For an example:
...
#pragma omp target map(tofrom: counter_N0)
#pragma omp simd
for (int i = 0 ; i < 1 ; i++ )
{
#pragma omp atomic update
counter_N0 = counter_N0 + 1 ;
}
...
I noticed that the result of the atomic update (%r30) is propagated:
...
@%r33 atom.add.u32 _, [%r29], 1;
shfl.sync.idx.b32 %r30, %r30, %r32, 31, 0xffffffff;
...
even though it is unused (which is why the bit bucket operand _ is used).
Fix this by not emitting the shuffle in this case, such that we have instead:
...
@%r33 atom.add.u32 _, [%r29], 1;
bar.warp.sync 0xffffffff;
...
Tested on nvptx.
gcc/ChangeLog:
2022-03-07 Tom de Vries <tdevries@suse.de>
* config/nvptx/nvptx.cc (nvptx_unisimt_handle_set): Handle unused
result.
gcc/testsuite/ChangeLog:
2022-03-07 Tom de Vries <tdevries@suse.de>
* gcc.target/nvptx/uniform-simt-4.c: New test.
nvptx_unisimt_handle_set (rtx set, rtx_insn *insn, rtx master)
{
rtx reg;
- if (GET_CODE (set) == SET && REG_P (reg = SET_DEST (set)))
+ if (GET_CODE (set) == SET
+ && REG_P (reg = SET_DEST (set))
+ && find_reg_note (insn, REG_UNUSED, reg) == NULL_RTX)
{
emit_insn_after (nvptx_gen_shuffle (reg, reg, master, SHUFFLE_IDX),
insn);
--- /dev/null
+/* { dg-do compile } */
+/* { dg-options "-O2 -muniform-simt -mptx=_" } */
+
+enum memmodel
+{
+ MEMMODEL_RELAXED = 0
+};
+
+unsigned long long int *p64;
+unsigned long long int v64;
+
+int
+main()
+{
+ __atomic_fetch_add (p64, v64, MEMMODEL_RELAXED);
+
+ return 0;
+}
+
+/* { dg-final { scan-assembler-times "atom.add.u64\[\t \]+_," 1 } } */
+/* { dg-final { scan-assembler-times "bar.warp.sync" 1 } } */
+/* { dg-final { scan-assembler-not "shfl.sync.idx" } } */