-
Notifications
You must be signed in to change notification settings - Fork 15.1k
[SPIRV] Expand spv_bitcast intrinsic during instruction selection #164884
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Conversation
The spv_bitcast intrinsic is currently replaced by an OpBitcast during prelegalization. This will cause a problem when we need to legalize the OpBitcast. The legalizer assumes that instruction already lowered to a target specific opcode is legal. We cannot lower it to a G_BITCAST because the bitcasts sometimes the LLT type will be the same, causing an error in the verifier, even if the SPIR-V types will be different. This commit keeps the intrinsic around until instructoin selection. We can create rules to legalize a G_INTRINISIC* instruction, and it does not create problem for the verifier.
|
@llvm/pr-subscribers-backend-spir-v Author: Steven Perron (s-perron) ChangesThe spv_bitcast intrinsic is currently replaced by an OpBitcast We cannot lower it to a G_BITCAST because the bitcasts sometimes This commit keeps the intrinsic around until instructoin selection. Full diff: https://github.com/llvm/llvm-project/pull/164884.diff 2 Files Affected:
diff --git a/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp b/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp
index 021353ab716f7..ccc2c0fc467fb 100644
--- a/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp
+++ b/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp
@@ -3119,6 +3119,14 @@ bool SPIRVInstructionSelector::selectIntrinsic(Register ResVReg,
return selectInsertElt(ResVReg, ResType, I);
case Intrinsic::spv_gep:
return selectGEP(ResVReg, ResType, I);
+ case Intrinsic::spv_bitcast: {
+ Register OpReg = I.getOperand(2).getReg();
+ SPIRVType *OpType =
+ OpReg.isValid() ? GR.getSPIRVTypeForVReg(OpReg) : nullptr;
+ if (!GR.isBitcastCompatible(ResType, OpType))
+ report_fatal_error("incompatible result and operand types in a bitcast");
+ return selectOpWithSrcs(ResVReg, ResType, I, {OpReg}, SPIRV::OpBitcast);
+ }
case Intrinsic::spv_unref_global:
case Intrinsic::spv_init_global: {
MachineInstr *MI = MRI->getVRegDef(I.getOperand(1).getReg());
diff --git a/llvm/lib/Target/SPIRV/SPIRVPreLegalizer.cpp b/llvm/lib/Target/SPIRV/SPIRVPreLegalizer.cpp
index db6f2d61e8f29..43ded6a71dd6d 100644
--- a/llvm/lib/Target/SPIRV/SPIRVPreLegalizer.cpp
+++ b/llvm/lib/Target/SPIRV/SPIRVPreLegalizer.cpp
@@ -192,31 +192,38 @@ static void buildOpBitcast(SPIRVGlobalRegistry *GR, MachineIRBuilder &MIB,
.addUse(OpReg);
}
-// We do instruction selections early instead of calling MIB.buildBitcast()
-// generating the general op code G_BITCAST. When MachineVerifier validates
-// G_BITCAST we see a check of a kind: if Source Type is equal to Destination
-// Type then report error "bitcast must change the type". This doesn't take into
-// account the notion of a typed pointer that is important for SPIR-V where a
-// user may and should use bitcast between pointers with different pointee types
-// (https://registry.khronos.org/SPIR-V/specs/unified1/SPIRV.html#OpBitcast).
-// It's important for correct lowering in SPIR-V, because interpretation of the
-// data type is not left to instructions that utilize the pointer, but encoded
-// by the pointer declaration, and the SPIRV target can and must handle the
-// declaration and use of pointers that specify the type of data they point to.
-// It's not feasible to improve validation of G_BITCAST using just information
-// provided by low level types of source and destination. Therefore we don't
-// produce G_BITCAST as the general op code with semantics different from
-// OpBitcast, but rather lower to OpBitcast immediately. As for now, the only
-// difference would be that CombinerHelper couldn't transform known patterns
-// around G_BUILD_VECTOR. See discussion
-// in https://github.com/llvm/llvm-project/pull/110270 for even more context.
-static void selectOpBitcasts(MachineFunction &MF, SPIRVGlobalRegistry *GR,
- MachineIRBuilder MIB) {
+// We lower G_BITCAST to OpBitcast here to avoid a MachineVerifier error.
