60b1967c39
Add the scratch wave offset to the scratch buffer descriptor (SRSrc) in the entry function prologue. This allows us to removes the scratch wave offset register from the calling convention ABI. As part of this change, allow the use of an inline constant zero for the SOffset of MUBUF instructions accessing the stack in entry functions when a frame pointer is not requested/required. Entry functions with calls still need to set up the calling convention ABI stack pointer register, and reference it in order to address arguments of called functions. The ABI stack pointer register remains unswizzled, but is now wave-relative instead of queue-relative. Non-entry functions also use an inline constant zero SOffset for wave-relative scratch access, but continue to use the stack and frame pointers as before. When the stack or frame pointer is converted to a swizzled offset it is now scaled directly, as the scratch wave offset no longer needs to be subtracted first. Update llvm/docs/AMDGPUUsage.rst to reflect these changes to the calling convention. Tags: #llvm Differential Revision: https://reviews.llvm.org/D75138
286 lines
11 KiB
YAML
286 lines
11 KiB
YAML
# RUN: llc -march=amdgcn -mcpu=gfx900 -amdgpu-vgpr-index-mode -run-pass=greedy -stress-regalloc=16 -o - %s | FileCheck -check-prefixes=GCN %s
|
|
|
|
# An interval for a register that was partially defined was split, creating
|
|
# a new use (a COPY) which was reached by the undef point. In particular,
|
|
# there was a subrange of the new register which was reached by an "undef"
|
|
# point. When the code in extendSegmentsToUses verified value numbers between
|
|
# the new and the old live ranges, it did not account for this kind of a
|
|
# situation and asserted expecting the old value to exist. For a PHI node
|
|
# it is legal to have a missing predecessor value as long as the end of
|
|
# the predecessor is jointly dominated by the undefs.
|
|
#
|
|
# A simplified form of this can be illustrated as
|
|
#
|
|
# bb.1:
|
|
# %0:vreg_64 = IMPLICIT_DEF
|
|
# ...
|
|
# S_CBRANCH_SCC1 %bb.2, implicit $vcc
|
|
# S_BRANCH %bb.3
|
|
#
|
|
# bb.2:
|
|
# ; predecessors: %bb.1, %bb.4
|
|
# dead %1:vreg_64 = COPY %0:vreg_64 ; This is the point of the inserted split
|
|
# ...
|
|
# S_BRANCH %bb.5
|
|
#
|
|
# bb.3:
|
|
# ; predecessors: %bb.1
|
|
# undef %0.sub0:vreg_64 = COPY %123:sreg_32 ; undef point for %0.sub1
|
|
# ...
|
|
# S_BRANCH %bb.4
|
|
#
|
|
# bb.4
|
|
# ; predecessors: %bb.4
|
|
# ...
|
|
# S_BRANCH %bb.2
|
|
#
|
|
# This test exposes this scenario which caused previously caused an assert
|
