[AMDGPU][MC][GFX9] Corrected parsing of v_cndmask_b32_sdwa
See https://bugs.llvm.org/show_bug.cgi?id=43607
Reviewers: arsenm, rampitec
Differential Revision: https://reviews.llvm.org/D69095
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@375231 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp b/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
index 75e0fb8..9dd511f 100644
--- a/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
+++ b/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
@@ -1423,9 +1423,12 @@
void cvtSdwaVOP1(MCInst &Inst, const OperandVector &Operands);
void cvtSdwaVOP2(MCInst &Inst, const OperandVector &Operands);
void cvtSdwaVOP2b(MCInst &Inst, const OperandVector &Operands);
+ void cvtSdwaVOP2e(MCInst &Inst, const OperandVector &Operands);
void cvtSdwaVOPC(MCInst &Inst, const OperandVector &Operands);
void cvtSDWA(MCInst &Inst, const OperandVector &Operands,
- uint64_t BasicInstType, bool skipVcc = false);
+ uint64_t BasicInstType,
+ bool SkipDstVcc = false,
+ bool SkipSrcVcc = false);
AMDGPUOperand::Ptr defaultBLGP() const;
AMDGPUOperand::Ptr defaultCBSZ() const;
@@ -6813,7 +6816,11 @@
}
void AMDGPUAsmParser::cvtSdwaVOP2b(MCInst &Inst, const OperandVector &Operands) {
- cvtSDWA(Inst, Operands, SIInstrFlags::VOP2, true);
+ cvtSDWA(Inst, Operands, SIInstrFlags::VOP2, true, true);
+}
+
+void AMDGPUAsmParser::cvtSdwaVOP2e(MCInst &Inst, const OperandVector &Operands) {
+ cvtSDWA(Inst, Operands, SIInstrFlags::VOP2, false, true);
}
void AMDGPUAsmParser::cvtSdwaVOPC(MCInst &Inst, const OperandVector &Operands) {
@@ -6821,11 +6828,14 @@
}
void AMDGPUAsmParser::cvtSDWA(MCInst &Inst, const OperandVector &Operands,
- uint64_t BasicInstType, bool skipVcc) {
+ uint64_t BasicInstType,
+ bool SkipDstVcc,
+ bool SkipSrcVcc) {
using namespace llvm::AMDGPU::SDWA;
OptionalImmIndexMap OptionalIdx;
- bool skippedVcc = false;
+ bool SkipVcc = SkipDstVcc || SkipSrcVcc;
+ bool SkippedVcc = false;
unsigned I = 1;
const MCInstrDesc &Desc = MII.get(Inst.getOpcode());
@@ -6835,19 +6845,21 @@
for (unsigned E = Operands.size(); I != E; ++I) {
AMDGPUOperand &Op = ((AMDGPUOperand &)*Operands[I]);
- if (skipVcc && !skippedVcc && Op.isReg() &&
+ if (SkipVcc && !SkippedVcc && Op.isReg() &&
(Op.getReg() == AMDGPU::VCC || Op.getReg() == AMDGPU::VCC_LO)) {
// VOP2b (v_add_u32, v_sub_u32 ...) sdwa use "vcc" token as dst.
// Skip it if it's 2nd (e.g. v_add_i32_sdwa v1, vcc, v2, v3)
// or 4th (v_addc_u32_sdwa v1, vcc, v2, v3, vcc) operand.
// Skip VCC only if we didn't skip it on previous iteration.
+ // Note that src0 and src1 occupy 2 slots each because of modifiers.
if (BasicInstType == SIInstrFlags::VOP2 &&
- (Inst.getNumOperands() == 1 || Inst.getNumOperands() == 5)) {
- skippedVcc = true;
+ ((SkipDstVcc && Inst.getNumOperands() == 1) ||
+ (SkipSrcVcc && Inst.getNumOperands() == 5))) {
+ SkippedVcc = true;
continue;
} else if (BasicInstType == SIInstrFlags::VOPC &&
Inst.getNumOperands() == 0) {
- skippedVcc = true;
+ SkippedVcc = true;
continue;
}
}
@@ -6859,7 +6871,7 @@
} else {
llvm_unreachable("Invalid operand type");
}
- skippedVcc = false;
+ SkippedVcc = false;
}
if (Inst.getOpcode() != AMDGPU::V_NOP_sdwa_gfx10 &&
diff --git a/lib/Target/AMDGPU/VOP2Instructions.td b/lib/Target/AMDGPU/VOP2Instructions.td
index 4b38a31..f91096b 100644
--- a/lib/Target/AMDGPU/VOP2Instructions.td
+++ b/lib/Target/AMDGPU/VOP2Instructions.td
@@ -224,7 +224,7 @@
foreach _ = BoolToList<P.HasExtSDWA>.ret in
def _sdwa : VOP2_SDWA_Pseudo <opName, P> {
- let AsmMatchConverter = "cvtSdwaVOP2b";
+ let AsmMatchConverter = "cvtSdwaVOP2e";
}
foreach _ = BoolToList<P.HasExtDPP>.ret in
diff --git a/test/MC/AMDGPU/vop_sdwa.s b/test/MC/AMDGPU/vop_sdwa.s
index 4f130ca..bfd0a45 100644
--- a/test/MC/AMDGPU/vop_sdwa.s
+++ b/test/MC/AMDGPU/vop_sdwa.s
@@ -594,6 +594,11 @@
// GFX89: v_cndmask_b32_sdwa v5, v1, sext(v2), vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD ; encoding: [0xf9,0x04,0x0a,0x00,0x01,0x06,0x06,0x0e]
v_cndmask_b32_sdwa v5, v1, sext(v2), vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+v_cndmask_b32_sdwa v5, vcc_lo, v2, vcc dst_sel:DWORD dst_unused:UNUSED_PRESERVE src0_sel:DWORD src1_sel:DWORD
+// NOSICI: error: not a valid operand
+// NOVI: error: invalid operand for instruction
+// NOGFX9: error: invalid operand (violates constant bus restrictions)
+
//===----------------------------------------------------------------------===//
// Check VOPC opcodes
//===----------------------------------------------------------------------===//