This patch introduces a new command-line option for clang, namely,
amdgpu-precise-mem-op (or precise-memory in the backend). When this option is specified, a waitcnt
instruction is generated after each memory load/store instruction. The
counter values are always 0, but which counters are involved depends on
the memory instruction.
---------
Co-authored-by: Jun Wang <jun.wang7@amd.com>
MIParser checks that implicit operands match the instruction definition,
so they have to be $vcc even in wave32 mode. Use the mirFileLoaded hook
to fix them after MIParser's checks, converting them to $vcc_lo which is
what that rest of CodeGen expects.
This is all just extending the fixImplicitOperands hack which was
introduced with GFX10, but at least it makes it possible to write a MIR
test which creates the same instructions that normal CodeGen would
generate.
This attempts to standardize and extend some of the insert vector
element lowering. Most notably:
- More types are handled by splitting illegal vectors.
- The index type for G_INSERT_VECTOR_ELT is canonicalized to
TLI.getVectorIdxTy(), similar to extact_vector_element.
- Some of the existing patterns now have the index type specified to
make sure they can apply to GISel too.
- The C++ selection code has been removed, relying on tablegen patterns.
- G_INSERT_VECTOR_ELT with small GPR input elements are pre-selected to
use a i32 type, allowing the existing patterns to apply.
- Variable index inserts are lowered in post-legalizer lowering,
expanding into a stack store and reload.
The conversion expansions did not properly handle bfloat types.
I'm not certain that these expansions are completely correct;
I don't have any experience with AMDGPU or the ability to run
anything to test it.
Note that it doesn't seem like AMDGPU with GlobalISel can
handle fptrunc of float to bfloat, which is needed for itofp.
I've omitted the GISEL run for the bfloat case.
This fixes#85379.
Call generateWaitcnt unconditionally at the end of
SIInsertWaitcnts::insertWaitcntInBlock. Even if we don't need to
generate a new waitcnt instruction it has the effect of combining or
removing redundant waitcnts that were already present. Tests show
various small improvements in waitcnt placement.
expand-large-divrem cannot handle vector types.
If overly large vector element types survive into
isel, they will likely be scalarized there, but since
isel cannot handle scalar integer types of that size,
it will assert.
Handle vector types in expand-large-divrem by
scalarizing them and then expanding the scalar type
operation. For large vectors, this results in a
*massive* code expansion, but it's better than
asserting.
The earlier implementation on AMDGPU used explicit token operands at
SI_CALL and SI_CALL_ISEL. This is now replaced with CONVERGENCECTRL_GLUE
operands, with the following effects:
- The treatment of tokens at call-like operations is now consistent with
the treatment at intrinsics.
- Support for tail calls using implicit tokens at SI_TCRETURN "just
works".
- The extra parameter at call-like instructions is eliminated, thus
restoring those instructions and their handling to the original state.
The new glue node is placed after the existing glue node for the
outgoing call parameters, which seems to not interfere with selection of
the call-like nodes.
Previously we wouldn't remove dead copies from basic blocks with
successors. The comment said we didn't want to trust the live-in lists.
The comment is very old so I'm not sure if that's still a concern today.
This patch checks the live-in lists and removes copies from
MaybeDeadCopies if they are referenced by any live-ins in any
successors. We only do this if the tracksLiveness property is set. If
that property is not set, we retain the old behavior.
The AMDGPUSimplifyLibCalls pass was lowering function calls with the
strictfp attribute to sequences that included function calls incorrectly
lacking the attribute. This patch corrects that.
The pass now also emits the correct constrained fp call instead of
normal FP instructions when in a function with the strictfp attribute.
Replacing non-constrained calls with constrained calls when required
is still on the IRBuilder's TODO list.
Reverts llvm/llvm-project#81394
This reverts commit 3ac243bc0d7922d083af2cf025247b5698556062.
It is not handling RSrc registers s0-s3 correctly. This leads to a
broken test, where it expects s0-s3 as function argument and uses it as
RSrc register as well.
We need to re-visit the patch, but apparently we only want to have s0-s3
as
argument registers if we don't need them as RSrc registers.
The IR for a double-to-i129 conversion looks like this in one of the
blocks in compiler-rt:
%cmp5.i = icmp ult i16 %3, -129, !dbg !24
But in ExpandLargeFpConvert, it looks like:
%13 = icmp ult i129 %12, 4294967167, !dbg !19
ExpandLargeFpConvert is wrong; the value should have been
signed before negating, but instead we get a very large
unsigned value. Another value in the same pass also has this
issue.
Rename the intrinsics to close to the instruction mnemonic names:
Use global_load_tr_b64 and global_load_tr_b128 instead of
global_load_tr.
