[PATCH] D84903: [NFC][AMDGPU] Improve fused fmul+fadd tests.

Thu Jul 30 09:55:57 PDT 2020

rampitec added inline comments.

================
Comment at: llvm/test/CodeGen/AMDGPU/fmuladd.f16.ll:10
 ; RUN: llc -march=amdgcn -mcpu=gfx1010 -denormal-fp-math=ieee -denormal-fp-math-f32=ieee -fp-contract=on -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GCN-STRICT,GCN-DENORM,GCN-DENORM-STRICT,GFX10-DENORM-STRICT,GFX10-DENORM,GFX10 %s
 ; RUN: llc -march=amdgcn -mcpu=gfx1010 -denormal-fp-math=ieee -denormal-fp-math-f32=ieee -fp-contract=fast -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GCN-CONTRACT,GCN-DENORM,GCN-DENORM-CONTRACT,GFX10-DENORM-CONTRACT,GFX10-DENORM,GFX10 %s

----------------
arsenm wrote:
> dfukalov wrote:
> > arsenm wrote:
> > > This probably should also gain gfx1030 lines
> > My understanding was that tests RUN lines are actually represent a different targets in terms of features set. I guess, gfx1030 has no difference with gfx1010 in fused f16 operations capabilities/features. So gfx1030 will test the same code as gfx1010?
> > 
> > Otherwise, if the tests are to cover all modes/features, we'll have to add extra RUN lines for each gfx90x, gfx10xx and, probably, older targets.
> No, the v_mad/v_mac instructions were removed in gfx1030 (but gfx1010 does have them), so the cases with f16 denormals disabled will differ
That is only f32 mad/mac were removed.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D84903/new/

https://reviews.llvm.org/D84903