[clang] e214f7a - [AMDGPU] Validate prefetch length and cache policy operands (#232143)
via cfe-commits
cfe-commits at lists.llvm.org
Thu Nov 5 15:50:54 PST 2026
Author: Gheorghe-Teodor Bercea
Date: 2026-11-05T23:50:49Z
New Revision: e214f7aa1d129c6d4358c619e3c91a0e11ef50de
URL: https://github.com/llvm/llvm-project/commit/e214f7aa1d129c6d4358c619e3c91a0e11ef50de
DIFF: https://github.com/llvm/llvm-project/commit/e214f7aa1d129c6d4358c619e3c91a0e11ef50de.diff
LOG: [AMDGPU] Validate prefetch length and cache policy operands (#232143)
Two operands of the AMDGPU prefetch intrinsics, the length and the cache
policy, are not validated. An out-of-range length is encoded differently
from what the assembly says, and an out-of-range cache policy is
silently miscompiled or rejected late by the backend.
Fixes:
- Length: Instruction selection now keeps only the low 5 bits of a
constant length, so the assembly shows the value that is encoded. The
object code is the same as before. As a safeguard for hand-written
assembly, including inline `asm`, the assembler on GFX12 and later now
rejects an immediate length that does not fit in 5 bits instead of
silently truncating it.
- Cache policy: clang and the IR verifier now reject a cache policy that
sets any bit other than th (bits 0-2), scope (bits 3-4) and nv (bit 5),
or that uses the reserved th value 7. The assembler needs no change: it
already rejects th 7.
Added:
Modified:
clang/include/clang/Basic/DiagnosticSemaKinds.td
clang/lib/Sema/SemaAMDGPU.cpp
clang/test/SemaOpenCL/builtins-amdgcn-error-gfx1250-param.cl
llvm/include/llvm/IR/IntrinsicsAMDGPU.td
llvm/lib/Target/AMDGPU/AMDGPUGISel.td
llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.h
llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
llvm/lib/Target/AMDGPU/SMInstructions.td
llvm/test/CodeGen/AMDGPU/llvm.amdgcn.s.buffer.prefetch.data.ll
llvm/test/CodeGen/AMDGPU/llvm.amdgcn.s.prefetch.data.ll
llvm/test/CodeGen/AMDGPU/llvm.amdgcn.s.prefetch.inst.ll
llvm/test/MC/AMDGPU/gfx12_asm_smem_err.s
llvm/test/Verifier/AMDGPU/intrinsic-prefetch.ll
Removed:
################################################################################
diff --git a/clang/include/clang/Basic/DiagnosticSemaKinds.td b/clang/include/clang/Basic/DiagnosticSemaKinds.td
index 941a58ae54a05e..4e3f729b4f36d4 100644
--- a/clang/include/clang/Basic/DiagnosticSemaKinds.td
+++ b/clang/include/clang/Basic/DiagnosticSemaKinds.td
@@ -14517,6 +14517,8 @@ def warn_amdgpu_av_requires_atomic : Warning<
def err_amdgcn_dmask_has_too_many_bits_set
: Error<"dmask argument cannot have more bits set than there are elements "
"in return type">;
+def err_amdgcn_prefetch_reserved_th
+ : Error<"cache policy uses the reserved th value 7">;
def warn_amdgpu_s_wait_event_mask_no_effect_target :
Warning<"event mask has no effect for target">,
diff --git a/clang/lib/Sema/SemaAMDGPU.cpp b/clang/lib/Sema/SemaAMDGPU.cpp
index fa68291d88ebcc..1d79fa881202dc 100644
--- a/clang/lib/Sema/SemaAMDGPU.cpp
+++ b/clang/lib/Sema/SemaAMDGPU.cpp
@@ -168,6 +168,22 @@ bool SemaAMDGPU::CheckAMDGCNBuiltinFunctionCall(const TargetInfo &TI,
case AMDGPU::BI__builtin_amdgcn_s_setreg:
return SemaRef.BuiltinConstantArgRange(TheCall, /*ArgNum=*/0, /*Low=*/0,
/*High=*/UINT16_MAX);
+ case AMDGPU::BI__builtin_amdgcn_flat_prefetch:
+ case AMDGPU::BI__builtin_amdgcn_global_prefetch: {
+ // Only th (bits 0-2, not 7), scope (bits 3-4) and nv (bit 5) may be set.
