AMDGPU gfx12: Add _dvgpr$ symbols for dynamic VGPRs #148251
Open
+118
−0
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
For each function with the AMDGPU_CS_Chain calling convention, with dynamic VGPRs enabled, add a _dvgpr$ symbol, with the value of the function symbol, plus an offset encoding one less than the number of VGPR blocks used by the function (16 VGPRs per block, no more than 128) in bits 5..3 of the symbol value. This is used by a front-end to have functions that are chained rather than called, and a dispatcher that dynamically resizes the VGPR count before dispatching to a function.
|
@llvm/pr-subscribers-backend-amdgpu Author: Tim Renouf (trenouf) ChangesFor each function with the AMDGPU_CS_Chain calling convention, with dynamic VGPRs enabled, add a _dvgpr$ symbol, with the value of the function symbol, plus an offset encoding one less than the number of VGPR blocks used by the function (16 VGPRs per block, no more than 128) in bits 5..3 of the symbol value. This is used by a front-end to have functions that are chained rather than called, and a dispatcher that dynamically resizes the VGPR count before dispatching to a function. Full diff: https://github.com/llvm/llvm-project/pull/148251.diff 2 Files Affected:
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp b/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp
index 749b9efc81378..00ed5f57967ce 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp
@@ -194,6 +194,32 @@ void AMDGPUAsmPrinter::emitFunctionBodyStart() {
return;
}
+ if (STM.isDynamicVGPREnabled() &&
+ MF->getFunction().getCallingConv() == CallingConv::AMDGPU_CS_Chain) {
+ // Add a _dvgpr$ symbol, with the value of the function symbol, plus an
+ // offset encoding one less than the number of VGPR blocks used by the
+ // function (16 VGPRs per block, no more than 128) in bits 5..3 of the
+ // symbol value. This is used by a front-end to have functions that are
+ // chained rather than called, and a dispatcher that dynamically resizes
+ // the VGPR count before dispatching to a function.
+ ResourceUsage = &getAnalysis<AMDGPUResourceUsageAnalysis>();
+ const AMDGPUResourceUsageAnalysis::SIFunctionResourceInfo &Info =
+ ResourceUsage->getResourceInfo();
+ MCContext &Ctx = MF->getContext();
+ unsigned EncodedNumVGPRs = (Info.NumVGPR - 1) >> 1 & 0x38;
+ MCSymbol *CurPCSym = Ctx.createTempSymbol();
+ OutStreamer->emitLabel(CurPCSym);
+ const MCExpr *DVgprFuncVal = MCBinaryExpr::createAdd(
+ MCSymbolRefExpr::create(CurPCSym, MCSymbolRefExpr::VK_None, Ctx),
+ MCConstantExpr::create(EncodedNumVGPRs, Ctx), Ctx);
+ MCSymbol *DVgprFuncSym =
+ Ctx.getOrCreateSymbol(Twine("_dvgpr$") + MF->getFunction().getName());
+ OutStreamer->emitAssignment(DVgprFuncSym, DVgprFuncVal);
+ cast<MCSymbolELF>(DVgprFuncSym)
+ ->setBinding(
+ cast<MCSymbolELF>(getSymbol(&MF->getFunction()))->getBinding());
+ }
+
if (!MFI.isEntryFunction())
return;
diff --git a/llvm/test/CodeGen/AMDGPU/dvgpr_sym.ll b/llvm/test/CodeGen/AMDGPU/dvgpr_sym.ll
new file mode 100644
index 0000000000000..992963d304ead
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/dvgpr_sym.ll
@@ -0,0 +1,12 @@
+; Test generation of _dvgpr$ symbol for an amdgpu_cs_chain function with +dynamic-vgpr.
+
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 -asm-verbose=0 < %s | FileCheck -check-prefixes=DVGPR %s
+
+; DVGPR-LABEL: func:
+; DVGPR: .Ltmp0:
+; DVGPR: .set _dvgpr$func, .Ltmp0+{{[0-9]+}}
+
+define amdgpu_cs_chain void @func() #0 {
+ ret void
+}
+attributes #0 = { "target-features"="+dynamic-vgpr" }
|
shiltian
reviewed
Jul 12, 2025
shiltian
reviewed
Jul 12, 2025
arsenm
reviewed
Jul 14, 2025
| MCSymbolRefExpr::create(CurPCSym, Ctx), | ||
| MCConstantExpr::create(EncodedNumVGPRs, Ctx), Ctx); | ||
| MCSymbol *DVgprFuncSym = | ||
| Ctx.getOrCreateSymbol(Twine("_dvgpr$") + MF->getFunction().getName()); |
There was a problem hiding this comment.
Is this the right prefix to use? Is this using the right visibility?
There was a problem hiding this comment.
You mean is "_dvgpr$" the right prefix? Visibility and linkage fixed.
* Use new func attr; * allow 16 or 32 block size; * put code in its own func; * enhance test, including anonymous func; * fix name, visibility and linkage
arsenm
reviewed
Jul 16, 2025
| if (!CurrentProgramInfo.NumVGPRsForWavesPerEU->evaluateAsRelocatable( | ||
| NumVGPRs, nullptr) || | ||
| !NumVGPRs.isAbsolute()) { | ||
| OutContext.reportError({}, "Unable to resolve _dvgpr$ symbol for '" + |
There was a problem hiding this comment.
Error messages should start with a lowercase letter
| NumVGPRs, nullptr) || | ||
| !NumVGPRs.isAbsolute()) { | ||
| OutContext.reportError({}, "Unable to resolve _dvgpr$ symbol for '" + | ||
| Twine(MF.getName()) + "'"); |
There was a problem hiding this comment.
Use the mangled symbol name, this breaks for anonymous functions
| BlockSize; | ||
| if (NumBlocks > 8) { | ||
| OutContext.reportError({}, | ||
| "Too many DVGPR blocks for _dvgpr$ symbol for '" + |
There was a problem hiding this comment.
Same as above. Also should test the error cases
shiltian
reviewed
Jul 16, 2025
| @@ -1768,6 +1768,10 @@ The AMDGPU backend supports the following LLVM IR attributes. | |||
| using dedicated instructions, but may not send the DEALLOC_VGPRS | |||
| message. If a shader has this attribute, then all its callees must | |||
| match its value. | |||
| An AMD_CS_Chain CC function with this enabled has an extra symbol | |||
There was a problem hiding this comment.
Suggested change
| An AMD_CS_Chain CC function with this enabled has an extra symbol | |
| An amd_cs_chain CC function with this enabled has an extra symbol |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
For each function with the AMDGPU_CS_Chain calling convention, with dynamic VGPRs enabled, add a _dvgpr$ symbol, with the value of the function symbol, plus an offset encoding one less than the number of VGPR blocks used by the function (16 VGPRs per block, no more than 128) in bits 5..3 of the symbol value. This is used by a front-end to have functions that are chained rather than called, and a dispatcher that dynamically resizes the VGPR count before dispatching to a function.