This patch also removes f16/bf16 versions of builtins/intrinsics. To
simplify the design, we should avoid enumerating all possible types in
implementing builtins. We can always use bitcast.
`ST.getMaxNumVGPRs(MF)` lowers to `AMDGPUBaseInfo.cpp:getTotalNumVGPRs`
which returns 512 for gfx90a. This is subsequently limited by
`AMDGPUBaseInfo:getAddressableNumVGPRs()`, which also returns 512 for
gfx90a. The ISA states we can have a total of 512 registers, but a
maximum of only 256 of each of AGPR and VGPR (gfx90a 3.6.4).
Therefore, in unified register file case, `ST.getMaxNumVGPRs(MF)`
calculates the maximum number of combined VGPR + AGPR. But, it is
currently used as the limit for accvgpr and as the limit for archvgpr.
This patch uses it as the combined limit, and accounts for the maximum addressable arch/acc VGPRs when calculating the per RegClass limits.
It is not unreasonable to think other clients of getTotalNumVGPRs are
using it in the wrong way.
buffer_load instructions that use TFE also need to zero initialize
return values similar to how the image instructions currently work. Add
support for this with standard zero init of all results + zero init of
just TFE flag when enable-prt-strict-null subtarget feature is disabled.
Here we introduce three new GMIR instructions to cover a set of trap
intrinsics. The idea behind it is that generic intrinsics shouldn't be
used with G_INTRINSIC opcode.
These new instructions can match perfectly with existing trap ISD nodes.
It allows X86, AArch64, RISCV and Mips to reuse SelectionDAG patterns for
selection and avoid manual selection. However AMDGPU is an exception. It
selects traps during legalization regardless SelectionDAG or GlobalISel.
Since there are not many places where traps are used, this change
attempts to clean up all the usages of G_INTRINSIC with trap intrinsics. So,
there is no stage when both G_TRAP and
G_INTRINSIC_W_SIDE_EFFECTS(@llvm.trap) are allowed.
Presently the atomic optimizer supports only 32-bit operations. Plan is
to extend the atomic optimizer for 64-bit operations for compute and
graphics. This patch extends support for double type for `uniform
values` only. Going forward, will extend the support for divergent
values. Adding support for divergent values requires
extending/legalizing readfirstlane, readlane, writelane, etc ops for
64-bit operations to avoid `bitcast` noise that we have currently.
---------
Authored-by: Pravin Jagtap <Pravin.Jagtap@amd.com>
Refactor the logic that checks if a module contains mixed
absolute/non-lowered LDS GVs.
The check now happens latter when the "worklists" are formed. This is
because in some cases (OpenMP) we can have non-lowered GVs in a lowered
module, and this is normal because those GVs are just unused and removed
from the list at some point before the end of `getUsesOfLDSByFunction`.
Doing the check later ensures that if a mixed module is spotted, then
it's a _real_ mixed module that needs rejection, not a module containing
an intentionally ignored GV.
SIMachineFunctionInfo has a scan of the function body for inline asm
which may use AGPRs, or callees in SIMachineFunctionInfo. Move this
into the attributor, so it actually works interprocedurally.
Could probably avoid most of the test churn if this bothered to avoid
adding this on subtargets without AGPRs. We should also probably
try to delete the MIR scan in usesAGPRs but it seems to be trickier
to eliminate.
When i1 true is used as an index, SExt extends it to i32 -1. This would
cause BitVector to overflow.
The language manual have specified that the index shall be treated as an
unsigned number, this patch fixes that.
(https://llvm.org/docs/LangRef.html#insertelement-instruction)
This patch fixes#85717
---------
Signed-off-by: Peter Rong <PeterRong96@gmail.com>
Added common check for DPP and Iterative strategies for uniform value
case since optimization applied is same.
Authored-by: Pravin Jagtap <Pravin.Jagtap@amd.com>
Update PromoteAllocaToVector so it considers the whole function before promoting allocas.
Allocas are scored & sorted so the highest value ones are seen first. The budget is now per function instead of per alloca.
Passed internal performance testing.
- Use computeMaxCallFrameSize() in PEI::calculateCallFrameInfo() instead of duplicating the code.
- Set AdjustsStack in FinalizeISel instead of in computeMaxCallFrameSize().
This change allows us to use `--lto-partitions` in some cases (not at
all guaranteed it works perfectly), as LDS is lowered before the module
is split for parallel codegen.
We must run LowerLDS before splitting modules as it needs to see all
callers of functions with LDS to properly lower them.
[GlobalISel] Implement convergence control tokens and intrinsics in GMIR
In the IR translator, convert the LLVM token type to LLT::token(), which is an
alias for the s0 type. These show up as implicit uses on convergent operations.
Differential Revision: https://reviews.llvm.org/D158147