+ llvm::APSInt CPol;
+ if (SemaRef.BuiltinConstantArgRange(TheCall, /*ArgNum=*/1, /*Low=*/0,
+ /*High=*/63) ||
+ SemaRef.BuiltinConstantArg(TheCall, /*ArgNum=*/1, CPol))
+ return true;
+ if ((CPol.getZExtValue() & 0x7) != 0x7)
+ return false;
+ Expr *ArgExpr = TheCall->getArg(1);
+ SemaRef.targetDiag(ArgExpr->getExprLoc(),
+ diag::err_amdgcn_prefetch_reserved_th)
+ << ArgExpr->getSourceRange();
+ return true;
+ }
case AMDGPU::BI__builtin_amdgcn_s_wait_event: {
llvm::APSInt Result;
if (SemaRef.BuiltinConstantArg(TheCall, 0, Result))
diff --git a/clang/test/SemaOpenCL/builtins-amdgcn-error-gfx1250-param.cl b/clang/test/SemaOpenCL/builtins-amdgcn-error-gfx1250-param.cl
index 0fdf8df507bec7..75b0bec29baffa 100644
--- a/clang/test/SemaOpenCL/builtins-amdgcn-error-gfx1250-param.cl
+++ b/clang/test/SemaOpenCL/builtins-amdgcn-error-gfx1250-param.cl
@@ -194,6 +194,11 @@ void test_amdgcn_tensor_load_store(v4u sg0, v8u sg1, v4u sg2, v4u sg3, v8u sg4,
void test_prefetch(generic void *fptr, global void *gptr, int cpol) {
__builtin_amdgcn_flat_prefetch(fptr, cpol); // expected-error {{'__builtin_amdgcn_flat_prefetch' must be a constant integer}}
__builtin_amdgcn_global_prefetch(gptr, cpol); // expected-error {{'__builtin_amdgcn_global_prefetch' must be a constant integer}}
+ __builtin_amdgcn_flat_prefetch(fptr, 64); // expected-error {{argument value 64 is outside the valid range [0, 63]}}
+ __builtin_amdgcn_global_prefetch(gptr, 2048); // expected-error {{argument value 2048 is outside the valid range [0, 63]}}
+ __builtin_amdgcn_flat_prefetch(fptr, 15); // expected-error {{cache policy uses the reserved th value 7}}
+ __builtin_amdgcn_global_prefetch(gptr, 7); // expected-error {{cache policy uses the reserved th value 7}}
+ __builtin_amdgcn_global_prefetch(gptr, 62);
}
void test_cvt_f32_fp8_e5m3(global int* out, int a)
diff --git a/llvm/include/llvm/IR/IntrinsicsAMDGPU.td b/llvm/include/llvm/IR/IntrinsicsAMDGPU.td
index a4abe8d2c4e0d8..4a5cc90c3038df 100644
--- a/llvm/include/llvm/IR/IntrinsicsAMDGPU.td
+++ b/llvm/include/llvm/IR/IntrinsicsAMDGPU.td
@@ -3436,12 +3436,18 @@ def int_amdgcn_ds_bpermute_fi_b32 :
Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty],
[IntrNoMem, IntrConvergent, IntrWillReturn, IntrNoCallback, IntrNoFree]>;
+// A prefetch cache policy sets th (bits 0-2) to anything but the reserved
+// value 7, scope (bits 3-4) and nv (bit 5).
