Fix simd shuffle operation on x86

Signed-off-by: Zoltan Herczeg zherczeg.u-szeged@partner.samsung.com
Samsung · May 23, 2024 · 3341dfe · 3341dfe
1 parent 29a1cd0
commit 3341dfe
Show file tree

Hide file tree

Showing 2 changed files with 54 additions and 26 deletions.
diff --git a/src/jit/SimdX86Inl.h b/src/jit/SimdX86Inl.h
@@ -1828,44 +1828,62 @@ static void emitShuffleSIMD(sljit_compiler* compiler, Instruction* instr)
     Operand* operands = instr->operands();
     sljit_s32 tmp1 = instr->requiredReg(2);
     sljit_s32 tmp2 = SLJIT_TMP_DEST_FREG;
+    const sljit_s32 type = SLJIT_SIMD_REG_128 | SLJIT_SIMD_ELEM_8;
+    I8X16Shuffle* shuffle = reinterpret_cast<I8X16Shuffle*>(instr->byteCode());
     JITArg args[3];
 
-    simdOperandToArg(compiler, operands, args[0], SLJIT_SIMD_ELEM_128, instr->requiredReg(0));
-    simdOperandToArg(compiler, operands + 1, args[1], SLJIT_SIMD_ELEM_128, instr->requiredReg(1));
-
     args[2].set(operands + 2);
     sljit_s32 dst = GET_TARGET_REG(args[2].arg, instr->requiredReg(1));
 
-    I8X16Shuffle* shuffle = reinterpret_cast<I8X16Shuffle*>(instr->byteCode());
-    const sljit_s32 type = SLJIT_SIMD_REG_128 | SLJIT_SIMD_ELEM_8;
-    sljit_emit_simd_mov(compiler, SLJIT_SIMD_LOAD | type, tmp1, SLJIT_MEM0(), reinterpret_cast<sljit_sw>(shuffle->value()));
+    if (operands[0].ref == operands[1].ref) {
+        simdOperandToArg(compiler, operands, args[0], SLJIT_SIMD_ELEM_128, dst);
+        sljit_emit_simd_mov(compiler, SLJIT_SIMD_LOAD | type, SLJIT_TMP_DEST_FREG, SLJIT_MEM0(), reinterpret_cast<sljit_sw>(shuffle->value()));
 
-    sljit_emit_simd_replicate(compiler, SLJIT_SIMD_REG_128 | SLJIT_SIMD_ELEM_8, tmp2, SLJIT_IMM, 0xf0);
-    simdEmitSSEOp(compiler, SimdOp::paddb, tmp1, tmp2);
+        if (args[0].arg != dst) {
+            if (sljit_has_cpu_feature(SLJIT_HAS_AVX)) {
+                simdEmitVexOp(compiler, SimdOp::pshufb, dst, args[1].arg, tmp1);
+            } else {
+                sljit_emit_simd_mov(compiler, SLJIT_SIMD_LOAD | type, dst, args[0].arg, 0);
+                args[1].arg = dst;
+            }
+        }
 
-    if (dst != args[1].arg) {
-        if (sljit_has_cpu_feature(SLJIT_HAS_AVX)) {
-            simdEmitVexOp(compiler, SimdOp::pshufb, dst, args[1].arg, tmp1);
-        } else {
-            sljit_emit_simd_mov(compiler, SLJIT_SIMD_LOAD | type, dst, args[1].arg, 0);
-            args[1].arg = dst;
+        if (dst == args[0].arg) {
+            simdEmitSSEOp(compiler, SimdOp::pshufb, dst, SLJIT_TMP_DEST_FREG);
         }
-    }
+    } else {
+        simdOperandToArg(compiler, operands, args[0], SLJIT_SIMD_ELEM_128, instr->requiredReg(0));
+        simdOperandToArg(compiler, operands + 1, args[1], SLJIT_SIMD_ELEM_128, instr->requiredReg(1));
 
-    if (dst == args[1].arg) {
-        simdEmitSSEOp(compiler, SimdOp::pshufb, dst, tmp1);
-    }
+        sljit_emit_simd_mov(compiler, SLJIT_SIMD_LOAD | type, tmp1, SLJIT_MEM0(), reinterpret_cast<sljit_sw>(shuffle->value()));
 
-    simdEmitSSEOp(compiler, SimdOp::pxor, tmp1, tmp2);
+        sljit_emit_simd_replicate(compiler, SLJIT_SIMD_REG_128 | SLJIT_SIMD_ELEM_8, tmp2, SLJIT_IMM, 0xf0);
+        simdEmitSSEOp(compiler, SimdOp::paddb, tmp1, tmp2);
 
-    if (sljit_has_cpu_feature(SLJIT_HAS_AVX)) {
-        simdEmitVexOp(compiler, SimdOp::pshufb, tmp2, args[0].arg, tmp1);
-    } else {
-        sljit_emit_simd_mov(compiler, SLJIT_SIMD_LOAD | type, tmp2, args[0].arg, 0);
-        simdEmitSSEOp(compiler, SimdOp::pshufb, tmp2, tmp1);
-    }
+        if (dst != args[1].arg) {
+            if (sljit_has_cpu_feature(SLJIT_HAS_AVX)) {
+                simdEmitVexOp(compiler, SimdOp::pshufb, dst, args[1].arg, tmp1);
+            } else {
+                sljit_emit_simd_mov(compiler, SLJIT_SIMD_LOAD | type, dst, args[1].arg, 0);
+                args[1].arg = dst;
+            }
+        }
+
+        if (dst == args[1].arg) {
+            simdEmitSSEOp(compiler, SimdOp::pshufb, dst, tmp1);
+        }
+
+        simdEmitSSEOp(compiler, SimdOp::pxor, tmp1, tmp2);
 
-    simdEmitSSEOp(compiler, SimdOp::por, dst, tmp2);
+        if (sljit_has_cpu_feature(SLJIT_HAS_AVX)) {
+            simdEmitVexOp(compiler, SimdOp::pshufb, tmp2, args[0].arg, tmp1);
+        } else {
+            sljit_emit_simd_mov(compiler, SLJIT_SIMD_LOAD | type, tmp2, args[0].arg, 0);
+            simdEmitSSEOp(compiler, SimdOp::pshufb, tmp2, tmp1);
+        }
+
+        simdEmitSSEOp(compiler, SimdOp::por, dst, tmp2);
+    }
 
     if (SLJIT_IS_MEM(args[2].arg)) {
         sljit_emit_simd_mov(compiler, SLJIT_SIMD_STORE | type, dst, args[2].arg, args[2].argw);

diff --git a/test/jit/simd-shuffle.wast b/test/jit/simd-shuffle.wast
@@ -28,6 +28,12 @@
   v128.not
   local.set 2
 )
+
+(func (export "test3") (param v128) (result v128)
+  local.get 0
+  local.get 0
+  i8x16.shuffle 30 14 28 12 26 10 24 8 22 6 20 4 18 2 16 0
+)
 )
 
 (assert_return (invoke "test1"
@@ -39,3 +45,7 @@
    (v128.const i64x2 0x0807060504030201 0x100f0e0d0c0b0a09)
    (v128.const i64x2 0x1817161514131211 0x201f1e1d1c1b1a19))
    (v128.const i64x2 0x19091b0b1d0d1f0f 0x1101130315051707))
+
+(assert_return (invoke "test3"
+   (v128.const i64x2 0x8182838485868788 0x2122232425262728))
+   (v128.const i64x2 0x2828262624242222 0x8888868684848282))