[AArch64] Switch to soft promoting half types. #80576
Conversation
|
@llvm/pr-subscribers-backend-aarch64 Author: Harald van Dijk (hvdijk) ChangesThe traditional promotion is known to generate wrong code. Like #80440 for ARM, except that far less is affected as on AArch64, hardware floating point support always includes FP16 support and is unaffected by these changes. This only affects Full diff: https://github.com/llvm/llvm-project/pull/80576.diff 2 Files Affected:
diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.h b/llvm/lib/Target/AArch64/AArch64ISelLowering.h
index 436b21fd13463..160eafab8c364 100644
--- a/llvm/lib/Target/AArch64/AArch64ISelLowering.h
+++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.h
@@ -1308,6 +1308,10 @@ class AArch64TargetLowering : public TargetLowering {
bool preferScalarizeSplat(SDNode *N) const override;
unsigned getMinimumJumpTableEntries() const override;
+
+ bool softPromoteHalfType() const override { return true; }
+
+ bool useFPRegsForHalfType() const override { return true; }
};
namespace AArch64 {
diff --git a/llvm/test/CodeGen/AArch64/strictfp_f16_abi_promote.ll b/llvm/test/CodeGen/AArch64/strictfp_f16_abi_promote.ll
index 37186cf22ccc7..a34f7abcc22a3 100644
--- a/llvm/test/CodeGen/AArch64/strictfp_f16_abi_promote.ll
+++ b/llvm/test/CodeGen/AArch64/strictfp_f16_abi_promote.ll
@@ -70,22 +70,20 @@ define void @v3f16_arg(<3 x half> %arg, ptr %ptr) #0 {
; NOFP16-NEXT: .cfi_offset w22, -32
; NOFP16-NEXT: .cfi_offset w30, -48
; NOFP16-NEXT: mov w21, w0
-; NOFP16-NEXT: and w0, w2, #0xffff
+; NOFP16-NEXT: and w0, w1, #0xffff
; NOFP16-NEXT: mov x19, x3
-; NOFP16-NEXT: mov w20, w1
+; NOFP16-NEXT: mov w20, w2
; NOFP16-NEXT: bl __gnu_h2f_ieee
; NOFP16-NEXT: mov w22, w0
; NOFP16-NEXT: and w0, w21, #0xffff
; NOFP16-NEXT: bl __gnu_h2f_ieee
-; NOFP16-NEXT: mov w21, w0
+; NOFP16-NEXT: mov w8, w0
; NOFP16-NEXT: and w0, w20, #0xffff
+; NOFP16-NEXT: orr x21, x8, x22, lsl #32
; NOFP16-NEXT: bl __gnu_h2f_ieee
-; NOFP16-NEXT: mov w8, w21
-; NOFP16-NEXT: // kill: def $w0 killed $w0 def $x0
-; NOFP16-NEXT: str w22, [x19, #8]
-; NOFP16-NEXT: orr x8, x8, x0, lsl #32
+; NOFP16-NEXT: str x21, [x19]
; NOFP16-NEXT: ldp x22, x21, [sp, #16] // 16-byte Folded Reload
-; NOFP16-NEXT: str x8, [x19]
+; NOFP16-NEXT: str w0, [x19, #8]
; NOFP16-NEXT: ldp x20, x19, [sp, #32] // 16-byte Folded Reload
; NOFP16-NEXT: ldr x30, [sp], #48 // 8-byte Folded Reload
; NOFP16-NEXT: ret
@@ -182,46 +180,17 @@ define void @v4f16_arg(<4 x half> %arg, ptr %ptr) #0 {
define void @outgoing_v4f16_return(ptr %ptr) #0 {
; NOFP16-LABEL: outgoing_v4f16_return:
; NOFP16: // %bb.0:
-; NOFP16-NEXT: stp x30, x23, [sp, #-48]! // 16-byte Folded Spill
