Currently machine LICM hoist PC_ADD_REL_OFFSET out of loops, causes
register pressure when function calls are deep in loops. This is a main
cause of sgpr spill for programs containing large number of function
calls in loops.
This patch marks PC_ADD_REL_OFFSET as rematerializable, which eliminates
sgpr spills due to function calls in loops.
Summary:
The old `s_memtime` instruction was supported until the GFX10
architecture. Although this instruction has a higher latency than the
new shader counter, it's much more usable as a processor clock as it is
a full 64-bit counter. The new shader counter is only a 20-bit counter,
which makes it difficult to use as a standard cycle counter as it will
overflow in a few milliseconds. This patch suggests preferring
`s_memtime` for this instrinsic if it is still available.
Extra space causes the checks generated by update_mir_test_checks to be
unavailable.
```
# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 4
# RUN: llc -mtriple=x86_64-- -o - %s -run-pass=none -verify-machineinstrs -simplify-mir | FileCheck %s
---
name: foo
body: |
; CHECK-LABEL: name: foo
; CHECK: bb.0:
; CHECK-NEXT: successors:
; CHECK-NEXT: {{ $}}
; CHECK-NEXT: {{ $}}
; CHECK-NEXT: bb.1:
; CHECK-NEXT: RET 0, $eax
bb.0:
successors:
bb.1:
RET 0, $eax
...
```
The failure log is as follows:
```
llvm/test/CodeGen/MIR/X86/unreachable-block-print.mir:9:16: error: CHECK-NEXT: is on the same line as previous match
; CHECK-NEXT: {{ $}}
^
<stdin>:21:13: note: 'next' match was here
successors:
^
<stdin>:21:13: note: previous match ended here
successors:
```
PAL uses ELF REL (not RELA) relocations which can only store a 32-bit
addend in the instruction, even for reloc types like R_AMDGPU_ABS32_HI
which require the upper 32 bits of a 64-bit address calculation to be
correct. This means that it is not safe to fold an arbitrary offset into
a GlobalAddressSDNode, so stop doing that.
In practice this is mostly a problem for small negative offsets which do
not work as expected because PAL treats the 32-bit addend as unsigned.
Make the wmma intrinsic type signatures to be canonical. We need
a type signature as long as the type is not fixed. However, when an
argument's type matches a previous argument's type, we do not need the
signature for this argument.
This patch fixes three general cases:
1. add missing signatures
2. remove signatures for matching arguments
3. reorer the signatures -- return type signature should always appear
first
MUL24 can now return a i64 for i32 operands, but the combine was never
updated to handle this case. Extend the operand when rewriting the ADD
to handle it.
Fixes SWDEV-436654
This commit extends separate-const-offset-from-gep to look at the
newly-added `disjoint` flag on `or` instructions so as to preserve
additional opportunities for optimization.
The tests were pre-committed in #76972.
This patch canonicalizes getelementptr instructions with constant
indices to use the `i8` source element type. This makes it easier for
optimizations to recognize that two GEPs are identical, because they
don't need to see past many different ways to express the same offset.
This is a first step towards
https://discourse.llvm.org/t/rfc-replacing-getelementptr-with-ptradd/68699.
This is limited to constant GEPs only for now, as they have a clear
canonical form, while we're not yet sure how exactly to deal with
variable indices.
The test llvm/test/Transforms/PhaseOrdering/switch_with_geps.ll gives
two representative examples of the kind of optimization improvement we
expect from this change. In the first test SimplifyCFG can now realize
that all switch branches are actually the same. In the second test it
can convert it into simple arithmetic. These are representative of
common optimization failures we see in Rust.
Fixes https://github.com/llvm/llvm-project/issues/69841.
One V_NOP or unrelated VALU instruction in between is required for
correctness when matrix A or B of current WMMA instruction overlaps with
matrix D of previous WMMA instruction.
Remaining cases of WMMA operand overlaps are handled by the hardware and
do not require handling in hazard recognizer.
Hardware may stall in cases where:
- matrix C of current WMMA instruction overlaps with matrix D of
previous WMMA instruction
- VALU instruction reads matrix D of previous WMMA instruction
- matrix A,B or C of WMMA instruction reads result of previous VALU
instruction
Implement PhiLoweringHelper for GlobalISel in DivergenceLoweringHelper.
Use machine uniformity analysis to find divergent i1 phis and select
them as lane mask phis in same way SILowerI1Copies select VReg_1 phis.
Note that divergent i1 phis include phis created by LCSSA and all cases
of uses outside of cycle are actually covered by "lowering LCSSA phis".
GlobalISel lane masks are registers with sgpr register class and S1 LLT.
TODO: General goal is that instructions created in this pass are fully
instruction-selected so that selection of lane mask phis is not split
across multiple passes.
patch 3 from: https://github.com/llvm/llvm-project/pull/73337
CSR SGPR spilling currently uses the early available physical VGPRs. It
currently imposes a high register pressure while trying to allocate
large VGPR tuples within the default register budget.
