c349d7f4ff
The main difference is that this preserves intermediate rounding steps, which the other route doesn't. This aligns bfloat16 more with half floats, which use this path on most targets. I didn't understand what the difference was between these softening approaches when I first added bfloat lowerings, would be nice if we only had one of them. Based on @pengfei 's D131502 Differential Revision: https://reviews.llvm.org/D133207
321 lines
10 KiB
LLVM
321 lines
10 KiB
LLVM
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
|
|
; RUN: llc < %s -mtriple=x86_64-linux-gnu | FileCheck %s
|
|
|
|
define void @add(ptr %pa, ptr %pb, ptr %pc) nounwind {
|
|
; CHECK-LABEL: add:
|
|
; CHECK: # %bb.0:
|
|
; CHECK-NEXT: pushq %rbx
|
|
; CHECK-NEXT: movq %rdx, %rbx
|
|
; CHECK-NEXT: movzwl (%rsi), %eax
|
|
; CHECK-NEXT: shll $16, %eax
|
|
; CHECK-NEXT: movd %eax, %xmm1
|
|
; CHECK-NEXT: movzwl (%rdi), %eax
|
|
; CHECK-NEXT: shll $16, %eax
|
|
; CHECK-NEXT: movd %eax, %xmm0
|
|
; CHECK-NEXT: addss %xmm1, %xmm0
|
|
; CHECK-NEXT: callq __truncsfbf2@PLT
|
|
; CHECK-NEXT: movd %xmm0, %eax
|
|
; CHECK-NEXT: movw %ax, (%rbx)
|
|
; CHECK-NEXT: popq %rbx
|
|
; CHECK-NEXT: retq
|
|
%a = load bfloat, ptr %pa
|
|
%b = load bfloat, ptr %pb
|
|
%add = fadd bfloat %a, %b
|
|
store bfloat %add, ptr %pc
|
|
ret void
|
|
}
|
|
|
|
define bfloat @add2(bfloat %a, bfloat %b) nounwind {
|
|
; CHECK-LABEL: add2:
|
|
; CHECK: # %bb.0:
|
|
; CHECK-NEXT: pushq %rax
|
|
; CHECK-NEXT: movd %xmm0, %eax
|
|
; CHECK-NEXT: movd %xmm1, %ecx
|
|
; CHECK-NEXT: shll $16, %ecx
|
|
; CHECK-NEXT: movd %ecx, %xmm1
|
|
; CHECK-NEXT: shll $16, %eax
|
|
; CHECK-NEXT: movd %eax, %xmm0
|
|
; CHECK-NEXT: addss %xmm1, %xmm0
|
|
; CHECK-NEXT: callq __truncsfbf2@PLT
|
|
; CHECK-NEXT: popq %rax
|
|
; CHECK-NEXT: retq
|
|
%add = fadd bfloat %a, %b
|
|
ret bfloat %add
|
|
}
|
|
|
|
define void @add_double(ptr %pa, ptr %pb, ptr %pc) nounwind {
|
|
; CHECK-LABEL: add_double:
|
|
; CHECK: # %bb.0:
|
|
; CHECK-NEXT: pushq %rbp
|
|
; CHECK-NEXT: pushq %r14
|
|
; CHECK-NEXT: pushq %rbx
|
|
; CHECK-NEXT: movq %rdx, %r14
|
|
; CHECK-NEXT: movq %rsi, %rbx
|
|
; CHECK-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
|
|
; CHECK-NEXT: callq __truncdfbf2@PLT
|
|
; CHECK-NEXT: movd %xmm0, %ebp
|
|
; CHECK-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
|
|
; CHECK-NEXT: callq __truncdfbf2@PLT
|
|
; CHECK-NEXT: movd %xmm0, %eax
|
|
; CHECK-NEXT: shll $16, %eax
|
|
; CHECK-NEXT: movd %eax, %xmm1
|
|
; CHECK-NEXT: shll $16, %ebp
|
|
; CHECK-NEXT: movd %ebp, %xmm0
|
|
; CHECK-NEXT: addss %xmm1, %xmm0
|
|
; CHECK-NEXT: callq __truncsfbf2@PLT
|
|
; CHECK-NEXT: movd %xmm0, %eax
|
|
; CHECK-NEXT: shll $16, %eax
|
|
; CHECK-NEXT: movd %eax, %xmm0
|
|
; CHECK-NEXT: cvtss2sd %xmm0, %xmm0
|
|
; CHECK-NEXT: movsd %xmm0, (%r14)
|
|
; CHECK-NEXT: popq %rbx
|
|
; CHECK-NEXT: popq %r14
|
|
; CHECK-NEXT: popq %rbp
|
|
; CHECK-NEXT: retq
|
|
%la = load double, ptr %pa