+// The verifier checks if the source and destination LLTs of a G_BITCAST are
+// different, but this check is too strict for SPIR-V's typed pointers, which
+// may have the same LLT but different SPIRVType (e.g. pointers to different
+// pointee types). By lowering to OpBitcast here, we bypass the verifier's
+// check. See discussion in https://github.com/llvm/llvm-project/pull/110270
+// for more context.
+//
+// We also handle the llvm.spv.bitcast intrinsic here. If the source and
+// destination SPIR-V types are the same, we lower it to a COPY to enable
+// further optimizations like copy propagation.
+static void lowerBitcasts(MachineFunction &MF, SPIRVGlobalRegistry *GR,
+ MachineIRBuilder MIB) {
SmallVector<MachineInstr *, 16> ToErase;
for (MachineBasicBlock &MBB : MF) {
for (MachineInstr &MI : MBB) {
+ if (isSpvIntrinsic(MI, Intrinsic::spv_bitcast)) {
+ Register DstReg = MI.getOperand(0).getReg();
+ Register SrcReg = MI.getOperand(2).getReg();
+ SPIRVType *DstType = GR->getSPIRVTypeForVReg(DstReg);
+ SPIRVType *SrcType = GR->getSPIRVTypeForVReg(SrcReg);
+ if (DstType && SrcType && DstType == SrcType) {
+ MIB.setInsertPt(*MI.getParent(), MI);
+ MIB.buildCopy(DstReg, SrcReg);
+ ToErase.push_back(&MI);
+ continue;
+ }
+ }
+
if (MI.getOpcode() != TargetOpcode::G_BITCAST)
continue;
+
MIB.setInsertPt(*MI.getParent(), MI);
buildOpBitcast(GR, MIB, MI.getOperand(0).getReg(),
MI.getOperand(1).getReg());
@@ -237,16 +244,11 @@ static void insertBitcasts(MachineFunction &MF, SPIRVGlobalRegistry *GR,
SmallVector<MachineInstr *, 10> ToErase;
for (MachineBasicBlock &MBB : MF) {
for (MachineInstr &MI : MBB) {
- if (!isSpvIntrinsic(MI, Intrinsic::spv_bitcast) &&
- !isSpvIntrinsic(MI, Intrinsic::spv_ptrcast))
+ if (!isSpvIntrinsic(MI, Intrinsic::spv_ptrcast))
continue;
assert(MI.getOperand(2).isReg());
MIB.setInsertPt(*MI.getParent(), MI);
ToErase.push_back(&MI);
- if (isSpvIntrinsic(MI, Intrinsic::spv_bitcast)) {
- MIB.buildBitcast(MI.getOperand(0).getReg(), MI.getOperand(2).getReg());
- continue;
- }
Register Def = MI.getOperand(0).getReg();
Register Source = MI.getOperand(2).getReg();
Type *ElemTy = getMDOperandAsType(MI.getOperand(3).getMetadata(), 0);
@@ -1089,7 +1091,7 @@ bool SPIRVPreLegalizer::runOnMachineFunction(MachineFunction &MF) {
removeImplicitFallthroughs(MF, MIB);
insertSpirvDecorations(MF, GR, MIB);
insertInlineAsm(MF, GR, ST, MIB);
- selectOpBitcasts(MF, GR, MIB);
+ lowerBitcasts(MF, GR, MIB);
return true;
}
|
|
I feel pretty confident about this one, so I will merge it. |
…vm#164884) The spv_bitcast intrinsic is currently replaced by an OpBitcast during prelegalization. This will cause a problem when we need to legalize the OpBitcast. The legalizer assumes that instruction already lowered to a target specific opcode is legal. We cannot lower it to a G_BITCAST because the bitcasts sometimes the LLT type will be the same, causing an error in the verifier, even if the SPIR-V types will be different. This commit keeps the intrinsic around until instruction selection. We can create rules to legalize a G_INTRINISIC* instruction, and it does not create problem for the verifier. No tests are updated because this change should be invisible to users.
The spv_bitcast intrinsic is currently replaced by an OpBitcast
during prelegalization. This will cause a problem when we need to
legalize the OpBitcast. The legalizer assumes that instruction
already lowered to a target specific opcode is legal.
We cannot lower it to a G_BITCAST because the bitcasts sometimes
the LLT type will be the same, causing an error in the verifier,
even if the SPIR-V types will be different.
This commit keeps the intrinsic around until instruction selection.
We can create rules to legalize a G_INTRINISIC* instruction, and
it does not create problem for the verifier.
No tests are updated because this change should be invisible to users.