-; NOFP16-NEXT: stp x22, x21, [sp, #16] // 16-byte Folded Spill
-; NOFP16-NEXT: stp x20, x19, [sp, #32] // 16-byte Folded Spill
-; NOFP16-NEXT: .cfi_def_cfa_offset 48
+; NOFP16-NEXT: stp x30, x19, [sp, #-16]! // 16-byte Folded Spill
+; NOFP16-NEXT: .cfi_def_cfa_offset 16
; NOFP16-NEXT: .cfi_offset w19, -8
-; NOFP16-NEXT: .cfi_offset w20, -16
-; NOFP16-NEXT: .cfi_offset w21, -24
-; NOFP16-NEXT: .cfi_offset w22, -32
-; NOFP16-NEXT: .cfi_offset w23, -40
-; NOFP16-NEXT: .cfi_offset w30, -48
+; NOFP16-NEXT: .cfi_offset w30, -16
; NOFP16-NEXT: mov x19, x0
; NOFP16-NEXT: bl v4f16_result
-; NOFP16-NEXT: and w0, w0, #0xffff
-; NOFP16-NEXT: mov w20, w1
-; NOFP16-NEXT: mov w21, w2
-; NOFP16-NEXT: mov w22, w3
-; NOFP16-NEXT: bl __gnu_h2f_ieee
-; NOFP16-NEXT: mov w23, w0
-; NOFP16-NEXT: and w0, w20, #0xffff
-; NOFP16-NEXT: bl __gnu_h2f_ieee
-; NOFP16-NEXT: mov w20, w0
-; NOFP16-NEXT: and w0, w21, #0xffff
-; NOFP16-NEXT: bl __gnu_h2f_ieee
-; NOFP16-NEXT: mov w21, w0
-; NOFP16-NEXT: and w0, w22, #0xffff
-; NOFP16-NEXT: bl __gnu_h2f_ieee
-; NOFP16-NEXT: bl __gnu_f2h_ieee
-; NOFP16-NEXT: strh w0, [x19, #6]
-; NOFP16-NEXT: mov w0, w21
-; NOFP16-NEXT: bl __gnu_f2h_ieee
-; NOFP16-NEXT: strh w0, [x19, #4]
-; NOFP16-NEXT: mov w0, w20
-; NOFP16-NEXT: bl __gnu_f2h_ieee
-; NOFP16-NEXT: strh w0, [x19, #2]
-; NOFP16-NEXT: mov w0, w23
-; NOFP16-NEXT: bl __gnu_f2h_ieee
+; NOFP16-NEXT: strh w2, [x19, #4]
+; NOFP16-NEXT: strh w3, [x19, #6]
+; NOFP16-NEXT: strh w1, [x19, #2]
; NOFP16-NEXT: strh w0, [x19]
-; NOFP16-NEXT: ldp x20, x19, [sp, #32] // 16-byte Folded Reload
-; NOFP16-NEXT: ldp x22, x21, [sp, #16] // 16-byte Folded Reload
-; NOFP16-NEXT: ldp x30, x23, [sp], #48 // 16-byte Folded Reload
+; NOFP16-NEXT: ldp x30, x19, [sp], #16 // 16-byte Folded Reload
; NOFP16-NEXT: ret
%val = call <4 x half> @v4f16_result()
store <4 x half> %val, ptr %ptr
@@ -231,82 +200,21 @@ define void @outgoing_v4f16_return(ptr %ptr) #0 {
define void @outgoing_v8f16_return(ptr %ptr) #0 {
; NOFP16-LABEL: outgoing_v8f16_return:
; NOFP16: // %bb.0:
-; NOFP16-NEXT: stp x30, x27, [sp, #-80]! // 16-byte Folded Spill
-; NOFP16-NEXT: stp x26, x25, [sp, #16] // 16-byte Folded Spill
-; NOFP16-NEXT: stp x24, x23, [sp, #32] // 16-byte Folded Spill
-; NOFP16-NEXT: stp x22, x21, [sp, #48] // 16-byte Folded Spill
-; NOFP16-NEXT: stp x20, x19, [sp, #64] // 16-byte Folded Spill
-; NOFP16-NEXT: .cfi_def_cfa_offset 80
+; NOFP16-NEXT: stp x30, x19, [sp, #-16]! // 16-byte Folded Spill
+; NOFP16-NEXT: .cfi_def_cfa_offset 16
; NOFP16-NEXT: .cfi_offset w19, -8
-; NOFP16-NEXT: .cfi_offset w20, -16
-; NOFP16-NEXT: .cfi_offset w21, -24
-; NOFP16-NEXT: .cfi_offset w22, -32
