llvm/test/Analysis/CostModel/AMDGPU/expanded-int-fp-casts.ll

; NOTE: Assertions have been autogenerated by utils/update_analyze_test_checks.py UTC_ARGS: --filter "instruction:.*(sitofp|uitofp|fptosi|fptoui)"
; RUN: opt -passes="print<cost-model>" -disable-output -mtriple=amdgpu6.01-amd-amdhsa < %s 2>&1 | FileCheck %s --check-prefixes=ALL,GFX6
; RUN: opt -passes="print<cost-model>" -disable-output -mtriple=amdgpu7.00-amd-amdhsa < %s 2>&1 | FileCheck %s --check-prefixes=ALL,GFX6
; RUN: opt -passes="print<cost-model>" -disable-output -mtriple=amdgpu8.03-amd-amdhsa < %s 2>&1 | FileCheck %s --check-prefixes=ALL,GFX9-THROUGHPUT
; RUN: opt -passes="print<cost-model>" -disable-output -mtriple=amdgpu9.00-amd-amdhsa < %s 2>&1 | FileCheck %s --check-prefixes=ALL,GFX9-THROUGHPUT
; RUN: opt -passes="print<cost-model>" -disable-output -mtriple=amdgpu9.00-amd-amdhsa -mattr=-gfx9-insts < %s 2>&1 | FileCheck %s --check-prefixes=ALL,GFX9-THROUGHPUT
; RUN: opt -passes="print<cost-model>" -disable-output -mtriple=amdgpu9.06-amd-amdhsa < %s 2>&1 | FileCheck %s --check-prefixes=ALL,HALF
; RUN: opt -passes="print<cost-model>" -disable-output -mtriple=amdgpu9.42-amd-amdhsa < %s 2>&1 | FileCheck %s --check-prefixes=ALL,FULL
; RUN: opt -passes="print<cost-model>" -disable-output -mtriple=amdgpu9.50-amd-amdhsa < %s 2>&1 | FileCheck %s --check-prefixes=ALL,FULL
; RUN: opt -passes="print<cost-model>" -disable-output -mtriple=amdgpu11.00-amd-amdhsa < %s 2>&1 | FileCheck %s --check-prefixes=ALL,GFX9-THROUGHPUT
; RUN: opt -passes="print<cost-model>" -disable-output -mtriple=amdgpu11.00-amd-amdhsa -mattr=-16-bit-insts < %s 2>&1 | FileCheck %s --check-prefixes=ALL,GFX6
; RUN: opt -passes="print<cost-model>" -disable-output -mtriple=amdgpu12.50-amd-amdhsa < %s 2>&1 | FileCheck %s --check-prefixes=ALL,GFX9-THROUGHPUT
; RUN: opt -passes="print<cost-model>" -cost-kind=code-size -disable-output -mtriple=amdgpu6.01-amd-amdhsa < %s 2>&1 | FileCheck %s --check-prefixes=ALL,GFX6
; RUN: opt -passes="print<cost-model>" -cost-kind=code-size -disable-output -mtriple=amdgpu9.00-amd-amdhsa < %s 2>&1 | FileCheck %s --check-prefixes=ALL,HALF
; RUN: opt -passes="print<cost-model>" -cost-kind=code-size -disable-output -mtriple=amdgpu9.42-amd-amdhsa < %s 2>&1 | FileCheck %s --check-prefixes=ALL,FULL
; RUN: opt -passes="print<cost-model>" -cost-kind=code-size -disable-output -mtriple=amdgpu9.50-amd-amdhsa < %s 2>&1 | FileCheck %s --check-prefixes=ALL,FULL
; RUN: opt -passes="print<cost-model>" -cost-kind=code-size -disable-output -mtriple=amdgpu12.50-amd-amdhsa < %s 2>&1 | FileCheck %s --check-prefixes=ALL,HALF

; Half fits in i32. Bfloat must extend to f32 and use the full i64 expansion.
