Implement PSLLQ

saiminhtet · May 31, 2020 · 03fd5bc · 03fd5bc
1 parent 45f3b38
commit 03fd5bc
Show file tree

Hide file tree

Showing 7 changed files with 60 additions and 1 deletion.
diff --git a/emu/decode.h b/emu/decode.h
@@ -283,7 +283,15 @@ __no_instrument DECODER_RET glue(DECODER_NAME, OP_SIZE)(DECODER_ARGS) {
                 case 0x73: READMODRM;
                            switch (modrm.opcode) {
                                case 0x02: TRACEI("psrlq imm, xmm");
-                                          READIMM8; VSHIFTR_IMM(imm, xmm_modrm_reg,64); break;
+                                          READIMM8;
+                                          VSHIFTR_IMM(imm, xmm_modrm_reg,64);
+                                          break;
+
+                               case 0x06: TRACEI("psllq imm, xmm");
+                                          READIMM8;
+                                          VSHIFTL_IMM(imm, xmm_modrm_reg,64);
+                                          break;
+
                                default: UNDEFINED;
                            }
                            break;

diff --git a/emu/vec.c b/emu/vec.c
@@ -58,6 +58,15 @@ void vec_merge128(NO_CPU, const void *src, void *dst) {
     memcpy(dst, src, 16);
 }
 
+void vec_imm_shiftl64(NO_CPU, const uint8_t amount, union xmm_reg *dst) {
+    if (amount > 63) {
+        zero_xmm(dst);
+    } else {
+        dst->qw[0] <<= amount;
+        dst->qw[1] <<= amount;
+    }
+}
+
 void vec_imm_shiftr64(NO_CPU, const uint8_t amount, union xmm_reg *dst) {
     if (amount > 63) {
         zero_xmm(dst);

diff --git a/emu/vec.h b/emu/vec.h
@@ -18,6 +18,7 @@ void vec_merge32(NO_CPU, const void *src, void *dst);
 void vec_merge64(NO_CPU, const void *src, void *dst);
 void vec_merge128(NO_CPU, const void *src, void *dst);
 
+void vec_imm_shiftl64(NO_CPU, const uint8_t amount, union xmm_reg *dst);
 void vec_imm_shiftr64(NO_CPU, const uint8_t amount, union xmm_reg *dst);
 void vec_xor128(NO_CPU, union xmm_reg *src, union xmm_reg *dst);
 

diff --git a/jit/gen.c b/jit/gen.c
@@ -531,6 +531,7 @@ static inline bool gen_vec(enum arg src, enum arg dst, void (*helper)(), gadget_
     }
 
 #define VCOMPARE(src, dst,z) v(compare, src, dst,z)
+#define VSHIFTL_IMM(src, dst, z) v(imm_shiftl, src, dst,z)
 #define VSHIFTR_IMM(src, dst, z) v(imm_shiftr, src, dst,z)
 #define V_OP(op, src, dst, z) v(op, src, dst, z)
 #define V_OP_IMM(op, src, dst, z) v_imm(op, src, dst, z)

diff --git a/tests/e2e/sse2/expected.txt b/tests/e2e/sse2/expected.txt
@@ -17,6 +17,10 @@ xorps
 7.05E-37 5.31E-37 1.46E-38 1.18E-38
 1.11E+01 2.22E+01 3.33E+01 4.44E+01
 0.00E+00 0.00E+00 0.00E+00 0.00E+00
+psllq
+01234 05678
+02468 11356
+00000 00000
 psrlq
 01234 05678
 00617 02839

diff --git a/tests/e2e/sse2/psllq.c b/tests/e2e/sse2/psllq.c
@@ -0,0 +1,32 @@
+#include <stdint.h>
+#include <stdio.h>
+#include <emmintrin.h>
+#include <xmmintrin.h>
+
+#define printout() printf("%05lld %05lld\n", (long long) out[0], (long long) out[1])
+
+void main(void) {
+	int64_t out[2] = { 0, 0 };
+	int64_t buf1234[2] = { 1234, 5678 };
+
+	// xmm1 Initially 1234
+	__m128i xmm1 = _mm_load_si128((__m128i*) buf1234);
+	_mm_store_si128((__m128i*) out, xmm1);
+	printout();
+
+	// xmm2 is just multiply by 2
+	asm volatile(	"psllq $1, %[vec1]\n\t"
+		: [vec1] "+x" (xmm1)
+		:);
+	_mm_store_si128((__m128i*) out, xmm1);
+	printout();
+
+	// xmm1 should be cleared
+	asm volatile(	"psllq $66, %[vec1]\n\t"
+		: [vec1] "+x" (xmm1)
+		:);
+	//shift_right(&xmm1, 66);
+	_mm_store_si128((__m128i*) out, xmm1);
+	printout();
+}
+
diff --git a/tests/e2e/sse2/test.sh b/tests/e2e/sse2/test.sh
@@ -11,6 +11,10 @@ echo xorps
 gcc -msse2 xorps.c -o test_xorps
 ./test_xorps
 
+echo psllq
+gcc -msse2 psllq.c -o test_psllq
+./test_psllq
+
 echo psrlq
 gcc -msse2 psrlq.c -o test_psrlq
 ./test_psrlq