|
|
%a = fptrunc double %la to bfloat
|
|
%lb = load double, ptr %pb
|
|
%b = fptrunc double %lb to bfloat
|
|
%add = fadd bfloat %a, %b
|
|
%dadd = fpext bfloat %add to double
|
|
store double %dadd, ptr %pc
|
|
ret void
|
|
}
|
|
|
|
define double @add_double2(double %da, double %db) nounwind {
|
|
; CHECK-LABEL: add_double2:
|
|
; CHECK: # %bb.0:
|
|
; CHECK-NEXT: pushq %rbx
|
|
; CHECK-NEXT: subq $16, %rsp
|
|
; CHECK-NEXT: movsd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
|
|
; CHECK-NEXT: callq __truncdfbf2@PLT
|
|
; CHECK-NEXT: movd %xmm0, %ebx
|
|
; CHECK-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 8-byte Folded Reload
|
|
; CHECK-NEXT: # xmm0 = mem[0],zero
|
|
; CHECK-NEXT: callq __truncdfbf2@PLT
|
|
; CHECK-NEXT: movd %xmm0, %eax
|
|
; CHECK-NEXT: shll $16, %eax
|
|
; CHECK-NEXT: movd %eax, %xmm1
|
|
; CHECK-NEXT: shll $16, %ebx
|
|
; CHECK-NEXT: movd %ebx, %xmm0
|
|
; CHECK-NEXT: addss %xmm1, %xmm0
|
|
; CHECK-NEXT: callq __truncsfbf2@PLT
|
|
; CHECK-NEXT: movd %xmm0, %eax
|
|
; CHECK-NEXT: shll $16, %eax
|
|
; CHECK-NEXT: movd %eax, %xmm0
|
|
; CHECK-NEXT: cvtss2sd %xmm0, %xmm0
|
|
; CHECK-NEXT: addq $16, %rsp
|
|
; CHECK-NEXT: popq %rbx
|
|
; CHECK-NEXT: retq
|
|
%a = fptrunc double %da to bfloat
|
|
%b = fptrunc double %db to bfloat
|
|
%add = fadd bfloat %a, %b
|
|
%dadd = fpext bfloat %add to double
|
|
ret double %dadd
|
|
}
|
|
|
|
define void @add_constant(ptr %pa, ptr %pc) nounwind {
|
|
; CHECK-LABEL: add_constant:
|
|
; CHECK: # %bb.0:
|
|
; CHECK-NEXT: pushq %rbx
|
|
; CHECK-NEXT: movq %rsi, %rbx
|
|
; CHECK-NEXT: movzwl (%rdi), %eax
|
|
; CHECK-NEXT: shll $16, %eax
|
|
; CHECK-NEXT: movd %eax, %xmm0
|
|
; CHECK-NEXT: addss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
|
|
; CHECK-NEXT: callq __truncsfbf2@PLT
|
|
; CHECK-NEXT: movd %xmm0, %eax
|
|
; CHECK-NEXT: movw %ax, (%rbx)
|
|
; CHECK-NEXT: popq %rbx
|
|
; CHECK-NEXT: retq
|
|
%a = load bfloat, ptr %pa
|
|
%add = fadd bfloat %a, 1.0
|
|
store bfloat %add, ptr %pc
|
|
ret void
|
|
}
|
|
|
|
define bfloat @add_constant2(bfloat %a) nounwind {
|
|
; CHECK-LABEL: add_constant2:
|
|
; CHECK: # %bb.0:
|
|
; CHECK-NEXT: pushq %rax
|
|
; CHECK-NEXT: movd %xmm0, %eax
|
|
; CHECK-NEXT: shll $16, %eax
|
|
; CHECK-NEXT: movd %eax, %xmm0
|
|
; CHECK-NEXT: addss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
|
|
; CHECK-NEXT: callq __truncsfbf2@PLT
|
|
; CHECK-NEXT: popq %rax
|
|
; CHECK-NEXT: retq
|
|
%add = fadd bfloat %a, 1.0
|
|
ret bfloat %add
|
|
}
|
|
|
|
define void @store_constant(ptr %pc) nounwind {
|
|
; CHECK-LABEL: store_constant:
|
|
; CHECK: # %bb.0:
|
|
; CHECK-NEXT: movw $16256, (%rdi) # imm = 0x3F80
|
|
; CHECK-NEXT: retq
|
|
store bfloat 1.0, ptr %pc
|
|
ret void
|
|
}
|
|
|
|
define void @fold_ext_trunc(ptr %pa, ptr %pc) nounwind {
|
|
; CHECK-LABEL: fold_ext_trunc:
|
|
; CHECK: # %bb.0:
|
|
; CHECK-NEXT: movzwl (%rdi), %eax
|
|
; CHECK-NEXT: movw %ax, (%rsi)
