[llvm] d0133bf - [NFC][AMDGPU] Pre-commit tests for D136432.
Thomas Symalla via llvm-commits
llvm-commits at lists.llvm.org
Wed Nov 9 02:24:35 PST 2022
Author: Thomas Symalla
Date: 2022-11-09T11:23:24+01:00
New Revision: d0133bf86dbc7e799e6adbf91c10133ffa2bddcf
URL: https://github.com/llvm/llvm-project/commit/d0133bf86dbc7e799e6adbf91c10133ffa2bddcf
DIFF: https://github.com/llvm/llvm-project/commit/d0133bf86dbc7e799e6adbf91c10133ffa2bddcf.diff
LOG: [NFC][AMDGPU] Pre-commit tests for D136432.
Added:
Modified:
llvm/test/CodeGen/AMDGPU/bfi_nested.ll
Removed:
################################################################################
diff --git a/llvm/test/CodeGen/AMDGPU/bfi_nested.ll b/llvm/test/CodeGen/AMDGPU/bfi_nested.ll
index 1ad70f01aaa58..15ba181f1aab3 100644
--- a/llvm/test/CodeGen/AMDGPU/bfi_nested.ll
+++ b/llvm/test/CodeGen/AMDGPU/bfi_nested.ll
@@ -1,15 +1,6 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -march=amdgcn -mcpu=gfx1030 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GFX10 %s
-; When translating sequences like
-; bitfieldInsert(bitfieldInsert(...)),
-; where one bitfieldInsert's result is the base for another one,
-; the SimplifyDemandedBits transform executed during InstCombine
-; can merge the inverted mask of the outer bitfieldInsert and the
-; inverted mask of the inner bitfieldInsert. When it is possible,
-; e. g. if the constants are disjoint and the original inverted mask of
-; the outer bitfieldInsert can be reconstructed, aim to generate multiple
-; v_bfi instructions.
define float @v_bfi_single_nesting_level(float %x, float %y, float %z) {
; GFX10-LABEL: v_bfi_single_nesting_level:
; GFX10: ; %bb.0: ; %.entry
@@ -44,6 +35,75 @@ define float @v_bfi_single_nesting_level(float %x, float %y, float %z) {
ret float %result
}
+define float @v_bfi_single_nesting_level_swapped_operands(float %x, float %y, float %z) {
+; GFX10-LABEL: v_bfi_single_nesting_level_swapped_operands:
+; GFX10: ; %bb.0: ; %.entry
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_mul_f32_e32 v2, 0x447fc000, v2
+; GFX10-NEXT: v_cvt_u32_f32_e32 v1, v1
+; GFX10-NEXT: v_mul_f32_e32 v0, 0x447fc000, v0
+; GFX10-NEXT: v_cvt_u32_f32_e32 v2, v2
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 10, v1
+; GFX10-NEXT: v_cvt_u32_f32_e32 v0, v0
+; GFX10-NEXT: v_and_b32_e32 v2, 0xc00003ff, v2
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 20, v0
+; GFX10-NEXT: v_and_or_b32 v1, 0xffc00, v1, v2
+; GFX10-NEXT: v_and_or_b32 v0, 0x3ff00000, v0, v1
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+.entry:
+ %mul.base = fmul reassoc nnan nsz arcp contract afn float %z, 1.023000e+03
+ %mul.base.i32 = fptoui float %mul.base to i32
+ %y.i32 = fptoui float %y to i32
+ %shl.inner.insert = shl i32 %y.i32, 10
+ %bfi1.and = and i32 1047552, %shl.inner.insert
+ %bfi1.andnot = and i32 -1073740801, %mul.base.i32
+ %bfi1.or = or i32 %bfi1.and, %bfi1.andnot
+ %mul.outer.insert = fmul reassoc nnan nsz arcp contract afn float %x, 1.023000e+03
+ %mul.outer.insert.i32 = fptoui float %mul.outer.insert to i32
+ %shl.outer.insert = shl i32 %mul.outer.insert.i32, 20
+ %and.outer = and i32 %shl.outer.insert, 1072693248
+ %or.outer = or i32 %and.outer, %bfi1.or
+ %result = bitcast i32 %or.outer to float
+ ret float %result
+}
+