define void @bf16_to_i64(bfloat %a, <2 x bfloat> %b, half %h) {
; ALL-LABEL: 'bf16_to_i64'
; ALL:  Cost Model: Found an estimated cost of 1 for instruction: %s = fptosi bfloat %a to i64
; ALL:  Cost Model: Found an estimated cost of 1 for instruction: %u = fptoui bfloat %a to i64
; ALL:  Cost Model: Found an estimated cost of 2 for instruction: %vs = fptosi <2 x bfloat> %b to <2 x i64>
; ALL:  Cost Model: Found an estimated cost of 2 for instruction: %vu = fptoui <2 x bfloat> %b to <2 x i64>
; ALL:  Cost Model: Found an estimated cost of 1 for instruction: %hs = fptosi half %h to i64
; ALL:  Cost Model: Found an estimated cost of 1 for instruction: %hu = fptoui half %h to i64
;
  %s = fptosi bfloat %a to i64
  %u = fptoui bfloat %a to i64
  %vs = fptosi <2 x bfloat> %b to <2 x i64>
  %vu = fptoui <2 x bfloat> %b to <2 x i64>
  %hs = fptosi half %h to i64
  %hu = fptoui half %h to i64
  ret void
}

; Conversion to bf16 goes through f32 and rounds the result when there are
; 16 bit instructions. Native rounding packs pairs, including the final
; unpaired element of an odd-sized vector.
define void @i64_to_bf16(i64 %a, <2 x i64> %b, <3 x i64> %c, i48 %d) {
; GFX6-LABEL: 'i64_to_bf16'
; GFX6:  Cost Model: Found an estimated cost of 1 for instruction: %s = sitofp i64 %a to bfloat
; GFX6:  Cost Model: Found an estimated cost of 1 for instruction: %u = uitofp i64 %a to bfloat
; GFX6:  Cost Model: Found an estimated cost of 2 for instruction: %vs = sitofp <2 x i64> %b to <2 x bfloat>
; GFX6:  Cost Model: Found an estimated cost of 2 for instruction: %vu = uitofp <2 x i64> %b to <2 x bfloat>
; GFX6:  Cost Model: Found an estimated cost of 9 for instruction: %v3s = sitofp <3 x i64> %c to <3 x bfloat>
; GFX6:  Cost Model: Found an estimated cost of 9 for instruction: %v3u = uitofp <3 x i64> %c to <3 x bfloat>
; GFX6:  Cost Model: Found an estimated cost of 1 for instruction: %s48 = sitofp i48 %d to bfloat
; GFX6:  Cost Model: Found an estimated cost of 1 for instruction: %u48 = uitofp i48 %d to bfloat
;
; GFX9-THROUGHPUT-LABEL: 'i64_to_bf16'
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 1 for instruction: %s = sitofp i64 %a to bfloat
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 1 for instruction: %u = uitofp i64 %a to bfloat
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 4 for instruction: %vs = sitofp <2 x i64> %b to <2 x bfloat>
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 4 for instruction: %vu = uitofp <2 x i64> %b to <2 x bfloat>
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 7 for instruction: %v3s = sitofp <3 x i64> %c to <3 x bfloat>
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 7 for instruction: %v3u = uitofp <3 x i64> %c to <3 x bfloat>
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 1 for instruction: %s48 = sitofp i48 %d to bfloat
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 1 for instruction: %u48 = uitofp i48 %d to bfloat
;
; HALF-LABEL: 'i64_to_bf16'
; HALF:  Cost Model: Found an estimated cost of 1 for instruction: %s = sitofp i64 %a to bfloat
; HALF:  Cost Model: Found an estimated cost of 1 for instruction: %u = uitofp i64 %a to bfloat
; HALF:  Cost Model: Found an estimated cost of 4 for instruction: %vs = sitofp <2 x i64> %b to <2 x bfloat>
; HALF:  Cost Model: Found an estimated cost of 4 for instruction: %vu = uitofp <2 x i64> %b to <2 x bfloat>
; HALF:  Cost Model: Found an estimated cost of 7 for instruction: %v3s = sitofp <3 x i64> %c to <3 x bfloat>
; HALF:  Cost Model: Found an estimated cost of 7 for instruction: %v3u = uitofp <3 x i64> %c to <3 x bfloat>
; HALF:  Cost Model: Found an estimated cost of 1 for instruction: %s48 = sitofp i48 %d to bfloat
; HALF:  Cost Model: Found an estimated cost of 1 for instruction: %u48 = uitofp i48 %d to bfloat
