LLVM ERROR: `floatFromInt` cast

[kassane]

Build and linking, using zig only:

./zig-x86_64-relsafe-espressif-linux-musl-baseline/zig build-exe start.zig -target xtensa-freestanding-none -mcpu=<cpuname>

• esp32
• esp32s3
• cnl (intel cannonlake variant for xtensa)

Skip error: add -fno-compiler-rt

LLVM Emit Object... LLVM ERROR: Cannot select: 0x7379c54b5200: f32 = fp16_to_fp 0x7379c5783cc0, float_from_int.zig:46:9 @[ floatsihf.zig:11:24 ]
  0x7379c5783cc0: i32 = or 0x7379c578d240, 0x7379c578d390, float_from_int.zig:46:9 @[ floatsihf.zig:11:24 ]
    0x7379c578d240: i32 = AssertZext 0x7379c54a17c0, ValueType:ch:i16, float_from_int.zig:46:9 @[ floatsihf.zig:11:24 ]
      0x7379c54a17c0: i32,ch = CopyFromReg 0x7379c5d20110, Register:i32 %1, float_from_int.zig:46:9 @[ floatsihf.zig:11:24 ]
        0x7379c54b60c0: i32 = Register %1
    0x7379c578d390: i32 = XtensaISD::PCREL_WRAPPER TargetConstantPool:i32<i32 31744> 0
      0x7379c5786680: i32 = TargetConstantPool<i32 31744> 0
In function: __floatsihf
[1]    3295 IOT instruction (core dumped)  ./zig-x86_64-relsafe-espressif-linux-musl-baseline/zig build-exe start.zig

zig-espressif-bootstrap/zig/lib/compiler_rt/float_from_int.zig

Lines 44 to 46 in 8d67986

	// Compute exponent
	if ((int_bits > max_exp) and (exp > max_exp)) // If exponent too large, overflow to infinity
	return @bitCast(sign_bit | @as(uT, @bitCast(inf)));

zig-espressif-bootstrap/zig/lib/compiler_rt/floatdihf.zig

Lines 10 to 12 in 8d67986

	fn __floatdihf(a: i64) callconv(.C) f16 {
	return floatFromInt(f16, a);
	}

• esp32s2 works!!

time report

===-------------------------------------------------------------------------===
                      Instruction Selection and Scheduling
===-------------------------------------------------------------------------===
  Total Execution Time: 0.0000 seconds (0.0000 wall clock)

   ---User Time---   --User+System--   ---Wall Time---  --- Name ---
   0.0000 ( 34.8%)   0.0000 ( 34.8%)   0.0000 ( 34.4%)  Instruction Selection
   0.0000 ( 15.2%)   0.0000 ( 15.2%)   0.0000 ( 15.4%)  Instruction Creation
   0.0000 ( 13.0%)   0.0000 ( 13.0%)   0.0000 ( 13.3%)  Instruction Scheduling
   0.0000 ( 10.9%)   0.0000 ( 10.9%)   0.0000 ( 11.3%)  Vector Legalization
   0.0000 ( 10.9%)   0.0000 ( 10.9%)   0.0000 ( 10.8%)  DAG Combining 1
   0.0000 (  4.3%)   0.0000 (  4.3%)   0.0000 (  4.6%)  DAG Legalization
   0.0000 (  4.3%)   0.0000 (  4.3%)   0.0000 (  4.1%)  Type Legalization
   0.0000 (  4.3%)   0.0000 (  4.3%)   0.0000 (  4.1%)  DAG Combining 2
   0.0000 (  2.2%)   0.0000 (  2.2%)   0.0000 (  2.1%)  Instruction Scheduling Cleanup
   0.0000 (100.0%)   0.0000 (100.0%)   0.0000 (100.0%)  Total

===-------------------------------------------------------------------------===
                          Pass execution timing report
===-------------------------------------------------------------------------===
  Total Execution Time: 0.0004 seconds (0.0004 wall clock)