This patch changes the spilling strategy by picking the VGPRs in the
reverse order, the highest available VGPR first and later after regalloc
shift them back to the lowest available range. With that, the initial
VGPRs would be available for allocation and possibility
of finding large number of contiguous registers will be more.
int_amdgcn_global_load_tr did not specify non-temporal load transpose,
thus we should
not genetrate the non-temporal hint for the load. We need to implement
getTgtMemIntrinsic
to create the corresponding MemSDNode. And we don't set the non-temporal
flag because
the intrinsic did not specify it.
NOTE: We need to implement getTgtMemIntrinsic for any memory intrinsics.
VGPRs used for spilling do not require explicit reservation with MRI.
freezeReservedRegs() executed before register allocation ensures these
are placed in the reserve set.
The only pass after SILowerSGPRSpills is SIPreAllocateWWMRegs which
explicitly tests for interference before register allocation so should
not reuse a WWM VGPR holding spill data. reserveReg prevents calculation
of correct liveness for physical registers which could be used to extend
SIPreAllocateWWMRegs.
This patch tweaks two AMDGPU passes to use iterators rather than
instruction pointers for expressing an insertion point. This is needed
to accurately support DPValues, the non-instruction storage object for
debug-info.
Two tests were sensitive to this change (variable assignments were being
put in the wrong place), and I've added extra run-lines with the "try
new debug-info..." flag. These get tested on our public buildbot to
ensure they continue to work accurately.
SIMemoryLegalizer tests were slow, with most of them taking 4.5 to 5.3s
to complete and that's on a fast machine. I also recall seeing them in
the slowest tests list on build bots.
This removes the verify-machineinstrs option from these tests to speed
them up, bringing the slowest test down to +-2s.
Verifier still runs in EXPENSIVE_CHECKS builds.
Named '.amdhsa_code_object_version'. This directive sets the
e_ident[ABIVERSION] in the ELF header, and should be used as the assumed
COV for the rest of the asm file.
This commit also weakens the --amdhsa-code-object-version CL flag.
Previously, the CL flag took precedence over the IR flag. Now the IR
flag/asm directive take precedence over the CL flag. This is implemented
by merging a few COV-checking functions in AMDGPUBaseInfo.h.
That instruction is not supported on GFX12.
Added a testcase which previously crashed without this change.
Co-authored-by: pvanhout <pierre.vanhoutryve@amd.com>
Add custom lowering for ctlz.i8 to avoid multiple add/sub operations.
---------
Co-authored-by: Leon Clark <leoclark@amd.com>
Co-authored-by: Matt Arsenault <Matthew.Arsenault@amd.com>
Support new amdgcn_global_load_tr instructions for load with transpose.
* MC layer support for GLOBAL_LOAD_TR_B64/GLOBAL_LOAD_TR_B128
* Intrinsic int_amdgcn_global_load_tr
* Clang builtins amdgcn_global_load_tr*
New pseudos were added for instructions that were natively VOP3 on
GFX11: V_ADD_F64_pseudo, V_MUL_F64_pseudo, V_MIN_NUM_F64, V_MAX_NUM_F64,
V_LSHLREV_B64_pseudo
---------
Co-authored-by: Mirko Brkusanin <Mirko.Brkusanin@amd.com>
Endoding is VOP3P. Tagged as deep/machine learning instructions. i32
type (v4fp8 or v4bf8 packed in i32) is used for src0 and src1. src0 and
src1 have no src_modifiers. src2 is f32 and has src_modifiers: f32
fneg(neg_lo[2]) and f32 fabs(neg_hi[2]).
---------
Co-authored-by: Petar Avramovic <Petar.Avramovic@amd.com>
Update SIMemoryLegalizer and SIInsertWaitcnts to use separate wait
instructions per counter (e.g. S_WAIT_LOADCNT) and split VMCNT into
separate LOADCNT, SAMPLECNT and BVHCNT counters.
https://github.com/llvm/llvm-project/pull/70634 has disabled use
of potentially negative scratch offsets, but we can use it on GFX12.
---------
Co-authored-by: Stanislav Mekhanoshin <Stanislav.Mekhanoshin@amd.com>
LDA DMA loads increase VMCNT and a load from the LDS stored must wait on
this counter to only read memory after it is written. Wait count
insertion pass does not track memory dependencies, it tracks register
dependencies. To model the LDS dependency a pseudo register is used in
the scoreboard, acting like if LDS DMA writes it and LDS load reads it.
This patch adds 8 more pseudo registers to use for independent LDS
locations if we can prove they are disjoint using alias analysis.
Fixes: SWDEV-433427
We currently force users to use a negative contant in the intrinsic
call. Changing it zext would break existing programs, so just sign
extend an argument.
