[x86] Fix some asm problems

A bag of assembler fixes, to be either compatible with GAS
(e.g. order of 'test' operands), accept more instructions,
count correct foo{bwlq} variants on x86_64, fix modrm/sib bytes
on x86_64 to not use %rip relative addressing mode, to not use
invalid insns in tests/asmtest.S for x86_64.

Result is that now output of GAS and of tcc on tests/asmtest.S
is mostly the same.

1 files changed, 26 insertions, 15 deletions

diff --git a/x86_64-asm.h b/x86_64-asm.h
index cc2c3a5..d7c0d77 100644
--- a/x86_64-asm.h
+++ b/x86_64-asm.h
@@ -5,8 +5,8 @@
      DEF_ASM_OP0(cmc, 0xf5)
      DEF_ASM_OP0(lahf, 0x9f)
      DEF_ASM_OP0(sahf, 0x9e)
-     DEF_ASM_OP0(pushfl, 0x9c)
-     DEF_ASM_OP0(popfl, 0x9d)
+     DEF_ASM_OP0(pushfq, 0x9c)
+     DEF_ASM_OP0(popfq, 0x9d)
      DEF_ASM_OP0(pushf, 0x9c)
      DEF_ASM_OP0(popf, 0x9d)
      DEF_ASM_OP0(stc, 0xf9)
@@ -34,14 +34,15 @@
      DEF_ASM_OP0(hlt, 0xf4)
      DEF_ASM_OP0(wait, 0x9b)
      DEF_ASM_OP0(nop, 0x90)
+     DEF_ASM_OP0(pause, 0xf390)
      DEF_ASM_OP0(xlat, 0xd7)
 
      /* strings */
 ALT(DEF_ASM_OP0L(cmpsb, 0xa6, 0, OPC_BWLQ))
 ALT(DEF_ASM_OP0L(scmpb, 0xa6, 0, OPC_BWLQ))
 
-ALT(DEF_ASM_OP0L(insb, 0x6c, 0, OPC_BWLQ))
-ALT(DEF_ASM_OP0L(outsb, 0x6e, 0, OPC_BWLQ))
+ALT(DEF_ASM_OP0L(insb, 0x6c, 0, OPC_BWL))
+ALT(DEF_ASM_OP0L(outsb, 0x6e, 0, OPC_BWL))
 
 ALT(DEF_ASM_OP0L(lodsb, 0xac, 0, OPC_BWLQ))
 ALT(DEF_ASM_OP0L(slodb, 0xac, 0, OPC_BWLQ))
@@ -90,8 +91,10 @@ ALT(DEF_ASM_OP2(btcw, 0x0fba, 7, OPC_MODRM | OPC_WLQ, OPT_IM8, OPT_REGW | OPT_EA
      DEF_ASM_OP0(ud2, 0x0f0b)
 
      /* NOTE: we took the same order as gas opcode definition order */
+/* Right now we can't express the fact that 0xa1/0xa3 can't use $eax and a 
+   32 bit moffset as operands.
 ALT(DEF_ASM_OP2(movb, 0xa0, 0, OPC_BWLQ, OPT_ADDR, OPT_EAX))
-ALT(DEF_ASM_OP2(movb, 0xa2, 0, OPC_BWLQ, OPT_EAX, OPT_ADDR))
+ALT(DEF_ASM_OP2(movb, 0xa2, 0, OPC_BWLQ, OPT_EAX, OPT_ADDR)) */
 ALT(DEF_ASM_OP2(movb, 0x88, 0, OPC_MODRM | OPC_BWLQ, OPT_REG, OPT_EA | OPT_REG))
 ALT(DEF_ASM_OP2(movb, 0x8a, 0, OPC_MODRM | OPC_BWLQ, OPT_EA | OPT_REG, OPT_REG))
 ALT(DEF_ASM_OP2(movb, 0xb0, 0, OPC_REG | OPC_BWLQ, OPT_IM, OPT_REG))
@@ -102,25 +105,28 @@ ALT(DEF_ASM_OP2(movw, 0x8e, 0, OPC_MODRM | OPC_WLQ, OPT_EA | OPT_REG, OPT_SEG))
 