   ---User Time---   --System Time--   --User+System--   ---Wall Time---  --- Name ---
   0.0002 ( 48.3%)   0.0000 (  0.0%)   0.0002 ( 45.2%)   0.0002 ( 44.6%)  Xtensa DAG->DAG Pattern Instruction Selection
   0.0001 ( 15.9%)   0.0000 (  0.0%)   0.0001 ( 14.9%)   0.0001 ( 14.3%)  Xtensa Assembly Printer
   0.0000 (  7.8%)   0.0000 (  0.0%)   0.0000 (  7.3%)   0.0000 (  7.0%)  Live DEBUG_VALUE analysis
   0.0000 (  3.6%)   0.0000 (  0.0%)   0.0000 (  3.4%)   0.0000 (  3.3%)  Prologue/Epilogue Insertion & Frame Finalization
   0.0000 (  3.0%)   0.0000 (  0.0%)   0.0000 (  2.8%)   0.0000 (  2.8%)  MachineDominator Tree Construction
   0.0000 (  0.0%)   0.0000 ( 13.0%)   0.0000 (  0.8%)   0.0000 (  1.4%)  Lower constant intrinsics
   0.0000 (  1.2%)   0.0000 (  0.0%)   0.0000 (  1.1%)   0.0000 (  1.1%)  Fast Register Allocator
   0.0000 (  0.9%)   0.0000 (  0.0%)   0.0000 (  0.8%)   0.0000 (  0.8%)  Free MachineFunction
   0.0000 (  0.9%)   0.0000 (  0.0%)   0.0000 (  0.8%)   0.0000 (  0.8%)  Machine Natural Loop Construction
   0.0000 (  0.0%)   0.0000 (  8.7%)   0.0000 (  0.6%)   0.0000 (  0.8%)  Remove unreachable blocks from the CFG #2
   0.0000 (  0.9%)   0.0000 (  0.0%)   0.0000 (  0.8%)   0.0000 (  0.8%)  Stack Frame Layout Analysis
   0.0000 (  0.9%)   0.0000 (  0.0%)   0.0000 (  0.8%)   0.0000 (  0.8%)  Finalize ISel and expand pseudo-instructions
   0.0000 (  0.9%)   0.0000 (  0.0%)   0.0000 (  0.8%)   0.0000 (  0.8%)  Xtensa instruction size reduction pass
   0.0000 (  0.6%)   0.0000 (  0.0%)   0.0000 (  0.6%)   0.0000 (  0.6%)  Two-Address instruction pass
   0.0000 (  0.3%)   0.0000 (  4.3%)   0.0000 (  0.6%)   0.0000 (  0.6%)  Safe Stack instrumentation pass
   0.0000 (  0.0%)   0.0000 (  4.3%)   0.0000 (  0.3%)   0.0000 (  0.6%)  Insert stack protectors
   0.0000 (  0.6%)   0.0000 (  0.0%)   0.0000 (  0.6%)   0.0000 (  0.6%)  Machine Natural Loop Construction #2
   0.0000 (  0.6%)   0.0000 (  0.0%)   0.0000 (  0.6%)   0.0000 (  0.6%)  Branch relaxation pass
   0.0000 (  0.0%)   0.0000 (  4.3%)   0.0000 (  0.3%)   0.0000 (  0.6%)  Scalarize Masked Memory Intrinsics
   0.0000 (  0.3%)   0.0000 (  0.0%)   0.0000 (  0.3%)   0.0000 (  0.6%)  Fixup Statepoint Caller Saved
   0.0000 (  0.0%)   0.0000 (  4.3%)   0.0000 (  0.3%)   0.0000 (  0.6%)  Expand vector predication intrinsics
   0.0000 (  0.3%)   0.0000 (  0.0%)   0.0000 (  0.3%)   0.0000 (  0.6%)  StackMap Liveness Analysis
   0.0000 (  0.3%)   0.0000 (  8.7%)   0.0000 (  0.8%)   0.0000 (  0.6%)  Shadow Stack GC Lowering
   0.0000 (  0.6%)   0.0000 (  0.0%)   0.0000 (  0.6%)   0.0000 (  0.6%)  MachineDominator Tree Construction #2
   0.0000 (  0.0%)   0.0000 (  8.7%)   0.0000 (  0.6%)   0.0000 (  0.6%)  Pre-ISel Intrinsic Lowering
   0.0000 (  0.6%)   0.0000 (  0.0%)   0.0000 (  0.6%)   0.0000 (  0.5%)  Eliminate PHI nodes for register allocation
   0.0000 (  0.0%)   0.0000 (  4.3%)   0.0000 (  0.3%)   0.0000 (  0.5%)  Expand large div/rem
   0.0000 (  0.6%)   0.0000 (  0.0%)   0.0000 (  0.6%)   0.0000 (  0.5%)  Machine Optimization Remark Emitter
   0.0000 (  0.3%)   0.0000 (  4.3%)   0.0000 (  0.6%)   0.0000 (  0.5%)  Expand large fp convert
   0.0000 (  0.3%)   0.0000 (  4.3%)   0.0000 (  0.6%)   0.0000 (  0.5%)  Expand Atomic instructions