|
|
|
---
|
|
name: _amdgpu_ps_main
|
|
tracksRegLiveness: true
|
|
machineFunctionInfo:
|
|
scratchRSrcReg: $sgpr0_sgpr1_sgpr2_sgpr3
|
|
stackPtrOffsetReg: $sgpr32
|
|
liveins:
|
|
- { reg: '$vgpr2', virtual-reg: '%0' }
|
|
- { reg: '$vgpr3', virtual-reg: '%1' }
|
|
- { reg: '$vgpr4', virtual-reg: '%2' }
|
|
body: |
|
|
bb.0:
|
|
successors: %bb.1(0x40000000), %bb.2(0x40000000)
|
|
liveins: $vgpr2, $vgpr3, $vgpr4
|
|
%2:vgpr_32 = COPY $vgpr4
|
|
%1:vgpr_32 = COPY $vgpr3
|
|
%0:vgpr_32 = COPY $vgpr2
|
|
S_CBRANCH_SCC0 %bb.2, implicit undef $scc
|
|
|
|
bb.1:
|
|
successors: %bb.5(0x80000000)
|
|
undef %3.sub0:vreg_128 = V_MOV_B32_e32 0, implicit $exec
|
|
%3.sub1:vreg_128 = COPY %3.sub0
|
|
%3.sub2:vreg_128 = COPY %3.sub0
|
|
S_BRANCH %bb.5
|
|
|
|
bb.2:
|
|
successors: %bb.3(0x40000000), %bb.4(0x40000000)
|
|
S_CBRANCH_SCC0 %bb.4, implicit undef $scc
|
|
|
|
bb.3:
|
|
successors: %bb.5(0x80000000)
|
|
undef %3.sub0:vreg_128 = V_MOV_B32_e32 0, implicit $exec
|
|
%3.sub1:vreg_128 = COPY %3.sub0
|
|
S_BRANCH %bb.5
|
|
|
|
bb.4:
|
|
successors: %bb.5(0x80000000)
|
|
%3:vreg_128 = IMPLICIT_DEF
|
|
|
|
bb.5:
|
|
successors: %bb.6(0x40000000), %bb.22(0x40000000)
|
|
%4:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
|
|
S_CBRANCH_SCC1 %bb.22, implicit undef $scc
|
|
S_BRANCH %bb.6
|
|
|
|
bb.6:
|
|
successors: %bb.8(0x40000000), %bb.11(0x40000000)
|
|
%5:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
|
|
dead %6:vgpr_32 = V_MUL_F32_e32 0, undef %7:vgpr_32, implicit $exec
|
|
dead %8:vgpr_32 = V_MUL_F32_e32 0, %2, implicit $exec
|
|
undef %9.sub1:vreg_64 = V_MUL_F32_e32 0, %1, implicit $exec
|
|
undef %10.sub0:vreg_128 = V_MUL_F32_e32 0, %0, implicit $exec
|
|
undef %11.sub0:sreg_256 = S_MOV_B32 0
|
|
%11.sub1:sreg_256 = COPY %11.sub0
|
|
%11.sub2:sreg_256 = COPY %11.sub0
|
|
%11.sub3:sreg_256 = COPY %11.sub0
|
|
%11.sub4:sreg_256 = COPY %11.sub0
|
|
%11.sub5:sreg_256 = COPY %11.sub0
|
|
%11.sub6:sreg_256 = COPY %11.sub0
|
|
%11.sub7:sreg_256 = COPY %11.sub0
|
|
%12:vreg_128 = IMAGE_SAMPLE_LZ_V4_V2 %9, %11, undef %13:sgpr_128, 15, 0, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16 from constant-pool, addrspace 4)
|
|
%14:vgpr_32 = V_MOV_B32_e32 -1, implicit $exec
|
|
%15:vreg_128 = IMPLICIT_DEF
|
|
S_CBRANCH_SCC1 %bb.8, implicit undef $scc
|
|
S_BRANCH %bb.11
|
|
|
|
bb.7:
|
|
successors: %bb.13(0x80000000)
|
|
|
|
; In reality we are checking that this code doesn't assert when splitting
|
|
; and inserting a spill. Here we just check that the point where the error
|
|
; occurs we see a correctly generated spill.