+defvar PrefetchCachePolicyRanges =
+ !foreach(i, !range(8), [!mul(i, 8), !add(!mul(i, 8), 6)]);
+
def int_amdgcn_flat_prefetch : ClangBuiltin<"__builtin_amdgcn_flat_prefetch">,
DefaultAttrsIntrinsic<[],
[llvm_ptr_ty, // Pointer
llvm_i32_ty], // cachepolicy(imm), bits [0-2] = th, bits [3-4] = scope
[IntrInaccessibleMemOrArgMemOnly, ReadOnly<ArgIndex<0>>,
- NoCapture<ArgIndex<0>>, ImmArg<ArgIndex<1>>],
+ NoCapture<ArgIndex<0>>, ImmArg<ArgIndex<1>>,
+ RangeSet<ArgIndex<1>, PrefetchCachePolicyRanges>],
"",
[SDNPMemOperand]
>;
@@ -3451,7 +3457,8 @@ def int_amdgcn_global_prefetch : ClangBuiltin<"__builtin_amdgcn_global_prefetch"
[LLVMQualPointerType<1>, // Pointer
llvm_i32_ty], // cachepolicy(imm), bits [0-2] = th, bits [3-4] = scope
[IntrInaccessibleMemOrArgMemOnly, ReadOnly<ArgIndex<0>>,
- NoCapture<ArgIndex<0>>, ImmArg<ArgIndex<1>>],
+ NoCapture<ArgIndex<0>>, ImmArg<ArgIndex<1>>,
+ RangeSet<ArgIndex<1>, PrefetchCachePolicyRanges>],
"",
[SDNPMemOperand]
>;
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUGISel.td b/llvm/lib/Target/AMDGPU/AMDGPUGISel.td
index 2a09ed982ac34c..3c9c596b9c6520 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUGISel.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPUGISel.td
@@ -465,6 +465,9 @@ def gi_as_i16timm : GICustomOperandRenderer<"renderTruncTImm">,
def gi_as_i8timm : GICustomOperandRenderer<"renderTruncTImm">,
GISDNodeXFormEquiv<as_i8timm>;
+def gi_as_prefetch_len : GICustomOperandRenderer<"renderPrefetchLen">,
+ GISDNodeXFormEquiv<as_prefetch_len>;
+
def gi_as_i1timm : GICustomOperandRenderer<"renderTruncTImm">,
GISDNodeXFormEquiv<as_i1timm>;
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp b/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
index 5eb817f034991f..c2331831d10180 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
@@ -7643,6 +7643,13 @@ void AMDGPUInstructionSelector::renderTruncTImm(MachineInstrBuilder &MIB,
MIB.addImm(Op.getImm());
}
+void AMDGPUInstructionSelector::renderPrefetchLen(MachineInstrBuilder &MIB,
+ const MachineInstr &MI,
+ int OpIdx) const {
+ Register Len = MI.getOperand(OpIdx).getReg();
+ MIB.addImm(*getIConstantVRegSExtVal(Len, *MRI) & 0x1f);
+}
+
void AMDGPUInstructionSelector::renderZextBoolTImm(MachineInstrBuilder &MIB,
const MachineInstr &MI,
int OpIdx) const {
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.h b/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.h
index 73c88ce4d845cc..b12ac778fd0856 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.h
@@ -364,6 +364,8 @@ class AMDGPUInstructionSelector final : public InstructionSelector {
void renderTruncTImm(MachineInstrBuilder &MIB, const MachineInstr &MI,
int OpIdx) const;
+ void renderPrefetchLen(MachineInstrBuilder &MIB, const MachineInstr &MI,
+ int OpIdx) const;
void renderZextBoolTImm(MachineInstrBuilder &MIB, const MachineInstr &MI,
int OpIdx) const;
diff --git a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
index 5f644a78383d7d..8ee55807ebb4b5 100644
--- a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
+++ b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
@@ -1614,6 +1614,7 @@ class AMDGPUAsmParser : public MCTargetAsmParser {
bool validateOffset(const MCInst &Inst, const OperandVector &Operands);
bool validateFlatOffset(const MCInst &Inst, const OperandVector &Operands);
bool validateSMEMOffset(const MCInst &Inst, const OperandVector &Operands);
+ bool validateSMEMSData(const MCInst &Inst, const OperandVector &Operands);
bool validateBF16InlineConst(const MCInst &Inst,
const OperandVector &Operands);
bool validateSOPLiteral(const MCInst &Inst, const OperandVector &Operands);
@@ -4675,6 +4676,22 @@ bool AMDGPUAsmParser::validateSMEMOffset(const MCInst &Inst,
return false;
}
+// GFX12+ SMEM prefetches and probes encode an immediate sdata in 5 bits.
