[TLI][AMDGPU] AMDPAL does not have tan function

[tpr]

Summary

The LLPC frontend for AMDPAL OS type does not have a tan library
function, so disable instcombine generating it. We are only just coming
across this issue in LLPC as we start setting fast math flags.

Change-Id: I02f907d3e64832117ea9800e9f9285282856e5df

[arsenm]

lib/Analysis/TargetLibraryInfo.cpp

117–118 ↗

(On Diff #219558)

This isn't a frontend specific property. Nothing has a library in the sense this means

[tpr]

lib/Analysis/TargetLibraryInfo.cpp

117–118 ↗

(On Diff #219558)

Well, in other respects, we're treating "PAL" as an OS, even though it clearly isn't. And it's a property of a real OS what functions the library has, right?

What alternative suggestion do you have to fix it up so that other amdgcn users can do this sin/cos->tan optimization but PAL can't? Should our frontend manually adjust the TLI?

Or should we disable it for all amdgcn users, given that there isn't actually a tan instruction?

[arsenm]

lib/Analysis/TargetLibraryInfo.cpp

117–118 ↗

(On Diff #219558)

We should just really loop through and setUnavailable on all TLI functions for amdgcn

[tpr]

V2: Disable all library functions, not just tan.

[arsenm]

LGTM

llvm/trunk/lib/Analysis/TargetLibraryInfo.cpp

static void initialize(TargetLibraryInfoImpl &TLI, const Triple &T,
   if (T.isMIPS()) {
     ShouldSignExtI32Param = true;
   }
   TLI.setShouldExtI32Param(ShouldExtI32Param);
   TLI.setShouldExtI32Return(ShouldExtI32Return);
   TLI.setShouldSignExtI32Param(ShouldSignExtI32Param);
 
   if (T.getArch() == Triple::r600 ||
-      T.getArch() == Triple::amdgcn) {
-    TLI.setUnavailable(LibFunc_ldexp);
-    TLI.setUnavailable(LibFunc_ldexpf);
-    TLI.setUnavailable(LibFunc_ldexpl);
-    TLI.setUnavailable(LibFunc_exp10);
-    TLI.setUnavailable(LibFunc_exp10f);
-    TLI.setUnavailable(LibFunc_exp10l);
-    TLI.setUnavailable(LibFunc_log10);
-    TLI.setUnavailable(LibFunc_log10f);
-    TLI.setUnavailable(LibFunc_log10l);
-    TLI.setUnavailable(LibFunc_printf);
-  }
+      T.getArch() == Triple::amdgcn)
+    TLI.disableAllFunctions();
 
   // There are no library implementations of memcpy and memset for AMD gpus and
   // these can be difficult to lower in the backend.
   if (T.getArch() == Triple::r600 ||
       T.getArch() == Triple::amdgcn) {
     TLI.setUnavailable(LibFunc_memcpy);
     TLI.setUnavailable(LibFunc_memset);
     TLI.setUnavailable(LibFunc_memset_pattern16);

llvm/trunk/test/Transforms/InstCombine/AMDGPU/tan.ll

+; RUN: opt -mtriple=amdgcn--amdpal -S -instcombine <%s | FileCheck --check-prefixes=GCN %s
+
+; Check that sin/cos is not folded to tan on amdgcn.
+
+; GCN-LABEL: define amdgpu_ps float @llpc.shader.FS.main
+; GCN: call float @llvm.sin.f32
+; GCN: call float @llvm.cos.f32
+
+declare float @llvm.sin.f32(float) #0
+declare float @llvm.cos.f32(float) #0
+
+define amdgpu_ps float @llpc.shader.FS.main(float %arg) {
+.entry:
+  %tmp32 = call float @llvm.sin.f32(float %arg)
+  %tmp33 = call float @llvm.cos.f32(float %arg)
+  %tmp34 = fdiv reassoc nnan nsz arcp contract float 1.000000e+00, %tmp33
+  %tmp35 = fmul reassoc nnan nsz arcp contract float %tmp32, %tmp34
+  ret float %tmp35
+}
+
+attributes #0 = { nounwind readnone speculatable willreturn }