 ALT(DEF_ASM_OP2(movw, 0x0f20, 0, OPC_MODRM | OPC_WLQ, OPT_CR, OPT_REG64))
 ALT(DEF_ASM_OP2(movw, 0x0f21, 0, OPC_MODRM | OPC_WLQ, OPT_DB, OPT_REG64))
-ALT(DEF_ASM_OP2(movw, 0x0f24, 0, OPC_MODRM | OPC_WLQ, OPT_TR, OPT_REG64))
 ALT(DEF_ASM_OP2(movw, 0x0f22, 0, OPC_MODRM | OPC_WLQ, OPT_REG64, OPT_CR))
 ALT(DEF_ASM_OP2(movw, 0x0f23, 0, OPC_MODRM | OPC_WLQ, OPT_REG64, OPT_DB))
-ALT(DEF_ASM_OP2(movw, 0x0f26, 0, OPC_MODRM | OPC_WLQ, OPT_REG64, OPT_TR))
 
 ALT(DEF_ASM_OP2(movsbl, 0x0fbe, 0, OPC_MODRM, OPT_REG8 | OPT_EA, OPT_REG32))
 ALT(DEF_ASM_OP2(movsbw, 0x0fbe, 0, OPC_MODRM | OPC_D16, OPT_REG8 | OPT_EA, OPT_REG16))
 ALT(DEF_ASM_OP2(movswl, 0x0fbf, 0, OPC_MODRM, OPT_REG16 | OPT_EA, OPT_REG32))
 ALT(DEF_ASM_OP2(movslq, 0x4863, 0, OPC_MODRM, OPT_REG32 | OPT_EA, OPT_REG))
-ALT(DEF_ASM_OP2(movzbw, 0x0fb6, 0, OPC_MODRM | OPC_WL, OPT_REG8 | OPT_EA, OPT_REGW))
+ALT(DEF_ASM_OP2(movzbw, 0x0fb6, 0, OPC_MODRM | OPC_WLQ, OPT_REG8 | OPT_EA, OPT_REGW))
 ALT(DEF_ASM_OP2(movzwl, 0x0fb7, 0, OPC_MODRM, OPT_REG16 | OPT_EA, OPT_REG32))
 
+ALT(DEF_ASM_OP1(pushq, 0x6a, 0, 0, OPT_IM8S))
+ALT(DEF_ASM_OP1(push, 0x6a, 0, 0, OPT_IM8S))
+ALT(DEF_ASM_OP1(pushw, 0x6a, 0, OPC_D16, OPT_IM8S))
 ALT(DEF_ASM_OP1(pushw, 0x50, 0, OPC_REG | OPC_WLQ, OPT_REG64))
+ALT(DEF_ASM_OP1(pushw, 0x50, 0, OPC_REG | OPC_WLQ, OPT_REG16))
 ALT(DEF_ASM_OP1(pushw, 0xff, 6, OPC_MODRM | OPC_WLQ, OPT_REG64 | OPT_EA))
+ALT(DEF_ASM_OP1(pushw, 0x68, 0, OPC_D16, OPT_IM16))
 ALT(DEF_ASM_OP1(pushw, 0x68, 0, OPC_WLQ, OPT_IM32))
 ALT(DEF_ASM_OP1(pushw, 0x06, 0, OPC_WLQ, OPT_SEG))
-    DEF_ASM_OP1(pushb, 0x6a, 0, OPC_B, OPT_IM8S)
 
-ALT(DEF_ASM_OP1(popw, 0x58, 0, OPC_REG | OPC_WLQ, OPT_REGW))
+ALT(DEF_ASM_OP1(popw, 0x58, 0, OPC_REG | OPC_WLQ, OPT_REG64))
+ALT(DEF_ASM_OP1(popw, 0x58, 0, OPC_REG | OPC_WLQ, OPT_REG16))
 ALT(DEF_ASM_OP1(popw, 0x8f, 0, OPC_MODRM | OPC_WLQ, OPT_REGW | OPT_EA))
 ALT(DEF_ASM_OP1(popw, 0x07, 0, OPC_WLQ, OPT_SEG))
 
