[PATCH] D106758: [AMDGPU] We would need FP if there is call and caller save VGPR spills

[Matt Arsenault via Phabricator via llvm-commits]

arsenm added a comment.

LGTM except the test is using dead code. I also think we ought to be able to remove VGPRReservedForSGPRSpill now that SGPR and VGPR allocation are split



================
Comment at: llvm/lib/Target/AMDGPU/SIFrameLowering.cpp:1310
+  const bool WillHaveFP = FrameInfo.hasCalls() &&
+                          (HaveAnyCSRVGPR || (TRI->spillSGPRToVGPR() &&
+                                              MFI->VGPRReservedForSGPRSpill));
----------------
Checking TRI->spillSGPRToVGPR()  should be redundant


================
Comment at: llvm/test/CodeGen/AMDGPU/need-fp-from-caller-save-vgpr-spill.ll:17
+  %retval = alloca i32, align 4, addrspace(5)
+  %retval.ascast = addrspacecast i32 addrspace(5)* %retval to i32*
+  ret i32 0
----------------
The cast/alloca are dead


================
Comment at: llvm/test/CodeGen/AMDGPU/need-fp-from-caller-save-vgpr-spill.ll:51-52
+entry:
+  %retval = alloca i32, align 4, addrspace(5)
+  %retval.ascast = addrspacecast i32 addrspace(5)* %retval to i32*
+  %call = call i32 @baz()
----------------
The cast/alloca are dead


================
Comment at: llvm/test/CodeGen/AMDGPU/need-fp-from-caller-save-vgpr-spill.ll:90-91
+entry:
+  %retval = alloca i32, align 4, addrspace(5)
+  %retval.ascast = addrspacecast i32 addrspace(5)* %retval to i32*
+  %call = call i32 @bar()
----------------
The cast/alloca are dead


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D106758/new/

https://reviews.llvm.org/D106758