+bool AMDGPUAsmParser::validateSMEMSData(const MCInst &Inst,
+ const OperandVector &Operands) {
+ if (!isGFX12Plus() || !SIInstrFlags::isSMRD(MII, Inst))
+ return true;
+
+ int OpNum =
+ AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::sdata);
+ if (OpNum == -1 || !Inst.getOperand(OpNum).isImm() ||
+ isUInt<5>(Inst.getOperand(OpNum).getImm()))
+ return true;
+
+ Error(getOperandLoc(Operands, OpNum), "expected a 5-bit unsigned immediate");
+ return false;
+}
+
// On subtargets with FeatureBF16InlineConstFromUpperFP32 the hardware generates
// a bf16 inline constant in the high half of the corresponding fp32 inline
// constant. A VOP1 bf16 opcode (v_cvt_f32_bf16 and the bf16 transcendentals)
@@ -5696,6 +5713,8 @@ bool AMDGPUAsmParser::validateInstruction(const MCInst &Inst, SMLoc IDLoc,
if (!validateOffset(Inst, Operands)) {
return false;
}
+ if (!validateSMEMSData(Inst, Operands))
+ return false;
if (!validateBF16InlineConst(Inst, Operands)) {
return false;
}
diff --git a/llvm/lib/Target/AMDGPU/SMInstructions.td b/llvm/lib/Target/AMDGPU/SMInstructions.td
index 5a54d13c1b6815..3eed14ef022fa9 100644
--- a/llvm/lib/Target/AMDGPU/SMInstructions.td
+++ b/llvm/lib/Target/AMDGPU/SMInstructions.td
@@ -1153,6 +1153,11 @@ defm : SMPrefetchPat<"INST", i32imm_zero>;
let AddedComplexity = 12 in // Prefer scalar prefetch over global for r/o case.
defm : SMPrefetchPat<"DATA", i32imm_one>;
+// The length field holds only 5 bits.
+def as_prefetch_len : SDNodeXForm<timm, [{
+ return CurDAG->getTargetConstant(N->getZExtValue() & 0x1f, SDLoc(N), MVT::i8);
+}]>;
+
let SubtargetPredicate = HasSmemPrefetchInsts in {
def : GCNPat <
(int_amdgcn_s_prefetch_data (SMRDImm i64:$sbase, i32:$offset), (i32 SReg_32:$len)),
@@ -1166,12 +1171,12 @@ let SubtargetPredicate = HasSmemPrefetchInsts in {
def : GCNPat <
(int_amdgcn_s_prefetch_data (SMRDImm i64:$sbase, i32:$offset), imm:$len),
- (S_PREFETCH_DATA $sbase, $offset, (i32 SGPR_NULL), (as_i8timm $len), 0)
+ (S_PREFETCH_DATA $sbase, $offset, (i32 SGPR_NULL), (as_prefetch_len $len), 0)
>;
def : GCNPat <
(int_amdgcn_s_prefetch_data (i64 SReg_64:$sbase), imm:$len),
- (S_PREFETCH_DATA $sbase, 0, (i32 SGPR_NULL), (as_i8timm $len), 0)
+ (S_PREFETCH_DATA $sbase, 0, (i32 SGPR_NULL), (as_prefetch_len $len), 0)
>;
@@ -1182,7 +1187,7 @@ let SubtargetPredicate = HasSmemPrefetchInsts in {
def : GCNPat <
(SIsbuffer_prefetch v4i32:$sbase, (SMRDBufferImm i32:$offset), imm:$len),
- (S_BUFFER_PREFETCH_DATA SReg_128:$sbase, i32imm:$offset, (i32 SGPR_NULL), (as_i8timm $len), 0)
+ (S_BUFFER_PREFETCH_DATA SReg_128:$sbase, i32imm:$offset, (i32 SGPR_NULL), (as_prefetch_len $len), 0)
>;
def : GCNPat <
@@ -1197,12 +1202,12 @@ let SubtargetPredicate = HasSmemPrefetchInsts in {
def : GCNPat <
(int_amdgcn_s_prefetch_inst (SMRDImm i64:$sbase, i32:$offset), imm:$len),
- (S_PREFETCH_INST $sbase, $offset, (i32 SGPR_NULL), (as_i8timm $len), 0)
+ (S_PREFETCH_INST $sbase, $offset, (i32 SGPR_NULL), (as_prefetch_len $len), 0)
>;
def : GCNPat <
(int_amdgcn_s_prefetch_inst (i64 SReg_64:$sbase), imm:$len),
