Wq %Vq";
       } else if (x->op.ild_f3) {
         return "movq2dq %Vdq %Nq";
@@ -862,7 +862,7 @@ const char *DisSpecMap1(struct XedDecodedInst *x, char *p) {
       }
       break;
     case 0x12:
-      switch (x->op.rep | Prefix66(x)) {
+      switch (Rep(x) | Osz(x)) {
         case 0:
           if (IsModrmRegister(x)) {
             return "movhlps %Vq %Uq";
@@ -881,14 +881,14 @@ const char *DisSpecMap1(struct XedDecodedInst *x, char *p) {
       }
       break;
     case 0x13:
-      if (Prefix66(x)) {
+      if (Osz(x)) {
         return "movlpd Mq %Vq";
       } else {
         return "movlps Mq %Vq";
       }
       break;
     case 0x16:
-      switch (x->op.rep | Prefix66(x)) {
+      switch (Rep(x) | Osz(x)) {
         case 0:
           if (IsModrmRegister(x)) {
             return "movlhps %Vq %Uq";
@@ -905,7 +905,7 @@ const char *DisSpecMap1(struct XedDecodedInst *x, char *p) {
       }
       break;
     case 0x17:
-      if (Prefix66(x)) {
+      if (Osz(x)) {
         return "movhpd Mq %Vq";
       } else {
         return "movhps Mq %Vq";
@@ -916,7 +916,7 @@ const char *DisSpecMap1(struct XedDecodedInst *x, char *p) {
         return "cvtsi2ss %Vss Edqp";
       } else if (x->op.ild_f2) {
         return "cvtsi2sd %Vsd Edqp";
-      } else if (Prefix66(x)) {
+      } else if (Osz(x)) {
         return "cvtpi2pd %Vpd Qpi";
       } else {
         return "cvtpi2ps %Vps Qpi";
@@ -927,7 +927,7 @@ const char *DisSpecMap1(struct XedDecodedInst *x, char *p) {
         return "cvttss2si %Gdqp Wss";
       } else if (x->op.ild_f2) {
         return "cvttsd2si %Gdqp Wsd";
-      } else if (Prefix66(x)) {
+      } else if (Osz(x)) {
         return "cvttpd2pi %Ppi Wpd";
       } else {
         return "cvttps2pi %Ppi Wpsq";
@@ -938,7 +938,7 @@ const char *DisSpecMap1(struct XedDecodedInst *x, char *p) {
         return "cvtss2si %Gdqp Wss";
       } else if (x->op.ild_f2) {
         return "cvtsd2si %Gdqp Wsd";
-      } else if (Prefix66(x)) {
+      } else if (Osz(x)) {
         return "cvtpd2pi %Ppi Wpd";
       } else {
         return "cvtps2pi %Ppi Wpsq";
@@ -949,7 +949,7 @@ const char *DisSpecMap1(struct XedDecodedInst *x, char *p) {
         return "cvtss2sd %Vsd Wss";
       } else if (x->op.ild_f2) {
         return "cvtsd2ss %Vss Wsd";
-      } else if (Prefix66(x)) {
+      } else if (Osz(x)) {
         return "cvtpd2ps %Vps Wpd";
       } else {
         return "cvtps2pd %Vpd Wps";
@@ -958,7 +958,7 @@ const char *DisSpecMap1(struct XedDecodedInst *x, char *p) {
     case 0x5b:
       if (x->op.ild_f3) {
         return "cvttps2dq %Vdq Wps";
-      } else if (Prefix66(x)) {
+      } else if (Osz(x)) {
         return "cvtps2dq %Vdq Wps";
       } else {
         return "cvtdq2ps %Vps Wdq";
@@ -969,14 +969,14 @@ const char *DisSpecMap1(struct XedDecodedInst *x, char *p) {
         return "sqrtss %Vss Wss";
       } else if (x->op.ild_f2) {
         return "sqrtsd %Vsd Wsd";
-      } else if (Prefix66(x)) {
+      } else if (Osz(x)) {
         return "sqrtpd %Vpd Wpd";
       } else {
         return "sqrtps %Vps Wps";
       }
       break;
     case 0x6E:
-      if (Prefix66(x)) {
+      if (Osz(x)) {
         if (Rexw(x)) {
           return "movq %Vdq Eqp";
         } else {
@@ -993,7 +993,7 @@ const char *DisSpecMap1(struct XedDecodedInst *x, char *p) {
     case 0x6F:
       if (x->op.ild_f3) {
         return "movdqu %Vdq Wdq";
-      } else if (Prefix66(x)) {
+      } else if (Osz(x)) {
         return "movdqa %Vdq Wdq";
       } else {
         return "movq %Pq Qq";
@@ -1002,7 +1002,7 @@ const char *DisSpecMap1(struct XedDecodedInst *x, char *p) {
     case 0x7E:
       if (x->op.ild_f3) {
         return "movq %Vq Wq";
-      } else if (Prefix66(x)) {
+      } else if (Osz(x)) {
         if (Rexw(x)) {
           return "movq Eqp %Vdq";
         } else {
@@ -1019,7 +1019,7 @@ const char *DisSpecMap1(struct XedDecodedInst *x, char *p) {
     case 0x7F:
       if (x->op.ild_f3) {
         return "movdqu Wdq %Vdq";
-      } else if (Prefix66(x)) {
+      } else if (Osz(x)) {
         return "movdqa Wdq %Vdq";
       } else {
         return "movq Qq %Pq";
@@ -1028,7 +1028,7 @@ const char *DisSpecMap1(struct XedDecodedInst *x, char *p) {
     case 0xE6:
       if (x->op.ild_f2) {
         return "cvtpd2dq %Vdq Wpd";
-      } else if (Prefix66(x)) {
+      } else if (Osz(x)) {
         return "cvttpd2dq %Vdq Wpd";
       } else if (x->op.ild_f3) {
         return "cvtdq2pd %Vpd Wdq";

