[llvm] [AMDGPU] Generalize MFMA VGPR->AGPR opcode mapping (NFC) (PR #213274)
via llvm-commits
llvm-commits at lists.llvm.org
Fri Jul 31 06:51:33 PDT 2026
llvmorg-github-actions[bot] wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-amdgpu
Author: K.Roy (MrRoy09)
<details>
<summary>Changes</summary>
Move the VGPR/AGPR pairing out of MFMATable into a generalized
AGPRFormTable class and rename getMFMASrcCVDstAGPROp to getAGPRFormOp.
MFMATable inherits AGPRFormTable, so the instantiation sites and the
generated table are unchanged.
Other instruction families with paired VGPR/AGPR pseudos, such as DS,
can now be tagged with AGPRFormTable.
Related to #<!-- -->168983
---
Full diff: https://github.com/llvm/llvm-project/pull/213274.diff
5 Files Affected:
- (modified) llvm/lib/Target/AMDGPU/AMDGPURewriteAGPRCopyMFMA.cpp (+3-4)
- (modified) llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp (+3-3)
- (modified) llvm/lib/Target/AMDGPU/SIInstrInfo.h (+3-3)
- (modified) llvm/lib/Target/AMDGPU/SIInstrInfo.td (+10-5)
- (modified) llvm/lib/Target/AMDGPU/VOP3PInstructions.td (+1-5)
``````````diff
diff --git a/llvm/lib/Target/AMDGPU/AMDGPURewriteAGPRCopyMFMA.cpp b/llvm/lib/Target/AMDGPU/AMDGPURewriteAGPRCopyMFMA.cpp
index 1d39b4f1bc52d..6510c07358e01 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPURewriteAGPRCopyMFMA.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPURewriteAGPRCopyMFMA.cpp
@@ -77,7 +77,7 @@ class AMDGPURewriteAGPRCopyMFMAImpl {
LIS(LIS), LSS(LSS), RegClassInfo(RegClassInfo) {}
bool isRewriteCandidate(const MachineInstr &MI) const {
- return TII.isMAI(MI) && AMDGPU::getMFMASrcCVDstAGPROp(MI.getOpcode()) != -1;
+ return TII.isMAI(MI) && AMDGPU::getAGPRFormOp(MI.getOpcode()) != -1;
}
/// Find AV_* registers assigned to AGPRs (or virtual registers which were
@@ -159,7 +159,7 @@ bool AMDGPURewriteAGPRCopyMFMAImpl::recomputeRegClassExceptRewritable(
// either AGPR or VGPR in src0/src1. We still need to check constraint
// effects for scale variant, which does not allow AGPR.
if (isRewriteCandidate(*MI)) {
- int AGPROp = AMDGPU::getMFMASrcCVDstAGPROp(MI->getOpcode());
+ int AGPROp = AMDGPU::getAGPRFormOp(MI->getOpcode());
const MCInstrDesc &AGPRDesc = TII.get(AGPROp);
const TargetRegisterClass *NewRC =
TII.getRegClass(AGPRDesc, MO.getOperandNo());
@@ -293,8 +293,7 @@ bool AMDGPURewriteAGPRCopyMFMAImpl::tryReassigningMFMAChain(
}
for (MachineInstr *RewriteCandidate : RewriteCandidates) {
- int NewMFMAOp =
- AMDGPU::getMFMASrcCVDstAGPROp(RewriteCandidate->getOpcode());
+ int NewMFMAOp = AMDGPU::getAGPRFormOp(RewriteCandidate->getOpcode());
RewriteCandidate->setDesc(TII.get(NewMFMAOp));
++NumMFMAsRewrittenToAGPR;
}
diff --git a/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp b/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp
index ff2ef307b1ea6..0ac656a3e02a7 100644
--- a/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp
+++ b/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp
@@ -2359,7 +2359,7 @@ void RewriteMFMAFormStage::resetRewriteCandsToVGPR(
bool RewriteMFMAFormStage::isRewriteCandidate(MachineInstr *MI) const {
if (!static_cast<const SIInstrInfo *>(DAG.TII)->isMAI(*MI))
return false;
- if (AMDGPU::getMFMASrcCVDstAGPROp(MI->getOpcode()) == -1)
+ if (AMDGPU::getAGPRFormOp(MI->getOpcode()) == -1)
return false;
// Reject candidates whose users force an unavoidable bridge copy.