- (S_PREFETCH_INST $sbase, 0, (i32 SGPR_NULL), (as_i8timm $len), 0)
+ (S_PREFETCH_INST $sbase, 0, (i32 SGPR_NULL), (as_prefetch_len $len), 0)
>;
} // End let SubtargetPredicate = HasSmemPrefetchInsts
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.s.buffer.prefetch.data.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.s.buffer.prefetch.data.ll
index 06b602386ddaf1..c42cea693b4d48 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.s.buffer.prefetch.data.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.s.buffer.prefetch.data.ll
@@ -24,6 +24,16 @@ entry:
ret void
}
+define amdgpu_ps void @buffer_prefetch_data_imm_offset_large_imm_len(ptr addrspace(8) inreg %rsrc) {
+; GCN-LABEL: buffer_prefetch_data_imm_offset_large_imm_len:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: s_buffer_prefetch_data s[0:3], 0x0, null, 8
+; GCN-NEXT: s_endpgm
+entry:
+ tail call void @llvm.amdgcn.s.buffer.prefetch.data(ptr addrspace(8) inreg %rsrc, i32 0, i32 40)
+ ret void
+}
+
define amdgpu_ps void @buffer_prefetch_data_imm_offset_vgpr_len(ptr addrspace(8) inreg %rsrc, i32 %len) {
; GCN-LABEL: buffer_prefetch_data_imm_offset_vgpr_len:
; GCN: ; %bb.0: ; %entry
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.s.prefetch.data.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.s.prefetch.data.ll
index 8535133bab9be9..75a99c4916bf00 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.s.prefetch.data.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.s.prefetch.data.ll
@@ -160,6 +160,83 @@ entry:
ret void
}
+define amdgpu_ps void @prefetch_data_sgpr_base_large_imm_len(ptr addrspace(4) inreg %ptr) {
+; GCN-LABEL: prefetch_data_sgpr_base_large_imm_len:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: s_prefetch_data s[0:1], 0x0, null, 8
+; GCN-NEXT: s_endpgm
+;
+; GFX1250-SDAG-LABEL: prefetch_data_sgpr_base_large_imm_len:
+; GFX1250-SDAG: ; %bb.0: ; %entry
+; GFX1250-SDAG-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
+; GFX1250-SDAG-NEXT: s_mov_b64 s[64:65], 0
+; GFX1250-SDAG-NEXT: v_nop
+; GFX1250-SDAG-NEXT: global_prefetch_b8 v0, s[64:65] scope:SCOPE_SE
+; GFX1250-SDAG-NEXT: s_prefetch_data s[0:1], 0x0, null, 8
+; GFX1250-SDAG-NEXT: s_endpgm
+;
+; GFX1250-GISEL-LABEL: prefetch_data_sgpr_base_large_imm_len:
+; GFX1250-GISEL: ; %bb.0: ; %entry
+; GFX1250-GISEL-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
+; GFX1250-GISEL-NEXT: s_mov_b64 s[64:65], 0
+; GFX1250-GISEL-NEXT: v_nop
+; GFX1250-GISEL-NEXT: global_prefetch_b8 v0, s[64:65] scope:SCOPE_SE
+; GFX1250-GISEL-NEXT: s_prefetch_data s[0:1], 0x0, null, 8
+; GFX1250-GISEL-NEXT: s_endpgm
+;
+; GFX1310-SDAG-LABEL: prefetch_data_sgpr_base_large_imm_len:
+; GFX1310-SDAG: ; %bb.0: ; %entry
+; GFX1310-SDAG-NEXT: s_prefetch_data s[0:1], 0x0, null, 8
+; GFX1310-SDAG-NEXT: s_endpgm
+;
+; GFX1310-GISEL-LABEL: prefetch_data_sgpr_base_large_imm_len:
+; GFX1310-GISEL: ; %bb.0: ; %entry
+; GFX1310-GISEL-NEXT: s_prefetch_data s[0:1], 0x0, null, 8
+; GFX1310-GISEL-NEXT: s_endpgm