tool/build/lib/machine.c

@@ -329,7 +329,7 @@ static void OpLoop(struct Machine *m, bool cond) {
 
 static void OpXlat(struct Machine *m) {
   int64_t v;
-  uint8_t al, *p;
+  uint8_t al;
   v = Read64(m->bx) + Read8(m->ax);
   if (Asz(m->xedd)) v &= 0xffffffff;
   SetReadAddr(m, v, 1);
@@ -613,7 +613,7 @@ static uint8_t pmovmskb(uint64_t x) {
 
 static void OpPmovmskbGdqpNqUdq(struct Machine *m) {
   uint64_t bitmask;
-  if (Prefix66(m->xedd)) {
+  if (Osz(m->xedd)) {
     bitmask = pmovmskb(Read64(XmmRexbRm(m) + 8)) << 8 |
               pmovmskb(Read64(XmmRexbRm(m)));
   } else {
@@ -895,7 +895,7 @@ static void OpMaskMovDiXmmRegXmmRm(struct Machine *m) {
   uint64_t v;
   unsigned i, n;
   uint8_t *mem, b[16];
-  n = Prefix66(m->xedd) ? 16 : 8;
+  n = Osz(m->xedd) ? 16 : 8;
   v = GetSegment() + Read64(m->di);
   if (Asz(m->xedd)) v &= 0xffffffff;
   mem = BeginStore(m, v, n, p, b);
@@ -1011,7 +1011,7 @@ static void OpMovapdWpdVpd(struct Machine *m) {
 
 static void OpMovWpsVps(struct Machine *m) {
   uint8_t *p, *r;
-  switch (m->xedd->op.rep | Prefix66(m->xedd)) {
+  switch (Rep(m->xedd) | Osz(m->xedd)) {
     case 0:
       OpMovupsWpsVps(m);
       break;
@@ -1030,7 +1030,7 @@ static void OpMovWpsVps(struct Machine *m) {
 }
 
 static void OpMov0f28(struct Machine *m) {
-  if (!Prefix66(m->xedd)) {
+  if (!Osz(m->xedd)) {
     OpMovapsVpsWps(m);
   } else {
     OpMovapdVpdWpd(m);
@@ -1038,7 +1038,7 @@ static void OpMov0f28(struct Machine *m) {
 }
 