|
|
; GCN-LABEL: bb.7:
|
|
; GCN: SI_SPILL_V128_SAVE %{{[0-9]+}}, %stack.1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, implicit $exec
|
|
|
|
undef %15.sub0:vreg_128 = V_MOV_B32_e32 0, implicit $exec
|
|
%15.sub1:vreg_128 = COPY %15.sub0
|
|
%15.sub2:vreg_128 = COPY %15.sub0
|
|
%5:vgpr_32 = IMPLICIT_DEF
|
|
S_BRANCH %bb.13
|
|
|
|
bb.8:
|
|
successors: %bb.9(0x40000000), %bb.10(0x40000000)
|
|
S_CBRANCH_SCC0 %bb.10, implicit undef $scc
|
|
|
|
bb.9:
|
|
successors: %bb.12(0x80000000)
|
|
|
|
; GCN-LABEL: bb.9:
|
|
; GCN: SI_SPILL_V128_SAVE %{{[0-9]+}}, %stack.1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, implicit $exec
|
|
|
|
undef %15.sub0:vreg_128 = V_MOV_B32_e32 0, implicit $exec
|
|
%15.sub1:vreg_128 = COPY %15.sub0
|
|
%15.sub2:vreg_128 = COPY %15.sub0
|
|
S_BRANCH %bb.12
|
|
|
|
bb.10:
|
|
successors: %bb.12(0x80000000)
|
|
|
|
; GCN-LABEL: bb.10:
|
|
; GCN: SI_SPILL_V128_SAVE %{{[0-9]+}}, %stack.1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 0, implicit $exec
|
|
|
|
undef %15.sub0:vreg_128 = V_MOV_B32_e32 2143289344, implicit $exec
|
|
%15.sub1:vreg_128 = COPY %15.sub0
|
|
%15.sub2:vreg_128 = COPY %15.sub0
|
|
S_BRANCH %bb.12
|
|
|
|
bb.11:
|
|
successors: %bb.7(0x40000000), %bb.13(0x40000000)
|
|
%16:sreg_64 = V_CMP_NE_U32_e64 0, %14, implicit $exec
|
|
%17:sreg_64 = S_AND_B64 $exec, %16, implicit-def dead $scc
|
|
$vcc = COPY %17
|
|
S_CBRANCH_VCCNZ %bb.7, implicit $vcc
|
|
S_BRANCH %bb.13
|
|
|
|
bb.12:
|
|
successors: %bb.11(0x80000000)
|
|
%14:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
|
|
%5:vgpr_32 = V_MOV_B32_e32 -1, implicit $exec
|
|
S_BRANCH %bb.11
|
|
|
|
bb.13:
|
|
successors: %bb.15(0x40000000), %bb.14(0x40000000)
|
|
|
|
%18:vgpr_32 = V_MAD_F32 0, %10.sub0, 0, target-flags(amdgpu-gotprel) 1073741824, 0, -1082130432, 0, 0, implicit $exec
|
|
%19:vgpr_32 = V_MAD_F32 0, %12.sub0, 0, target-flags(amdgpu-gotprel) 0, 0, 0, 0, 0, implicit $exec
|
|
%20:sgpr_128 = S_BUFFER_LOAD_DWORDX4_IMM undef %21:sgpr_128, 1040, 0, 0 :: (dereferenceable invariant load 16)
|
|
%22:vgpr_32 = V_ADD_F32_e32 0, %19, implicit $exec
|
|
%23:vgpr_32 = V_MAD_F32 0, %18, 0, 0, 0, 0, 0, 0, implicit $exec
|
|
%24:vgpr_32 = COPY %20.sub3
|
|
%25:vgpr_32 = V_MUL_F32_e64 0, target-flags(amdgpu-gotprel32-lo) 0, 0, %20.sub1, 0, 0, implicit $exec
|
|
%26:sgpr_128 = S_BUFFER_LOAD_DWORDX4_IMM undef %27:sgpr_128, 1056, 0, 0 :: (dereferenceable invariant load 16)
|
|
%28:vgpr_32 = V_MAD_F32 0, %18, 0, %26.sub0, 0, 0, 0, 0, implicit $exec
