[AMDGPU] Check for denormal flushing when selecting V_FMA/MAD_MIX*

[matejam]

Summary

Tests that were affected by this change did not have denormal flushing enabled.

[matejam]

I could add a function attribute containing denormal flushing or add more run lines to these tests.

[matejam]

llvm/test/CodeGen/AMDGPU/fdot2.ll

138 ↗

(On Diff #538554)

I will remove this change.

[arsenm]

In D154824#4484524, @matejam wrote:

I could add a function attribute containing denormal flushing or add more run lines to these tests.

Ideally we would have checks for at least ieee and preserve-sign modes for everything touching mad selection

[arsenm]

May need a rebase after fbe4ff8149a2b656a66cc0c64a38a5302c051c6e

[glehmann]

Why do you need to denorm flushing enabled to use v_fma_mix? I know that v_mad_mix always flushes denorms regardless of MODE register settings, but as far as I understand v_fma_mix doesn't have that issue.

[arsenm]

In D154824#4490827, @glehmann wrote:

Why do you need to denorm flushing enabled to use v_fma_mix? I know that v_mad_mix always flushes denorms regardless of MODE register settings, but as far as I understand v_fma_mix doesn't have that issue.

Right this is only a problem with v_mad* v_mad_mix*

[matejam]

Since it doesn't affect any tests, should I add this line in D153544?

[foad]

In D154824#4496498, @matejam wrote:

Since it doesn't affect any tests, should I add this line in D153544?

It would be better to add some tests here (or maybe just add some RUN lines to existing tests) that do show the effect.

llvm/lib/Target/AMDGPU/VOP3PInstructions.td

251

Update this comment.

[matejam]

Added tests without denormal flushing as a function attribute.
In all cases except GFX900 it should have the same result as with denormal flushing.

[matejam]

llvm/test/CodeGen/AMDGPU/mad-mix-lo.ll

13 ↗

(On Diff #539916)

I wasn't sure should I add all the RUN lines with and without denormal flushing.
It would have 20 RUN lines, I'm not sure is that too much?

[arsenm]

llvm/test/CodeGen/AMDGPU/mad-mix-lo.ll

59 ↗

(On Diff #539916)

typo siml. Also "s/no_flush/ieee"?

llvm/lib/Target/AMDGPU/VOP3PInstructions.td

(v2f16 (mixhi_inst $hi_src0_modifiers, $hi_src0,
                        (mixlo_inst $lo_src0_modifiers, $lo_src0,
                                    $lo_src1_modifiers, $lo_src1,
                                    $lo_src2_modifiers, $lo_src2,
                                    DSTCLAMP.ENABLE,
                                    (i32 (IMPLICIT_DEF)))))
   >;
 }
 
-let SubtargetPredicate = HasMadMixInsts in {
+let SubtargetPredicate = HasMadMixInsts, OtherPredicates = [NoFP32Denormals] in {
 
 // These are VOP3a-like opcodes which accept no omod.
 // Size of src arguments (16/32) is controlled by op_sel.
 // For 16-bit src arguments their location (hi/lo) are controlled by op_sel_hi.
 let isCommutable = 1, mayRaiseFPException = 0 in {
 let isReMaterializable = 1 in
 defm V_MAD_MIX_F32 : VOP3_VOP3PInst<"v_mad_mix_f32", VOP3P_Mix_Profile<VOP_F32_F16_F16_F16, VOP3_OPSEL>>;
 
 let FPDPRounding = 1 in {
 // Clamp modifier is applied after conversion to f16.
 defm V_MAD_MIXLO_F16 : VOP3_VOP3PInst<"v_mad_mixlo_f16", VOP3P_Mix_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL, 1>>;
 
 let ClampLo = 0, ClampHi = 1 in {
 defm V_MAD_MIXHI_F16 : VOP3_VOP3PInst<"v_mad_mixhi_f16", VOP3P_Mix_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL, 1>>;
 }
 } // End FPDPRounding = 1
 }
 
 defm : MadFmaMixPats<fmad, V_MAD_MIX_F32, V_MAD_MIXLO_F16, V_MAD_MIXHI_F16>;
 } // End SubtargetPredicate = HasMadMixInsts
 
 
 // Essentially the same as the mad_mix versions
 let SubtargetPredicate = HasFmaMixInsts in {
 let isCommutable = 1 in {
 
 let isReMaterializable = 1 in
 defm V_FMA_MIX_F32 : VOP3_VOP3PInst<"v_fma_mix_f32", VOP3P_Mix_Profile<VOP_F32_F16_F16_F16, VOP3_OPSEL>>;