 static void OpMov0f6e(struct Machine *m) {
-  if (Prefix66(m->xedd)) {
+  if (Osz(m->xedd)) {
     if (Rexw(m->xedd)) {
       OpMovqVdqEqp(m);
     } else {
@@ -1054,7 +1054,7 @@ static void OpMov0f6e(struct Machine *m) {
 }
 
 static void OpMov0f6f(struct Machine *m) {
-  if (Prefix66(m->xedd)) {
+  if (Osz(m->xedd)) {
     OpMovdqaVdqWdq(m);
   } else if (m->xedd->op.ild_f3) {
     OpMovdquVdqWdq(m);
@@ -1064,7 +1064,7 @@ static void OpMov0f6f(struct Machine *m) {
 }
 
 static void OpMov0fE7(struct Machine *m) {
-  if (!Prefix66(m->xedd)) {
+  if (!Osz(m->xedd)) {
     OpMovntqMqPq(m);
   } else {
     OpMovntdqMdqVdq(m);
@@ -1074,7 +1074,7 @@ static void OpMov0fE7(struct Machine *m) {
 static void OpMov0f7e(struct Machine *m) {
   if (m->xedd->op.ild_f3) {
     OpMovqVqWq(m);
-  } else if (Prefix66(m->xedd)) {
+  } else if (Osz(m->xedd)) {
     if (Rexw(m->xedd)) {
       OpMovqEqpVdq(m);
     } else {
@@ -1092,7 +1092,7 @@ static void OpMov0f7e(struct Machine *m) {
 static void OpMov0f7f(struct Machine *m) {
   if (m->xedd->op.ild_f3) {
     OpMovdquWdqVdq(m);
-  } else if (Prefix66(m->xedd)) {
+  } else if (Osz(m->xedd)) {
     OpMovdqaWdqVdq(m);
   } else {
     OpMovqQqPq(m);
@@ -1101,7 +1101,7 @@ static void OpMov0f7f(struct Machine *m) {
 
 static void OpMov0f10(struct Machine *m) {
   uint8_t *p, *r;
-  switch (m->xedd->op.rep | Prefix66(m->xedd)) {
+  switch (Rep(m->xedd) | Osz(m->xedd)) {
     case 0:
       OpMovupsVpsWps(m);
       break;
@@ -1120,7 +1120,7 @@ static void OpMov0f10(struct Machine *m) {
 }
 
 static void OpMov0f29(struct Machine *m) {
-  if (!Prefix66(m->xedd)) {
+  if (!Osz(m->xedd)) {
     OpMovapsWpsVps(m);
   } else {
     OpMovapdWpdVpd(m);
@@ -1128,7 +1128,7 @@ static void OpMov0f29(struct Machine *m) {
 }
 
 static void OpMov0f2b(struct Machine *m) {
-  if (!Prefix66(m->xedd)) {
+  if (!Osz(m->xedd)) {
     OpMovntpsMpsVps(m);
   } else {
     OpMovntpdMpdVpd(m);
@@ -1136,7 +1136,7 @@ static void OpMov0f2b(struct Machine *m) {
 }
 
 static void OpMov0f12(struct Machine *m) {
-  switch (m->xedd->op.rep | Prefix66(m->xedd)) {
+  switch (Rep(m->xedd) | Osz(m->xedd)) {
     case 0:
       if (IsModrmRegister(m->xedd)) {
         OpMovhlpsVqUq(m);
@@ -1159,7 +1159,7 @@ static void OpMov0f12(struct Machine *m) {
 }
 
 static void OpMov0f13(struct Machine *m) {
-  if (Prefix66(m->xedd)) {
+  if (Osz(m->xedd)) {
     OpMovlpdMqVq(m);
   } else {
     OpMovlpsMqVq(m);
@@ -1167,7 +1167,7 @@ static void OpMov0f13(struct Machine *m) {
 }
 
 static void OpMov0f16(struct Machine *m) {
-  switch (m->xedd->op.rep | Prefix66(m->xedd)) {
+  switch (Rep(m->xedd) | Osz(m->xedd)) {
     case 0:
       if (IsModrmRegister(m->xedd)) {
         OpMovlhpsVqUq(m);
@@ -1188,7 +1188,7 @@ static void OpMov0f16(struct Machine *m) {
 }
 
