src/hotspot/cpu/aarch64/stubGenerator_aarch64.cpp

*** 263,273 ****
  #ifdef ASSERT
      // make sure we have no pending exceptions
      {
        Label L;
        __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset())));
!       __ cmp(rscratch1, (unsigned)NULL_WORD);
        __ br(Assembler::EQ, L);
        __ stop("StubRoutines::call_stub: entered with pending exception");
        __ BIND(L);
      }
  #endif
--- 263,273 ----
  #ifdef ASSERT
      // make sure we have no pending exceptions
      {
        Label L;
        __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset())));
!       __ cmp(rscratch1, (unsigned char)NULL_WORD);
        __ br(Assembler::EQ, L);
        __ stop("StubRoutines::call_stub: entered with pending exception");
        __ BIND(L);
      }
  #endif
*** 4104,4114 ****
      __ sub(cnt1, strU, cnt2, __ LSL, 1); // strU pointer to load from
  
      __ ldr(tmp3, Address(__ post(cnt1, 8)));
  
      if (SoftwarePrefetchHintDistance >= 0) {
!       __ cmp(cnt2, prefetchLoopExitCondition);
        __ br(__ LT, SMALL_LOOP);
        __ bind(LARGE_LOOP_PREFETCH);
          __ prfm(Address(tmp2, SoftwarePrefetchHintDistance));
          __ mov(tmp4, 2);
          __ prfm(Address(cnt1, SoftwarePrefetchHintDistance));
--- 4104,4114 ----
      __ sub(cnt1, strU, cnt2, __ LSL, 1); // strU pointer to load from
  
      __ ldr(tmp3, Address(__ post(cnt1, 8)));
  
      if (SoftwarePrefetchHintDistance >= 0) {
!       __ subs(rscratch2, cnt2, prefetchLoopExitCondition);
        __ br(__ LT, SMALL_LOOP);
        __ bind(LARGE_LOOP_PREFETCH);
          __ prfm(Address(tmp2, SoftwarePrefetchHintDistance));
          __ mov(tmp4, 2);
          __ prfm(Address(cnt1, SoftwarePrefetchHintDistance));
*** 4121,4131 ****
          __ bind(LARGE_LOOP_PREFETCH_REPEAT2);
            compare_string_16_x_LU(tmpL, tmpU, DIFF1, DIFF2);
            __ subs(tmp4, tmp4, 1);
            __ br(__ GT, LARGE_LOOP_PREFETCH_REPEAT2);
            __ sub(cnt2, cnt2, 64);
!           __ cmp(cnt2, prefetchLoopExitCondition);
            __ br(__ GE, LARGE_LOOP_PREFETCH);
      }
      __ cbz(cnt2, LOAD_LAST); // no characters left except last load
      __ subs(cnt2, cnt2, 16);
      __ br(__ LT, TAIL);
--- 4121,4131 ----
          __ bind(LARGE_LOOP_PREFETCH_REPEAT2);
            compare_string_16_x_LU(tmpL, tmpU, DIFF1, DIFF2);
            __ subs(tmp4, tmp4, 1);
            __ br(__ GT, LARGE_LOOP_PREFETCH_REPEAT2);
            __ sub(cnt2, cnt2, 64);
!           __ subs(rscratch2, cnt2, prefetchLoopExitCondition); 
            __ br(__ GE, LARGE_LOOP_PREFETCH);
      }
      __ cbz(cnt2, LOAD_LAST); // no characters left except last load
      __ subs(cnt2, cnt2, 16);
      __ br(__ LT, TAIL);
*** 4238,4248 ****
          __ prfm(Address(str2, SoftwarePrefetchHintDistance));
          compare_string_16_bytes_same(DIFF, DIFF2);
          compare_string_16_bytes_same(DIFF, DIFF2);
          __ sub(cnt2, cnt2, isLL ? 64 : 32);
          compare_string_16_bytes_same(DIFF, DIFF2);
!         __ cmp(cnt2, largeLoopExitCondition);
          compare_string_16_bytes_same(DIFF, DIFF2);
          __ br(__ GT, LARGE_LOOP_PREFETCH);
          __ cbz(cnt2, LAST_CHECK_AND_LENGTH_DIFF); // no more chars left?
          // less than 16 bytes left?
          __ subs(cnt2, cnt2, isLL ? 16 : 8);
--- 4238,4248 ----
          __ prfm(Address(str2, SoftwarePrefetchHintDistance));
          compare_string_16_bytes_same(DIFF, DIFF2);
          compare_string_16_bytes_same(DIFF, DIFF2);
          __ sub(cnt2, cnt2, isLL ? 64 : 32);
          compare_string_16_bytes_same(DIFF, DIFF2);
!         __ subs(rscratch2, cnt2, largeLoopExitCondition);
          compare_string_16_bytes_same(DIFF, DIFF2);
          __ br(__ GT, LARGE_LOOP_PREFETCH);
          __ cbz(cnt2, LAST_CHECK_AND_LENGTH_DIFF); // no more chars left?
          // less than 16 bytes left?
          __ subs(cnt2, cnt2, isLL ? 16 : 8);
*** 4657,4675 ****
      __ sub(octetCounter, octetCounter, 2);
      __ zip1(v1, __ T16B, v1, v0);
      __ zip1(v2, __ T16B, v2, v0);
      __ st1(v1, v2, __ T16B, __ post(dst, 32));
      __ ld1(v3, v4, v5, v6, __ T16B, Address(__ post(src, 64)));
!     __ cmp(octetCounter, large_loop_threshold);
      __ br(__ LE, LOOP_START);
      __ b(LOOP_PRFM_START);
      __ bind(LOOP_PRFM);
        __ ld1(v3, v4, v5, v6, __ T16B, Address(__ post(src, 64)));
      __ bind(LOOP_PRFM_START);
        __ prfm(Address(src, SoftwarePrefetchHintDistance));
        __ sub(octetCounter, octetCounter, 8);
!       __ cmp(octetCounter, large_loop_threshold);
        inflate_and_store_2_fp_registers(true, v3, v4);
        inflate_and_store_2_fp_registers(true, v5, v6);
        __ br(__ GT, LOOP_PRFM);
        __ cmp(octetCounter, 8);
        __ br(__ LT, DONE);
--- 4657,4675 ----
      __ sub(octetCounter, octetCounter, 2);
      __ zip1(v1, __ T16B, v1, v0);
      __ zip1(v2, __ T16B, v2, v0);
      __ st1(v1, v2, __ T16B, __ post(dst, 32));
      __ ld1(v3, v4, v5, v6, __ T16B, Address(__ post(src, 64)));
!     __ subs(rscratch1, octetCounter, large_loop_threshold);
      __ br(__ LE, LOOP_START);
      __ b(LOOP_PRFM_START);
      __ bind(LOOP_PRFM);
        __ ld1(v3, v4, v5, v6, __ T16B, Address(__ post(src, 64)));
      __ bind(LOOP_PRFM_START);
        __ prfm(Address(src, SoftwarePrefetchHintDistance));
        __ sub(octetCounter, octetCounter, 8);
!       __ subs(rscratch1, octetCounter, large_loop_threshold);
        inflate_and_store_2_fp_registers(true, v3, v4);
        inflate_and_store_2_fp_registers(true, v5, v6);
        __ br(__ GT, LOOP_PRFM);
        __ cmp(octetCounter, 8);
        __ br(__ LT, DONE);