AMDGPU/SI: Fold operands with sub-registers

Summary:
Multi-dword constant loads generated unnecessary moves from SGPRs into VGPRs,
increasing the code size and VGPR pressure. These moves are now folded away.

Note that this lack of operand folding was not a problem for VMEM loads,
because COPY nodes from VReg_Nnn to VGPR32 are eliminated by the register
coalescer.

Some tests are updated, note that the fsub.ll test explicitly checks that
the move is elided.

With the IR generated by current Mesa, the changes are obviously relatively
minor:

7063 shaders in 3531 tests
Totals:
SGPRS: 351872 -> 352560 (0.20 %)
VGPRS: 199984 -> 200732 (0.37 %)
Code Size: 9876968 -> 9881112 (0.04 %) bytes
LDS: 91 -> 91 (0.00 %) blocks
Scratch: 1779712 -> 1767424 (-0.69 %) bytes per wave
Wait states: 295164 -> 295337 (0.06 %)

Totals from affected shaders:
SGPRS: 65784 -> 66472 (1.05 %)
VGPRS: 38064 -> 38812 (1.97 %)
Code Size: 1993828 -> 1997972 (0.21 %) bytes
LDS: 42 -> 42 (0.00 %) blocks
Scratch: 795648 -> 783360 (-1.54 %) bytes per wave
Wait states: 54026 -> 54199 (0.32 %)

Reviewers: tstellarAMD, arsenm, mareko

Subscribers: arsenm, llvm-commits

Differential Revision: http://reviews.llvm.org/D15875

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@257074 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/AMDGPU/SIFixSGPRCopies.cpp b/lib/Target/AMDGPU/SIFixSGPRCopies.cpp
index 96e37c56624084da42aea8fd3d9e5878a6314b38..f59d9948f98ed2e27789fb9b2e67a3dee67ac6a2 100644 (file)
--- a/lib/Target/AMDGPU/SIFixSGPRCopies.cpp
+++ b/lib/Target/AMDGPU/SIFixSGPRCopies.cpp
@@ -215,7 +215,7 @@ static bool foldVGPRCopyIntoRegSequence(MachineInstr &MI,
 
   for (unsigned I = 1, N = MI.getNumOperands(); I != N; I += 2) {
     unsigned SrcReg = MI.getOperand(I).getReg();
-    unsigned SrcSubReg = MI.getOperand(I).getReg();
+    unsigned SrcSubReg = MI.getOperand(I).getSubReg();
 
     const TargetRegisterClass *SrcRC = MRI.getRegClass(SrcReg);
     assert(TRI->isSGPRClass(SrcRC) &&

diff --git a/lib/Target/AMDGPU/SIFoldOperands.cpp b/lib/Target/AMDGPU/SIFoldOperands.cpp
index 02a39307e74e793c7cf6f9a4510f5940fd3d1b1e..ccbf7c80f2a8d33ca41954fb5a35aa5821b114ba 100644 (file)
--- a/lib/Target/AMDGPU/SIFoldOperands.cpp
+++ b/lib/Target/AMDGPU/SIFoldOperands.cpp
@@ -334,13 +334,10 @@ bool SIFoldOperands::runOnMachineFunction(MachineFunction &MF) {
           !MRI.hasOneUse(MI.getOperand(0).getReg()))
         continue;
 
-      // FIXME: Fold operands with subregs.
       if (OpToFold.isReg() &&
-          (!TargetRegisterInfo::isVirtualRegister(OpToFold.getReg()) ||
-           OpToFold.getSubReg()))
+          !TargetRegisterInfo::isVirtualRegister(OpToFold.getReg()))
         continue;
 
-
       // We need mutate the operands of new mov instructions to add implicit
       // uses of EXEC, but adding them invalidates the use_iterator, so defer
       // this.

diff --git a/lib/Target/AMDGPU/SIInstrInfo.cpp b/lib/Target/AMDGPU/SIInstrInfo.cpp
index a08a5a8fed3612ab0c53868a9d3fc40be99caeaf..bd80756d5c8fb790a5aec60f0f80563313ac676b 100644 (file)
--- a/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -1777,6 +1777,10 @@ bool SIInstrInfo::isLegalRegOperand(const MachineRegisterInfo &MRI,
     MRI.getRegClass(Reg) :
     RI.getPhysRegClass(Reg);
 
+  const SIRegisterInfo *TRI =
+      static_cast<const SIRegisterInfo*>(MRI.getTargetRegisterInfo());
+  RC = TRI->getSubRegClass(RC, MO.getSubReg());
+
   // In order to be legal, the common sub-class must be equal to the
   // class of the current operand.  For example:
   //

diff --git a/lib/Target/AMDGPU/SIRegisterInfo.cpp b/lib/Target/AMDGPU/SIRegisterInfo.cpp
index 017664f094caaaa6061aa788337ba187f2d475bc..2d68b8b5df0e2e9c4fe2eb69e93ff234a58a6b96 100644 (file)
--- a/lib/Target/AMDGPU/SIRegisterInfo.cpp
+++ b/lib/Target/AMDGPU/SIRegisterInfo.cpp
@@ -464,12 +464,38 @@ const TargetRegisterClass *SIRegisterInfo::getSubRegClass(
   if (SubIdx == AMDGPU::NoSubRegister)
     return RC;
 