 static void OpMov0f17(struct Machine *m) {
-  if (Prefix66(m->xedd)) {
+  if (Osz(m->xedd)) {
     OpMovhpdMqVq(m);
   } else {
     OpMovhpsMqVq(m);
@@ -1200,7 +1200,7 @@ static void OpMov0fD6(struct Machine *m) {
     OpMovq2dqVdqNq(m);
   } else if (m->xedd->op.ild_f2) {
     OpMovdq2qPqUq(m);
-  } else if (Prefix66(m->xedd)) {
+  } else if (Osz(m->xedd)) {
     OpMovqWqVq(m);
   } else {
     OpUd(m);
@@ -1211,7 +1211,7 @@ static void OpUnpcklpsd(struct Machine *m) {
   uint8_t *a, *b;
   a = XmmRexrReg(m);
   b = GetModrmRegisterXmmPointerRead8(m);
-  if (Prefix66(m->xedd)) {
+  if (Osz(m->xedd)) {
     memcpy(a + 8, b, 8);
   } else {
     memcpy(a + 4 * 3, b + 4, 4);
@@ -1224,7 +1224,7 @@ static void OpUnpckhpsd(struct Machine *m) {
   uint8_t *a, *b;
   a = XmmRexrReg(m);
   b = GetModrmRegisterXmmPointerRead16(m);
-  if (Prefix66(m->xedd)) {
+  if (Osz(m->xedd)) {
     memcpy(a + 0, b + 8, 8);
     memcpy(a + 8, b + 8, 8);
   } else {
@@ -1238,14 +1238,14 @@ static void OpUnpckhpsd(struct Machine *m) {
 static void OpPextrwGdqpUdqIb(struct Machine *m) {
   uint8_t i;
   i = m->xedd->op.uimm0;
-  i &= Prefix66(m->xedd) ? 7 : 3;
+  i &= Osz(m->xedd) ? 7 : 3;
   Write16(RegRexrReg(m), Read16(XmmRexbRm(m) + i * 2));
 }
 
 static void OpPinsrwVdqEwIb(struct Machine *m) {
   uint8_t i;
   i = m->xedd->op.uimm0;
-  i &= Prefix66(m->xedd) ? 7 : 3;
+  i &= Osz(m->xedd) ? 7 : 3;
   Write16(XmmRexrReg(m) + i * 2, Read16(GetModrmRegisterWordPointerRead2(m)));
 }
 
@@ -1253,7 +1253,7 @@ static void OpShuffle(struct Machine *m) {
   int16_t q16[4];
   int16_t x16[8];
   int32_t x32[4];
-  switch (m->xedd->op.rep | Prefix66(m->xedd)) {
+  switch (Rep(m->xedd) | Osz(m->xedd)) {
     case 0:
       memcpy(q16, GetModrmRegisterXmmPointerRead8(m), 8);
       (pshufw)(q16, q16, m->xedd->op.uimm0);
@@ -1282,7 +1282,7 @@ static void OpShuffle(struct Machine *m) {
 static void OpShufpsd(struct Machine *m) {
   float s[4];
   double d[2];
-  if (Prefix66(m->xedd)) {
+  if (Osz(m->xedd)) {
     memcpy(d, GetModrmRegisterXmmPointerRead16(m), 16);
     (shufpd)(d, d, m->xedd->op.uimm0);
     memcpy(XmmRexrReg(m), d, 16);
@@ -1297,7 +1297,7 @@ static void OpSqrtpsd(struct Machine *m) {
   long i;
   float_v xf;
   double_v xd;
-  switch (m->xedd->op.rep | Prefix66(m->xedd)) {
+  switch (Rep(m->xedd) | Osz(m->xedd)) {
     case 0:
       memcpy(&xf, GetModrmRegisterXmmPointerRead16(m), 16);
       for (i = 0; i < 4; ++i) xf[i] = sqrtf(xf[i]);
@@ -1375,13 +1375,13 @@ static void OpVpsdWpsd(struct Machine *m,
 static void OpVpsdWpsd66(struct Machine *m,
                          float_v opf(struct Machine *, float_v, float_v),
                          double_v opd(struct Machine *, double_v, double_v)) {
-  OpVpsdWpsd(m, opf, opd, !Prefix66(m->xedd), Prefix66(m->xedd));
+  OpVpsdWpsd(m, opf, opd, !Osz(m->xedd), Osz(m->xedd));
 }
 