|
|
; CHECK-NEXT: retq
|
|
%a = load bfloat, ptr %pa
|
|
%ext = fpext bfloat %a to float
|
|
%trunc = fptrunc float %ext to bfloat
|
|
store bfloat %trunc, ptr %pc
|
|
ret void
|
|
}
|
|
|
|
define bfloat @fold_ext_trunc2(bfloat %a) nounwind {
|
|
; CHECK-LABEL: fold_ext_trunc2:
|
|
; CHECK: # %bb.0:
|
|
; CHECK-NEXT: retq
|
|
%ext = fpext bfloat %a to float
|
|
%trunc = fptrunc float %ext to bfloat
|
|
ret bfloat %trunc
|
|
}
|
|
|
|
define <8 x bfloat> @addv(<8 x bfloat> %a, <8 x bfloat> %b) nounwind {
|
|
; CHECK-LABEL: addv:
|
|
; CHECK: # %bb.0:
|
|
; CHECK-NEXT: pushq %rbp
|
|
; CHECK-NEXT: pushq %r15
|
|
; CHECK-NEXT: pushq %r14
|
|
; CHECK-NEXT: pushq %r13
|
|
; CHECK-NEXT: pushq %r12
|
|
; CHECK-NEXT: pushq %rbx
|
|
; CHECK-NEXT: subq $56, %rsp
|
|
; CHECK-NEXT: movq %xmm0, %rcx
|
|
; CHECK-NEXT: movq %rcx, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
|
|
; CHECK-NEXT: movq %rcx, %rax
|
|
; CHECK-NEXT: shrq $32, %rax
|
|
; CHECK-NEXT: movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
|
|
; CHECK-NEXT: movq %xmm1, %rdx
|
|
; CHECK-NEXT: movq %rdx, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
|
|
; CHECK-NEXT: movq %rdx, %rax
|
|
; CHECK-NEXT: shrq $32, %rax
|
|
; CHECK-NEXT: movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
|
|
; CHECK-NEXT: movq %rcx, %rax
|
|
; CHECK-NEXT: shrq $48, %rax
|
|
; CHECK-NEXT: movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
|
|
; CHECK-NEXT: movq %rdx, %rax
|
|
; CHECK-NEXT: shrq $48, %rax
|
|
; CHECK-NEXT: movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
|
|
; CHECK-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]
|
|
; CHECK-NEXT: movq %xmm0, %rbx
|
|
; CHECK-NEXT: movq %rbx, %rax
|
|
; CHECK-NEXT: shrq $32, %rax
|
|
; CHECK-NEXT: movq %rax, (%rsp) # 8-byte Spill
|
|
; CHECK-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
|
|
; CHECK-NEXT: movq %xmm0, %rbp
|
|
; CHECK-NEXT: movq %rbp, %r15
|
|
; CHECK-NEXT: shrq $32, %r15
|
|
; CHECK-NEXT: movq %rbx, %r13
|
|
; CHECK-NEXT: shrq $48, %r13
|
|
; CHECK-NEXT: movq %rbp, %r12
|
|
; CHECK-NEXT: shrq $48, %r12
|
|
; CHECK-NEXT: movl %ebp, %eax
|
|
; CHECK-NEXT: andl $-65536, %eax # imm = 0xFFFF0000
|
|
; CHECK-NEXT: movd %eax, %xmm1
|
|
; CHECK-NEXT: movl %ebx, %eax
|
|
; CHECK-NEXT: andl $-65536, %eax # imm = 0xFFFF0000
|
|
; CHECK-NEXT: movd %eax, %xmm0
|
|
; CHECK-NEXT: addss %xmm1, %xmm0
|
|
; CHECK-NEXT: callq __truncsfbf2@PLT
|
|
; CHECK-NEXT: movd %xmm0, %r14d
|
|
; CHECK-NEXT: shll $16, %r14d
|
|
; CHECK-NEXT: shll $16, %ebp
|
|
; CHECK-NEXT: movd %ebp, %xmm1
|
|
; CHECK-NEXT: shll $16, %ebx
|
|
; CHECK-NEXT: movd %ebx, %xmm0
|
|
; CHECK-NEXT: addss %xmm1, %xmm0
|
|
; CHECK-NEXT: callq __truncsfbf2@PLT
|
|
; CHECK-NEXT: movd %xmm0, %eax
|
|
; CHECK-NEXT: movzwl %ax, %ebx
|
|
; CHECK-NEXT: orl %r14d, %ebx
|
|
; CHECK-NEXT: shll $16, %r12d
|
|
; CHECK-NEXT: movd %r12d, %xmm1
|
|
; CHECK-NEXT: shll $16, %r13d
|
|