|
|
%29:vgpr_32 = V_ADD_F32_e32 %28, %19, implicit $exec
|
|
%30:vgpr_32 = V_RCP_F32_e32 %29, implicit $exec
|
|
%25:vgpr_32 = V_MAC_F32_e32 0, %18, %25, implicit $exec
|
|
%31:vgpr_32 = V_MAD_F32 0, target-flags(amdgpu-gotprel) 0, 0, %12.sub0, 0, %24, 0, 0, implicit $exec
|
|
%32:vgpr_32 = V_ADD_F32_e32 %25, %31, implicit $exec
|
|
%33:vgpr_32 = V_MUL_F32_e32 %22, %30, implicit $exec
|
|
%34:vgpr_32 = V_MUL_F32_e32 %23, %30, implicit $exec
|
|
%35:vgpr_32 = V_MUL_F32_e32 %32, %30, implicit $exec
|
|
%36:vgpr_32 = V_MUL_F32_e32 0, %34, implicit $exec
|
|
%36:vgpr_32 = V_MAC_F32_e32 0, %33, %36, implicit $exec
|
|
%37:vgpr_32 = V_MAD_F32 0, %35, 0, 0, 0, 0, 0, 0, implicit $exec
|
|
%38:sreg_64_xexec = V_CMP_NE_U32_e64 0, %5, implicit $exec
|
|
%39:vgpr_32 = V_CNDMASK_B32_e64 0, 0, 0, 1, %38, implicit $exec
|
|
V_CMP_NE_U32_e32 1, %39, implicit-def $vcc, implicit $exec
|
|
$vcc = S_AND_B64 $exec, $vcc, implicit-def dead $scc
|
|
%40:vgpr_32 = V_ADD_F32_e32 %36, %37, implicit $exec
|
|
S_CBRANCH_VCCZ %bb.15, implicit $vcc
|
|
|
|
bb.14:
|
|
successors: %bb.17(0x80000000)
|
|
S_BRANCH %bb.17
|
|
|
|
bb.15:
|
|
successors: %bb.16(0x40000000), %bb.18(0x40000000)
|
|
%41:vgpr_32 = V_MAD_F32 0, %40, 0, 0, 0, 0, 0, 0, implicit $exec
|
|
%42:sreg_64 = V_CMP_LE_F32_e64 0, 0, 0, %41, 0, implicit $exec
|
|
%43:sreg_64 = V_CMP_GE_F32_e64 0, 1065353216, 0, %41, 0, implicit $exec
|
|
%44:sreg_64 = S_AND_B64 %43, %43, implicit-def dead $scc
|
|
%45:sreg_64 = S_AND_B64 %42, %42, implicit-def dead $scc
|
|
%46:sreg_64 = S_AND_B64 %45, %44, implicit-def dead $scc
|
|
%47:sreg_64 = COPY $exec, implicit-def $exec
|
|
%48:sreg_64 = S_AND_B64 %47, %46, implicit-def dead $scc
|
|
$exec = S_MOV_B64_term %48
|
|
SI_MASK_BRANCH %bb.18, implicit $exec
|
|
S_BRANCH %bb.16
|
|
|
|
bb.16:
|
|
successors: %bb.18(0x80000000)
|
|
S_BRANCH %bb.18
|
|
|
|
bb.17:
|
|
successors: %bb.21(0x40000000), %bb.23(0x40000000)
|
|
%49:sreg_64 = V_CMP_NE_U32_e64 0, %5, implicit $exec
|
|
%50:sreg_64 = S_AND_B64 $exec, %49, implicit-def dead $scc
|
|
%51:vreg_128 = IMPLICIT_DEF
|
|
$vcc = COPY %50
|
|
S_CBRANCH_VCCNZ %bb.21, implicit $vcc
|
|
S_BRANCH %bb.23
|
|
|
|
bb.18:
|
|
successors: %bb.20(0x40000000), %bb.19(0x40000000)
|
|
$exec = S_OR_B64 $exec, %47, implicit-def $scc
|
|
%52:vgpr_32 = V_MAD_F32 0, %3.sub1, 0, target-flags(amdgpu-gotprel32-lo) 0, 1, %3.sub0, 0, 0, implicit $exec
|
|
%53:vgpr_32 = V_MUL_F32_e32 -2147483648, %3.sub1, implicit $exec
