From d259668731f4a7d6f477cefff102fe1e0b86f461 Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault@amd.com>
Date: Tue, 26 May 2020 21:56:53 -0400
Subject: [PATCH] AMDGPU: Set mayRaiseFPException

This may be missing a few overrides to set it off still in some
special cases. Since the flags set during selection should now be
reliably preserved, this should not change codegen for non-strictfp
functions.
---
 llvm/lib/Target/AMDGPU/VOP1Instructions.td |  3 +--
 llvm/lib/Target/AMDGPU/VOP2Instructions.td |  9 +++++++--
 llvm/lib/Target/AMDGPU/VOP3Instructions.td |  1 +
 llvm/lib/Target/AMDGPU/VOPInstructions.td  | 12 +++---------
 4 files changed, 12 insertions(+), 13 deletions(-)

diff --git a/llvm/lib/Target/AMDGPU/VOP1Instructions.td b/llvm/lib/Target/AMDGPU/VOP1Instructions.td
index d11e798..c83442a 100644
--- a/llvm/lib/Target/AMDGPU/VOP1Instructions.td
+++ b/llvm/lib/Target/AMDGPU/VOP1Instructions.td
@@ -50,8 +50,7 @@ class VOP1_Pseudo <string opName, VOPProfile P, list<dag> pattern=[], bit VOP1On
 
   let ReadsModeReg = !or(isFloatType<P.DstVT>.ret, isFloatType<P.Src0VT>.ret);
 
-  // FIXME
-  // let mayRaiseFPException = ReadsModeReg;
+  let mayRaiseFPException = ReadsModeReg;
 
   let VOP1 = 1;
   let VALU = 1;
diff --git a/llvm/lib/Target/AMDGPU/VOP2Instructions.td b/llvm/lib/Target/AMDGPU/VOP2Instructions.td
index 86dc179..74ae812 100644
--- a/llvm/lib/Target/AMDGPU/VOP2Instructions.td
+++ b/llvm/lib/Target/AMDGPU/VOP2Instructions.td
@@ -71,8 +71,7 @@ class VOP2_Pseudo <string opName, VOPProfile P, list<dag> pattern=[], string suf
 
   let ReadsModeReg = !or(isFloatType<P.DstVT>.ret, isFloatType<P.Src0VT>.ret);
 
-  // FIXME: Set this
-  // let mayRaiseFPException = ReadsModeReg;
+  let mayRaiseFPException = ReadsModeReg;
 
   let VOP2 = 1;
   let VALU = 1;
@@ -489,12 +488,14 @@ defm V_AND_B32 : VOP2Inst <"v_and_b32", VOP_PAT_GEN<VOP_I32_I32_I32>, and>;
 defm V_OR_B32 : VOP2Inst <"v_or_b32", VOP_PAT_GEN<VOP_I32_I32_I32>, or>;
 defm V_XOR_B32 : VOP2Inst <"v_xor_b32", VOP_PAT_GEN<VOP_I32_I32_I32>, xor>;
 
+let mayRaiseFPException = 0 in {
 let Constraints = "$vdst = $src2", DisableEncoding="$src2",
     isConvertibleToThreeAddress = 1 in {
 defm V_MAC_F32 : VOP2Inst <"v_mac_f32", VOP_MAC_F32>;
 }
 