; CHECK-NEXT: movd %r13d, %xmm0
|
|
; CHECK-NEXT: addss %xmm1, %xmm0
|
|
; CHECK-NEXT: callq __truncsfbf2@PLT
|
|
; CHECK-NEXT: movd %xmm0, %ebp
|
|
; CHECK-NEXT: shll $16, %ebp
|
|
; CHECK-NEXT: shll $16, %r15d
|
|
; CHECK-NEXT: movd %r15d, %xmm1
|
|
; CHECK-NEXT: movq (%rsp), %rax # 8-byte Reload
|
|
; CHECK-NEXT: shll $16, %eax
|
|
; CHECK-NEXT: movd %eax, %xmm0
|
|
; CHECK-NEXT: addss %xmm1, %xmm0
|
|
; CHECK-NEXT: callq __truncsfbf2@PLT
|
|
; CHECK-NEXT: movd %xmm0, %eax
|
|
; CHECK-NEXT: movzwl %ax, %r14d
|
|
; CHECK-NEXT: orl %ebp, %r14d
|
|
; CHECK-NEXT: shlq $32, %r14
|
|
; CHECK-NEXT: orq %rbx, %r14
|
|
; CHECK-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %r15 # 8-byte Reload
|
|
; CHECK-NEXT: movl %r15d, %eax
|
|
; CHECK-NEXT: andl $-65536, %eax # imm = 0xFFFF0000
|
|
; CHECK-NEXT: movd %eax, %xmm1
|
|
; CHECK-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rbx # 8-byte Reload
|
|
; CHECK-NEXT: movl %ebx, %eax
|
|
; CHECK-NEXT: andl $-65536, %eax # imm = 0xFFFF0000
|
|
; CHECK-NEXT: movd %eax, %xmm0
|
|
; CHECK-NEXT: addss %xmm1, %xmm0
|
|
; CHECK-NEXT: callq __truncsfbf2@PLT
|
|
; CHECK-NEXT: movd %xmm0, %ebp
|
|
; CHECK-NEXT: shll $16, %ebp
|
|
; CHECK-NEXT: movq %r15, %rax
|
|
; CHECK-NEXT: shll $16, %eax
|
|
; CHECK-NEXT: movd %eax, %xmm1
|
|
; CHECK-NEXT: movq %rbx, %rax
|
|
; CHECK-NEXT: shll $16, %eax
|
|
; CHECK-NEXT: movd %eax, %xmm0
|
|
; CHECK-NEXT: addss %xmm1, %xmm0
|
|
; CHECK-NEXT: callq __truncsfbf2@PLT
|
|
; CHECK-NEXT: movd %xmm0, %eax
|
|
; CHECK-NEXT: movzwl %ax, %ebx
|
|
; CHECK-NEXT: orl %ebp, %ebx
|
|
; CHECK-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload
|
|
; CHECK-NEXT: shll $16, %eax
|
|
; CHECK-NEXT: movd %eax, %xmm1
|
|
; CHECK-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload
|
|
; CHECK-NEXT: shll $16, %eax
|
|
; CHECK-NEXT: movd %eax, %xmm0
|
|
; CHECK-NEXT: addss %xmm1, %xmm0
|
|
; CHECK-NEXT: callq __truncsfbf2@PLT
|
|
; CHECK-NEXT: movd %xmm0, %ebp
|
|
; CHECK-NEXT: shll $16, %ebp
|
|
; CHECK-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload
|
|
; CHECK-NEXT: shll $16, %eax
|
|
; CHECK-NEXT: movd %eax, %xmm1
|
|
; CHECK-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload
|
|
; CHECK-NEXT: shll $16, %eax
|
|
; CHECK-NEXT: movd %eax, %xmm0
|
|
; CHECK-NEXT: addss %xmm1, %xmm0
|
|
; CHECK-NEXT: callq __truncsfbf2@PLT
|
|
; CHECK-NEXT: movd %xmm0, %eax
|
|
; CHECK-NEXT: movzwl %ax, %eax
|
|
; CHECK-NEXT: orl %ebp, %eax
|
|
; CHECK-NEXT: shlq $32, %rax
|
|
; CHECK-NEXT: orq %rbx, %rax
|
|
; CHECK-NEXT: movq %rax, %xmm0
|
|
; CHECK-NEXT: movq %r14, %xmm1
|
|
; CHECK-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
|
|
; CHECK-NEXT: addq $56, %rsp
|
|
; CHECK-NEXT: popq %rbx
|
|
; CHECK-NEXT: popq %r12
|
|
; CHECK-NEXT: popq %r13
|
|
; CHECK-NEXT: popq %r14
|
|
; CHECK-NEXT: popq %r15
|
|
; CHECK-NEXT: popq %rbp
|
|
; CHECK-NEXT: retq
|
|
%add = fadd <8 x bfloat> %a, %b
|
|
ret <8 x bfloat> %add
|
|
}
|