[PATCH] D30209: AMDGPU: Fold omod into instructions

[Artem Tamazov via Phabricator via llvm-commits]

artem.tamazov requested changes to this revision.
artem.tamazov added a comment.
This revision now requires changes to proceed.

Comments need to be updated at least.



================
Comment at: lib/Target/AMDGPU/SIFoldOperands.cpp:788-790
+// FIXME: Does this really not support denormals with f16?
+// FIXME: Does this need to check IEEE mode bit? SNaNs are generally not
+// handled, so will anything other than that break?
----------------
Output modifiers are not compatible with output denorms, i.e.:
- If output denorms are allowed (in the HW MODe register), then any output modifier is ignored
- If output denorms are not allowed, then denorms will be flushed to +/-0 first. Then, if output modifier is non-zero, -0 will be forced to +0 prior applying the modifier.
Output modifiers are not IEEE compliant (-0*x=+0). Output modifiers are ignored by hardware if ieee bit is set in the HW MODE register.

The above applies to all supported floating types, including f16, f32, f64.



================
Comment at: lib/Target/AMDGPU/SIFoldOperands.cpp:827
+  case AMDGPU::V_ADD_F16_e64: {
+    // XXX - Do denormals work with f16 omod?
+    if ((ST->hasFP32Denormals() && Op == AMDGPU::V_ADD_F32_e64) ||
----------------
Nope. Pls. see comment above.


================
Comment at: lib/Target/AMDGPU/SIFoldOperands.cpp:851
+
+// FIXME: Does this need to check IEEE bit on function?
+bool SIFoldOperands::tryFoldOMod(MachineInstr &MI) {
----------------
Yes. If IEEE is set, OMOD does not work.


https://reviews.llvm.org/D30209