Register DstReg = MI->getOperand(0).getReg();
@@ -2397,7 +2397,7 @@ bool RewriteMFMAFormStage::initHeuristics(
if (!isRewriteCandidate(&MI))
continue;
- int ReplacementOp = AMDGPU::getMFMASrcCVDstAGPROp(MI.getOpcode());
+ int ReplacementOp = AMDGPU::getAGPRFormOp(MI.getOpcode());
assert(ReplacementOp != -1);
RewriteCands.push_back({&MI, MI.getOpcode()});
@@ -2657,7 +2657,7 @@ bool RewriteMFMAFormStage::rewrite(
}
for (auto &[MI, OriginalOpcode] : RewriteCands) {
- int ReplacementOp = AMDGPU::getMFMASrcCVDstAGPROp(MI->getOpcode());
+ int ReplacementOp = AMDGPU::getAGPRFormOp(MI->getOpcode());
if (ReplacementOp == -1)
continue;
MI->setDesc(TII->get(ReplacementOp));
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.h b/llvm/lib/Target/AMDGPU/SIInstrInfo.h
index 4c8641a6091d7..491b3aa58ac15 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.h
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.h
@@ -1894,10 +1894,10 @@ namespace AMDGPU {
LLVM_READONLY
int32_t getMFMAEarlyClobberOp(uint32_t Opcode);
- /// \returns Version of an MFMA instruction which uses AGPRs for srcC and
- /// vdst, given an \p Opcode of an MFMA which uses VGPRs for srcC/vdst.
+ /// \returns Version of an instruction which uses AGPRs for coupled operands
+ /// given an \p Opcode which uses VGPRs for coupled operands.
LLVM_READONLY
- int32_t getMFMASrcCVDstAGPROp(uint32_t Opcode);
+ int32_t getAGPRFormOp(uint32_t Opcode);
/// \returns v_cmpx version of a v_cmp instruction.
LLVM_READONLY
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.td b/llvm/lib/Target/AMDGPU/SIInstrInfo.td
index 31f6043a146ef..0c22373c019db 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.td
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.td
@@ -3243,6 +3243,11 @@ class Commutable_REV <string revOp, bit isOrig> {
bit IsOrig = isOrig;
}
+class AGPRFormTable<string Kind, string Name = NAME> {
+ string AGPRFormOp = Name;
+ string AGPRFormKind = Kind;
+}
+
//===----------------------------------------------------------------------===//
// Interpolation opcodes
//===----------------------------------------------------------------------===//
@@ -3493,11 +3498,11 @@ def getMFMAEarlyClobberOp : InstrMapping {
let ValueCols = [["0"]];
}
-// Map from an mfma using VGPRs to one using AGPRs.
-def getMFMASrcCVDstAGPROp : InstrMapping {
- let FilterClass = "MFMATable";
- let RowFields = ["AGPROp"];
- let ColFields = ["MFMAKind"];
+// Map from an instruction using VGPRs to one using AGPRs.
+def getAGPRFormOp : InstrMapping {
+ let FilterClass = "AGPRFormTable";
+ let RowFields = ["AGPRFormOp"];
+ let ColFields = ["AGPRFormKind"];
let KeyCol = ["VGPR"];
let ValueCols = [["AGPR"]];
}
diff --git a/llvm/lib/Target/AMDGPU/VOP3PInstructions.td b/llvm/lib/Target/AMDGPU/VOP3PInstructions.td
index 8728e70ea351f..87793bf71f337 100644
--- a/llvm/lib/Target/AMDGPU/VOP3PInstructions.td
+++ b/llvm/lib/Target/AMDGPU/VOP3PInstructions.td
@@ -1103,13 +1103,9 @@ def VOPProfileMAI_I32_V4I32_X16_VCD : VOPProfileMAI<VOP_V16I32_V4I32_V4I32_V16I3
class MFMATable <bit is_mac, string Kind, string Name,
- string AGPROpName = NAME> {
+ string AGPROpName = NAME> : AGPRFormTable<Kind, AGPROpName> {
bit IsMac = is_mac;
string FMAOp = Name;
- string AGPROp = AGPROpName;
-
- // Does this MFMA use "AGPR" or "VGPR" for srcC/vdst
- string MFMAKind = Kind;
}
class MFMA_F8F6F4_WithSizeTable<int A, int B, Instruction ThisVariant, Instruction F8F8Variant> {
``````````
</details>
https://github.com/llvm/llvm-project/pull/213274
More information about the llvm-commits
mailing list