 static void OpVpsdWpsd66f2(struct Machine *m,
                            float_v opf(struct Machine *, float_v, float_v),
                            double_v opd(struct Machine *, double_v, double_v)) {
-  OpVpsdWpsd(m, opf, opd, m->xedd->op.ild_f2, Prefix66(m->xedd));
+  OpVpsdWpsd(m, opf, opd, m->xedd->op.ild_f2, Osz(m->xedd));
 }
 
 static void OpVspsdWspsd(struct Machine *m,
@@ -1389,7 +1389,7 @@ static void OpVspsdWspsd(struct Machine *m,
                          double_v opd(struct Machine *, double_v, double_v)) {
   float_v xf, yf;
   double_v xd, yd;
-  switch (m->xedd->op.rep | Prefix66(m->xedd)) {
+  switch (Rep(m->xedd) | Osz(m->xedd)) {
     case 0:
       memcpy(&yf, GetModrmRegisterXmmPointerRead16(m), 16);
       memcpy(&xf, XmmRexrReg(m), 16);
@@ -1423,7 +1423,7 @@ static void OpComissVsWs(struct Machine *m) {
   float xf, yf;
   double xd, yd;
   uint8_t zf, cf, pf, ie;
-  if (!Prefix66(m->xedd)) {
+  if (!Osz(m->xedd)) {
     memcpy(&xf, XmmRexrReg(m), 4);
     memcpy(&yf, GetModrmRegisterXmmPointerRead4(m), 4);
     if (!isnan(xf) && !isnan(yf)) {
@@ -2366,6 +2366,7 @@ static void SaveStash(struct Machine *m) {
 }
 
 void ExecuteInstruction(struct Machine *m) {
+  uint8_t *p;
   m->ip += m->xedd->length;
   switch (m->xedd->op.map) {
     CASE(XED_ILD_MAP0, ExecuteInstructionMap0(m));

tool/build/lib/machine.h

@@ -1,6 +1,7 @@
 #ifndef COSMOPOLITAN_TOOL_BUILD_LIB_MACHINE_H_
 #define COSMOPOLITAN_TOOL_BUILD_LIB_MACHINE_H_
 #include "libc/elf/struct/ehdr.h"
+#include "libc/runtime/runtime.h"
 #include "third_party/xed/x86.h"
 #include "tool/build/lib/pml4t.h"
 

tool/build/lib/modrm.h

@@ -27,7 +27,7 @@ COSMOPOLITAN_C_START_
 #define Rexb(x)             (((x)->op.rde & 000010000000) >> 21)
 #define Rex(x)              (((x)->op.rde & 000000000020) >> 4)
 #define Osz(x)              (((x)->op.rde & 000000000040) >> 5)
-#define Prefix66(x)         (((x)->op.rde & 010000000000) >> 30)
+#define Rep(x)              (((x)->op.rde & 030000000000) >> 30)
 #define ByteRexrReg(m)      m->beg[(m->xedd->op.rde & 00000000037) >> 0]
 #define ByteRexbRm(m)       m->beg[(m->xedd->op.rde & 00000003700) >> 6]
 #define ByteRexbSrm(m)      m->beg[(m->xedd->op.rde & 00000370000) >> 12]

tool/build/lib/sse.c

@@ -119,7 +119,7 @@ void OpSse(struct Machine *m, enum OpSseKernel kernel) {
   uint8_t *p;
   union MachineVector x, y, t;
   p = GetModrmRegisterXmmPointerRead16(m);
-  if (Prefix66(m->xedd)) {
+  if (Osz(m->xedd)) {
     memcpy(&y, p, 16);
   } else {
     memset(&t, 0, 16);
@@ -204,7 +204,7 @@ void OpSse(struct Machine *m, enum OpSseKernel kernel) {
     default:
       unreachable;
   }
-  if (Prefix66(m->xedd)) {
+  if (Osz(m->xedd)) {
     memcpy(XmmRexrReg(m), &x, 16);
   } else {
     memcpy(XmmRexrReg(m), &x, 8);
@@ -230,7 +230,7 @@ void OpSseUdqIb(struct Machine *m, enum OpSseUdqIbKernel kernel) {
     default:
       unreachable;
   }
-  if (Prefix66(m->xedd)) {
+  if (Osz(m->xedd)) {
     memcpy(XmmRexbRm(m), &x, 16);
   } else {
     memcpy(XmmRexbRm(m), &x, 8);
@@ -246,7 +246,7 @@ static void OpSsePalignrMmx(struct Machine *m) {
 }
 