-; NOFP16-NEXT: .cfi_offset w23, -40
-; NOFP16-NEXT: .cfi_offset w24, -48
-; NOFP16-NEXT: .cfi_offset w25, -56
-; NOFP16-NEXT: .cfi_offset w26, -64
-; NOFP16-NEXT: .cfi_offset w27, -72
-; NOFP16-NEXT: .cfi_offset w30, -80
+; NOFP16-NEXT: .cfi_offset w30, -16
; NOFP16-NEXT: mov x19, x0
; NOFP16-NEXT: bl v8f16_result
-; NOFP16-NEXT: and w0, w0, #0xffff
-; NOFP16-NEXT: mov w21, w1
-; NOFP16-NEXT: mov w22, w2
-; NOFP16-NEXT: mov w23, w3
-; NOFP16-NEXT: mov w24, w4
-; NOFP16-NEXT: mov w25, w5
-; NOFP16-NEXT: mov w26, w6
-; NOFP16-NEXT: mov w27, w7
-; NOFP16-NEXT: bl __gnu_h2f_ieee
-; NOFP16-NEXT: mov w20, w0
-; NOFP16-NEXT: and w0, w21, #0xffff
-; NOFP16-NEXT: bl __gnu_h2f_ieee
-; NOFP16-NEXT: mov w21, w0
-; NOFP16-NEXT: and w0, w22, #0xffff
-; NOFP16-NEXT: bl __gnu_h2f_ieee
-; NOFP16-NEXT: mov w22, w0
-; NOFP16-NEXT: and w0, w23, #0xffff
-; NOFP16-NEXT: bl __gnu_h2f_ieee
-; NOFP16-NEXT: mov w23, w0
-; NOFP16-NEXT: and w0, w24, #0xffff
-; NOFP16-NEXT: bl __gnu_h2f_ieee
-; NOFP16-NEXT: mov w24, w0
-; NOFP16-NEXT: and w0, w25, #0xffff
-; NOFP16-NEXT: bl __gnu_h2f_ieee
-; NOFP16-NEXT: mov w25, w0
-; NOFP16-NEXT: and w0, w26, #0xffff
-; NOFP16-NEXT: bl __gnu_h2f_ieee
-; NOFP16-NEXT: mov w26, w0
-; NOFP16-NEXT: and w0, w27, #0xffff
-; NOFP16-NEXT: bl __gnu_h2f_ieee
-; NOFP16-NEXT: bl __gnu_f2h_ieee
-; NOFP16-NEXT: strh w0, [x19, #14]
-; NOFP16-NEXT: mov w0, w26
-; NOFP16-NEXT: bl __gnu_f2h_ieee
-; NOFP16-NEXT: strh w0, [x19, #12]
-; NOFP16-NEXT: mov w0, w25
-; NOFP16-NEXT: bl __gnu_f2h_ieee
-; NOFP16-NEXT: strh w0, [x19, #10]
-; NOFP16-NEXT: mov w0, w24
-; NOFP16-NEXT: bl __gnu_f2h_ieee
-; NOFP16-NEXT: strh w0, [x19, #8]
-; NOFP16-NEXT: mov w0, w23
-; NOFP16-NEXT: bl __gnu_f2h_ieee
-; NOFP16-NEXT: strh w0, [x19, #6]
-; NOFP16-NEXT: mov w0, w22
-; NOFP16-NEXT: bl __gnu_f2h_ieee
-; NOFP16-NEXT: strh w0, [x19, #4]
-; NOFP16-NEXT: mov w0, w21
-; NOFP16-NEXT: bl __gnu_f2h_ieee
-; NOFP16-NEXT: strh w0, [x19, #2]
-; NOFP16-NEXT: mov w0, w20
-; NOFP16-NEXT: bl __gnu_f2h_ieee
+; NOFP16-NEXT: strh w5, [x19, #10]
+; NOFP16-NEXT: strh w7, [x19, #14]
+; NOFP16-NEXT: strh w6, [x19, #12]
+; NOFP16-NEXT: strh w4, [x19, #8]
+; NOFP16-NEXT: strh w3, [x19, #6]
+; NOFP16-NEXT: strh w2, [x19, #4]
+; NOFP16-NEXT: strh w1, [x19, #2]
; NOFP16-NEXT: strh w0, [x19]
-; NOFP16-NEXT: ldp x20, x19, [sp, #64] // 16-byte Folded Reload
-; NOFP16-NEXT: ldp x22, x21, [sp, #48] // 16-byte Folded Reload
-; NOFP16-NEXT: ldp x24, x23, [sp, #32] // 16-byte Folded Reload
-; NOFP16-NEXT: ldp x26, x25, [sp, #16] // 16-byte Folded Reload
-; NOFP16-NEXT: ldp x30, x27, [sp], #80 // 16-byte Folded Reload
+; NOFP16-NEXT: ldp x30, x19, [sp], #16 // 16-byte Folded Reload
