diff options
author | Ilia Mirkin <[email protected]> | 2016-02-04 22:58:48 -0500 |
---|---|---|
committer | Ilia Mirkin <[email protected]> | 2016-02-05 00:41:33 -0500 |
commit | 1a0fde1f52c59e0dbada03f387c8e25d9475ebbe (patch) | |
tree | d038162846770c289cecf0ab7f17200fc6282658 /src/gallium | |
parent | 2fed18b8a54319f22888e3761a6ed5cd85f9688c (diff) |
nvc0/ir: fix converting between predicate and gpr
The spill logic will insert convert ops when moving between files. It
seems like the emission logic wasn't quite ready for these converts.
Tested on fermi, and visually looked at nvdisasm output for maxwell.
Signed-off-by: Ilia Mirkin <[email protected]>
Cc: [email protected]
Diffstat (limited to 'src/gallium')
3 files changed, 41 insertions, 11 deletions
diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_gk110.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_gk110.cpp index 17cb484d2ba..0c7cd1d8137 100644 --- a/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_gk110.cpp +++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_gk110.cpp @@ -1947,10 +1947,16 @@ CodeEmitterGK110::emitInstruction(Instruction *insn) case OP_CEIL: case OP_FLOOR: case OP_TRUNC: - case OP_CVT: case OP_SAT: emitCVT(insn); break; + case OP_CVT: + if (insn->def(0).getFile() == FILE_PREDICATE || + insn->src(0).getFile() == FILE_PREDICATE) + emitMOV(insn); + else + emitCVT(insn); + break; case OP_RSQ: emitSFnOp(insn, 5 + 2 * insn->subOp); break; diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_gm107.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_gm107.cpp index 1fa0eb6da6d..dee26225b7e 100644 --- a/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_gm107.cpp +++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_gm107.cpp @@ -673,7 +673,12 @@ CodeEmitterGM107::emitMOV() (insn->sType != TYPE_F32 && !longIMMD(insn->src(0)))) { switch (insn->src(0).getFile()) { case FILE_GPR: - emitInsn(0x5c980000); + if (insn->def(0).getFile() == FILE_PREDICATE) { + emitInsn(0x5b6a0000); + emitGPR (0x08); + } else { + emitInsn(0x5c980000); + } emitGPR (0x14, insn->src(0)); break; case FILE_MEMORY_CONST: @@ -684,18 +689,32 @@ CodeEmitterGM107::emitMOV() emitInsn(0x38980000); emitIMMD(0x14, 19, insn->src(0)); break; + case FILE_PREDICATE: + emitInsn(0x50880000); + emitPRED(0x0c, insn->src(0)); + emitPRED(0x1d); + emitPRED(0x27); + break; default: assert(!"bad src file"); break; } - emitField(0x27, 4, insn->lanes); + if (insn->def(0).getFile() != FILE_PREDICATE && + insn->src(0).getFile() != FILE_PREDICATE) + emitField(0x27, 4, insn->lanes); } else { emitInsn (0x01000000); emitIMMD (0x14, 32, insn->src(0)); emitField(0x0c, 4, insn->lanes); } - emitGPR(0x00, insn->def(0)); + if (insn->def(0).getFile() == FILE_PREDICATE) { + emitPRED(0x27); + emitPRED(0x03, insn->def(0)); + emitPRED(0x00); + } else { + emitGPR(0x00, insn->def(0)); + } } void @@ -2684,11 +2703,7 @@ CodeEmitterGM107::emitInstruction(Instruction *i) emitRAM(); break; case OP_MOV: - if (insn->def(0).getFile() == FILE_GPR && - insn->src(0).getFile() != FILE_PREDICATE) - emitMOV(); - else - assert(!"R2P/P2R"); + emitMOV(); break; case OP_RDSV: emitS2R(); @@ -2700,7 +2715,10 @@ CodeEmitterGM107::emitInstruction(Instruction *i) case OP_CEIL: case OP_TRUNC: case OP_CVT: - if (isFloatType(insn->dType)) { + if (insn->op == OP_CVT && (insn->def(0).getFile() == FILE_PREDICATE || + insn->src(0).getFile() == FILE_PREDICATE)) { + emitMOV(); + } else if (isFloatType(insn->dType)) { if (isFloatType(insn->sType)) emitF2F(); else diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_nvc0.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_nvc0.cpp index 6cf69e5339f..8637db91521 100644 --- a/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_nvc0.cpp +++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_emit_nvc0.cpp @@ -2435,10 +2435,16 @@ CodeEmitterNVC0::emitInstruction(Instruction *insn) case OP_CEIL: case OP_FLOOR: case OP_TRUNC: - case OP_CVT: case OP_SAT: emitCVT(insn); break; + case OP_CVT: + if (insn->def(0).getFile() == FILE_PREDICATE || + insn->src(0).getFile() == FILE_PREDICATE) + emitMOV(insn); + else + emitCVT(insn); + break; case OP_RSQ: emitSFnOp(insn, 5 + 2 * insn->subOp); break; |