@@ -154,8 +160,8 @@ ALT(DEF_ASM_OP2(addb, 0x04, 0, OPC_ARITH | OPC_BWLQ, OPT_IMNO64, OPT_EAX))
 ALT(DEF_ASM_OP2(addb, 0x80, 0, OPC_ARITH | OPC_MODRM | OPC_BWLQ, OPT_IMNO64, OPT_EA | OPT_REG))
 ALT(DEF_ASM_OP2(addw, 0x83, 0, OPC_ARITH | OPC_MODRM | OPC_WLQ, OPT_IM8S, OPT_EA | OPT_REG))
 
-ALT(DEF_ASM_OP2(testb, 0x84, 0, OPC_MODRM | OPC_BWLQ, OPT_EA | OPT_REG, OPT_REG))
 ALT(DEF_ASM_OP2(testb, 0x84, 0, OPC_MODRM | OPC_BWLQ, OPT_REG, OPT_EA | OPT_REG))
+ALT(DEF_ASM_OP2(testb, 0x84, 0, OPC_MODRM | OPC_BWLQ, OPT_EA | OPT_REG, OPT_REG))
 ALT(DEF_ASM_OP2(testb, 0xa8, 0, OPC_BWLQ, OPT_IMNO64, OPT_EAX))
 ALT(DEF_ASM_OP2(testb, 0xf6, 0, OPC_MODRM | OPC_BWLQ, OPT_IMNO64, OPT_EA | OPT_REG))
 
@@ -206,7 +212,9 @@ ALT(DEF_ASM_OP1(setob, 0x0f90, 0, OPC_MODRM | OPC_TEST, OPT_REG8 | OPT_EA))
     DEF_ASM_OP2(enter, 0xc8, 0, 0, OPT_IM16, OPT_IM8)
     DEF_ASM_OP0(leave, 0xc9)
     DEF_ASM_OP0(ret, 0xc3)
+    DEF_ASM_OP0(retq, 0xc3)
 ALT(DEF_ASM_OP1(ret, 0xc2, 0, 0, OPT_IM16))
+ALT(DEF_ASM_OP1(retq, 0xc2, 0, 0, OPT_IM16))
     DEF_ASM_OP0(lret, 0xcb)
 ALT(DEF_ASM_OP1(lret, 0xca, 0, 0, OPT_IM16))
 
@@ -216,7 +224,7 @@ ALT(DEF_ASM_OP1(jo, 0x70, 0, OPC_SHORTJMP | OPC_JMP | OPC_TEST, OPT_ADDR))
     DEF_ASM_OP1(loope, 0xe1, 0, OPC_SHORTJMP, OPT_ADDR)
     DEF_ASM_OP1(loopz, 0xe1, 0, OPC_SHORTJMP, OPT_ADDR)
     DEF_ASM_OP1(loop, 0xe2, 0, OPC_SHORTJMP, OPT_ADDR)
-    DEF_ASM_OP1(jecxz, 0xe3, 0, OPC_SHORTJMP, OPT_ADDR)
+    DEF_ASM_OP1(jecxz, 0x67e3, 0, OPC_SHORTJMP, OPT_ADDR)
 
      /* float */
      /* specific fcomp handling */
@@ -224,6 +232,8 @@ ALT(DEF_ASM_OP0L(fcomp, 0xd8d9, 0, 0))
 