 void OpSsePalignr(struct Machine *m) {
-  if (Prefix66(m->xedd)) {
+  if (Osz(m->xedd)) {
     palignr(XmmRexrReg(m), XmmRexrReg(m), GetModrmRegisterXmmPointerRead8(m),
             m->xedd->op.uimm0);
   } else {

tool/build/lib/string.c

@@ -78,7 +78,7 @@ void OpString(struct Machine *m, int op) {
   lg2 = RegLog2(m->xedd);
   n = 1 << lg2;
   for (;;) {
-    if (m->xedd->op.rep && !Read64(m->cx)) break;
+    if (Rep(m->xedd) && !Read64(m->cx)) break;
     v = 0;
     *p = NULL;
     compare = false;
@@ -125,11 +125,11 @@ void OpString(struct Machine *m, int op) {
         abort();
     }
     EndStore(m, v, n, p, s[0]);
-    if (!m->xedd->op.rep) break;
+    if (!Rep(m->xedd)) break;
     Write64(m->cx, Read64(m->cx) - 1);
     if (compare) {
-      if (m->xedd->op.rep == 2 && GetFlag(m->flags, FLAGS_ZF)) break;
+      if (Rep(m->xedd) == 2 && GetFlag(m->flags, FLAGS_ZF)) break;
-      if (m->xedd->op.rep == 3 && !GetFlag(m->flags, FLAGS_ZF)) break;
+      if (Rep(m->xedd) == 3 && !GetFlag(m->flags, FLAGS_ZF)) break;
     }
   }
 }
@@ -211,7 +211,7 @@ void OpRepStosbEnhanced(struct Machine *m) {
 }
 
 void OpMovsb(struct Machine *m) {
-  if (m->xedd->op.rep && !GetFlag(m->flags, FLAGS_DF)) {
+  if (Rep(m->xedd) && !GetFlag(m->flags, FLAGS_DF)) {
     OpRepMovsbEnhanced(m);
   } else {
     OpString(m, STRING_MOVS);
@@ -219,7 +219,7 @@ void OpMovsb(struct Machine *m) {
 }
 
 void OpStosb(struct Machine *m) {
-  if (m->xedd->op.rep && !GetFlag(m->flags, FLAGS_DF)) {
+  if (Rep(m->xedd) && !GetFlag(m->flags, FLAGS_DF)) {
     OpRepStosbEnhanced(m);
   } else {
     OpString(m, STRING_STOS);

tool/build/lib/throw.c

@@ -19,6 +19,7 @@
 ╚─────────────────────────────────────────────────────────────────────────────*/
 #include "libc/log/check.h"
 #include "libc/runtime/runtime.h"
+#include "libc/str/str.h"
 #include "tool/build/lib/throw.h"
 
 static bool IsHaltingInitialized(struct Machine *m) {

tool/decode/x86opinfo.c

@@ -189,7 +189,6 @@ int main(int argc, char *argv[]) {
   SHOWOP(amd3dnow);
   SHOWOP(bcrc);
   SHOWOP(disp_width);
-  SHOWOP(first_f2f3);
   SHOWOP(hint);
   SHOWOP(ild_f2);
   SHOWOP(ild_f3);
@@ -197,10 +196,8 @@ int main(int argc, char *argv[]) {
   SHOWOP(imm1_bytes);
   SHOWOP(imm_width);
   SHOWOP(imm_signed);
-  SHOWOP(last_f2f3);
   SHOWOP(llrc);
   SHOWOP(mask);
-  SHOWOP(prefix66);
   SHOWOP(max_bytes);
   SHOWOP(mode_first_prefix);
   SHOWOP(nprefixes);