+entry:
+ tail call void @llvm.amdgcn.s.prefetch.data.p4(ptr addrspace(4) %ptr, i32 40)
+ ret void
+}
+
+define amdgpu_ps void @prefetch_data_sgpr_imm_base_large_imm_len(ptr addrspace(4) inreg %ptr) {
+; GCN-LABEL: prefetch_data_sgpr_imm_base_large_imm_len:
+; GCN: ; %bb.0: ; %entry
+; GCN-NEXT: s_prefetch_data s[0:1], 0x200, null, 8
+; GCN-NEXT: s_endpgm
+;
+; GFX1250-SDAG-LABEL: prefetch_data_sgpr_imm_base_large_imm_len:
+; GFX1250-SDAG: ; %bb.0: ; %entry
+; GFX1250-SDAG-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
+; GFX1250-SDAG-NEXT: s_mov_b64 s[64:65], 0
+; GFX1250-SDAG-NEXT: v_nop
+; GFX1250-SDAG-NEXT: global_prefetch_b8 v0, s[64:65] scope:SCOPE_SE
+; GFX1250-SDAG-NEXT: s_prefetch_data s[0:1], 0x200, null, 8
+; GFX1250-SDAG-NEXT: s_endpgm
+;
+; GFX1250-GISEL-LABEL: prefetch_data_sgpr_imm_base_large_imm_len:
+; GFX1250-GISEL: ; %bb.0: ; %entry
+; GFX1250-GISEL-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
+; GFX1250-GISEL-NEXT: s_mov_b64 s[64:65], 0
+; GFX1250-GISEL-NEXT: v_nop
+; GFX1250-GISEL-NEXT: global_prefetch_b8 v0, s[64:65] scope:SCOPE_SE
+; GFX1250-GISEL-NEXT: s_prefetch_data s[0:1], 0x200, null, 8
+; GFX1250-GISEL-NEXT: s_endpgm
+;
+; GFX1310-SDAG-LABEL: prefetch_data_sgpr_imm_base_large_imm_len:
+; GFX1310-SDAG: ; %bb.0: ; %entry
+; GFX1310-SDAG-NEXT: s_prefetch_data s[0:1], 0x200, null, 8
+; GFX1310-SDAG-NEXT: s_endpgm
+;
+; GFX1310-GISEL-LABEL: prefetch_data_sgpr_imm_base_large_imm_len:
+; GFX1310-GISEL: ; %bb.0: ; %entry
+; GFX1310-GISEL-NEXT: s_prefetch_data s[0:1], 0x200, null, 8
+; GFX1310-GISEL-NEXT: s_endpgm
+entry:
+ %gep = getelementptr i32, ptr addrspace(4) %ptr, i32 128
+ tail call void @llvm.amdgcn.s.prefetch.data.p4(ptr addrspace(4) %gep, i32 40)
+ ret void
+}
+
define amdgpu_ps void @prefetch_data_vgpr_base_sgpr_len(ptr addrspace(4) %ptr, i32 inreg %len) {
; GCN-LABEL: prefetch_data_vgpr_base_sgpr_len:
; GCN: ; %bb.0: ; %entry
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.s.prefetch.inst.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.s.prefetch.inst.ll
index 3b47afdda009d6..26fb43764265f4 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.s.prefetch.inst.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.s.prefetch.inst.ll
@@ -78,6 +78,43 @@ define amdgpu_ps void @prefetch_inst_sgpr_imm_base_imm_len(ptr addrspace(4) inre
ret void
}
+define amdgpu_ps void @prefetch_inst_sgpr_base_large_imm_len(ptr addrspace(4) inreg %ptr) {
+; GFX12-LABEL: prefetch_inst_sgpr_base_large_imm_len:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_prefetch_inst s[0:1], 0x0, null, 8
+; GFX12-NEXT: s_endpgm
+;
+; GFX1250-LABEL: prefetch_inst_sgpr_base_large_imm_len:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
+; GFX1250-NEXT: s_mov_b64 s[64:65], 0
+; GFX1250-NEXT: v_nop
+; GFX1250-NEXT: global_prefetch_b8 v0, s[64:65] scope:SCOPE_SE
+; GFX1250-NEXT: s_prefetch_inst s[0:1], 0x0, null, 8
+; GFX1250-NEXT: s_endpgm
+ tail call void @llvm.amdgcn.s.prefetch.inst.p4(ptr addrspace(4) %ptr, i32 40)
+ ret void