 ALT(DEF_ASM_OP1(fadd, 0xd8c0, 0, OPC_FARITH | OPC_REG, OPT_ST))
 ALT(DEF_ASM_OP2(fadd, 0xd8c0, 0, OPC_FARITH | OPC_REG, OPT_ST, OPT_ST0))
+ALT(DEF_ASM_OP2(fadd, 0xdcc0, 0, OPC_FARITH | OPC_REG, OPT_ST0, OPT_ST))
+ALT(DEF_ASM_OP2(fmul, 0xdcc8, 0, OPC_FARITH | OPC_REG, OPT_ST0, OPT_ST))
 ALT(DEF_ASM_OP0L(fadd, 0xdec1, 0, OPC_FARITH))
 ALT(DEF_ASM_OP1(faddp, 0xdec0, 0, OPC_FARITH | OPC_REG, OPT_ST))
 ALT(DEF_ASM_OP2(faddp, 0xdec0, 0, OPC_FARITH | OPC_REG, OPT_ST, OPT_ST0))
@@ -334,7 +344,7 @@ ALT(DEF_ASM_OP1(fstsw, 0xdd, 7, OPC_MODRM | OPC_FWAIT, OPT_EA ))
     DEF_ASM_OP1(lidt, 0x0f01, 3, OPC_MODRM, OPT_EA)
     DEF_ASM_OP1(lldt, 0x0f00, 2, OPC_MODRM, OPT_EA | OPT_REG)
     DEF_ASM_OP1(lmsw, 0x0f01, 6, OPC_MODRM, OPT_EA | OPT_REG)
-ALT(DEF_ASM_OP2(lslw, 0x0f03, 0, OPC_MODRM | OPC_WL, OPT_EA | OPT_REG, OPT_REG))
+ALT(DEF_ASM_OP2(lslw, 0x0f03, 0, OPC_MODRM | OPC_WLQ, OPT_EA | OPT_REG, OPT_REG))
     DEF_ASM_OP1(ltr, 0x0f00, 3, OPC_MODRM, OPT_EA | OPT_REG)
     DEF_ASM_OP1(sgdt, 0x0f01, 0, OPC_MODRM, OPT_EA)
     DEF_ASM_OP1(sidt, 0x0f01, 1, OPC_MODRM, OPT_EA)
@@ -346,8 +356,8 @@ ALT(DEF_ASM_OP2(lslw, 0x0f03, 0, OPC_MODRM | OPC_WL, OPT_EA | OPT_REG, OPT_REG))
 
     /* 486 */
     DEF_ASM_OP1(bswap, 0x0fc8, 0, OPC_REG, OPT_REG32 )
-ALT(DEF_ASM_OP2(xaddb, 0x0fc0, 0, OPC_MODRM | OPC_BWL, OPT_REG, OPT_REG | OPT_EA ))
-ALT(DEF_ASM_OP2(cmpxchgb, 0x0fb0, 0, OPC_MODRM | OPC_BWL, OPT_REG, OPT_REG | OPT_EA ))
+ALT(DEF_ASM_OP2(xaddb, 0x0fc0, 0, OPC_MODRM | OPC_BWLQ, OPT_REG, OPT_REG | OPT_EA ))
+ALT(DEF_ASM_OP2(cmpxchgb, 0x0fb0, 0, OPC_MODRM | OPC_BWLQ, OPT_REG, OPT_REG | OPT_EA ))
     DEF_ASM_OP1(invlpg, 0x0f01, 7, OPC_MODRM, OPT_EA )
 
     DEF_ASM_OP2(boundl, 0x62, 0, OPC_MODRM, OPT_REG32, OPT_EA)
@@ -383,6 +393,7 @@ ALT(DEF_ASM_OP2(cmova, 0x0f47, 0, OPC_MODRM | OPC_TEST, OPT_REG32 | OPT_EA, OPT_
     /* mmx */
     DEF_ASM_OP0(emms, 0x0f77) /* must be last OP0 */
     DEF_ASM_OP2(movd, 0x0f6e, 0, OPC_MODRM, OPT_EA | OPT_REG32, OPT_MMX )
+ALT(DEF_ASM_OP2(movd, 0x0f6e, 0, OPC_MODRM, OPT_EA | OPT_REG64, OPT_MMX ))
 ALT(DEF_ASM_OP2(movd, 0x0f7e, 0, OPC_MODRM, OPT_MMX, OPT_EA | OPT_REG32 ))
 ALT(DEF_ASM_OP2(movq, 0x0f6f, 0, OPC_MODRM, OPT_EA | OPT_MMX, OPT_MMX ))
 ALT(DEF_ASM_OP2(movq, 0x0f7f, 0, OPC_MODRM, OPT_MMX, OPT_EA | OPT_MMX ))