;
; FULL-LABEL: 'i64_to_bf16'
; FULL:  Cost Model: Found an estimated cost of 1 for instruction: %s = sitofp i64 %a to bfloat
; FULL:  Cost Model: Found an estimated cost of 1 for instruction: %u = uitofp i64 %a to bfloat
; FULL:  Cost Model: Found an estimated cost of 4 for instruction: %vs = sitofp <2 x i64> %b to <2 x bfloat>
; FULL:  Cost Model: Found an estimated cost of 4 for instruction: %vu = uitofp <2 x i64> %b to <2 x bfloat>
; FULL:  Cost Model: Found an estimated cost of 7 for instruction: %v3s = sitofp <3 x i64> %c to <3 x bfloat>
; FULL:  Cost Model: Found an estimated cost of 7 for instruction: %v3u = uitofp <3 x i64> %c to <3 x bfloat>
; FULL:  Cost Model: Found an estimated cost of 1 for instruction: %s48 = sitofp i48 %d to bfloat
; FULL:  Cost Model: Found an estimated cost of 1 for instruction: %u48 = uitofp i48 %d to bfloat
;
  %s = sitofp i64 %a to bfloat
  %u = uitofp i64 %a to bfloat
  %vs = sitofp <2 x i64> %b to <2 x bfloat>
  %vu = uitofp <2 x i64> %b to <2 x bfloat>
  %v3s = sitofp <3 x i64> %c to <3 x bfloat>
  %v3u = uitofp <3 x i64> %c to <3 x bfloat>
  %s48 = sitofp i48 %d to bfloat
  %u48 = uitofp i48 %d to bfloat
  ret void
}

; Subword extraction and bf16 rounding are independent costs. Native packed
; rounding still needs one integer-to-f32 conversion for each input lane.
define void @narrow_to_bf16(<4 x i8> %a, <2 x i16> %b, <3 x i32> %c) {
; GFX6-LABEL: 'narrow_to_bf16'
; GFX6:  Cost Model: Found an estimated cost of 4 for instruction: %bs = sitofp <4 x i8> %a to <4 x bfloat>
; GFX6:  Cost Model: Found an estimated cost of 4 for instruction: %bu = uitofp <4 x i8> %a to <4 x bfloat>
; GFX6:  Cost Model: Found an estimated cost of 2 for instruction: %hs = sitofp <2 x i16> %b to <2 x bfloat>
; GFX6:  Cost Model: Found an estimated cost of 2 for instruction: %hu = uitofp <2 x i16> %b to <2 x bfloat>
; GFX6:  Cost Model: Found an estimated cost of 9 for instruction: %ws = sitofp <3 x i32> %c to <3 x bfloat>
; GFX6:  Cost Model: Found an estimated cost of 9 for instruction: %wu = uitofp <3 x i32> %c to <3 x bfloat>
;
; GFX9-THROUGHPUT-LABEL: 'narrow_to_bf16'
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 4 for instruction: %bs = sitofp <4 x i8> %a to <4 x bfloat>
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 4 for instruction: %bu = uitofp <4 x i8> %a to <4 x bfloat>
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 4 for instruction: %hs = sitofp <2 x i16> %b to <2 x bfloat>
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 4 for instruction: %hu = uitofp <2 x i16> %b to <2 x bfloat>
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 7 for instruction: %ws = sitofp <3 x i32> %c to <3 x bfloat>
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 7 for instruction: %wu = uitofp <3 x i32> %c to <3 x bfloat>
;
; HALF-LABEL: 'narrow_to_bf16'
; HALF:  Cost Model: Found an estimated cost of 4 for instruction: %bs = sitofp <4 x i8> %a to <4 x bfloat>
; HALF:  Cost Model: Found an estimated cost of 4 for instruction: %bu = uitofp <4 x i8> %a to <4 x bfloat>
; HALF:  Cost Model: Found an estimated cost of 4 for instruction: %hs = sitofp <2 x i16> %b to <2 x bfloat>
; HALF:  Cost Model: Found an estimated cost of 4 for instruction: %hu = uitofp <2 x i16> %b to <2 x bfloat>
; HALF:  Cost Model: Found an estimated cost of 7 for instruction: %ws = sitofp <3 x i32> %c to <3 x bfloat>
; HALF:  Cost Model: Found an estimated cost of 7 for instruction: %wu = uitofp <3 x i32> %c to <3 x bfloat>
;
; FULL-LABEL: 'narrow_to_bf16'
; FULL:  Cost Model: Found an estimated cost of 4 for instruction: %bs = sitofp <4 x i8> %a to <4 x bfloat>
; FULL:  Cost Model: Found an estimated cost of 4 for instruction: %bu = uitofp <4 x i8> %a to <4 x bfloat>
