[llvm] AMDGPU: Break vop3p handling out of vop3 base patterns (PR #77472)
Matt Arsenault via llvm-commits
llvm-commits at lists.llvm.org
Tue Jan 9 05:41:08 PST 2024
https://github.com/arsenm created https://github.com/llvm/llvm-project/pull/77472
Add the vop3p op_sel fields in getInsVOP3P instead of getInsVOP3Base. Also start using defvar for some of the intermediate fields. let overrides of all the visible fields are really difficult to follow.
>From 10f12f86b5f2ea7cc742544b7e82163a9c342f65 Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault at amd.com>
Date: Tue, 9 Jan 2024 18:38:08 +0700
Subject: [PATCH] AMDGPU: Break vop3p handling out of vop3 base patterns
Add the vop3p op_sel fields in getInsVOP3P instead of
getInsVOP3Base. Also start using defvar for some of the
intermediate fields. let overrides of all the visible fields
are really difficult to follow.
---
llvm/lib/Target/AMDGPU/SIInstrInfo.td | 33 +++++++++++++---------
llvm/lib/Target/AMDGPU/VOP2Instructions.td | 2 +-
2 files changed, 20 insertions(+), 15 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.td b/llvm/lib/Target/AMDGPU/SIInstrInfo.td
index f07b8fa0ea4cd8..04c92155f5aada 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.td
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.td
@@ -1773,28 +1773,27 @@ class getIns64 <RegisterOperand Src0RC, RegisterOperand Src1RC,
class getInsVOP3Base<RegisterOperand Src0RC, RegisterOperand Src1RC,
RegisterOperand Src2RC, int NumSrcArgs,
bit HasClamp, bit HasModifiers, bit HasSrc2Mods, bit HasOMod,
- Operand Src0Mod, Operand Src1Mod, Operand Src2Mod, bit HasOpSel,
- bit IsVOP3P> {
+ Operand Src0Mod, Operand Src1Mod, Operand Src2Mod, bit HasOpSel> {
// getInst64 handles clamp and omod. implicit mutex between vop3p and omod
dag base = getIns64 <Src0RC, Src1RC, Src2RC, NumSrcArgs,
HasClamp, HasModifiers, HasSrc2Mods, HasOMod,
Src0Mod, Src1Mod, Src2Mod>.ret;
dag opsel = (ins op_sel0:$op_sel);
- dag vop3pOpsel = (ins op_sel_hi0:$op_sel_hi);
- dag vop3pFields = !con(!if(HasOpSel, vop3pOpsel, (ins)), (ins neg_lo0:$neg_lo, neg_hi0:$neg_hi));
-
- dag ret = !con(base,
- !if(HasOpSel, opsel,(ins)),
- !if(IsVOP3P, vop3pFields,(ins)));
+ dag ret = !con(base, !if(HasOpSel, opsel, (ins)));
}
class getInsVOP3P <RegisterOperand Src0RC, RegisterOperand Src1RC,
RegisterOperand Src2RC, int NumSrcArgs, bit HasClamp, bit HasOpSel,
Operand Src0Mod, Operand Src1Mod, Operand Src2Mod> {
- dag ret = getInsVOP3Base<Src0RC, Src1RC, Src2RC, NumSrcArgs,
+ dag base = getInsVOP3Base<Src0RC, Src1RC, Src2RC, NumSrcArgs,
HasClamp, 1/*HasModifiers*/, 1/*HasSrc2Mods*/,
- 0/*HasOMod*/, Src0Mod, Src1Mod, Src2Mod,
- HasOpSel, 1/*IsVOP3P*/>.ret;