 def V_MADAK_F32 : VOP2_Pseudo <"v_madak_f32", VOP_MADAK_F32, []>;
+}
 
 // No patterns so that the scalar instructions are always selected.
 // The scalar versions will be replaced with vector when needed later.
@@ -633,7 +634,11 @@ defm V_ADD_F16 : VOP2Inst <"v_add_f16", VOP_F16_F16_F16, fadd>;
 defm V_SUB_F16 : VOP2Inst <"v_sub_f16", VOP_F16_F16_F16, fsub>;
 defm V_SUBREV_F16 : VOP2Inst <"v_subrev_f16", VOP_F16_F16_F16, null_frag, "v_sub_f16">;
 defm V_MUL_F16 : VOP2Inst <"v_mul_f16", VOP_F16_F16_F16, fmul>;
+
+let mayRaiseFPException = 0 in {
 def V_MADAK_F16 : VOP2_Pseudo <"v_madak_f16", VOP_MADAK_F16, [], "">;
+}
+
 } // End FPDPRounding = 1
 defm V_ADD_U16 : VOP2Inst <"v_add_u16", VOP_I16_I16_I16_ARITH, add>;
 defm V_SUB_U16 : VOP2Inst <"v_sub_u16" , VOP_I16_I16_I16_ARITH, sub>;
diff --git a/llvm/lib/Target/AMDGPU/VOP3Instructions.td b/llvm/lib/Target/AMDGPU/VOP3Instructions.td
index 66a4e62..e9fb289 100644
--- a/llvm/lib/Target/AMDGPU/VOP3Instructions.td
+++ b/llvm/lib/Target/AMDGPU/VOP3Instructions.td
@@ -228,6 +228,7 @@ def VOP3b_I64_I1_I32_I32_I64 : VOPProfile<[i64, i32, i32, i64]> {
 class VOP3Interp<string OpName, VOPProfile P, list<dag> pattern = []> :
                  VOP3_Pseudo<OpName, P, pattern> {
   let AsmMatchConverter = "cvtVOP3Interp";
+  let mayRaiseFPException = 0;
 }
 
 def VOP3_INTERP : VOPProfile<[f32, f32, i32, untyped]> {
diff --git a/llvm/lib/Target/AMDGPU/VOPInstructions.td b/llvm/lib/Target/AMDGPU/VOPInstructions.td
index d52ad7f..f8a83e5 100644
--- a/llvm/lib/Target/AMDGPU/VOPInstructions.td
+++ b/llvm/lib/Target/AMDGPU/VOPInstructions.td
@@ -122,9 +122,7 @@ class VOP3_Pseudo <string opName, VOPProfile P, list<dag> pattern = [],
 
   let ReadsModeReg = !or(isFloatType<P.DstVT>.ret, isFloatType<P.Src0VT>.ret);
 
-  // FIXME: Set this. Right now it seems regular IR operations don't
-  // automatically imply no FP exceptions.
-  // let mayRaiseFPException = ReadsModeReg;
+  let mayRaiseFPException = ReadsModeReg;
   let Uses = !if(ReadsModeReg, [MODE, EXEC], [EXEC]);
 
   let AsmVariantName = AMDGPUAsmVariants.VOP3;
@@ -500,9 +498,7 @@ class VOP_SDWA_Pseudo <string opName, VOPProfile P, list<dag> pattern=[]> :
 
   let ReadsModeReg = !or(isFloatType<P.DstVT>.ret, isFloatType<P.Src0VT>.ret);
 
-  // FIXME: Set this. Right now it seems regular IR operations don't
-  // automatically imply no FP exceptions.
-  // let mayRaiseFPException = ReadsModeReg;
+  let mayRaiseFPException = ReadsModeReg;
   let Uses = !if(ReadsModeReg, [MODE, EXEC], [EXEC]);
 
   let SubtargetPredicate = HasSDWA;
@@ -623,9 +619,7 @@ class VOP_DPP_Pseudo <string OpName, VOPProfile P, list<dag> pattern=[]> :
 
   let ReadsModeReg = !or(isFloatType<P.DstVT>.ret, isFloatType<P.Src0VT>.ret);
 
-  // FIXME: Set this. Right now it seems regular IR operations don't
-  // automatically imply no FP exceptions.
-  // let mayRaiseFPException = ReadsModeReg;
+  let mayRaiseFPException = ReadsModeReg;
   let Uses = !if(ReadsModeReg, [MODE, EXEC], [EXEC]);
   let isConvergent = 1;
 
-- 
2.7.4