   0.0000 (  0.6%)   0.0000 (  0.0%)   0.0000 (  0.6%)   0.0000 (  0.5%)  Analyze Machine Code For Garbage Collection
   0.0000 (  0.6%)   0.0000 (  0.0%)   0.0000 (  0.6%)   0.0000 (  0.5%)  Insert fentry calls
   0.0000 (  0.3%)   0.0000 (  0.0%)   0.0000 (  0.3%)   0.0000 (  0.5%)  Insert XRay ops
   0.0000 (  0.6%)   0.0000 (  0.0%)   0.0000 (  0.6%)   0.0000 (  0.5%)  Implement the 'patchable-function' attribute
   0.0000 (  0.6%)   0.0000 (  0.0%)   0.0000 (  0.6%)   0.0000 (  0.5%)  Xtensa bool reg fixup pass
   0.0000 (  0.3%)   0.0000 (  4.3%)   0.0000 (  0.6%)   0.0000 (  0.5%)  Assignment Tracking Analysis
   0.0000 (  0.0%)   0.0000 (  0.0%)   0.0000 (  0.0%)   0.0000 (  0.5%)  Lower invoke and unwind, for unwindless code generators
   0.0000 (  0.6%)   0.0000 (  0.0%)   0.0000 (  0.6%)   0.0000 (  0.5%)  Machine Optimization Remark Emitter #2
   0.0000 (  0.6%)   0.0000 (  0.0%)   0.0000 (  0.6%)   0.0000 (  0.5%)  Lazy Machine Block Frequency Analysis #2
   0.0000 (  0.6%)   0.0000 (  0.0%)   0.0000 (  0.6%)   0.0000 (  0.5%)  Contiguously Lay Out Funclets
   0.0000 (  0.0%)   0.0000 ( 13.0%)   0.0000 (  0.8%)   0.0000 (  0.5%)  Remove unreachable blocks from the CFG
   0.0000 (  0.0%)   0.0000 (  4.3%)   0.0000 (  0.3%)   0.0000 (  0.5%)  Lower Garbage Collection Instructions
   0.0000 (  0.0%)   0.0000 (  4.3%)   0.0000 (  0.3%)   0.0000 (  0.5%)  Prepare callbr
   0.0000 (  0.6%)   0.0000 (  0.0%)   0.0000 (  0.6%)   0.0000 (  0.5%)  Xtensa Hardware Loop Fixup
   0.0000 (  0.3%)   0.0000 (  0.0%)   0.0000 (  0.3%)   0.0000 (  0.3%)  Assumption Cache Tracker
   0.0000 (  0.3%)   0.0000 (  0.0%)   0.0000 (  0.3%)   0.0000 (  0.3%)  Local Stack Slot Allocation
   0.0000 (  0.3%)   0.0000 (  0.0%)   0.0000 (  0.3%)   0.0000 (  0.3%)  Xtensa fix PSRAM cache issue in the ESP32 chips
   0.0000 (  0.0%)   0.0000 (  4.3%)   0.0000 (  0.3%)   0.0000 (  0.3%)  Expand reduction intrinsics
   0.0000 (  0.3%)   0.0000 (  0.0%)   0.0000 (  0.3%)   0.0000 (  0.3%)  Xtensa Hardware Loops
   0.0000 (  0.6%)   0.0000 (  0.0%)   0.0000 (  0.6%)   0.0000 (  0.3%)  Post-RA pseudo instruction expansion pass
   0.0000 (  0.3%)   0.0000 (  0.0%)   0.0000 (  0.3%)   0.0000 (  0.3%)  Lazy Machine Block Frequency Analysis
   0.0000 (  0.3%)   0.0000 (  0.0%)   0.0000 (  0.3%)   0.0000 (  0.3%)  Create Garbage Collector Module Metadata
   0.0000 (  0.3%)   0.0000 (  0.0%)   0.0000 (  0.3%)   0.0000 (  0.3%)  Target Library Information
   0.0000 (  0.3%)   0.0000 (  0.0%)   0.0000 (  0.3%)   0.0000 (  0.3%)  Target Transform Information
   0.0000 (  0.3%)   0.0000 (  0.0%)   0.0000 (  0.3%)   0.0000 (  0.3%)  Remove Redundant DEBUG_VALUE analysis
   0.0000 (  0.3%)   0.0000 (  0.0%)   0.0000 (  0.3%)   0.0000 (  0.3%)  Profile summary info
   0.0000 (  0.3%)   0.0000 (  0.0%)   0.0000 (  0.3%)   0.0000 (  0.3%)  Machine Sanitizer Binary Metadata
   0.0000 (  0.3%)   0.0000 (  0.0%)   0.0000 (  0.3%)   0.0000 (  0.2%)  Xtensa Constant Islands
   0.0000 (  0.3%)   0.0000 (  0.0%)   0.0000 (  0.3%)   0.0000 (  0.2%)  Machine Branch Probability Analysis
   0.0000 (  0.0%)   0.0000 (  0.0%)   0.0000 (  0.0%)   0.0000 (  0.0%)  Target Pass Configuration
   0.0000 (  0.3%)   0.0000 (  0.0%)   0.0000 (  0.3%)   0.0000 (  0.0%)  Machine Module Information
   0.0003 (100.0%)   0.0000 (100.0%)   0.0004 (100.0%)   0.0004 (100.0%)  Total