; FULL:  Cost Model: Found an estimated cost of 4 for instruction: %hs = sitofp <2 x i16> %b to <2 x bfloat>
; FULL:  Cost Model: Found an estimated cost of 4 for instruction: %hu = uitofp <2 x i16> %b to <2 x bfloat>
; FULL:  Cost Model: Found an estimated cost of 7 for instruction: %ws = sitofp <3 x i32> %c to <3 x bfloat>
; FULL:  Cost Model: Found an estimated cost of 7 for instruction: %wu = uitofp <3 x i32> %c to <3 x bfloat>
;
  %bs = sitofp <4 x i8> %a to <4 x bfloat>
  %bu = uitofp <4 x i8> %a to <4 x bfloat>
  %hs = sitofp <2 x i16> %b to <2 x bfloat>
  %hu = uitofp <2 x i16> %b to <2 x bfloat>
  %ws = sitofp <3 x i32> %c to <3 x bfloat>
  %wu = uitofp <3 x i32> %c to <3 x bfloat>
  ret void
}

; Unsigned lanes wider than a byte are zero extended before the conversion.
define void @narrow_unsigned_lanes(<4 x i12> %a) {
; GFX6-LABEL: 'narrow_unsigned_lanes'
; GFX6:  Cost Model: Found an estimated cost of 1 for instruction: %u = uitofp <4 x i12> %a to <4 x float>
; GFX6:  Cost Model: Found an estimated cost of 4 for instruction: %uh = uitofp <4 x i12> %a to <4 x half>
;
; GFX9-THROUGHPUT-LABEL: 'narrow_unsigned_lanes'
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 1 for instruction: %u = uitofp <4 x i12> %a to <4 x float>
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 10 for instruction: %uh = uitofp <4 x i12> %a to <4 x half>
;
; HALF-LABEL: 'narrow_unsigned_lanes'
; HALF:  Cost Model: Found an estimated cost of 1 for instruction: %u = uitofp <4 x i12> %a to <4 x float>
; HALF:  Cost Model: Found an estimated cost of 10 for instruction: %uh = uitofp <4 x i12> %a to <4 x half>
;
; FULL-LABEL: 'narrow_unsigned_lanes'
; FULL:  Cost Model: Found an estimated cost of 1 for instruction: %u = uitofp <4 x i12> %a to <4 x float>
; FULL:  Cost Model: Found an estimated cost of 10 for instruction: %uh = uitofp <4 x i12> %a to <4 x half>
;
  %u = uitofp <4 x i12> %a to <4 x float>
  %uh = uitofp <4 x i12> %a to <4 x half>
  ret void
}

; FP64 rates vary independently of native trunc/floor support and code size.
define void @i64_double(i64 %a, <2 x i64> %b, double %c, <2 x double> %d, i48 %e) {
; ALL-LABEL: 'i64_double'
; ALL:  Cost Model: Found an estimated cost of 1 for instruction: %s = sitofp i64 %a to double
; ALL:  Cost Model: Found an estimated cost of 1 for instruction: %u = uitofp i64 %a to double
; ALL:  Cost Model: Found an estimated cost of 2 for instruction: %vs = sitofp <2 x i64> %b to <2 x double>
; ALL:  Cost Model: Found an estimated cost of 2 for instruction: %vu = uitofp <2 x i64> %b to <2 x double>
; ALL:  Cost Model: Found an estimated cost of 1 for instruction: %rs = fptosi double %c to i64
; ALL:  Cost Model: Found an estimated cost of 1 for instruction: %ru = fptoui double %c to i64
; ALL:  Cost Model: Found an estimated cost of 2 for instruction: %rvs = fptosi <2 x double> %d to <2 x i64>
; ALL:  Cost Model: Found an estimated cost of 2 for instruction: %rvu = fptoui <2 x double> %d to <2 x i64>
; ALL:  Cost Model: Found an estimated cost of 1 for instruction: %s48 = sitofp i48 %e to double
; ALL:  Cost Model: Found an estimated cost of 1 for instruction: %u48 = uitofp i48 %e to double
; ALL:  Cost Model: Found an estimated cost of 1 for instruction: %ru48 = fptoui double %c to i48
;
  %s = sitofp i64 %a to double
  %u = uitofp i64 %a to double
  %vs = sitofp <2 x i64> %b to <2 x double>
  %vu = uitofp <2 x i64> %b to <2 x double>
  %rs = fptosi double %c to i64
  %ru = fptoui double %c to i64
  %rvs = fptosi <2 x double> %d to <2 x i64>
  %rvu = fptoui <2 x double> %d to <2 x i64>
  %s48 = sitofp i48 %e to double
  %u48 = uitofp i48 %e to double
  %ru48 = fptoui double %c to i48
  ret void
}

; Integers of 33 to 63 bits are converted through i64.