+}
+
+define amdgpu_ps void @prefetch_inst_sgpr_imm_base_large_imm_len(ptr addrspace(4) inreg %ptr) {
+; GFX12-LABEL: prefetch_inst_sgpr_imm_base_large_imm_len:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_prefetch_inst s[0:1], 0x200, null, 8
+; GFX12-NEXT: s_endpgm
+;
+; GFX1250-LABEL: prefetch_inst_sgpr_imm_base_large_imm_len:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
+; GFX1250-NEXT: s_mov_b64 s[64:65], 0
+; GFX1250-NEXT: v_nop
+; GFX1250-NEXT: global_prefetch_b8 v0, s[64:65] scope:SCOPE_SE
+; GFX1250-NEXT: s_prefetch_inst s[0:1], 0x200, null, 8
+; GFX1250-NEXT: s_endpgm
+ %gep = getelementptr i32, ptr addrspace(4) %ptr, i32 128
+ tail call void @llvm.amdgcn.s.prefetch.inst.p4(ptr addrspace(4) %gep, i32 40)
+ ret void
+}
+
define amdgpu_ps void @prefetch_inst_vgpr_base_sgpr_len(ptr addrspace(4) %ptr, i32 inreg %len) {
; GFX12-LABEL: prefetch_inst_vgpr_base_sgpr_len:
; GFX12: ; %bb.0:
diff --git a/llvm/test/MC/AMDGPU/gfx12_asm_smem_err.s b/llvm/test/MC/AMDGPU/gfx12_asm_smem_err.s
index d26cc9bdca9944..5f0e1208fd7fb1 100644
--- a/llvm/test/MC/AMDGPU/gfx12_asm_smem_err.s
+++ b/llvm/test/MC/AMDGPU/gfx12_asm_smem_err.s
@@ -1,6 +1,9 @@
// NOTE: Assertions have been autogenerated by utils/update_mc_test_checks.py UTC_ARGS: --unique --sort --version 6
// RUN: not llvm-mc -triple=amdgpu12.00 %s -filetype=null 2>&1 | FileCheck --check-prefixes=NOGFX12 --implicit-check-not=error: %s
+s_atc_probe 32, s[4:5], 0x64
+// NOGFX12: :[[@LINE-1]]:13: error: expected a 5-bit unsigned immediate
+
s_atc_probe_buffer 7, null, s2
// NOGFX12: :[[@LINE-1]]:23: error: invalid operand for instruction
@@ -48,3 +51,9 @@ s_buffer_load_u8 s5, null, s0
s_buffer_prefetch_data null, 100, s10, 7
// NOGFX12: :[[@LINE-1]]:24: error: invalid operand for instruction
+
+s_prefetch_data s[18:19], 100, null, 32
+// NOGFX12: :[[@LINE-1]]:38: error: expected a 5-bit unsigned immediate
+
+s_prefetch_inst s[18:19], 100, null, 0x100
+// NOGFX12: :[[@LINE-1]]:38: error: expected a 5-bit unsigned immediate
diff --git a/llvm/test/Verifier/AMDGPU/intrinsic-prefetch.ll b/llvm/test/Verifier/AMDGPU/intrinsic-prefetch.ll
index 2c7f23a34fb10c..83cb16c5863123 100644
--- a/llvm/test/Verifier/AMDGPU/intrinsic-prefetch.ll
+++ b/llvm/test/Verifier/AMDGPU/intrinsic-prefetch.ll
@@ -6,3 +6,17 @@ entry:
tail call void @llvm.amdgcn.s.prefetch.data.p3(ptr addrspace(3) %ptr, i32 31)
ret void
}
+
+define amdgpu_ps void @global_prefetch_reserved_th(ptr addrspace(1) %ptr) {
+ ; CHECK: immarg value 7 for arg 1 out of range set
+ ; CHECK-NEXT: call void @llvm.amdgcn.global.prefetch(ptr addrspace(1) %ptr, i32 7)
+ call void @llvm.amdgcn.global.prefetch(ptr addrspace(1) %ptr, i32 7)
+ ret void
+}
+
+define amdgpu_ps void @flat_prefetch_extra_cpol_bits(ptr %ptr) {
+ ; CHECK: immarg value 2048 for arg 1 out of range set
+ ; CHECK-NEXT: call void @llvm.amdgcn.flat.prefetch(ptr %ptr, i32 2048)
+ call void @llvm.amdgcn.flat.prefetch(ptr %ptr, i32 2048)
+ ret void
+}
More information about the cfe-commits
mailing list