===-------------------------------------------------------------------------===
                                 DWARF Emission
===-------------------------------------------------------------------------===
  Total Execution Time: 0.0002 seconds (0.0002 wall clock)

   ---User Time---   --System Time--   --User+System--   ---Wall Time---  --- Name ---
   0.0001 (100.0%)   0.0001 (100.0%)   0.0002 (100.0%)   0.0002 (100.0%)  Debug Info Emission
   0.0001 (100.0%)   0.0001 (100.0%)   0.0002 (100.0%)   0.0002 (100.0%)  Total

===-------------------------------------------------------------------------===
                        Analysis execution timing report
===-------------------------------------------------------------------------===
  Total Execution Time: 0.0000 seconds (0.0000 wall clock)

   --System Time--   --User+System--   ---Wall Time---  --- Name ---
   0.0000 ( 60.0%)   0.0000 ( 60.0%)   0.0000 ( 61.9%)  TargetLibraryAnalysis
   0.0000 ( 20.0%)   0.0000 ( 20.0%)   0.0000 ( 19.0%)  InnerAnalysisManagerProxy<FunctionAnalysisManager, Module>
   0.0000 ( 20.0%)   0.0000 ( 20.0%)   0.0000 ( 19.0%)  ProfileSummaryAnalysis
   0.0000 (100.0%)   0.0000 (100.0%)   0.0000 (100.0%)  Total

===-------------------------------------------------------------------------===
                          Pass execution timing report
===-------------------------------------------------------------------------===
  Total Execution Time: 0.0000 seconds (0.0000 wall clock)

   ---User Time---   --System Time--   --User+System--   ---Wall Time---  --- Name ---
   0.0000 ( 20.0%)   0.0000 ( 34.8%)   0.0000 ( 32.1%)   0.0000 ( 38.5%)  AlwaysInlinerPass
   0.0000 ( 40.0%)   0.0000 ( 34.8%)   0.0000 ( 35.7%)   0.0000 ( 34.4%)  AnnotationRemarksPass
   0.0000 ( 40.0%)   0.0000 ( 30.4%)   0.0000 ( 32.1%)   0.0000 ( 27.0%)  CoroConditionalWrapper
   0.0000 (100.0%)   0.0000 (100.0%)   0.0000 (100.0%)   0.0000 (100.0%)  Total

===-------------------------------------------------------------------------===
                          Pass execution timing report
===-------------------------------------------------------------------------===
  Total Execution Time: 0.0000 seconds (0.0000 wall clock)

   ---User Time---   --System Time--   --User+System--   ---Wall Time---  --- Name ---
   0.0000 ( 20.0%)   0.0000 ( 34.8%)   0.0000 ( 32.1%)   0.0000 ( 38.5%)  AlwaysInlinerPass
   0.0000 ( 40.0%)   0.0000 ( 34.8%)   0.0000 ( 35.7%)   0.0000 ( 34.4%)  AnnotationRemarksPass
   0.0000 ( 40.0%)   0.0000 ( 30.4%)   0.0000 ( 32.1%)   0.0000 ( 27.0%)  CoroConditionalWrapper
   0.0000 (100.0%)   0.0000 (100.0%)   0.0000 (100.0%)   0.0000 (100.0%)  Total

===-------------------------------------------------------------------------===
                        Analysis execution timing report
===-------------------------------------------------------------------------===
  Total Execution Time: 0.0000 seconds (0.0000 wall clock)