; NOFP16-NEXT: ret
%val = call <8 x half> @v8f16_result()
store <8 x half> %val, ptr %ptr
|
hvdijk
force-pushed
the
softpromotehalf-fpreg-aarch64
branch
from
6e43d1d to
d563b51
Compare
|
Given this is a ABI change should we document this somewhere in the release notes? Do we need to worry about backwards compatibility with older versions of clang? |
Hi David, this is not intended to be an ABI change; code compiled with old and new Clang should be fully interoperable (provided, of course, neither or both are using |
Ah sorry, my bad! I'm not sure why I thought it was an ABI break. Looking at how your patch changes the tests I see now it doesn't look like it. However, is this something still worth mentioning in release notes somewhere? |
Possibly, yes. Ideally I would like to get multiple arches moved over (ARM already done, a few other ones I'll be able to submit PRs for next week), so maybe it should be in the general release notes with a list of affected arches? AArch64 is probably the one it will have the least impact on: the default configuration is unaffected because it supports FP16 natively. Not sure what's best here, I would welcome some input. |
hvdijk
force-pushed
the
softpromotehalf-fpreg-aarch64
branch
from
d563b51 to
b4fe6e5
Compare
|
@david-arm I had another look at how to document and have amended in a way that easily extends to the other architectures when I create more PRs for them, does it look good to you like this? |
|
Hi - I don't personally thing this is worth a release note. It seems like just a standard backend codegen change/bugfix that we would not usually release note, especially when -fp-armv8 is less encouraged for AArch64 and this effects fp16. A general release note for the better support of -fp-armv8 might be good in the llvm 19 time frame, but that can be left for Olivers other work. Happy to hear alternative thoughts though. |
The traditional promotion is known to generate wrong code.
hvdijk
force-pushed
the
softpromotehalf-fpreg-aarch64
branch
from
b4fe6e5 to
7abbd96
Compare
|
Ha, okay, back out it goes, back to the first version :) I don't mind either way. |
The traditional promotion is known to generate wrong code.
Like #80440 for ARM, except that far less is affected as on AArch64, hardware floating point support always includes FP16 support and is unaffected by these changes. This only affects
-mgeneral-regs-only(Clang) /-mattr=-fp-armv8(LLVM).Because this only affects a configuration where no FP support is available at all,
useFPRegsForHalfType()has no effect and is not specified:f32was getting legalized as a parameter and return type to an integer anyway.