+ 0/*HasOMod*/, Src0Mod, Src1Mod, Src2Mod, HasOpSel>.ret;
+
+ dag vop3pOpsel = (ins op_sel_hi0:$op_sel_hi);
+ dag vop3p_neg = (ins neg_lo0:$neg_lo, neg_hi0:$neg_hi);
+
+ dag vop3pFields = !con(!if(HasOpSel, vop3pOpsel, (ins)), vop3p_neg);
+ dag ret = !con(base, vop3pFields);
}
class getInsVOP3OpSel <RegisterOperand Src0RC, RegisterOperand Src1RC,
@@ -1804,7 +1803,7 @@ class getInsVOP3OpSel <RegisterOperand Src0RC, RegisterOperand Src1RC,
dag ret = getInsVOP3Base<Src0RC, Src1RC,
Src2RC, NumSrcArgs,
HasClamp, 1/*HasModifiers*/, 1/*HasSrc2Mods*/, HasOMod,
- Src0Mod, Src1Mod, Src2Mod, 1/*HasOpSel*/, 0>.ret;
+ Src0Mod, Src1Mod, Src2Mod, /*HasOpSel=*/1>.ret;
}
class getInsDPPBase <RegisterOperand OldRC, RegisterClass Src0RC, RegisterClass Src1RC,
@@ -2390,9 +2389,15 @@ class VOPProfile <list<ValueType> _ArgVT, bit _EnableClamp = 0> {
field dag InsDPP8 = getInsDPP8<DstRCDPP, Src0DPP, Src1DPP, Src2DPP,
NumSrcArgs, HasModifiers,
Src0ModDPP, Src1ModDPP, Src2ModDPP>.ret;
- field dag InsVOP3Base = getInsVOP3Base<Src0VOP3DPP, Src1VOP3DPP,
+ defvar InsVOP3DPPBase = getInsVOP3Base<Src0VOP3DPP, Src1VOP3DPP,
Src2VOP3DPP, NumSrcArgs, HasClamp, HasModifiers, HasSrc2Mods, HasOMod,
- Src0ModVOP3DPP, Src1ModVOP3DPP, Src2ModVOP3DPP, HasOpSel, IsVOP3P>.ret;
+ Src0ModVOP3DPP, Src1ModVOP3DPP, Src2ModVOP3DPP, HasOpSel>.ret;
+ defvar InsVOP3PDPPBase = getInsVOP3P<Src0VOP3DPP, Src1VOP3DPP,
+ Src2VOP3DPP, NumSrcArgs, HasClamp, HasOpSel,
+ Src0ModVOP3DPP, Src1ModVOP3DPP, Src2ModVOP3DPP>.ret;
+
+ field dag InsVOP3Base = !if(IsVOP3P, InsVOP3PDPPBase, InsVOP3DPPBase);
+
field dag InsVOP3DPP = getInsVOP3DPP<InsVOP3Base, DstRCVOP3DPP, NumSrcArgs>.ret;
field dag InsVOP3DPP16 = getInsVOP3DPP16<InsVOP3Base, DstRCVOP3DPP, NumSrcArgs>.ret;
field dag InsVOP3DPP8 = getInsVOP3DPP8<InsVOP3Base, DstRCVOP3DPP, NumSrcArgs>.ret;
diff --git a/llvm/lib/Target/AMDGPU/VOP2Instructions.td b/llvm/lib/Target/AMDGPU/VOP2Instructions.td
index ecee61daa1c877..3e66b5550cce2b 100644
--- a/llvm/lib/Target/AMDGPU/VOP2Instructions.td
+++ b/llvm/lib/Target/AMDGPU/VOP2Instructions.td
@@ -437,7 +437,7 @@ class VOP_MAC <ValueType vt0, ValueType vt1=vt0> : VOPProfile <[vt0, vt1, vt1, v
let InsDPP16 = !con(InsDPP, (ins FI:$fi));
let InsVOP3Base = getInsVOP3Base<Src0VOP3DPP, Src1VOP3DPP, RegisterOperand<VGPR_32>, 3,
0, HasModifiers, HasModifiers, HasOMod,
- Src0ModVOP3DPP, Src1ModVOP3DPP, Src2Mod, HasOpSel, 0/*IsVOP3P*/>.ret;
+ Src0ModVOP3DPP, Src1ModVOP3DPP, Src2Mod, HasOpSel>.ret;
// We need a dummy src2 tied to dst to track the use of that register for s_delay_alu
let InsVOPDX = (ins Src0RC32:$src0X, Src1RC32:$vsrc1X, VGPRSrc_32:$src2X);
let InsVOPDXDeferred =
More information about the llvm-commits
mailing list