define void @i48(i48 %a, float %b, bfloat %c) {
; ALL-LABEL: 'i48'
; ALL:  Cost Model: Found an estimated cost of 1 for instruction: %s = sitofp i48 %a to float
; ALL:  Cost Model: Found an estimated cost of 1 for instruction: %u = uitofp i48 %a to float
; ALL:  Cost Model: Found an estimated cost of 1 for instruction: %rs = fptosi float %b to i48
; ALL:  Cost Model: Found an estimated cost of 1 for instruction: %ru = fptoui float %b to i48
; ALL:  Cost Model: Found an estimated cost of 1 for instruction: %bu = fptoui bfloat %c to i48
;
  %s = sitofp i48 %a to float
  %u = uitofp i48 %a to float
  %rs = fptosi float %b to i48
  %ru = fptoui float %b to i48
  %bu = fptoui bfloat %c to i48
  ret void
}

; i128 and fp128 casts keep the generic estimates.
define void @other_widths(i128 %a, double %b, <2 x i128> %c, <2 x double> %d, fp128 %e, i64 %f) {
; GFX6-LABEL: 'other_widths'
; GFX6:  Cost Model: Found an estimated cost of 1 for instruction: %s = sitofp i128 %a to double
; GFX6:  Cost Model: Found an estimated cost of 1 for instruction: %u = uitofp i128 %a to bfloat
; GFX6:  Cost Model: Found an estimated cost of 2 for instruction: %vs = sitofp <2 x i128> %c to <2 x double>
; GFX6:  Cost Model: Found an estimated cost of 6 for instruction: %vu = uitofp <2 x i128> %c to <2 x bfloat>
; GFX6:  Cost Model: Found an estimated cost of 1 for instruction: %rs = fptosi double %b to i128
; GFX6:  Cost Model: Found an estimated cost of 1 for instruction: %ru = fptoui double %b to i128
; GFX6:  Cost Model: Found an estimated cost of 2 for instruction: %rvs = fptosi <2 x double> %d to <2 x i128>
; GFX6:  Cost Model: Found an estimated cost of 2 for instruction: %rvu = fptoui <2 x double> %d to <2 x i128>
; GFX6:  Cost Model: Found an estimated cost of 1 for instruction: %qs = fptosi fp128 %e to i64
; GFX6:  Cost Model: Found an estimated cost of 1 for instruction: %qu = fptoui fp128 %e to i64
; GFX6:  Cost Model: Found an estimated cost of 1 for instruction: %rqs = sitofp i64 %f to fp128
; GFX6:  Cost Model: Found an estimated cost of 1 for instruction: %rqu = uitofp i64 %f to fp128
;
; GFX9-THROUGHPUT-LABEL: 'other_widths'
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 1 for instruction: %s = sitofp i128 %a to double
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 1 for instruction: %u = uitofp i128 %a to bfloat
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 2 for instruction: %vs = sitofp <2 x i128> %c to <2 x double>
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 2 for instruction: %vu = uitofp <2 x i128> %c to <2 x bfloat>
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 1 for instruction: %rs = fptosi double %b to i128
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 1 for instruction: %ru = fptoui double %b to i128
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 2 for instruction: %rvs = fptosi <2 x double> %d to <2 x i128>
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 2 for instruction: %rvu = fptoui <2 x double> %d to <2 x i128>
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 1 for instruction: %qs = fptosi fp128 %e to i64
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 1 for instruction: %qu = fptoui fp128 %e to i64