|
|
%53:vgpr_32 = V_MAC_F32_e32 target-flags(amdgpu-gotprel32-hi) 1065353216, %3.sub2, %53, implicit $exec
|
|
%54:vgpr_32 = V_MUL_F32_e32 %53, %53, implicit $exec
|
|
%54:vgpr_32 = V_MAC_F32_e32 %52, %52, %54, implicit $exec
|
|
%55:vgpr_32 = V_SQRT_F32_e32 %54, implicit $exec
|
|
%5:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
|
|
%56:vgpr_32 = V_MOV_B32_e32 981668463, implicit $exec
|
|
%57:sreg_64 = V_CMP_NGT_F32_e64 0, %55, 0, %56, 0, implicit $exec
|
|
%58:sreg_64 = S_AND_B64 $exec, %57, implicit-def dead $scc
|
|
$vcc = COPY %58
|
|
S_CBRANCH_VCCZ %bb.20, implicit $vcc
|
|
|
|
bb.19:
|
|
successors: %bb.17(0x80000000)
|
|
S_BRANCH %bb.17
|
|
|
|
bb.20:
|
|
successors: %bb.17(0x80000000)
|
|
S_BRANCH %bb.17
|
|
|
|
bb.21:
|
|
successors: %bb.23(0x80000000)
|
|
%59:sreg_32 = S_MOV_B32 0
|
|
undef %51.sub0:vreg_128 = COPY %59
|
|
S_BRANCH %bb.23
|
|
|
|
bb.22:
|
|
successors: %bb.24(0x80000000)
|
|
S_BRANCH %bb.24
|
|
|
|
bb.23:
|
|
successors: %bb.22(0x80000000)
|
|
undef %60.sub1:vreg_64 = V_CVT_I32_F32_e32 %1, implicit $exec
|
|
%60.sub0:vreg_64 = V_CVT_I32_F32_e32 %0, implicit $exec
|
|
undef %61.sub0:sreg_256 = S_MOV_B32 0
|
|
%61.sub1:sreg_256 = COPY %61.sub0
|
|
%61.sub2:sreg_256 = COPY %61.sub0
|
|
%61.sub3:sreg_256 = COPY %61.sub0
|
|
%61.sub4:sreg_256 = COPY %61.sub0
|
|
%61.sub5:sreg_256 = COPY %61.sub0
|
|
%61.sub6:sreg_256 = COPY %61.sub0
|
|
%61.sub7:sreg_256 = COPY %61.sub0
|
|
%62:vgpr_32 = V_MOV_B32_e32 1033100696, implicit $exec
|
|
%63:vgpr_32 = V_MUL_F32_e32 1060575065, %15.sub1, implicit $exec
|
|
%63:vgpr_32 = V_MAC_F32_e32 1046066128, %15.sub0, %63, implicit $exec
|
|
%64:vgpr_32 = IMAGE_LOAD_V1_V2 %60, %61, 1, -1, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16 from constant-pool, addrspace 4)
|
|
%64:vgpr_32 = V_MAC_F32_e32 target-flags(amdgpu-gotprel) 0, %51.sub0, %64, implicit $exec
|
|
%65:vgpr_32 = V_MUL_F32_e32 0, %64, implicit $exec
|
|
%66:vgpr_32 = V_MUL_F32_e32 0, %65, implicit $exec
|
|
%67:vgpr_32 = V_MAD_F32 0, %66, 0, %62, 0, 0, 0, 0, implicit $exec
|
|
%63:vgpr_32 = V_MAC_F32_e32 %15.sub2, %62, %63, implicit $exec
|
|
%4:vgpr_32 = V_ADD_F32_e32 %63, %67, implicit $exec
|
|
S_BRANCH %bb.22
|
|
|
|
bb.24:
|
|
%68:vgpr_32 = V_MUL_F32_e32 0, %4, implicit $exec
|
|
%69:vgpr_32 = V_CVT_PKRTZ_F16_F32_e64 0, undef %70:vgpr_32, 0, %68, 0, 0, implicit $exec
|
|
EXP 0, undef %71:vgpr_32, %69, undef %72:vgpr_32, undef %73:vgpr_32, -1, -1, 15, implicit $exec
|
|
S_ENDPGM 0
|
|
...
|