-  // If this register has a sub-register, we can safely assume it is a 32-bit
-  // register, because all of SI's sub-registers are 32-bit.
+  // We can assume that each lane corresponds to one 32-bit register.
+  unsigned Count = countPopulation(getSubRegIndexLaneMask(SubIdx));
   if (isSGPRClass(RC)) {
-    return &AMDGPU::SGPR_32RegClass;
+    switch (Count) {
+    case 1:
+      return &AMDGPU::SGPR_32RegClass;
+    case 2:
+      return &AMDGPU::SReg_64RegClass;
+    case 4:
+      return &AMDGPU::SReg_128RegClass;
+    case 8:
+      return &AMDGPU::SReg_256RegClass;
+    case 16: /* fall-through */
+    default:
+      llvm_unreachable("Invalid sub-register class size");
+    }
   } else {
-    return &AMDGPU::VGPR_32RegClass;
+    switch (Count) {
+    case 1:
+      return &AMDGPU::VGPR_32RegClass;
+    case 2:
+      return &AMDGPU::VReg_64RegClass;
+    case 3:
+      return &AMDGPU::VReg_96RegClass;
+    case 4:
+      return &AMDGPU::VReg_128RegClass;
+    case 8:
+      return &AMDGPU::VReg_256RegClass;
+    case 16: /* fall-through */
+    default:
+      llvm_unreachable("Invalid sub-register class size");
+    }
   }
 }
 

diff --git a/test/CodeGen/AMDGPU/fmin_legacy.ll b/test/CodeGen/AMDGPU/fmin_legacy.ll
index 52fc3d0d251a497ba035e311eb9cf2374bc4ef9a..69a0a520a476c53741d728a61bc96f731435fcbe 100644 (file)
--- a/test/CodeGen/AMDGPU/fmin_legacy.ll
+++ b/test/CodeGen/AMDGPU/fmin_legacy.ll
@@ -8,8 +8,8 @@ declare i32 @llvm.r600.read.tidig.x() #1
 
 ; FUNC-LABEL: @test_fmin_legacy_f32
 ; EG: MIN *
-; SI-SAFE: v_min_legacy_f32_e32
-; SI-NONAN: v_min_f32_e32
+; SI-SAFE: v_min_legacy_f32_e64
+; SI-NONAN: v_min_f32_e64
 define void @test_fmin_legacy_f32(<4 x float> addrspace(1)* %out, <4 x float> inreg %reg0) #0 {
    %r0 = extractelement <4 x float> %reg0, i32 0
    %r1 = extractelement <4 x float> %reg0, i32 1

diff --git a/test/CodeGen/AMDGPU/fsub.ll b/test/CodeGen/AMDGPU/fsub.ll
index dfe41cb5b11110fb06ed394ec7642d3b34429400..38d573258a5e29461210219798a869df04beebb5 100644 (file)
--- a/test/CodeGen/AMDGPU/fsub.ll
+++ b/test/CodeGen/AMDGPU/fsub.ll
@@ -32,9 +32,8 @@ declare void @llvm.AMDGPU.store.output(float, i32)
 ; R600-DAG: ADD {{\** *}}T{{[0-9]+\.[XYZW]}}, KC0[3].X, -KC0[3].Z
 ; R600-DAG: ADD {{\** *}}T{{[0-9]+\.[XYZW]}}, KC0[2].W, -KC0[3].Y
 
-; FIXME: Should be using SGPR directly for first operand
-; SI: v_subrev_f32_e32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
-; SI: v_subrev_f32_e32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
+; SI: v_subrev_f32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
+; SI: v_subrev_f32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
 define void @fsub_v2f32(<2 x float> addrspace(1)* %out, <2 x float> %a, <2 x float> %b) {
   %sub = fsub <2 x float> %a, %b
   store <2 x float> %sub, <2 x float> addrspace(1)* %out, align 8
@@ -60,13 +59,11 @@ define void @v_fsub_v4f32(<4 x float> addrspace(1)* %out, <4 x float> addrspace(
   ret void
 }
 
-; FIXME: Should be using SGPR directly for first operand
-
 ; FUNC-LABEL: {{^}}s_fsub_v4f32:
-; SI: v_subrev_f32_e32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
-; SI: v_subrev_f32_e32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
-; SI: v_subrev_f32_e32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
-; SI: v_subrev_f32_e32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}
+; SI: v_subrev_f32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
+; SI: v_subrev_f32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
+; SI: v_subrev_f32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
+; SI: v_subrev_f32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
 ; SI: s_endpgm
 define void @s_fsub_v4f32(<4 x float> addrspace(1)* %out, <4 x float> %a, <4 x float> %b) {
   %result = fsub <4 x float> %a, %b

diff --git a/test/CodeGen/AMDGPU/llvm.round.f64.ll b/test/CodeGen/AMDGPU/llvm.round.f64.ll
index 6b365dc09e2a9e21da81501a57ba59122f0b509c..98afbeee93e6cf9d58ccc569d0954ae6986e2eed 100644 (file)
--- a/test/CodeGen/AMDGPU/llvm.round.f64.ll
+++ b/test/CodeGen/AMDGPU/llvm.round.f64.ll
@@ -21,7 +21,7 @@ define void @round_f64(double addrspace(1)* %out, double %x) #0 {
 ; SI-DAG: v_cmp_eq_i32
 
 ; SI-DAG: s_mov_b32 [[BFIMASK:s[0-9]+]], 0x7fffffff
-; SI-DAG: v_cmp_gt_i32_e32
+; SI-DAG: v_cmp_gt_i32
 ; SI-DAG: v_bfi_b32 [[COPYSIGN:v[0-9]+]], [[BFIMASK]]
 
 ; SI: buffer_store_dwordx2