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 1 for instruction: %rqs = sitofp i64 %f to fp128
; GFX9-THROUGHPUT:  Cost Model: Found an estimated cost of 1 for instruction: %rqu = uitofp i64 %f to fp128
;
; HALF-LABEL: 'other_widths'
; HALF:  Cost Model: Found an estimated cost of 1 for instruction: %s = sitofp i128 %a to double
; HALF:  Cost Model: Found an estimated cost of 1 for instruction: %u = uitofp i128 %a to bfloat
; HALF:  Cost Model: Found an estimated cost of 2 for instruction: %vs = sitofp <2 x i128> %c to <2 x double>
; HALF:  Cost Model: Found an estimated cost of 2 for instruction: %vu = uitofp <2 x i128> %c to <2 x bfloat>
; HALF:  Cost Model: Found an estimated cost of 1 for instruction: %rs = fptosi double %b to i128
; HALF:  Cost Model: Found an estimated cost of 1 for instruction: %ru = fptoui double %b to i128
; HALF:  Cost Model: Found an estimated cost of 2 for instruction: %rvs = fptosi <2 x double> %d to <2 x i128>
; HALF:  Cost Model: Found an estimated cost of 2 for instruction: %rvu = fptoui <2 x double> %d to <2 x i128>
; HALF:  Cost Model: Found an estimated cost of 1 for instruction: %qs = fptosi fp128 %e to i64
; HALF:  Cost Model: Found an estimated cost of 1 for instruction: %qu = fptoui fp128 %e to i64
; HALF:  Cost Model: Found an estimated cost of 1 for instruction: %rqs = sitofp i64 %f to fp128
; HALF:  Cost Model: Found an estimated cost of 1 for instruction: %rqu = uitofp i64 %f to fp128
;
; FULL-LABEL: 'other_widths'
; FULL:  Cost Model: Found an estimated cost of 1 for instruction: %s = sitofp i128 %a to double
; FULL:  Cost Model: Found an estimated cost of 1 for instruction: %u = uitofp i128 %a to bfloat
; FULL:  Cost Model: Found an estimated cost of 2 for instruction: %vs = sitofp <2 x i128> %c to <2 x double>
; FULL:  Cost Model: Found an estimated cost of 2 for instruction: %vu = uitofp <2 x i128> %c to <2 x bfloat>
; FULL:  Cost Model: Found an estimated cost of 1 for instruction: %rs = fptosi double %b to i128
; FULL:  Cost Model: Found an estimated cost of 1 for instruction: %ru = fptoui double %b to i128
; FULL:  Cost Model: Found an estimated cost of 2 for instruction: %rvs = fptosi <2 x double> %d to <2 x i128>
; FULL:  Cost Model: Found an estimated cost of 2 for instruction: %rvu = fptoui <2 x double> %d to <2 x i128>
; FULL:  Cost Model: Found an estimated cost of 1 for instruction: %qs = fptosi fp128 %e to i64
; FULL:  Cost Model: Found an estimated cost of 1 for instruction: %qu = fptoui fp128 %e to i64
; FULL:  Cost Model: Found an estimated cost of 1 for instruction: %rqs = sitofp i64 %f to fp128
; FULL:  Cost Model: Found an estimated cost of 1 for instruction: %rqu = uitofp i64 %f to fp128
;
  %s = sitofp i128 %a to double
  %u = uitofp i128 %a to bfloat
  %vs = sitofp <2 x i128> %c to <2 x double>
  %vu = uitofp <2 x i128> %c to <2 x bfloat>
  %rs = fptosi double %b to i128
  %ru = fptoui double %b to i128
  %rvs = fptosi <2 x double> %d to <2 x i128>
  %rvu = fptoui <2 x double> %d to <2 x i128>
  %qs = fptosi fp128 %e to i64
  %qu = fptoui fp128 %e to i64
  %rqs = sitofp i64 %f to fp128
  %rqu = uitofp i64 %f to fp128
  ret void
}