+define float @v_bfi_single_nesting_level_unbalanced_subtree(float %x, float %y, float %z) {
+; GFX10-LABEL: v_bfi_single_nesting_level_unbalanced_subtree:
+; GFX10: ; %bb.0: ; %.entry
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_mul_f32_e32 v2, 0x447fc000, v2
+; GFX10-NEXT: v_cvt_u32_f32_e32 v1, v1
+; GFX10-NEXT: v_mul_f32_e32 v0, 0x447fc000, v0
+; GFX10-NEXT: v_cvt_u32_f32_e32 v2, v2
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 10, v1
+; GFX10-NEXT: v_cvt_u32_f32_e32 v0, v0
+; GFX10-NEXT: v_and_b32_e32 v3, 0x3e0, v2
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 20, v0
+; GFX10-NEXT: v_and_or_b32 v1, 0xffc00, v1, v3
+; GFX10-NEXT: v_and_or_b32 v1, 0xc000001f, v2, v1
+; GFX10-NEXT: v_and_or_b32 v0, 0x3ff00000, v0, v1
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+.entry:
+ %mul.base = fmul reassoc nnan nsz arcp contract afn float %z, 1.023000e+03
+ %mul.base.i32 = fptoui float %mul.base to i32
+ %y.i32 = fptoui float %y to i32
+ %shl.inner.2.insert = shl i32 %y.i32, 10
+ %bfi.inner.2.and.1 = and i32 %shl.inner.2.insert, 1047552
+ %bfi.inner.2.and.2 = and i32 %mul.base.i32, 992
+ %bfi.inner.2 = or i32 %bfi.inner.2.and.1, %bfi.inner.2.and.2
+ %mul.inner.1.insert = fmul reassoc nnan nsz arcp contract afn float %x, 1.023000e+03
+ %mul.inner.1.insert.1.i32 = fptoui float %mul.inner.1.insert to i32
+ %shl.inner.1.insert.1 = shl i32 %mul.inner.1.insert.1.i32, 20
+ %bfi.inner.1.and.1 = and i32 %shl.inner.1.insert.1, 1072693248
+ %bfi.inner.1.and.2 = and i32 %mul.base.i32, -1073741793
+ %bfi.inner.1 = or i32 %bfi.inner.1.and.2, %bfi.inner.2
+ %bfi.outer = or i32 %bfi.inner.1.and.1, %bfi.inner.1
+ %result = bitcast i32 %bfi.outer to float
+ ret float %result
+}
+
define float @v_bfi_single_nesting_level_inner_use(float %x, float %y, float %z) {
; GFX10-LABEL: v_bfi_single_nesting_level_inner_use:
; GFX10: ; %bb.0: ; %.entry
@@ -145,3 +205,32 @@ define float @v_bfi_two_levels(float %x, float %y, float %z) {
%result = bitcast i32 %or.outer to float
ret float %result
}
+
+define float @v_bfi_single_constant_as_partition(float %x, float %y, float %z) {
+; GFX10-LABEL: v_bfi_single_constant_as_partition:
+; GFX10: ; %bb.0: ; %.entry
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT: v_mul_f32_e32 v0, 0x447fc000, v0
+; GFX10-NEXT: v_mul_f32_e32 v2, 0x447fc000, v2
+; GFX10-NEXT: v_cvt_u32_f32_e32 v1, v1
+; GFX10-NEXT: v_cvt_u32_f32_e32 v0, v0
+; GFX10-NEXT: v_cvt_u32_f32_e32 v2, v2
+; GFX10-NEXT: v_lshlrev_b32_e32 v1, 10, v1
+; GFX10-NEXT: v_lshlrev_b32_e32 v0, 20, v0
+; GFX10-NEXT: v_or3_b32 v0, v1, v2, v0
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+.entry:
+ %mul.base = fmul reassoc nnan nsz arcp contract afn float %z, 1.023000e+03
+ %mul.base.i32 = fptoui float %mul.base to i32
+ %y.i32 = fptoui float %y to i32
+ %shl.inner.insert = shl i32 %y.i32, 10
+ %bfi1.or = or i32 %shl.inner.insert, %mul.base.i32
+ %mul.outer.insert = fmul reassoc nnan nsz arcp contract afn float %x, 1.023000e+03
+ %mul.outer.insert.i32 = fptoui float %mul.outer.insert to i32
+ %shl.outer.insert = shl i32 %mul.outer.insert.i32, 20
+ %and.outer = and i32 %shl.outer.insert, -1
+ %or.outer = or i32 %bfi1.or, %and.outer
+ %result = bitcast i32 %or.outer to float
+ ret float %result
+}
More information about the llvm-commits
mailing list