   --System Time--   --User+System--   ---Wall Time---  --- Name ---
   0.0000 ( 60.0%)   0.0000 ( 60.0%)   0.0000 ( 61.9%)  TargetLibraryAnalysis
   0.0000 ( 20.0%)   0.0000 ( 20.0%)   0.0000 ( 19.0%)  InnerAnalysisManagerProxy<FunctionAnalysisManager, Module>
   0.0000 ( 20.0%)   0.0000 ( 20.0%)   0.0000 ( 19.0%)  ProfileSummaryAnalysis
   0.0000 (100.0%)   0.0000 (100.0%)   0.0000 (100.0%)  Total

Reference

• https://gist.github.com/kassane/7bdb782a1984d0c6581ae7b44e1fc0c2?permalink_comment_id=4919689#gistcomment-4919689
• espressif/llvm-project@bc0d57e
• FFT HW Acceleration esp-rs/rust#190
• Optimize floating point arithmetic elkel53930/micromouse-std-esp32#1
• FPU notes on zig kubo39/esp32-resource#1
• https://www.cadence.com/content/dam/cadence-www/global/en_US/documents/tools/ip/tensilica-ip/isa-summary.pdf
• https://dl.espressif.com/github_assets/espressif/xtensa-isa-doc/releases/download/latest/Xtensa.pdf
• No, the ESP32-S2 is not faster at floating point

[kassane]

Maybe here's where the problem lies:

zig-espressif-bootstrap/zig/lib/compiler_rt/common.zig

Lines 31 to 64 in 4dcafc2

	/// This governs whether to use these symbol names for f16/f32 conversions
	/// rather than the standard names:
	/// * __gnu_f2h_ieee
	/// * __gnu_h2f_ieee
	/// Known correct configurations:
	/// x86_64-freestanding-none => true
	/// x86_64-linux-none => true
	/// x86_64-linux-gnu => true
	/// x86_64-linux-musl => true
	/// x86_64-linux-eabi => true
	/// arm-linux-musleabihf => true
	/// arm-linux-gnueabihf => true
	/// arm-linux-eabihf => false
	/// wasm32-wasi-musl => false
	/// wasm32-freestanding-none => false
	/// x86_64-windows-gnu => true
	/// x86_64-windows-msvc => true
	/// any-macos-any => false
	pub const gnu_f16_abi = switch (builtin.cpu.arch) {
	.wasm32,
	.wasm64,
	.riscv64,
	.riscv32,
	=> false,
	.x86, .x86_64 => true,
	.arm, .armeb, .thumb, .thumbeb => switch (builtin.abi) {
	.eabi, .eabihf => false,
	else => true,
	},
	else => !builtin.os.tag.isDarwin(),
	};

zig-espressif-bootstrap/zig/lib/compiler_rt/common.zig

Lines 80 to 104 in 4dcafc2

	/// AArch64 is the only ABI (at the moment) to support f16 arguments without the
	/// need for extending them to wider fp types.
	/// TODO remove this; do this type selection in the language rather than
	/// here in compiler-rt.
	pub fn F16T(comptime OtherType: type) type {
	return switch (builtin.cpu.arch) {
	.arm, .armeb, .thumb, .thumbeb => if (std.Target.arm.featureSetHas(builtin.cpu.features, .has_v8))
	switch (builtin.abi.floatAbi()) {
	.soft => u16,
	.hard => f16,
	}
	else
	u16,
	.aarch64, .aarch64_be, .aarch64_32 => f16,
	.riscv64 => if (builtin.zig_backend == .stage1) u16 else f16,
	.x86, .x86_64 => if (builtin.target.isDarwin()) switch (OtherType) {
	// Starting with LLVM 16, Darwin uses different abi for f16
	// depending on the type of the other return/argument..???
	f32, f64 => u16,
	f80, f128 => f16,
	else => unreachable,
	} else f16,
	else => u16,
	};
	}

[kassane]

Based on https://github.com/kassane/zig-espressif-bootstrap/blob/xtensa/zig/lib/std/Target/xtensa.zig esp32s2 not use S32C1I and Single FP:

$> ./zig-x86_64-relsafe-espressif-linux-musl-baseline/zig build-exe start.zig \
    -target xtensa-freestanding-none -mcpu=<esp32|esp32s3>-fp
LLVM Emit Object... error: <unknown>:0: Undefined temporary symbol .LBB385_2

$> ./zig-x86_64-relsafe-espressif-linux-musl-baseline/zig build-exe start.zig \
    -target xtensa-freestanding-none -mcpu=<esp32|esp32s3>-fp-s32c1i

[kassane]

Fix:

• [Xtensa] Fix fp16 conversion espressif/llvm-project#92